Steel bridge weld joint multi-axial fatigue damage evaluation method and device considering welding residual stress

Abstract

The invention provides a multi-axial fatigue damage evaluation method and device for a steel bridge welding seam considering welding residual stress, and relates to the technical field of fatigue strength analysis and service life prediction of a steel bridge structure. The method comprises the following steps: measuring three-dimensional residual stress in a geometric transition area of a steel bridge welding seam by adopting X rays, acquiring heat distribution and surface topography data in combination with infrared thermal imaging and structured light scanning, and realizing multi-source data space alignment based on a unified coordinate system; secondly, random disturbance is applied to the residual stress, stress fluctuation in service is simulated, and a robustness weight risk factor is constructed; the method comprises the following steps: constructing a bimodal deep neural network, adaptively fusing information of an infrared thermal imaging image and a welding seam three-dimensional topography image through an attention mechanism, training a fatigue risk regression model by taking a robustness weight risk factor as a supervision signal, completing prediction of risk factors at all positions, dividing low, medium and high risk levels based on a preset threshold value, and obtaining a fatigue risk prediction result. And accurate grading evaluation of the fatigue damage of the welding seam is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of fatigue strength analysis and life prediction of steel bridge structures, and in particular to a multi-axial fatigue damage assessment method and device for steel bridge welds taking welding residual stress into consideration. Background Art

[0002] Currently, methods based on nominal stress, hot spot stress, or local stress / strain are commonly used to evaluate the fatigue performance of Q345 steel in bridge welds. However, under multiaxial stress conditions, especially when complex residual stress distributions and geometric discontinuities exist at weld toes and grooves, traditional methods often fail to accurately reflect the fatigue damage risk under actual service conditions. This is particularly true during the service life of steel bridges, where welding residual stress, a key driver of fatigue crack initiation, exhibits significant spatial heterogeneity and uncertainty.

[0003] Existing assessment methods typically treat residual stress as a deterministic input, ignoring the effects of measurement error disturbances, local material heterogeneity, and stress relaxation after service loading on fatigue response. This single-point deterministic value assumption can easily lead to large fluctuations and error sensitivity in the assessment results of risk factors in actual structures, thereby affecting the accurate prediction of fatigue life and the reliability of risk grading. In addition, due to the significant multiaxial stress state and dynamic load effects in the weld area, current fatigue damage assessment methods still have applicability limitations in integrating multiaxial fatigue criteria with complex residual stress fields, making it difficult to cover the fatigue vulnerability identification needs under all working conditions.

[0004] Therefore, there is an urgent need for a multi-axial fatigue assessment method that can fully consider the spatial variation characteristics of welding residual stress and its uncertainty influence, so as to improve the fatigue risk identification accuracy of steel bridge weld areas and the robustness of engineering applications.

[0005] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not form the prior art that is already known to a person of ordinary skill in the art. Summary of the Invention

[0006] The object of the present invention is to provide a multi-axial fatigue damage assessment method and device for steel bridge welds taking into account welding residual stress, so as to solve the problems raised in the above background technology.

[0007] To achieve the above object, the present invention provides the following technical solutions: The multiaxial fatigue damage assessment method for steel bridge welds considering welding residual stress includes the following steps: Step 1: Measure the three-dimensional residual stress in the geometric transition area of the steel bridge weld using X-rays to generate infrared thermal images and three-dimensional topography of the weld surface, and perform normalization processing. The geometric transition area includes the weld toe tip, the center point of the groove bottom, and the base metal reference point. Step 2: Apply zero-mean Gaussian perturbation to the residual stress, construct a perturbation sample set through multiple rounds of Monte Carlo calculations, perform multiple rounds of risk factor calculations on the perturbation samples using the Dang Van criterion, and construct a robust weighted risk factor based on the risk factor; Step 3: Using the infrared thermal image and the weld 3D topography as dual-modal input, a deep convolutional neural network is used to extract the weld thermal distribution characteristics and geometric topography characteristics respectively. The shared attention mechanism is used to fuse the features. The robustness weighted risk factor of the geometric transition region is used as the supervisory signal to construct a fatigue risk nonlinear regression model. Step 4: In the weld area, infrared thermal imaging images and weld surface three-dimensional topography images are obtained through image window sliding sampling as input to the fatigue risk nonlinear regression model. The risk factors of each window position are output to classify the fatigue damage in the weld area.

[0008] Furthermore, the method for measuring the three-dimensional residual stress in the geometric transition area of the steel bridge weld area by X-ray is as follows: With the intersection of the weld section geometric centerline and the parent material as the origin O, the longitudinal direction of the weld as the x-axis, the direction perpendicular to the plate thickness as the y-axis, and the z-axis perpendicular to the weld cross section as the z-axis, a spatial rectangular coordinate system is established to define the geometric transition area: the weld toe tip point is within 0.5mm of the geometric tip of the fusion line between the weld and the parent material, the bottom center point of the groove is within 0.3mm of the weld root, that is, the lowest point of contact between the fusion zone and the parent material, and the parent material surface 30mm away from the weld edge along the z-axis is the parent material reference point; The measurement point grid at the weld toe tip is divided into: within 0.5mm from the weld toe tip, a 5×5 equally spaced grid is used, where the point spacing is 0.5mm, and there are 25 measurement points in total; the measurement point grid at the center point of the groove bottom is divided into: with the lowest point of contact between the fusion zone and the base material as the center, a 3×3 equally spaced grid is used, where the point spacing is 0.3mm, and there are 9 measurement points in total; the measurement point grid at the base material reference point is divided into: in this area, a 2×3 equally spaced grid is used, where the point spacing is 5mm, and there are 6 measurement points in total; an X-ray diffractometer is used to measure the three-dimensional residual stress at each measurement point to obtain the normal stress and shear stress in MPa. Diffraction data at three offset angles of -20°, 0°, and +20° are collected at each measurement point to construct the stress tensor matrix: ; Where, represents the stress tensor matrix, represents the normal stress along the x-axis, represents the normal stress along the y-axis, represents the normal stress along the z-axis, represents the shear stress in the xy plane, represents the shear stress in the yz plane, represents the shear stress in the xz plane.

