Lightweight dynamic weighing method based on bridge monitoring response

Abstract

The invention discloses a lightweight dynamic weighing method based on bridge monitoring response, and the method comprises the steps: capturing a monitoring video of a random driving vehicle through a high-definition camera and an image processing module, classifying the vehicle, and obtaining the deflection response data of a bridge through a visual displacement sensor when the vehicle passes; a video image signal and a bridge deflection response data signal are aligned and matched through a time synchronization module, a corresponding data set is established, a Bi-LSTM model is utilized to train the corresponding data set under a special vehicle, and a specific vehicle load-response mapping relation learned by the model is utilized to determine the deflection response of the bridge. And analyzing the structural response data under the action of the random vehicle, and outputting a load value of the random vehicle by using the trained Bi-LSTM model. According to the method, the dynamic response data of the bridge can be effectively collected, the dynamic load characteristics of the random heavy-load traffic flow can be accurately recognized in combination with the load information of a specific vehicle, and technical support is provided for load recognition and safety management of small and medium-span bridges.

Description

Technical Field

[0001] The present invention belongs to the field of bridge health monitoring and vehicle dynamic weighing, and specifically relates to a lightweight dynamic weighing method based on bridge monitoring response. By deploying a lightweight structural response monitoring device on the bridge and combining the load information of specific vehicles on the bridge deck, dynamic weighing of random heavy-loaded vehicle flow can be achieved. Background Art

[0002] Traffic flow load is the main variable load borne by highway bridge structures, which is highly random in time and space. With the rapid development of transportation, the actual traffic flow load borne by bridges has far exceeded the design load standard. Vehicle overload and repeated effects of traffic flow load have become the main factors affecting the safety of bridge operation and significantly shortening its service life. Therefore, accurate identification of traffic flow load is crucial to bridge structure design and safety management.

[0003] In traffic flow load, the key vehicle characteristic parameter is load information. The introduction of the Weigh-In-Motion (WIM) system has brought revolutionary changes to vehicle load detection and has now become an important tool for load measurement. The WIM system collects key data such as vehicle axle weight, total weight, driving speed and arrival time through piezoelectric sensors installed on the road surface. The technology is mature and widely used in actual engineering. However, due to limiting factors such as high cost and complex construction, the WIM system is difficult to popularize in small and medium span bridges. Therefore, the method of vehicle load identification based on bridge dynamic response urgently needs to break through the existing technical bottleneck.

[0004] The present invention aims to achieve dynamic weighing of vehicles on small and medium span bridges based on bridge monitoring response through a lightweight bridge health monitoring system. The system can not only effectively collect bridge dynamic response data, but also accurately identify the dynamic load characteristics of random heavy-loaded traffic in combination with the load information of specific vehicles. The purpose of the invention is to provide a low-cost, easy-to-deploy and reliable weighing system to provide technical support for load identification and safety management of small and medium span bridges. Summary of the invention

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

[0006] The technical solution adopted by the present invention is as follows: A lightweight dynamic weighing method based on bridge monitoring response includes the following steps: 1) A video surveillance system is arranged on the bridge deck. The video surveillance system includes a high-definition camera and an image processing module. The high-definition camera captures 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 for transporting dangerous goods. Special vehicles are further accurately subdivided according to the types of dangerous goods transported. The loads of different types of special vehicles are all known fixed values; 2) A high-precision visual displacement sensor is deployed at the mid-span position of the bridge superstructure. The visual displacement sensor uses an industrial camera to capture high-resolution images of the target area deployed at the mid-span of the bridge superstructure under the action of random vehicle loads, and uses DIC technology to calculate the dynamic response data of the bridge structure, that is, the data of the bridge deflection response; 3) The random traffic flow monitoring video image signal obtained by the video surveillance system and the bridge structure dynamic response data signal obtained by the visual displacement sensor are aligned and matched on the preset time axis through the time synchronization module. For the random traffic flow on the bridge, the vehicle load is calibrated as w j , the corresponding bridge structure dynamic response data is calibrated as x j , establish a total spatiotemporal correlation data set for building mapping relationships; Identify and extract special vehicles from random traffic monitoring video images, and calibrate the load of special vehicles as w i And the corresponding bridge structure dynamic response data is calibrated as x i , establish a spatiotemporal correlation sub-dataset for learning the characteristic values ​​of the knot response; 4) A bidirectional long short-term memory network Bi-LSTM model is established. The spatiotemporal correlation sub-dataset for learning the response eigenvalues ​​established in step 3) is input into the model for training. Then the spatiotemporal correlation total data set for building the mapping relationship is input into the trained Bi-LSTM model. The specific vehicle load-response mapping relationship learned by the model is used to parse the structural response data under the action of the random vehicle, and the load value of the random vehicle is output.

