Bridge bearing visual inspection system and method based on artificial intelligence

Abstract

The application discloses a bridge support visual detection system and method based on artificial intelligence, and relates to the technical field of intelligent detection.The system comprises the following steps: collecting bridge support data and performing pretreatment on the bridge support data to obtain an image sequence and a load event time axis; the bridge support data comprises real-time video stream under the action of dynamic vehicle load, static images, real-time displacement distance of the surface of the bridge support, and temperature and humidity of the bridge support; a space-time double-flow network is constructed based on the image sequence and the load event time axis, time feature maps and space feature maps are extracted and fused; the fused features are subjected to time sequence enhancement, decoding and up-sampling to generate a dynamic mask matrix.The application improves the precision of deformation separation, realizes seamless connection from damage identification to mechanical property evaluation, can intuitively show the influence of damage on the bearing capacity and stability of the bridge support, and improves the scientificity and effectiveness of bridge management decision.

Description

Technical Field

[0001] This invention relates to the field of intelligent inspection technology, and in particular to a visual inspection system and method for bridge bearings based on artificial intelligence. Background Technology

[0002] With the increasing demands for bridge operation safety monitoring and maintenance management, bridge bearings, as crucial components in bridge structures that transfer loads and adapt to deformation, directly impact the overall service performance of the bridge. In recent years, visual inspection technology has been gradually applied to bridge bearing condition identification. By using cameras to capture images of the bearing surface and combining them with image processing or deep learning models to identify cracks, corrosion, and displacement anomalies, this has become an important technological development direction in the field of bridge health monitoring.

[0003] Existing methods have some shortcomings. Vibration disturbances caused by vehicle traffic on the bridge deck and actual structural deformation of the bearings often appear simultaneously in the image sequence. Traditional time-averaging filtering or frequency domain filtering methods are difficult to effectively distinguish between instantaneous dynamic disturbances and continuous structural deformations, which can easily weaken or even omit effective deformation information. In addition, most existing methods are limited to the level of apparent damage identification. The extraction of cracks or damage directions usually relies on texture features or ordinary edge features, and lacks a direction-sensitive analysis mechanism that combines with the stress characteristics of bridge bearings. Therefore, it is difficult to establish a reliable relationship between visual recognition results and bearing capacity and stability changes, resulting in insufficient engineering interpretability and subsequent evaluation value of the detection results. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an artificial intelligence-based visual inspection method for bridge bearings, which solves the problems of insufficient deformation separation accuracy under dynamic load interference and the disconnect between damage direction identification and mechanical assessment.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides an artificial intelligence-based visual inspection method for bridge bearings, comprising: acquiring bridge bearing data and performing preprocessing on the bridge bearing data to obtain an image sequence and a load event timeline, wherein the bridge bearing data includes real-time video streams under dynamic vehicle loads, static images, real-time displacement distances of the bridge bearing surface, and temperature and humidity of the bridge bearings; constructing a spatiotemporal dual-stream network based on the image sequence and the load event timeline, extracting and fusing temporal feature maps and spatial feature maps; performing temporal enhancement, decoding, and upsampling on the fused features to generate a dynamic mask matrix, and marking and removing dynamic interference regions according to the dynamic mask matrix to obtain a separated structural deformation image sequence; calculating the response intensity of each directional convolution kernel to the structural deformation image using a pre-constructed directional convolution kernel group based on the separated structural deformation image sequence, generating a multi-channel response map and comparing directional responses to obtain a vector damage distribution map; correlating and comparing the vector damage distribution map with the design parameters of the bridge bearings to evaluate the impact of damage on the bearing capacity and stability of the bridge bearings, and generating a bridge bearing structural inspection report.

[0007] As a preferred embodiment of the artificial intelligence-based bridge bearing visual inspection method of the present invention, the preprocessing includes image denoising and illumination correction of real-time video streams and static images under dynamic vehicle loads. The real-time video frames corrected by illumination are time-stamped and synchronized with the real-time displacement distance of the bridge bearing surface and the temperature and humidity of the bridge bearing. Convert the aligned data to a multi-channel format.