[0009] Furthermore, the method for generating infrared thermal imaging images and weld surface three-dimensional topography images and performing normalization processing is as follows: The preset heating temperature is ,and After heating the steel bridge weld to the preset heating temperature, the spatial resolution of the high-resolution infrared thermal imager is set to 0.1mm. A vertical scan along the z-axis is performed at a distance of 0.3m from the weld surface to collect infrared radiation data in the 7.5-13μm band. A 640×512 pixel infrared thermal image is generated. The coordinate axis corresponds to the physical spatial position of the weld area, and the pixel value corresponds to the surface temperature. The temperature is normalized and the output is an infrared thermal image in the form of a two-dimensional heat distribution matrix. The formula for normalizing the temperature is as follows: ; Where, Indicates the surface temperature, represents the normalized surface temperature; A structured light 3D scanner is used to scan the weld surface. The scanner is translated perpendicular to the weld surface along the z-axis. The scanning range covers the entire length of the weld. The scanning resolution is set to 0.1 mm / point to obtain the weld surface point cloud data and convert it into a 640×512 pixel weld 3D topography map. The pixel coordinates correspond to the actual spatial coordinates of the weld. The x-axis and z-axis in the coordinate system are set to correspond to the row and column directions of the weld 3D topography matrix respectively. The grid is divided according to the spatial resolution of 0.1 mm to construct the coordinate system of the weld 3D topography map. Then, each pixel point in the weld 3D topography map is The corresponding physical space coordinates are: , and the coordinate system uses the same resolution and alignment as the infrared thermal image, so that the pixel positions of the image in space can correspond one to one, and the height of each coordinate is normalized: Where, represents the normalized height, Indicates the weld surface height value, Indicates the maximum value of all weld surface height values, Indicates the minimum value of all weld surface height values.

[0010] Furthermore, a zero-mean Gaussian perturbation is applied to the residual stress, and the method for constructing a perturbation sample set through multiple rounds of Monte Carlo calculation is as follows: A zero-mean Gaussian perturbation is introduced into the stress tensor of the residual stress at each measuring point to generate a perturbation tensor: ; Where, represents the perturbation tensor, represents the perturbation matrix, where the elements of the perturbation matrix are generated by the following method: Generate standard normal random variables using the Mersenne Twister algorithm , generate the perturbation matrix elements: ; Where, It represents the disturbance intensity coefficient, which is 5%-10% of the yield strength of the steel bridge material. Indicates the The perturbation matrix elements generated by the simulation, represents the elements in the stress tensor matrix, represents the row index of the stress tensor matrix, represents the column index of the stress tensor matrix, represents the row index of the perturbation matrix, Represents the column index of the perturbation matrix, which is generated repeatedly for each measurement point Group perturbation tensors to form a set of perturbation samples ,in, is a positive integer greater than 100.

[0011] Furthermore, the Dang Van criterion is used to perform multiple rounds of risk factor calculations on the disturbed samples, and the method of constructing a robust weighted risk factor based on the risk factor is as follows: For each set of perturbation tensors in the perturbation sample set, after applying a sinusoidal external load along the x-axis, the Dang Van criterion is used to calculate risk factors, where the risk factor obtained by the k-th simulation is expressed as , and sort in ascending order, take the Robust weighted risk factors .

[0012] Furthermore, the method of extracting the thermal distribution characteristics and geometric morphology characteristics of the weld seam respectively by using the infrared thermal imaging image and the weld seam 3D morphology image as dual-modal input is as follows: A 64×64 pixel sliding window with a sliding step of 16 pixels is set. The window slides 40 times along the x-axis and 32 times along the z-axis. Zero values are used to fill the edge window to complete it. A total of 1280 windows are generated. For each window, the physical space coordinates corresponding to its center pixel are extracted and it is determined whether it is located in the geometric transition area of the weld area. Only windows whose center points fall into the geometric transition area are retained as valid training samples. For each valid sample, the nearest neighbor interpolation method is used to find the robustness weight risk factor corresponding to the nearest measurement point based on the physical coordinates of the window center, and use it as the supervision label value of the window. , then the 64×64 pixel matrix intercepted in the infrared thermal imaging image is recorded as , the 64×64 pixel matrix of the weld 3D topography is recorded as ; A dual-branch parallel convolutional neural network is used to extract features from infrared thermal images and three-dimensional topography images respectively. The specific structure is as follows: Input layer: receives a single-channel image of 64×64×1; the convolution layer of the first convolution block: 32 3×3 convolution kernels, padding mode "same", activation function ReLU, batch normalization layer is used to standardize the channel mean and variance, the maximum pooling layer uses a 2×2 pooling kernel, stride 2, and outputs a 32×32×32 feature map; the convolution layer of the second convolution block: 64 3×3 convolution kernels, padding mode "same", activation function ReLU, batch normalization layer is used to standardize the channel mean and variance, the maximum pooling layer uses a 2×2 pooling kernel, stride 2, and outputs a 16×16×64 feature map; flattening layer: flattens the three-dimensional feature map into a 16384-dimensional feature vector.

[0013] Furthermore, the shared attention mechanism is used to fuse features and the robust weighted risk factor of the geometric transition region is used as the supervisory signal to construct a fatigue risk nonlinear regression model: Shared attention mechanism, concatenating the 16384-dimensional vector of the two branches into 32768 dimensions, fully connected layer 1: 1024 neurons, activation function ReLU; fully connected layer 2: 512 neurons, activation function ReLU; fully connected layer 3: 2 neurons, activation function Sigmoid, output thermal weight and shape weight ,and ; A regression model is constructed using a fully connected network and a Dropout layer to output the predicted risk factors. The steps for constructing a regression model using a fully connected network and a Dropout layer are as follows: Set the fully connected layer 1 to 2048 neurons and the activation function ReLU; Dropout layer: dropout rate 25%; fully connected layer 2 to 1024 neurons and the activation function ReLU; Dropout layer: dropout rate 25%; fully connected layer 3 to 1 neuron and linear activation function; The training learning rate is , the optimizer is Adam, the training rounds are 200 rounds, the Dropout is 0.25, and the batch size is 32, that is, 32 samples are randomly selected from the valid training samples. The main loss function in the current batch is: ; Where, represents the main loss function, express The output predicted risk factors for samples, Indicates the The supervised label value of each sample, Indicates the serial number index of 32 randomly selected samples, .

[0014] Furthermore, the risk factor of each window position is output, and the fatigue damage of the weld area is graded as follows: For steel bridge welds that require damage assessment, extract the risk factors output by each sliding window and construct The two-dimensional risk matrix has a one-to-one correspondence with the physical coordinates of the weld, and the risk factor threshold is pre-set and ,in, , for each risk factor output by the sliding window , when satisfied When the damage area is judged as low risk, When the damage area is determined to be medium risk, When the damage is detected, it is judged as a high-risk damage area, where Indicates the serial index of the sliding window, .