[0007] Furthermore, in step 1), the image processing module uses multi-view regression detection technology to realize automatic recognition and classification of vehicle information in the monitoring video. First, it roughly divides them into small vehicles, heavy-loaded vehicles and special vehicles, and further detects the special vehicles in the rough classification. By identifying the side photo of the vehicle, it detects whether there is a dangerous goods transportation sign. After confirming the existence of the sign, the content of the chemical hazardous goods vehicle sign is analyzed to identify the specific type of the sign, and the accurate classification of special vehicles is realized by combining the vehicle appearance and the sign information.

[0008] Furthermore, in step 2), the process of calculating the dynamic response data of the bridge structure using the DIC technology is to use an image processing algorithm with sub-pixel accuracy 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 the dynamic response data of the bridge structure for subsequent load identification and structural performance evaluation.

[0009] Furthermore, the specific processing process 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 the calibration coefficient in mm / pixel. S2: Next, the target displacement video is collected by the camera and the image is converted into a grayscale image, the pixel coordinates of the target point to be measured on the measurement target are selected, and the data is preprocessed and outliers are eliminated. The preprocessing includes the processing steps of data interception and reference value extraction, that is, firstly intercepting the time history data of the bridge deflection response before and after the vehicle passes, that is, the change of the pixel coordinates of the target point to be measured over time, and taking the initial position pixel coordinates of the target point to be measured before the vehicle passes as the reference value; the outlier elimination adopts the interquartile range method to eliminate the extreme deviation data in the original number; S3: The positional relationship of the target point on the image is calculated using the cross-correlation function, and the pixel coordinate difference between the target point to be measured and its initial position when the vehicle passes by is calculated, which is the pixel displacement, in units of 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 to be measured is obtained as the extracted feature value; S4: Finally, the extracted pixel displacement characteristic value of the target point to be measured is multiplied by the calibration coefficient and converted into the actual physical displacement, that is, the displacement time history of the target point to be measured is obtained.

[0010] Furthermore, in step S3, an image processing algorithm with sub-pixel level accuracy is used to search the target point twice, namely, an integer pixel search and a sub-pixel search. The integer pixel search searches for the integer pixel position of the target point in the target image, that is, the integer pixel coordinates of the target point are obtained; the sub-pixel search is performed on the basis of the integer pixel search to obtain the precise position of the pixel point, and the obtained pixel coordinates of the target point are decimals. This process improves the accuracy of the monitoring data.

[0011] Furthermore, the time-space correlation total / sub-dataset established in step 3) specifically refers to matching the vehicle information and structural response data obtained by multiple sensors at a time of vehicle passing through by corresponding to the acquisition time, and storing them in the data set. The time-space correlation total / sub-dataset specifically includes: the video surveillance image acquisition time, license plate, vehicle type, vehicle weight, and structural deflection response data of each vehicle passing through each time. The time-space correlation total data set includes data information of all vehicles passing through the bridge; while the time-space correlation sub-dataset only includes vehicle data information of vehicles whose vehicle type is a specific special vehicle.