[0008] As a preferred embodiment of the artificial intelligence-based bridge bearing visual detection method of the present invention, the construction of the spatiotemporal dual-stream network includes collecting spatiotemporal data of bridge bearings and generating a training sample library, and using 3D convolutional neural networks and 2D convolutional neural networks as core networks. Multi-frame image data and single-frame image data are extracted from the training sample library. The multi-frame image data is input into a 3D convolutional neural network to construct a temporal flow network, and the single-frame image data is input into the 2D convolutional neural network to construct a spatial flow network.

[0009] As a preferred embodiment of the artificial intelligence-based bridge bearing visual inspection method of the present invention, the generation of the dynamic mask matrix includes extracting temporal feature maps and spatial feature maps using a spatiotemporal dual-stream network, and performing channel stitching on the temporal feature maps and spatial feature maps to obtain a fused feature map. The fused feature maps corresponding to consecutive frames are input into a gated recurrent unit for temporal correlation enhancement to obtain temporally enhanced feature maps. Upsampling decoding is performed on the temporal enhanced feature map to obtain a dynamic mask matrix corresponding to the size of the input image.

[0010] As a preferred embodiment of the artificial intelligence-based bridge bearing visual detection method of the present invention, the obtained separated structural deformation image sequence includes identifying the periodic pattern of vibration noise using time feature maps, and determining the pixel weight threshold based on the correspondence statistics between the pixel prediction probability distribution in the validation set and the actual annotation results. Regions below the pixel weight threshold are marked as dynamic interference regions and removed to obtain a preliminary structural deformation image sequence. Morphological smoothing is performed on the initial structural deformation image sequence to output the separated structural deformation image sequence.

[0011] As a preferred embodiment of the artificial intelligence-based visual inspection method for bridge bearings described in this invention, the method for obtaining the vector damage distribution map includes constructing a group of directional convolution kernels based on the stress characteristics of the bridge bearings and using ANSYS finite element analysis software. The stress characteristics of the bridge bearings include transverse shear, longitudinal compression, and oblique composite stress. The separated structural deformation image sequence is input into a group of directional convolution kernels to generate a response map; Calculate the average pixel intensity of each response map and arrange them according to the direction of the convolution kernel to obtain the directional response vector; Construct a direction discriminator and input the direction response vector into the direction discriminator to obtain a vector damage distribution map.

[0012] As a preferred embodiment of the bridge bearing visual inspection method based on artificial intelligence described in this invention, the generation of the bridge bearing structure inspection report includes determining the design parameters of the bridge bearing and comparing the design parameters with the vector damage distribution map to obtain the comparison result; Based on the comparison results, a support stress model is constructed. The support stress model is used to evaluate the impact of damage on the bearing capacity and stability of the support, and the comprehensive results of the support status are output. A random forest model was used as a classifier to classify the comprehensive results of the bearing status, and multi-frame consistency was used to verify the classification results to obtain a bridge bearing structure inspection report.

[0013] Secondly, this invention provides an artificial intelligence-based visual inspection system for bridge bearings, comprising: a data acquisition module, which acquires bridge bearing data and performs preprocessing on the data to obtain image sequences and load event timelines, wherein the bridge bearing data includes real-time video streams under dynamic vehicle loads, static images, real-time displacement distances of the bridge bearing surfaces, and temperature and humidity of the bridge bearings; and an image separation module, which constructs a spatiotemporal dual-stream network based on the image sequences and load event timelines, extracts and fuses temporal feature maps and spatial feature maps; and performs temporal enhancement, decoding, and uploading on the fused features. The system generates a dynamic mask matrix and marks and removes dynamic interference regions based on the dynamic mask matrix to obtain a sequence of separated structural deformation images. The distribution map generation module calculates the response intensity of each directional convolution kernel to the structural deformation image using a pre-constructed set of directional convolution kernels, generates a multi-channel response map, and compares the directional responses to obtain a vector damage distribution map. The report integration module correlates and compares the vector damage distribution map with the design parameters of the bridge bearing to evaluate the impact of damage on the bearing capacity and stability of the bridge bearing and generates a bridge bearing structural inspection report.

[0014] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein the computer program, when executed by the processor, implements any step of the artificial intelligence-based visual inspection method for bridge bearings as described in the first aspect of the present invention.

[0015] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the artificial intelligence-based visual inspection method for bridge bearings as described in the first aspect of the present invention.