[0015] In addition, a multi-axial fatigue damage assessment device for steel bridge welds taking into account welding residual stress is provided, which is characterized in that: the device is used to perform the above-mentioned multi-axial fatigue damage assessment method for steel bridge welds taking into account welding residual stress, comprising: An image generation module is used to measure the three-dimensional residual stress of the geometric transition area of the steel bridge weld area using X-rays, generate infrared thermal images and three-dimensional topography images of the weld surface, and perform normalization processing. The geometric transition area includes the weld toe tip, the center point of the groove bottom, and the base metal reference point; The robust risk factor module is used to apply zero-mean Gaussian perturbations to residual stresses, construct a perturbation sample set through multiple rounds of Monte Carlo calculations, perform multiple rounds of risk factor calculations on the perturbation samples using the Dang Van criterion, and construct robust weighted risk factors based on the risk factors. The fatigue risk regression module uses infrared thermal imaging and weld 3D topography as dual-modal inputs, extracts weld thermal distribution characteristics and geometric topography features through a deep convolutional neural network, fuses these features through a shared attention mechanism, and uses the robustness weighted risk factor of the geometric transition region as a supervisory signal to construct a fatigue risk nonlinear regression model. The weld risk grading module is used to obtain infrared thermal images and three-dimensional topography images of the weld surface through image window sliding sampling in the weld area as inputs of the fatigue risk nonlinear regression model, output the risk factors of each window position, and grade the fatigue damage in the weld area.

[0016] Compared with the prior art, the present invention has the following beneficial effects: The present invention uses X-ray diffraction to measure the three-dimensional residual stress in the geometric transition area of the steel bridge weld, and dynamically modifies the input parameters of the Dang Van criterion through Gaussian perturbation and multiple rounds of Monte Carlo simulation to calculate the robustness weight risk factor. It is used as the supervision signal of the deep convolutional neural network to construct a fatigue risk nonlinear regression model for the thermal distribution characteristics and geometric morphology characteristics, avoiding the fatigue damage prediction deviation caused by the traditional Dang Van criterion based on the stable load assumption, and also making up for the limitations of single stress or geometric data in fatigue identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 Schematic diagram of the overall method flow of the present invention; Figure 2 This is a fitting diagram of disturbance, normal stress and risk factor of the present invention; Figure 3 This is the D parameter and risk factor fitting diagram of the present invention; Figure 4 This is a risk factor prediction and assessment diagram of the present invention; Figure 5 This is a schematic diagram of the prediction accuracy of the present invention; Figure 6 This is the morphology weight output map of the present invention; Figure 7 This is the thermal weight output diagram of the present invention; Figure 8 It is a schematic diagram of the overall structure of the device of the present invention. DETAILED DESCRIPTION

[0018] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to specific embodiments.

[0019] It should be noted that, unless otherwise defined, the technical or scientific terms used in the present invention should have the usual meanings understood by people with ordinary skills in the field to which the present invention belongs. The "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative position relationships. When the absolute position of the object being described changes, the relative position relationship may also change accordingly.

[0020] Example: See also Figures 1 to 7 , the present invention provides a technical solution: The multiaxial fatigue damage assessment method for steel bridge welds considering welding residual stress includes the following steps: Step 1: Measure the three-dimensional residual stress in the geometric transition area of the steel bridge weld using X-rays to generate infrared thermal images and three-dimensional topography of the weld surface, and perform normalization processing. The geometric transition area includes the weld toe tip, the center point of the groove bottom, and the base metal reference point. With the intersection of the weld section geometric centerline and the parent material as the origin O, the longitudinal direction of the weld as the x-axis, the direction perpendicular to the plate thickness as the y-axis, and the vertical weld cross section as the z-axis, a spatial rectangular coordinate system is established, so that data from different sources such as residual stress measured by X-rays, infrared thermal imaging, and three-dimensional morphology scanning can be aligned based on the same coordinate system, thereby assessing fatigue damage of steel structures and defining a geometric transition area: the 0.5mm range of the geometric tip of the fusion line between the weld and the parent material is taken as the weld toe tip point. The weld toe is the geometric transition point between the weld and the parent material, with a significant shape mutation, and is the location where fatigue cracks are most likely to initiate. Studies have shown that the area within 0.5mm from the weld toe tip is affected by the stress concentration system. The number is high and the residual stress gradient is large. Its stress state has a significant impact on fatigue life. The root of the weld, that is, the lowest point of contact between the fusion zone and the base material, within 0.3mm is used as the center point of the groove bottom. The weld root is prone to defects such as lack of fusion and slag inclusion during welding, and the geometric shape, such as blunt edges and root gaps, leads to stress concentration. The 0.3mm range can better capture the stress mutation corresponding to the tiny geometric defects at the root. The base material surface 30mm away from the weld edge along the z-axis is used as the base material reference point. This is because engineering experience generally believes that the stress state of the base material 20 to 50mm away from the weld edge is close to the original state, which is a better effect for comparing stress anomalies in the weld area. The measurement point grid at the weld toe tip is divided into a 5×5 equally spaced grid with a point spacing of 0.5mm within 0.5mm of the weld toe tip, for a total of 25 measurement points. The geometric mutation at the weld toe leads to a high stress concentration factor, which can reach 3-5. The 0.5mm point spacing can capture 2-3 stress gradient changes within a single grid, avoiding distortion of stress characteristics and leading to optimistic fatigue life assessment. The measuring point grid at the center point of the groove bottom is divided into a 3×3 equally spaced grid with a point spacing of 0.3mm, centered at the lowest point where the fusion zone contacts the base metal. There are 9 measuring points in total. The stress concentration factor at the root is approximately 2-3, and the stress gradient is lower than that at the weld toe. However, due to the possible presence of a lack of fusion defect, a 0.3mm point spacing is required to capture the stress mutation within a 0.5mm range around the defect. Divide the measuring point grid of the parent metal reference point: a 2×3 equally spaced grid is used in this area, where the point spacing is 5mm, with a total of 6 measuring points. The residual stress at 30mm away from the weld has decayed to a very small value. The 5mm point spacing is sufficient to characterize the uniform stress field and reduce redundant data. Use an X-ray diffractometer, such as Rigaku SmartLab, to measure the three-dimensional residual stress at each measuring point. , the spot size is set to The tube voltage was set to 40 kV, the tube current was 100 mA, and the exposure time was 30 s per angle. Normal stress and shear stress were obtained in MPa. The scanning mode of each measuring point was azimuth angles φ = 0°, 45°, and 90°. Nine sets of diffraction patterns were collected at tilt angles ψ = -20°, 0°, and +20°. The diffraction patterns were imported into the PDXL3 software. The stress tensor matrix was output through the three-dimensional stress tensor mode of the stress calculation wizard: ; Where, represents the stress tensor matrix, represents the normal stress along the x-axis, represents the normal stress along the y-axis, represents the normal stress along the z-axis, represents the shear stress in the xy plane, represents the shear stress in the yz plane, represents the shear stress in the xz plane.