[0012] Furthermore, step 4) before the Bi-LSTM model trains the data set, it also includes a feature extraction step, the specific steps are: M1: First, extract the Mel-frequency cepstral coefficient features in the main beam deflection monitoring response signal when different vehicles pass by. First, denoise the signal to remove high-frequency noise; then, divide the signal into frames and apply the Hanning window function for smoothing, and then convert the signal from the time domain to the frequency domain through the fast Fourier transform FFT; then, use the Mel filter bank to convert the spectrum into the Mel frequency scale, calculate the energy of each Mel filter channel, and obtain the Mel spectrum; finally, use the discrete cosine transform DCT to convert the Mel spectrum into Mel frequency cepstral coefficients MFCCs, and select the first m MFCCs coefficients as the feature vector, and further calculate the first-order and second-order differences of each selected MFCCs coefficient to enhance the dynamic change information, and concatenate the MFCCs, first-order difference and second-order difference features to form the final feature vector X i ,These eigenvectors effectively characterize the dynamic response of the main beam deflection when the vehicle passes, and are suitable for subsequent signal analysis and classification; The formulas for the above Mel frequency conversion and discrete cosine transform DCT are as follows: Mel frequency conversion: .

[0013] in, f is the linear frequency, M(f) is the Mel frequency; DCT transform: .

[0014] in C k are the 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, combined with the chi-square test, the features that are highly relevant to the classification task are screened out. The specific process is as follows: First, the MFCCs feature values ​​are classified according to the labels of vehicle type, load level, driving speed and number of axles, and the expected frequency is calculated according to the following formula: .

[0015] in, O j MFCCs and first-order and second-order differences, a total of 3m features X i The total number of samples with the value j, m is the number of selected MFCCs coefficients, and the total number of selected MFCCs coefficients and their first-order difference and second-order difference features is 3m. O k Label k The total number of samples, O is the total number of samples; E jk is the expected frequency of samples with value j and label k; Then for the eigenvalue X i The chi-square statistic x is calculated as follows 2 ,in O jk is the number of samples with value j and label k, that is, the actual frequency, E jk is the expected frequency, a It is a feature X i The different values ​​of b is the number of label categories; .

[0016] Then, according to the selected significance level and degrees of freedom, query the chi-square distribution table to find the corresponding chi-square critical value χ critical 2 ; Keep the calculated chi-square value x 2 Greater than the critical value χ critical 2 This step can screen out highly correlated features and reduce redundant features.

[0017] Finally, in order to verify the accuracy of the vehicle load dynamic identification method of the present invention, 500 groups of random vehicle sample data were tested under the test conditions of vehicle speed 30-100 km / h and vehicle gross weight 10-60 tons. The verification results are shown in Figure 5The Bi-LSTM model trained by the present invention is used to analyze the structural response of the random vehicle, and the load value of the random vehicle is output as Figure 5 The results show that the relative error between the identified load value and the true load value is less than 10% in 90% of the samples, and the overall recognition error of the system is stably controlled within the range of ±10%.

[0018] Compared with the prior art, the present invention has the following beneficial effects: Traditional dynamic weighing systems mainly rely on dedicated sensor arrays and complex hardware equipment to monitor vehicle loads, and have three significant defects: first, the procurement and deployment of high-precision weighing sensors and supporting equipment leads to high overall system costs; second, when installing them on an already operating bridge, the bridge deck pavement needs to be removed for embedded installation, which poses engineering challenges such as long construction periods and large traffic interference; third, sensors are prone to fatigue damage due to long-term vehicle rolling, which limits system reliability and service life. In contrast, the lightweight dynamic weighing system based on structural response monitoring of the present invention innovatively constructs a "vehicle-bridge" dynamic coupling inversion model by integrating random vehicle information on the bridge deck and structural deflection response data. While significantly reducing the reliance on dedicated sensors, it uses non-invasive monitoring equipment to achieve lightweight deployment of the system. The system not only reduces hardware costs by about 90%, but also avoids destructive construction of the bridge deck structure. At the same time, by optimizing the stress transfer path, the durability of the equipment is increased by more than 3 times, and finally achieves a dynamic identification accuracy of 90% for vehicle loads under a low-cost operation and maintenance framework. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a technical flow chart of a lightweight dynamic weighing monitoring method based on the monitoring response of a bridge under a specific vehicle according to the present invention; Figure 2 It is the structural diagram of the bridge lightweight monitoring system; Figure 3 This is the network structure diagram of the bidirectional long short-term memory network (BiLSTM); Figure 4 Schematic diagram of the chemical hazardous goods vehicle sign.