[0016] The beneficial effects of this invention are as follows: By constructing a spatiotemporal dual-stream network that combines the time axis of load events, the image data of bridge bearings acquired under dynamic vehicle loads are jointly analyzed to distinguish and separate dynamic interference areas from real structural deformation areas, thereby improving the accuracy and stability of structural deformation extraction and providing a more reliable data foundation for subsequent damage identification. By utilizing directional convolution kernel groups corresponding to the stress characteristics of bridge bearings, directional response analysis is performed on the separated structural deformation image sequence to achieve directional identification of crack and damage propagation directions. Furthermore, the bearing capacity and stability are assessed in conjunction with bearing design parameters, which connects visual inspection results with mechanical performance analysis, thereby improving the engineering applicability, risk assessment capability, and maintenance decision reference value of bridge bearing inspection results. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of the AI-based visual inspection method for bridge bearings in Example 1.

[0019] Figure 2 This is a diagram of the spatiotemporal dual-stream network architecture in Example 1.

[0020] Figure 3 This is a diagram showing the response analysis of the directional convolution kernel in Example 1.

[0021] Figure 4 This is a diagram showing the mechanical performance evaluation and report generation in Example 1. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Example 1, referring to Figures 1-4 This embodiment provides an artificial intelligence-based visual inspection method for bridge bearings, including the following steps: S1. Collect bridge bearing data and perform preprocessing on the bridge bearing data to obtain image sequences and load event timelines.

[0026] Includes the following steps, S1.1 First, four high-resolution industrial cameras are deployed, installed on the top, front, and 45° angles on both sides of the bridge bearing to achieve 360° coverage without blind spots. Simultaneously, a high-precision laser rangefinder (measurement accuracy ±0.1mm), an infrared camera (thermal resolution 320×240), and a set of temperature and humidity sensors are provided to acquire displacement information, surface temperature distribution information, and ambient temperature and humidity information of the bridge bearing surface, respectively. The collected multi-source bridge bearing data includes real-time video streams (60 frames per second) under dynamic vehicle loads, static images (automatically captured every 5 seconds, totaling 1000 frames), real-time displacement distance of the bridge bearing surface, and temperature and humidity of the bridge bearing.

[0027] S1.2 Next, the multi-source bridge bearing data needs to be preprocessed. Specifically, for the real-time video stream and static images under dynamic vehicle loads, image denoising is performed first. An adaptive median filtering method is used, which dynamically adjusts the filter window size (default 5×5) and preserves edge details by analyzing the gray value distribution in the neighborhood of each pixel. Next, illumination correction is performed on the denoised static image. Adaptive gamma correction technology is used to dynamically adjust the gamma value according to the overall average brightness of the denoised static image (if the average is below 50, the gamma value is set to 1.2; if it is above 200, the gamma value is set to 0.8) to ensure balanced contrast in the illumination-corrected image and avoid uneven illumination. For the illumination-corrected real-time video frames, a timestamp synchronization operation is performed. The built-in clock is used to align each frame of the denoised real-time video frame with the real-time displacement distance of the bridge bearing surface and the temperature and humidity of the bridge bearing. The error needs to be controlled within ±10 milliseconds.

[0028] Finally, all images after denoising, illumination correction, and synchronization timestamps are converted into a multi-channel format. The RGB channels retain visual information, while a depth channel (generated by a laser rangefinder) and a thermal imaging channel (generated by an infrared camera) are added, forming a five-channel data structure. The resolution of each frame is normalized to 1920×1080, and the pixel values ​​are normalized to the range of 0-1. The output is a multi-channel image sequence (60 frames per second) and the corresponding load event timeline (marked with timestamps and load intensity).

[0029] S2. Construct a spatiotemporal dual-stream network based on image sequences and load event timelines, extract and fuse temporal and spatial feature maps; perform temporal enhancement, decoding and upsampling on the fused features to generate a dynamic mask matrix, and mark and remove dynamic interference regions based on the dynamic mask matrix to obtain the separated structural deformation image sequence.