[0021] The preset heating temperature is ,and , by stimulating the thermal response of the weld area through low-temperature heating, the internal stress state of the material is coupled with different temperatures, providing a temperature difference basis for subsequent infrared thermal imaging detection, while avoiding high temperature damage to the microstructure of the steel bridge material. The recrystallization temperature of Q345 steel is about 600℃, and 40-60℃ is far below its phase transition temperature, which will not cause a decrease in material strength or lattice distortion. Engineering experience shows that the thermal expansion effect of steel bridges in this temperature range is significant, which makes it easy to capture small temperature differences, thereby avoiding the decrease in weld hardness or residual stress release caused by high-temperature heating; after heating the steel bridge weld to the preset heating temperature, the spatial resolution of the high-resolution infrared thermal imager is set to 0.1mm, and it is scanned vertically along the z-axis at 0.3m from the weld surface to collect infrared radiation data in the 7.5-13μm band. , where the spatial resolution of the high-resolution infrared thermal imager is set to 0.1mm in order to make the temperature data correspond to the subsequent three-dimensional morphology data in physical space. The diffraction effect of infrared light in the 7.5-13μm band is weak, and sub-millimeter resolution can be achieved at a distance of 0.3m, which meets the needs of detecting micro defects in welds, such as microcracks and temperature anomalies in stress concentration areas. A 640×512 pixel infrared thermal image is generated, where the coordinate axis corresponds to the physical spatial position of the weld area, and the pixel value corresponds to the surface temperature. The temperature is normalized to map the temperature data to the [0,1] interval, eliminating the influence of the temperature dimension, facilitating the comparison of multiple sets of data and the input of machine learning models. The output is an infrared thermal image in the form of a two-dimensional heat distribution matrix, where the formula for normalizing the temperature is: ; Where, Indicates the surface temperature, represents the normalized surface temperature; A structured light 3D scanner is used to scan the weld surface. The scanner moves perpendicular to the weld surface along the z-axis, and the scanning range covers the entire length of the weld. The scanning resolution is set to 0.1mm / point. The weld height of Q345 steel bridge is usually ≤3mm, and surface defects such as undercut depth will be higher than 0.5mm. The 0.1mm resolution can capture the defect size and meet the engineering detection accuracy requirements. The weld surface point cloud data is obtained and converted into a 640×512 pixel weld 3D topography map. The pixel coordinates correspond to the actual spatial coordinates of the weld one-to-one, so that the temperature and height data of the same physical position can be analyzed synchronously. The x-axis and z-axis in the coordinate system are set to correspond to the row and column directions of the weld 3D topography matrix respectively. The grid is divided according to the spatial resolution of 0.1mm to construct the coordinate system of the weld 3D topography map. Then, each pixel point in the weld 3D topography map The corresponding physical space coordinates are: , and the coordinate system uses the same resolution and alignment as the infrared thermal image, so that the pixel positions of the image in space can correspond one to one, and the height of each coordinate is normalized: Where, represents the normalized height, Indicates the weld surface height value, Indicates the maximum value of all weld surface height values, It represents the minimum value of all weld surface height values. The overall height of different welds varies greatly. After normalization, the height data is unified into the interval [0,1] to eliminate the influence of temperature dimension, which is convenient for comparison of multiple groups of data and input of machine learning models.

[0022] Step 2: Apply zero-mean Gaussian perturbation to the residual stress, construct a perturbation sample set through multiple rounds of Monte Carlo calculations, perform multiple rounds of risk factor calculations on the perturbation samples using the Dang Van criterion, and construct a robust weighted risk factor based on the risk factor; The residual stress of steel bridge welds is affected by measurement error, material discreteness, and service disturbances (vehicle load fluctuations). From historical data or experimental simulations, static experimental simulations usually use multiple measurements at a single point, which cannot reproduce the three-dimensional stress field non-uniformity along the entire length of the weld. Traditional experiments only use fixed-amplitude sinusoidal loading, lacking random processes. Zero-mean Gaussian perturbations are introduced into the stress tensor of the residual stress at each measuring point. Through zero-mean Gaussian perturbations, the above uncertainties can be converted into random corrections to the stress tensor, turning fatigue risk assessment into a probabilistic analysis and avoiding risk misjudgments caused by ignoring non-systematic deviations. The formula for generating the perturbation tensor is as follows: ; Where, represents the perturbation tensor, represents the perturbation matrix, where the elements of the perturbation matrix are generated by the following method: By using MATLAB's randn function, which has a built-in Mersenne Twister algorithm, an industrial-grade pseudo-random number algorithm, it avoids simulation bias caused by random number repetition, thereby generating a standard normal random variable , the formula for generating the perturbation matrix elements is: ; Where, It represents the disturbance intensity coefficient. Based on engineering experience, the value is 5%-10% of the yield strength of the steel bridge material. For example, the yield strength of Q345 steel bridge material is 345MPa. The value is in the range of 17.25-34.5MPa, set The value is 20MPa, Indicates the The perturbation matrix elements generated by the simulation, represents the row index of the perturbation matrix, represents the column index of the perturbation matrix, for example, if , ,but , then by For the perturbation tensor synthesis, for each measurement point, repeatedly generate Group perturbation tensors to form a set of perturbation samples ,in, is a positive integer greater than 100.