[0020] Figure 5 Vehicle weight calculation results during the testing phase. DETAILED DESCRIPTION

[0021] A lightweight dynamic weighing system based on monitoring the response of a bridge under a specific vehicle includes the following three modules: (1) a structural response monitoring module based on a visual displacement sensor; (2) a bridge deck vehicle recognition and classification module based on multi-view regression technology; and (3) a load-response mapping model establishment module based on a bidirectional long short-term memory network.

[0022] In order to make the objectives, technical solutions and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0023] like Figure 1 The lightweight dynamic weighing monitoring method based on the monitoring response of a bridge under a specific vehicle described in this embodiment includes the following steps: Step 1: Build a lightweight monitoring technology system for small and medium span bridges based on visual displacement sensors and high-definition cameras to monitor the structural response under random vehicle loads and capture the dynamics of the moving traffic; For details, see Figure 2 , by deploying high-precision visual displacement sensors at the mid-span position of the bridge superstructure to monitor the dynamic displacement response of the structure. At the same time, high-definition cameras are arranged at both ends of the bridge deck to capture the appearance characteristics and driving trajectory of the vehicle, so as to achieve the synchronous collection of vehicle data such as license plate, driving lane, model and structural response data such as dynamic deflection in the mid-span caused by the vehicle. The high-definition camera is connected to the image processing module signal.

[0024] In order to suppress environmental interference, visual displacement sensors and high-definition cameras are installed with anti-interference designs and are equipped with sunshades or waterproof casings to adapt to harsh conditions.

[0025] Step 2: The visual displacement sensor realizes real-time dynamic monitoring of the mid-span deflection of the bridge based on the digital image correlation technology (DIC).

[0026] Specifically, the DIC technology uses high-resolution images of the target area in the middle of the bridge span captured by an industrial camera to calculate the deflection change based on the correlation analysis of the image grayscale distribution. The visual displacement sensor records the vertical displacement response of the target point in the middle of the bridge span under the action of random vehicle loads in real time, and uses an image processing algorithm with sub-pixel accuracy 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 the bridge mid-span deflection for subsequent load identification and structural performance evaluation.

[0027] Specifically, the DIC technology first performs camera calibration to eliminate distortion, establishes the relationship between image pixels and actual size, and obtains the calibration coefficient (unit: mm / pixel); then, the target displacement video is collected by the camera and the image is converted into a grayscale image. The pixel coordinates of the target point to be measured on the measurement target are selected, and the data is preprocessed and outliers are eliminated. The preprocessing includes the processing steps of data interception and reference value extraction.

[0028] Specifically, the data preprocessing mainly includes the steps of data interception and reference value extraction. First, the time history data of the bridge deflection response 15s before and after the vehicle passes is intercepted, that is, the change of the pixel coordinates of the target point to be measured over time, and the initial position pixel coordinates of the target point to be measured before the vehicle passes are taken as the reference value.

[0029] Outlier elimination uses IQR (interquartile range method) to eliminate extremely deviant data in the original numbers.

[0030] The cross-correlation function is used to calculate the positional relationship of the target point on the image, and the pixel coordinate difference between the target point to be measured and its initial position when the vehicle passes by is calculated, which is the pixel displacement (unit: pixel). The pixel displacement is calculated for each frame image in the monitoring video of the target point to be measured, and the pixel displacement time history of the target point is obtained. The absolute maximum value of the pixel displacement of the target point to be measured that deviates from its initial position is taken as the extracted feature value.