[0030] Includes the following steps, S2.1 Collect spatiotemporal data of bridge bearings (specifically including dynamic video sequences for 30 consecutive days and corresponding load records). To enhance the diversity of spatiotemporal data, data augmentation techniques are used to randomly superimpose Gaussian noise (noise standard deviation range of 0.01-0.05) into the dynamic video sequences. The load intensity of different vehicles is simulated by adjusting the displacement between video frames (displacement amplitude range of 0.1-5mm). At the same time, image segmentation tools are used to mark vibration noise areas (such as temporary shaking caused by vehicles passing by) and structural deformation areas (such as cracks or permanent displacements in bridge bearings) frame by frame, generating a training sample library containing 100,000 frames.

[0031] S2.2 Construct and train a spatiotemporal dual-stream network, which consists of a temporal stream network and a spatial stream network. Specifically, a 3D convolutional neural network (i.e., 3DCNN) model is selected as the core architecture. Sixteen consecutive frames of image data (including RGB, depth information, and thermal imaging information) are extracted from a training sample library of 100,000 frames. The temporal stream network contains five convolutional layers, each using 3×3×3 convolutional kernels with strides of 1, 2, 1, 2, and 1, progressively downsampling the temporal and spatial dimensions to capture the dynamic changes between video frames. The first layer of the temporal stream network has 64 convolutional kernels, which are then increased to 512 layer by layer. Non-linearity is introduced through the ReLU activation function. The first four layers are followed by a 3D max-pooling layer (with 2×2×2 pooling kernels, which reduces computational complexity and enhances feature representation). The temporal flow network retains the downsampled temporal feature map output at the end, which is used to characterize the temporal distribution characteristics of vibration noise in consecutive frames. Instead of using global average pooling to compress the temporal feature map into a single vector, it retains spatial location information for subsequent pixel-level dynamic mask generation.

[0032] S2.3 Next is the process of constructing the spatial flow network. Specifically, a 2D convolutional neural network (2DCNN) model is selected as the core architecture. The input is single-frame image data extracted from a training sample library of 100,000 frames. The spatial flow network contains a total of 6 convolutional layers. The first three layers use 3×3 convolutional kernels (stride of 1, kernel numbers of 64, 128, and 256 respectively), and the last three layers use 5×5 convolutional kernels (stride of 2, kernel numbers of 256, 512, and 512 respectively). This multi-scale convolutional kernel combination captures the continuous deformation features of bridge supports, such as crack edges or deformation contours. After each convolutional operation, batch normalization is used as a condition to accelerate convergence and stabilize training. The ReLU activation function is then used to enhance the nonlinear expressive power of the spatial flow network. The spatial flow network retains the downsampled spatial feature map output at the end, which is used to characterize the spatial distribution details of the deformation of the bridge bearing structure. Instead of using global average pooling to compress the spatial feature map into a single vector, it ensures the positional correspondence required for subsequent dynamic mask generation.

[0033] S2.4. Train the temporal and spatial flow networks using a training sample library of 100,000 frames. Specifically, the training sample library is first divided into a training set (80%), a validation set (10%), and a test set (10%) in an 8:1:1 ratio. The training set is used for parameter optimization, the validation set for hyperparameter tuning, and the test set for evaluating the final performance. The temporal and spatial flow networks are trained independently, with the input data determined according to the network requirements: the temporal flow network receives 16 frames of image data, while the spatial flow network receives single-frame image data. Training uses batch gradient descent with a batch size of 32 and an initial learning rate of 0.001. The Adam optimizer is used to adjust the weights. The loss function is designed as a weighted combination of cross-entropy loss and Dice loss, where cross-entropy loss measures classification accuracy and Dice loss optimizes the overlap of feature regions. The weight ratio is set to 0.4:0.6 to balance the needs of classification and segmentation tasks.

[0034] The training iterations were set to 100 epochs, with performance evaluated on the validation set every 10 epochs. The loss values ​​and feature extraction effects of the temporal and spatial flow networks were recorded. When the loss on the validation set no longer decreased after 5 consecutive evaluations, an early stopping mechanism was triggered, and the current optimal parameters were saved. During training, for the temporal flow network, the focus was on optimizing the ability to extract the temporal variation features of vibration noise in consecutive frames. For the spatial flow network, the focus was on optimizing the ability to extract the edges of cracks and local deformation details in bridge supports, so as to improve the ability of subsequent fused features to distinguish between dynamic interference areas and structural deformation areas. After training was completed, the test set was used for final validation. The temporal flow network was required to maintain stable extraction results of vibration noise temporal features, and the spatial flow network was required to have a high crossover-union ratio (CUI) for segmenting structural deformation areas.