[0023] For each set of perturbation tensors in the perturbation sample set, after applying a sinusoidal external load along the x-axis, the Dang Van criterion is used to calculate A risk factor is taken. Based on material mechanics, 35% of the yield strength of Q345 steel is taken as the external load amplitude, which can keep the principal stress of the composite stress within the elastic range and simulate the alternating stress caused by vehicle loads. That is, the external load amplitude is 120 MPa. Statistics from the bridge dynamic weighing system show that the longitudinal stress amplitude of the typical steel bridge deck weld is 80-150 MPa, 80 MPa corresponding to the passage of light vehicles and 150 MPa corresponding to the overloaded heavy truck condition. The middle value of 120 MPa can cover 85% of conventional traffic loads. The load frequency is set as 0.5 Hz, corresponding to the first-order vertical vibration frequency of the bridge. The time history length is 0-4 s, corresponding to 2 complete cycles, and the time step is 0.02 s, which satisfies the Nyquist sampling theorem and captures the load peak. The sinusoidal external load along the x-axis is expressed as a three-dimensional stress tensor: ; Where, represents the time-varying external load stress tensor, Indicates the external load amplitude, It represents the frequency of the sinusoidal external load, and its value is 0.5Hz. The first-order vertical vibration frequency of typical steel bridges in engineering measurement data is mostly in the range of 0.1-1.0Hz. 0.5Hz is the middle value of this range, which can cover the typical working conditions of most steel bridges and is the typical frequency of the first-order vertical vibration of steel bridges. For example, hour, , at this time the three-dimensional stress tensor is 0, hour, , at this time the three-dimensional stress tensor is 120 MPa, where for the k-th perturbation tensor, the resultant stress at time t is: ; Where, represents the resultant stress at time t, represents the k-th perturbation tensor; for example, ,when When , the resultant stress is: ; right Solve the characteristic equation: ; Where, represents the resultant stress, represents the identity matrix, that is , we get three principal stresses , the three principal stresses reflect the most dangerous stress state after the residual stress is coupled with the external load; Calculate the mean hydrostatic pressure: ; Where, It represents the average hydrostatic pressure, which is essentially the linear weighted average of the principal stresses. It converts the complex three-dimensional stress state into a single quantitative index and characterizes the average volume stress state of the material microelement under the three-dimensional principal stress field. The larger this value is, The value of increases accordingly. In practice, when the average hydrostatic pressure is greater than 0, dislocation movement and crack extension will be promoted. The increase of any principal stress will lead to an increase in hydrostatic pressure. The three are positively correlated with the average hydrostatic pressure. Calculate the maximum shear stress: ; Where, Indicates the maximum shear stress, which is directly determined by the difference between the maximum and minimum principal stresses, and reflects the maximum shear failure trend that may occur inside the material. In material mechanics, when an object is subjected to a complex stress state, shear stress will cause the material to slip or shear fracture. Its physical meaning is one of the critical indicators of the material's resistance to shear failure. The value of the maximum shear stress increases with The difference between Positively correlated with Inversely correlated; Q345 steel is fitted through fatigue test, and the standard recommended value of hydrostatic pressure sensitivity coefficient is 0.15-0.2. The hydrostatic pressure sensitivity coefficient is set to , after a large number of experiments, the shear fatigue limit value is recommended , calculate the Dang Van parameter: ; Where, Represents Dang Van parameters; Calculate risk factors: ; Where, The risk factor is expressed as: , and sorted in ascending order. For example, after sorting the principal stress of a certain instantaneous stress in ascending order, , , , calculate the mean hydrostatic pressure: ; Calculate the maximum shear stress: ; Calculate Dang Van parameters: ; Calculate risk factors: ; Then, for the risk factors sorted in ascending order, take the Robust weighted risk factors ,In the random perturbation simulation, 95% of the hazard factor values are less than this value, and only 5% of the values may be larger, thus providing a reasonable safety margin for structural fatigue design.,Table 1 shows the hazard factor output table.,By applying zero-mean Gaussian perturbations to the stress tensor and then superimposing the principal stresses of the composite stresses obtained by sinusoidal external loads, the coupling effect of residual stress fluctuations and vehicle loads is,simulated.

[0024]

[0025] like Figure 2-Figure 3 As shown, Dang Van parameters and risk factors The fitting relationship shows a linear relationship, and the scatter points closely fit the fitting line. The coupling effect of shear stress and hydrostatic pressure can be quantified by the Dang Van parameter. After applying zero-mean Gaussian perturbation to the x-axis, as the x-axis normal stress increases, the principal stress of the composite stress and the hydrostatic pressure increase, and the Dang Van parameter also increases accordingly. The simulated vehicle alternating load is closer to the actual level, providing a data basis for the quantification and extraction of risk factors.

[0026] Step 3: Using infrared thermal images and weld 3D topography as dual-modal inputs, a deep convolutional neural network is used to extract weld thermal distribution features and geometric topography features, respectively. A shared attention mechanism is used to fuse these features, and a robust weighted risk factor of the geometric transition region is used as a supervisory signal to construct a fatigue risk nonlinear regression model. A 64×64 pixel sliding window with a sliding step of 16 pixels was set, sliding 40 times along the x-axis and 32 times along the z-axis. Zero-value padding was used to complete the edge window, generating a total of 1280 windows to capture the microscopic features of stress concentration. For each window, the physical space coordinates corresponding to its central pixel were extracted, and it was determined whether it was located in the geometric transition area of the weld area. Only windows with central points falling in the geometric transition area were retained as valid training samples. The sliding window was used to segment the continuous thermal images and topography images into local feature blocks, allowing the model to focus on subtle anomalies in high-risk areas such as weld toes and grooves, avoiding the global feature averaging that masks local defects, filtering out the majority of low-risk areas, and improving the correlation between samples and fatigue risk. For each valid sample, the nearest neighbor interpolation method is used to find the robustness weight risk factor corresponding to the nearest measurement point based on the physical coordinates of the window center, and use it as the supervision label value of the window. , solve the spatial matching problem between image pixels and discrete measurement points, and form a mapping relationship between thermal distribution, morphology and fatigue risk. The 64×64 pixel matrix intercepted in the infrared thermal imaging image is recorded as , the 64×64 pixel matrix of the weld 3D topography is recorded as A dual-branch parallel convolutional neural network is used to extract features from infrared thermal images and 3D topography images respectively. The two branches independently learn complementary features and can find correlations with fatigue risk by monitoring the temperature rise of thermoelastic effects and geometric defects in stress concentration areas. The specific structure is as follows: Input layer: receives a single-channel image of 64×64×1; the convolution layer of the first convolution block: 32 3×3 convolution kernels, padding mode "same", activation function ReLU, 3×3 convolution kernel covers 3×3 pixels to capture small defects such as microcracks, batch normalization layer is used to standardize channel mean and variance, the maximum pooling layer uses 2×2 pooling kernel, step size 2, and outputs 32×32×32 feature map; the convolution layer of the second convolution block: 64 3×3 convolution kernels, padding mode "same", activation function ReLU, batch normalization layer is used to standardize channel mean and variance, the maximum pooling layer uses 2×2 pooling kernel, step size 2, and outputs 16×16×64 feature map; flattening layer: flattens the three-dimensional feature map into a 16384-dimensional feature vector, The shared attention mechanism is used to adaptively allocate the weights of bimodal features. For example, when there is a clear temperature gradient in the thermal image, the model automatically improves , strengthen the contribution of thermal distribution to fatigue risk, concatenate the 16384-dimensional vector of the two branches into 32768 dimensions, fully connected layer 1: 1024 neurons, activation function ReLU; fully connected layer 2: 512 neurons, activation function ReLU; fully connected layer 3: 2 neurons, activation function Sigmoid, output thermal weight and shape weight ,and , through Sigmoid activation to ensure that the weight is in the range of [0,1], characterizing the importance ratio of thermal imaging and morphological features, for example, when morphological features dominate, it may be ; When the temperature anomaly characteristics dominate, it may ; A regression model is constructed using a fully connected network and a Dropout layer to output the predicted risk factors. The steps for constructing a regression model using a fully connected network and a Dropout layer are as follows: Set the fully connected layer 1 to 2048 neurons and the activation function ReLU; the dropout layer to a dropout rate of 25%; the fully connected layer 2 to 1024 neurons and the activation function ReLU; the dropout layer to a dropout rate of 25%; the fully connected layer 3 to 1 neuron and a linear activation function. The dropout layer has a dropout rate of 25% to randomly drop 25% of neurons to alleviate overfitting caused by the limited sample size in steel bridge inspection. The linear activation is selected so that the risk factor can be greater than 1. The training learning rate is The optimizer is Adam, which balances convergence speed and accuracy. The training rounds are 200 rounds to avoid overfitting. The batch size is 32 to reduce the influence of noise. That is, 32 samples are randomly selected from the valid training samples. The main loss function in the current batch is: ; Where, represents the main loss function, express The output predicted risk factors for samples, Indicates the The supervised label value of each sample, Indicates the serial number index of 32 randomly selected samples, It directly measures the deviation between the predicted risk factor and the actual measurement, is suitable for regression tasks, and is insensitive to extreme values. It is suitable for the requirement of steel bridge fatigue risk assessment that requires attention to the average error.