[0031] Finally, the extracted pixel displacement characteristic value of the target point to be measured is multiplied by the calibration coefficient and converted into the actual physical displacement, that is, the displacement time history of the target point to be measured is obtained.

[0032] The cross-correlation function expression is as follows: .

[0033] in f(i , j) is the grayscale function of the target point area to be measured, (i , j) It is the pixel coordinate of the target point area to be measured. g(i+y , j+x) is the grayscale function of the search image sub-area at the distance from the target sub-area to be measured, (i+y , j+x) is the pixel coordinate of the search image sub-area. The pixel coordinate difference between the search image sub-area and the target point sub-area to be measured is (x, y).

[0034] w is the current search window size, C CC is the mutual correlation coefficient between the two sub-areas of the current search window, (x,y) is the displacement between the two sub-areas, when C CC When the maximum value is obtained in the search area, the corresponding (x,y) It is the maximum displacement between the two sub-areas, that is, the absolute maximum value of the pixel displacement of the target point to be measured deviating from its initial position.

[0035] The sub-pixel level precision image processing algorithm, that is, based on the calculation method of the above-mentioned cross-correlation function, searches the target point twice (that is, using the above-mentioned cross-correlation function, the method of calculating the absolute maximum value of the pixel displacement of the target point to be measured from its initial position is used for searching, and the coordinates (x, y) corresponding to the integer pixel search are integers, and the sub-pixels are decimals), which are integer pixel search and sub-pixel search. The integer pixel search searches for the integer pixel position of the target point in the target image, that is, the integer pixel coordinates of the target point are obtained; the sub-pixel search searches on the basis of the integer pixel search to obtain the precise position of the pixel point, and the pixel coordinates of the target point obtained are decimals. This process improves the accuracy of the monitoring data.

[0036] Step 3: Automatically identify and classify traffic information in surveillance videos based on multi-view regression detection technology. First, roughly classify vehicle types (such as small vehicles, heavy-loaded vehicles, and special vehicles, etc.); then, further detect the special vehicles in the rough classification to detect whether there are dangerous goods transportation signs (such as dangerous chemicals signs affixed to the vehicle body), and parse the content of the chemical dangerous goods vehicle signs to identify their specific types (such as explosives, compressed gases, and liquefied gases, etc.), so as to achieve accurate classification of special vehicles. The schematic diagram of the chemical dangerous goods vehicle sign is shown below. Figure 4 , and listed the following hazardous chemical signs: 1. Explosives, 2. Compressed gases and liquefied gases, 3. Flammable liquids, 4. Flammable solids, 5. Oxidizers and organic peroxides, 6. Toxic and infectious substances, 7. Radioactive substances.

[0037] Step 4: Based on the time series correlation of vehicle-bridge monitoring data, a spatiotemporal correlation dataset of vehicle load-structure response is established.

[0038] The random traffic flow monitoring video image signal obtained by the video monitoring system and the dynamic response data signal of the bridge structure obtained by the visual displacement sensor are aligned and matched on the preset time axis through the time synchronization module. For the random traffic flow on the bridge, the vehicle load is calibrated as w j , the corresponding bridge structure dynamic response data is calibrated as x j , establish a total spatiotemporal correlation data set for building mapping relationships; Using target detection technology, special vehicles are identified and extracted from random traffic monitoring video images, and the load of special vehicles is calibrated as w i And the corresponding bridge structure dynamic response data is calibrated as x i , establish a spatiotemporal correlation sub-dataset for learning the knot response eigenvalues.

[0039] Since the load of the specific vehicle type selected in this technical solution is relatively constant, especially the chemical hazardous goods transport vehicle has strict transportation standards, its load is basically fixed, and the corresponding structural response is known, so when training the model, the specific vehicle load can be input as a known value w i , and with the known structural response x i Perform feature learning.

[0040] Step 5: Based on the bidirectional long short-term memory network (Bi-LSTM), the time-history characteristics of the structural response under the action of a specific vehicle load are extracted through autonomous learning, and a mapping relationship model between vehicle load and structural response is constructed.