[0027] Step 4: In the weld area, infrared thermal images and three-dimensional topography images of the weld surface are obtained through image window sliding sampling as input to the fatigue risk nonlinear regression model. The risk factors at each window position are output and the fatigue damage of the weld area is graded. Use an electric heating plate to evenly heat the weld area to the heating temperature , using a thermal imager, set the spatial resolution to 0.1mm, scan vertically at a distance of 0.3m from the weld surface, with a wavelength of 7.5-13μm, and generate a 640×512 pixel thermal image, normalizing the temperature: ,Use a structured light scanner with a scanning resolution of 0.1mm / point, a translation speed of 5mm / s along the z-axis, covering the full length of the weld, and generating 640×512 pixel point cloud data. Taking the intersection point of the weld section centerline and the base material as the origin, ensuring that the x / z axis is consistent with the thermal image coordinate system, the pixel physical coordinate mapping is ; Set a 64×64 pixel sliding window with a sliding step of 16 pixels, slide 40 times along the x-axis and 32 times along the z-axis, and use zero value filling to make the edge window complete, generating a total of 1280 windows. For each window, extract the physical space coordinates corresponding to its central pixel point, and use the constructed dual-branch CNN model to input a 64×64×1 matrix 、 , the shared attention mechanism uses adaptive weight fusion to output thermal weights and morphology , the regression layer outputs the predicted risk factors under each window, and pre-sets the risk factor threshold and ,in, , for each risk factor output by the sliding window , when satisfied When the damage area is judged as low risk, When the damage area is determined to be medium risk, When the damage is detected, it is judged as a high-risk damage area, where Indicates the serial index of the sliding window, In the Dang Van criterion, the shear fatigue limit value , when risk factors When it exceeds 1, it indicates that the Dang Van parameter D under the current stress state has exceeded the fatigue limit of the material. At this time, the risk of fatigue crack initiation and growth increases significantly. For example, Set to 1, when When the damage is detected, it is determined to be a high-risk damage area, and local stress optimization or structural reinforcement is required immediately. In actual service of steel bridges, factors such as vehicle load and ambient temperature will cause stress amplitude fluctuations. When the safety margin of 50% is reserved, it can avoid the low-risk area being misjudged as medium-risk due to accidental factors. When the risk of damage is determined to be medium, it should be included in the key monitoring targets; when When the damage area is judged as low-risk, it can be inspected according to the regular cycle, thereby reducing maintenance costs. Table 2 shows the prediction result verification table of the dual-modal fatigue risk regression model, which is used to quantify the model's predictive ability for fatigue risk in each area. After the model is trained with Monte Carlo Gaussian perturbation, the center point of the window is located in the geometric transition area. After collecting data from 40 groups of window IDs and summarizing them into a table, the predicted risk factors are sorted from small to large. The overall error is less than 5%, with high accuracy. The model has a good predictive ability for risk factors.

[0028]

[0029] like Figure 4-Figure 7 As shown, in Figure 4 Among them, the predicted risk factors are The supervised label values of the samples are almost the same, and the learning of the heat-shape-risk mapping relationship is sufficient without overfitting. Figure 5 It can be seen that the overall error is small, especially in high-risk areas, which not only meets the accuracy requirements, but also has a higher recognition ability for areas with higher fatigue damage. Figure 6-Figure 7 It is the numerical change of thermal weight and morphology weight with the change of window ID. The color of the point in the figure indicates the risk level of the corresponding window, red for risk level 1, green for risk level 2, and blue for risk level 3. Due to stress concentration, the local temperature rise gradient is large. Between window ID 30 and 40, thermal imaging is more sensitive and the weight occupies a dominant position; while the structural induction effect of morphological features such as excess height mutation and biting edge on fatigue is more prominent. Therefore, between window ID 10 and 20, the weight of morphology occupies a dominant position, which not only has the ability to adaptively identify complex defects, but also can ensure prediction accuracy.

[0030] See also Figure 8The present invention further provides a multi-axial fatigue damage assessment device for a steel bridge weld considering welding residual stress, which is used to perform the multi-axial fatigue damage assessment method for a steel bridge weld considering welding residual stress, comprising: An image generation module is used to measure the three-dimensional residual stress of the geometric transition area of the steel bridge weld area using X-rays, generate infrared thermal images and three-dimensional topography images of the weld surface, and perform normalization processing. The geometric transition area includes the weld toe tip, the center point of the groove bottom, and the base metal reference point; The robust risk factor module is used to apply zero-mean Gaussian perturbations to residual stresses, construct a perturbation sample set through multiple rounds of Monte Carlo calculations, perform multiple rounds of risk factor calculations on the perturbation samples using the Dang Van criterion, and construct robust weighted risk factors based on the risk factors. The fatigue risk regression module uses infrared thermal imaging and weld 3D topography as dual-modal inputs, extracts weld thermal distribution characteristics and geometric topography features through a deep convolutional neural network, fuses these features through a shared attention mechanism, and uses the robustness weighted risk factor of the geometric transition region as a supervisory signal to construct a fatigue risk nonlinear regression model. The weld risk grading module is used to obtain infrared thermal images and three-dimensional topography images of the weld surface through image window sliding sampling in the weld area as inputs of the fatigue risk nonlinear regression model, output the risk factors of each window position, and grade the fatigue damage in the weld area.