[0041] Specifically, the Mel-Frequency Cepstral Coefficients (MFCCs) features in the main beam deflection monitoring response signal when different vehicles pass by are first extracted, and the features that are highly relevant to the classification task are screened out in combination with the Chi-square test. Subsequently, a Bi-LSTM model is constructed, and the Bi-LSTM is used to capture the time dependency of the signal. The accuracy and robustness of the monitoring signal and vehicle load mapping model are improved by introducing an attention mechanism. Through model training, the load parameters of specific vehicles are accurately restored, thereby achieving efficient inversion of the load parameters of heavy-loaded vehicles.

[0042] The extraction of Mel frequency cepstral coefficients (MFCCs) features first uses VMD variational mode decomposition to perform signal denoising to remove high-frequency noise; then, the signal is divided into frames and each frame of the signal is smoothed with a Hanning window function to reduce spectrum leakage, and then the signal is converted from the time domain to the frequency domain through a fast Fourier transform (FFT); then, the linear spectrum is converted to a Mel frequency scale using a Mel filter bank, and the energy of each Mel filter channel is calculated to obtain a Mel spectrum; finally, the Mel spectrum is converted to a Mel frequency cepstral coefficient (MFCC) through a discrete cosine transform (DCT), and the first 13 MFCCs coefficients are selected as feature vectors, and the first and second order differences (Δ and ΔΔ) of each selected MFCCs coefficient are further calculated to enhance dynamic change information. The MFCCs, first order difference, and second order difference features are concatenated to form the final feature vector X i , a total of 39 features. 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 formulas of the key steps involved are as follows: Mel filter bank, Mel frequency conversion: .

[0043] in, f is the linear frequency, M(f) is the mel frequency.

[0044] DCT transform: .

[0045] in C k are the cepstral coefficients, E(f n ) is the Mel spectrum (computed from the energy of each Mel filter channel), k is the index of the MFCC. N represents the signal length (or block size), n Indicates the sampling point index of the time domain signal (the value range is 0≤ n ≤ N -1). In the formula of 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 range of n is 0 ≤n ≤N-1, corresponding to the nth sampling point of the signal.

[0046] The Chi-square test first classifies the MFCCs feature values ​​according to vehicle type, load level, driving speed, and number of axles, and calculates the expected frequency according to the following formula. O j Features X i The total number of samples with the value j, j is the vector set of the eigenvalues, O k Label k The total number of samples, O is the total number of all samples.

[0047] .

[0048] O j There are 39 features in total for MFCCs and first-order and second-order differences X i The total number of samples with a value of j (where the j value range of MFCCs is [-50,50], and the j value range of the first-order and second-order differences is [-20,20]).

[0049] Then for the eigenvalue X i The chi-square statistic x is calculated as follows 2 ,in Ojk is the number of samples with value j and label k, that is, the actual frequency, E jk is the expected frequency, a It is a feature X i The different values ​​of b is the number of label categories.

[0050] .

[0051] Then, according to the selected significance level (generally 0.05) and degrees of freedom, the corresponding chi-square critical value χ is found in the table. critical 2 . Keep the calculated chi-square value (x 2 ) is greater than the critical value (χ critical 2 ), this step can screen out highly correlated features and reduce redundant features.

[0052] The bidirectional long short-term memory network (Bi-LSTM) model comprehensively captures the context information of the input feature value sequence by combining the forward and reverse LSTM networks. The network structure diagram of BiLSTM is shown in Figure 3 . First, the input time series data sequence is input into the Bi-LSTM network, and then the input sequence is processed from left to right by the forward LSTM unit to capture the forward dependency; at the same time, the input sequence is also processed from right to left by the reverse LSTM unit to capture the backward dependency. Subsequently, at each time step, the Bi-LSTM combines the output results of the forward and reverse LSTM to enhance the expression of time series data. Finally, the output combining the forward and reverse information is used for subsequent tasks.

[0053] In step 6, the structural response data under the action of random vehicles is input into the Bi-LSTM model established in step 5. The structural response of the random vehicle is analyzed using the load-response mapping relationship of the specific vehicle that the model has learned, and the load value of the random vehicle is output. The experimental results show that the model shows significant differences in identifying vehicles of different load levels: for heavy-loaded vehicles, the recognition accuracy is significantly better than that of medium-loaded and light-loaded vehicles, and the accuracy rate can reach more than 95%. Through this process, the dynamic identification of random vehicle loads on the bridge 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 High-definition cameras and image processing modules are used to capture surveillance videos of randomly moving vehicles and classify the vehicles. Special vehicles are identified and extracted from random traffic monitoring video images. Visual displacement sensors are used to obtain data on the deflection response of the bridge when the vehicle passes by. The video image signal and the data signal of the bridge deflection response are aligned and matched through the time synchronization module to establish the corresponding data set. The Bi-LSTM model is used to train the corresponding data set under special vehicles. The model is used to learn the load-response mapping relationship of special vehicles. The trained model then parses the structural response data under the action of random vehicles and outputs the load value of the random vehicle.

2. A lightweight dynamic weighing method based on bridge monitoring response as claimed in claim 1, characterized in that The following steps are involved: Step 1: A video surveillance system is arranged on the bridge deck. The video surveillance system includes a high-definition camera and an image processing module. The high-definition camera captures 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 for transporting dangerous goods. Special vehicles are further accurately subdivided according to the types of dangerous goods transported. The loads of different types of special vehicles are all known fixed values; Step 2: A high-precision visual displacement sensor is deployed at the mid-span position of the bridge superstructure. The visual displacement sensor uses an industrial camera to capture a high-resolution image of the target area deployed at the mid-span of the bridge superstructure under the action of random vehicle loads, and uses DIC technology to calculate the dynamic response data of the bridge structure, that is, the data of the bridge deflection response; Step 3: Align and match the random traffic flow monitoring video image signal obtained by the video surveillance system and the bridge structure dynamic response data signal obtained by the visual displacement sensor on the preset time axis through the time synchronization module. For the 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 total spatiotemporal correlation data set for building mapping relationships; Identify and extract special vehicles from random traffic monitoring video images, and calibrate the load of special vehicles as w i And the corresponding bridge structure dynamic response data is calibrated as x i , establish a spatiotemporal correlation sub-dataset for learning the characteristic values ​​of the knot response; Step 4: Establish a bidirectional long short-term memory network Bi-LSTM model. The spatiotemporal correlation sub-dataset established in step 3 for learning the response eigenvalues ​​is input into the model for training. Then the spatiotemporal correlation total data set used to construct the mapping relationship is input into the trained Bi-LSTM model. The specific vehicle load-response mapping relationship learned by the model is used to parse the structural response data under the action of random vehicles and output the load value of the random vehicle.

3. A lightweight dynamic weighing method based on bridge monitoring response as claimed in claim 2, characterized in that In step 1, the image processing module uses multi-view regression detection technology to realize automatic recognition and classification of vehicle information in the monitoring video. First, it roughly divides the vehicle information into small vehicles, heavy-loaded vehicles and special vehicles, and further detects the special vehicles in the rough classification. By identifying the side photo of the vehicle, it detects whether there is a dangerous goods transport sign. After confirming the existence of the sign, the content of the chemical hazardous goods vehicle sign is analyzed to identify the specific type of the sign, and the vehicle appearance and sign information are combined to realize accurate classification of special vehicles.

4. A lightweight dynamic weighing method based on bridge monitoring response as claimed in claim 2, characterized in that The process of calculating the dynamic response data of the bridge structure using DIC technology in step 2 is to use an image processing algorithm with sub-pixel accuracy 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 the dynamic response data of the bridge structure for subsequent load identification and structural performance evaluation.