[0031] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters in the formulas are set by technicians in this field according to actual conditions.

[0032] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed by hardware or software depends on the specific application and design constraints of the technical solution.

[0033] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, and may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment as needed.

[0034] The above is only a specific implementation method of the present application, but the scope of protection of the present application is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed in this application, which should be covered by the scope of protection of the present application.

Claims

1. A multiaxial fatigue damage assessment method for steel bridge welds considering welding residual stress, characterized in that: The specific steps include: Step 1: Measure the three-dimensional residual stress in the geometric transition area of the steel bridge weld using X-rays to generate infrared thermal images and three-dimensional topography of the weld surface, and perform normalization processing. The geometric transition area includes the weld toe tip, the center point of the groove bottom, and the base metal reference point. Step 2: Apply zero-mean Gaussian perturbation to the residual stress, construct a perturbation sample set through multiple rounds of Monte Carlo calculations, perform multiple rounds of risk factor calculations on the perturbation samples using the Dang Van criterion, and construct a robust weighted risk factor based on the risk factor; Step 3: Using the infrared thermal image and the weld 3D topography as dual-modal input, a deep convolutional neural network is used to extract the weld thermal distribution characteristics and geometric topography characteristics respectively. The shared attention mechanism is used to fuse the features. The robustness weighted risk factor of the geometric transition region is used as the supervisory signal to construct a fatigue risk nonlinear regression model. Step 4: In the weld area, infrared thermal imaging images and weld surface three-dimensional topography images are obtained through image window sliding sampling as input to the fatigue risk nonlinear regression model. The risk factors of each window position are output to classify the fatigue damage in the weld area.

2. The multiaxial fatigue damage assessment method for steel bridge welds considering welding residual stress according to claim 1, characterized in that: The method for measuring the three-dimensional residual stress in the geometric transition area of the steel bridge weld area by X-ray is as follows: With the intersection of the weld section geometric centerline and the parent material as the origin O, the longitudinal direction of the weld as the x-axis, the direction perpendicular to the plate thickness as the y-axis, and the z-axis perpendicular to the weld cross section as the z-axis, a spatial rectangular coordinate system is established to define the geometric transition area: the weld toe tip point is within 0.5mm of the geometric tip of the fusion line between the weld and the parent material, the bottom center point of the groove is within 0.3mm of the weld root, that is, the lowest point of contact between the fusion zone and the parent material, and the parent material surface 30mm away from the weld edge along the z-axis is the parent material reference point; The measurement point grid at the weld toe tip is divided into: within 0.5mm from the weld toe tip, a 5×5 equally spaced grid is used, where the point spacing is 0.5mm, and there are 25 measurement points in total; the measurement point grid at the center point of the groove bottom is divided into: with the lowest point of contact between the fusion zone and the base material as the center, a 3×3 equally spaced grid is used, where the point spacing is 0.3mm, and there are 9 measurement points in total; the measurement point grid at the base material reference point is divided into: in this area, a 2×3 equally spaced grid is used, where the point spacing is 5mm, and there are 6 measurement points in total; an X-ray diffractometer is used to measure the three-dimensional residual stress at each measurement point to obtain the normal stress and shear stress in MPa. Diffraction data at three offset angles of -20°, 0°, and +20° are collected at each measurement point to construct the stress tensor matrix: ； Where, represents the stress tensor matrix, represents the normal stress along the x-axis, represents the normal stress along the y-axis, represents the normal stress along the z-axis, represents the shear stress in the xy plane, represents the shear stress in the yz plane, represents the shear stress in the xz plane.

3. The multiaxial fatigue damage assessment method for steel bridge welds considering welding residual stress according to claim 1, characterized in that: The method for generating infrared thermal imaging images and weld surface three-dimensional topography images and performing normalization processing is as follows: The preset heating temperature is ,and After heating the steel bridge weld to the preset heating temperature, the spatial resolution of the high-resolution infrared thermal imager is set to 0.1mm. A vertical scan along the z-axis is performed at a distance of 0.3m from the weld surface to collect infrared radiation data in the 7.5-13μm band. A 640×512 pixel infrared thermal image is generated. The coordinate axis corresponds to the physical spatial position of the weld area, and the pixel value corresponds to the surface temperature. The temperature is normalized and the output is an infrared thermal image in the form of a two-dimensional heat distribution matrix. The formula for normalizing the temperature is as follows: ； Where, Indicates the surface temperature, represents the normalized surface temperature; A structured light 3D scanner is used to scan the weld surface. The scanner is translated perpendicular to the weld surface along the z-axis. The scanning range covers the entire length of the weld. The scanning resolution is set to 0.1 mm / point to obtain the weld surface point cloud data and convert it into a 640×512 pixel weld 3D topography map. The pixel coordinates correspond to the actual spatial coordinates of the weld. The x-axis and z-axis in the coordinate system are set to correspond to the row and column directions of the weld 3D topography matrix respectively. The grid is divided according to the spatial resolution of 0.1 mm to construct the coordinate system of the weld 3D topography map. Then, each pixel point in the weld 3D topography map is The corresponding physical space coordinates are: , and the coordinate system uses the same resolution and alignment as the infrared thermal image, so that the pixel positions of the image in space can correspond one to one, and the height of each coordinate is normalized: Where, represents the normalized height, Indicates the weld surface height value, Indicates the maximum value of all weld surface height values, Indicates the minimum value of all weld surface height values.