5. A lightweight dynamic weighing method based on bridge monitoring response as claimed in claim 4, characterized in that The specific processing process 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 the calibration coefficient in mm / pixel. S2: Next, the target displacement video is collected by the camera and the image is converted into a grayscale image, the pixel coordinates of the target point to be measured on the measurement target are selected, and the data is preprocessed and outliers are eliminated. The preprocessing includes the processing steps of data interception and reference value extraction, that is, firstly intercepting the time history data of the bridge deflection response before and after the vehicle passes, that is, the change of the pixel coordinates of the target point to be measured over time, and taking the initial position pixel coordinates of the target point to be measured before the vehicle passes as the reference value; the outlier elimination adopts the interquartile range method to eliminate the extreme deviation data in the original number; S3: The positional relationship of the target point on the image is calculated using the cross-correlation function, and the pixel coordinate difference between the target point to be measured and its initial position when the vehicle passes by is calculated, which is the pixel displacement, in units of 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 to be measured is obtained, and the absolute maximum value of the pixel displacement of the target point to be measured that deviates from its initial position is taken as the extracted feature value; S4: Finally, the extracted pixel displacement characteristic value of the target point to be measured is multiplied by the calibration coefficient and converted into the actual physical displacement, that is, the displacement time history of the target point to be measured is obtained.

6. A lightweight dynamic weighing method based on bridge monitoring response as claimed in claim 5, characterized in that In step S3, an image processing algorithm with sub-pixel level precision is used to search the target point twice, namely, an integer pixel search and a sub-pixel search. The integer pixel search searches for the integer pixel position of the target point in the target image, that is, the integer pixel coordinates of the target point are obtained; the sub-pixel search is performed on the basis of the integer pixel search to obtain the precise position of the pixel point. The obtained pixel coordinates of the target point are decimals. This process improves the accuracy of the monitoring data.

7. A lightweight dynamic weighing method based on bridge monitoring response as claimed in claim 2, characterized in that In step 4, before the Bi-LSTM model trains the data set, it also includes the step of feature extraction. The specific steps are as follows: M1: First, extract the Mel frequency cepstral coefficient features in the main beam deflection monitoring response signal when different vehicles pass by, and first denoise the signal to remove high-frequency noise; Subsequently, the signal is divided into frames and smoothed by the Hanning window function, and then the signal is converted from the time domain to the frequency domain by the fast Fourier transform FFT; then, the spectrum is converted into the Mel frequency scale by the Mel filter bank, and the energy of each Mel filter channel is calculated to obtain the Mel spectrum; finally, the Mel spectrum is converted into the Mel frequency cepstrum coefficients MFCCs by discrete cosine transform DCT, and the first m MFCCs coefficients are selected as feature vectors. The first and second order differences of each selected MFCCs coefficient are further calculated to enhance the dynamic change information, and the MFCCs, first order difference and second order difference features are concatenated to form the final feature vector X i ,These eigenvectors 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 the Mel frequency scale using the Mel filter bank. The conversion formula is as follows: ； in, f is the linear frequency, M(f) is the Mel frequency; The formula for DCT transformation is as follows: ； in C k are the 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, combined with the chi-square test, the features that are highly relevant to the classification task are screened out. The specific process is as follows: First, the MFCCs feature values ​​are classified according to the labels of vehicle type, load level, driving speed and number of axles, and the expected frequency is calculated according to the following formula: ； in, O j MFCCs and first-order and second-order differences, a total of 3m features X i The total number of samples with the value j, m is the number of selected MFCCs coefficients, and the total number of selected MFCCs coefficients and their first-order difference and second-order difference features is 3m. O k Label k The total number of samples, O is the total number of samples; E jk is the expected frequency of samples with value j and label k; Then for the eigenvalue X i The chi-square statistic x is calculated as follows 2 ,in O jk is the number of samples with value j and label k, that is, the actual frequency, E jk is the expected frequency, a It is a feature X i The different values ​​of b is the number of label categories; ； Then, according to the selected significance level and degrees of freedom, query the chi-square distribution table to find the corresponding chi-square critical value χ critical 2 ; Keep the calculated chi-square value x 2 Greater than the critical value χ critical 2 This step can screen out highly correlated features and reduce redundant features.

Images