4. The multiaxial fatigue damage assessment method for steel bridge welds considering welding residual stress according to claim 1, characterized in that: The method of applying zero-mean Gaussian perturbation to the residual stress and constructing the perturbation sample set through multiple rounds of Monte Carlo calculation is as follows: A zero-mean Gaussian perturbation is introduced into the stress tensor of the residual stress at each measuring point to generate a perturbation tensor: ； Where, represents the perturbation tensor, represents the perturbation matrix, where the elements of the perturbation matrix are generated by the following method: Generate standard normal random variables using the Mersenne Twister algorithm , generate the perturbation matrix elements: ； Where, It represents the disturbance intensity coefficient, which is 5%-10% of the yield strength of the steel bridge material. Indicates the The perturbation matrix elements generated by the simulation, represents the elements in the stress tensor matrix, represents the row index of the stress tensor matrix, represents the column index of the stress tensor matrix, represents the row index of the perturbation matrix, Represents the column index of the perturbation matrix, which is generated repeatedly for each measurement point Group perturbation tensors to form a set of perturbation samples ,in, is a positive integer greater than 100.

5. The multiaxial fatigue damage assessment method for steel bridge welds considering welding residual stress according to claim 4, characterized in that: The Dang Van criterion is used to calculate multiple rounds of risk factors for disturbed samples, and the method of constructing robust weighted risk factors based on risk factors is as follows: For each set of perturbation tensors in the perturbation sample set, after applying a sinusoidal external load along the x-axis, the Dang Van criterion is used to calculate risk factors, where the risk factor obtained by the k-th simulation is expressed as , and sort in ascending order, take the Robust weighted risk factors .

6. The multiaxial fatigue damage assessment method for steel bridge welds considering welding residual stress according to claim 4, characterized in that: The method of using infrared thermal imaging images and weld 3D topography images as dual-modal inputs and extracting weld thermal distribution characteristics and geometric morphology characteristics respectively through deep convolutional neural networks is as follows: A 64×64 pixel sliding window with a sliding step of 16 pixels is set. The window slides 40 times along the x-axis and 32 times along the z-axis. Zero values are used to fill the edge window to complete it. A total of 1280 windows are generated. For each window, the physical space coordinates corresponding to its center pixel are extracted and it is determined whether it is located in the geometric transition area of the weld area. Only windows whose center points fall into the geometric transition area are retained as valid training samples. For each valid sample, the nearest neighbor interpolation method is used to find the robustness weight risk factor corresponding to the nearest measurement point based on the physical coordinates of the window center, and use it as the supervision label value of the window. , then the 64×64 pixel matrix intercepted in the infrared thermal imaging image is recorded as , the 64×64 pixel matrix of the weld 3D topography is recorded as ; A dual-branch parallel convolutional neural network is used to extract features from infrared thermal images and three-dimensional topography images respectively. The specific structure is as follows: Input layer: receives a single-channel image of 64×64×1; the convolution layer of the first convolution block: 32 3×3 convolution kernels, padding mode "same", activation function ReLU, batch normalization layer is used to standardize the channel mean and variance, the maximum pooling layer uses a 2×2 pooling kernel, stride 2, and outputs a 32×32×32 feature map; the convolution layer of the second convolution block: 64 3×3 convolution kernels, padding mode "same", activation function ReLU, batch normalization layer is used to standardize the channel mean and variance, the maximum pooling layer uses a 2×2 pooling kernel, stride 2, and outputs a 16×16×64 feature map; flattening layer: flattens the three-dimensional feature map into a 16384-dimensional feature vector.

7. The multiaxial fatigue damage assessment method for steel bridge welds considering welding residual stress according to claim 6, characterized in that: The shared attention mechanism fuses features and uses the robust weighted risk factor of the geometric transition region as a supervisory signal to construct a fatigue risk nonlinear regression model: Shared attention mechanism, concatenating the 16384-dimensional vector of the two branches into 32768 dimensions, fully connected layer 1: 1024 neurons, activation function ReLU; fully connected layer 2: 512 neurons, activation function ReLU; fully connected layer 3: 2 neurons, activation function Sigmoid, output thermal weight and shape weight ,and ; A regression model is constructed using a fully connected network and a Dropout layer to output the predicted risk factors. The steps for constructing a regression model using a fully connected network and a Dropout layer are as follows: Set the fully connected layer 1 to 2048 neurons and the activation function ReLU; Dropout layer: dropout rate 25%; fully connected layer 2 to 1024 neurons and the activation function ReLU; Dropout layer: dropout rate 25%; fully connected layer 3 to 1 neuron and linear activation function; The training learning rate is , the optimizer is Adam, the training rounds are 200 rounds, the Dropout is 0.25, and the batch size is 32, that is, 32 samples are randomly selected from the valid training samples. The main loss function in the current batch is: ； Where, represents the main loss function, express The output predicted risk factors for samples, Indicates the The supervised label value of each sample, Indicates the serial number index of 32 randomly selected samples, .

8. The multiaxial fatigue damage assessment method for steel bridge welds considering welding residual stress according to claim 7, characterized in that: Output the risk factor of each window position and classify the fatigue damage of the weld area as follows: For steel bridge welds that require damage assessment, extract the risk factors output by each sliding window and construct The two-dimensional risk matrix has a one-to-one correspondence with the physical coordinates of the weld, and the risk factor threshold is pre-set and ,in, , for each risk factor output by the sliding window , when satisfied When the risk of damage is low, it is judged as a low-risk damage area. When the damage area is determined to be medium risk, When the damage is detected, it is judged as a high-risk damage area, where Indicates the serial index of the sliding window, .

9. A multiaxial fatigue damage assessment device for steel bridge welds taking into account welding residual stress, characterized by: The device is used to perform the multiaxial fatigue damage assessment method for a steel bridge weld considering welding residual stress according to any one of claims 1 to 8, comprising: An image generation module is used to measure the three-dimensional residual stress of the geometric transition area of the steel bridge weld area using X-rays, generate infrared thermal images and three-dimensional topography images of the weld surface, and perform normalization processing. The geometric transition area includes the weld toe tip, the center point of the groove bottom, and the base metal reference point; The robust risk factor module is used to apply zero-mean Gaussian perturbations to residual stresses, construct a perturbation sample set through multiple rounds of Monte Carlo calculations, perform multiple rounds of risk factor calculations on the perturbation samples using the Dang Van criterion, and construct robust weighted risk factors based on the risk factors. The fatigue risk regression module uses infrared thermal imaging and weld 3D topography as dual-modal inputs, extracts weld thermal distribution characteristics and geometric topography features through a deep convolutional neural network, fuses these features through a shared attention mechanism, and uses the robustness weighted risk factor of the geometric transition region as a supervisory signal to construct a fatigue risk nonlinear regression model. The weld risk grading module is used to obtain infrared thermal images and three-dimensional topography images of the weld surface through image window sliding sampling in the weld area as inputs of the fatigue risk nonlinear regression model, output the risk factors of each window position, and grade the fatigue damage in the weld area.

Images