[0035] After the spatiotemporal dual-stream network is trained, 16 consecutive frames of image data are input into the temporal stream network to obtain the corresponding temporal feature map; single-frame image data are input into the spatial stream network to obtain the corresponding spatial feature map. The temporal feature map is used to characterize the variation of vibration noise in consecutive frames, and the spatial feature map is used to characterize the spatial distribution details of the deformation of the bridge support structure. Subsequently, the temporal and spatial feature maps are fed into the fusion, temporal enhancement, and decoding reconstruction processes to generate a dynamic mask matrix corresponding to the size of the input image.

[0036] S2.5. Pixel-level separation of vibration noise and structural deformation is achieved using temporal and spatial feature maps. Specifically, the temporal and spatial feature maps are concatenated along the channel dimension to obtain a fused feature map. The fused feature map corresponding to consecutive frames is input into the gated recurrent unit to extract the temporal correlation information between adjacent frames, resulting in a temporal enhancement feature map. The labeled mask in the training samples is used as a supervision signal to jointly train the gated recurrent unit and subsequent decoding layers. The loss function adopts a weighted combination of binary cross-entropy loss and Dice loss as the optimization objective, which is used to simultaneously constrain pixel classification accuracy and target region integrity. During training, parameter updates and early stopping control are performed based on the changes in the validation set loss. When the validation set loss no longer decreases after multiple consecutive iterations, the current optimal parameters are saved. In this way, the gated recurrent unit undertakes the functions of continuous frame feature fusion and temporal enhancement. The dynamic mask matrix is ​​directly given by the decoding results after supervised training, thereby avoiding the problems of infeasibility in constructing Q-tables in high-dimensional continuous states and the mixing of multiple training logics.

[0037] S2.6. Generate a dynamic mask matrix based on the temporal enhanced feature map, and use the dynamic mask matrix to mark dynamic interference regions. Specifically, perform convolutional decoding and progressive upsampling processing on the temporal enhanced feature map in sequence to restore the feature resolution to be consistent with the input image, and obtain a dynamic mask matrix corresponding to the size of the input image; the value of each pixel position in the dynamic mask matrix is ​​used to represent the confidence level of the current pixel belonging to the structural deformation region; according to the determined pixel weight threshold, perform pixel classification on the dynamic mask matrix: mark the pixel positions below the pixel weight threshold as dynamic interference regions, and retain the pixel positions not below the pixel weight threshold as structural deformation candidate regions, thereby generating a preliminary structural deformation image sequence.

[0038] S2.7. Post-processing is performed on the preliminary structural deformation image sequence to generate the final separated structural deformation image sequence. Specifically, morphological smoothing is performed on the structural deformation candidate regions corresponding to the dynamic mask matrix, and opening and closing operations are performed using 3×3 structuring elements to remove isolated noise points and fill local holes; the consistency between the mask area change and the load triggering change between consecutive frames is checked in conjunction with the load event time axis; when an abnormally high proportion of dynamic interference markers still exist in several consecutive frames after the load event ends, the abnormal frames and adjacent frames are re-input into the separation network for updating to improve the stability and continuity of the separation results, and the separated structural deformation image sequence is output.

[0039] S3. Based on the separated structural deformation image sequence, the response intensity of each directional convolution kernel to the structural deformation image is calculated using a pre-constructed directional convolution kernel group, a multi-channel response map is generated, and the directional responses are compared to obtain a vector damage distribution map.

[0040] Includes the following steps, S3.1 First, it is necessary to construct a directional convolution kernel group based on the stress characteristics of the bridge bearings (including transverse shear, such as horizontal displacement or cracks caused by vehicle lateral forces or wind loads; longitudinal compression, such as vertical deformation or extrusion damage caused by the bridge deck's own weight and vehicle vertical loads; and oblique composite stress, such as cracks formed by multi-directional force coupling, manifested as 45° or 135° angles). The specific construction process involves using ANSYS finite element analysis software to input material properties (e.g., rubber hardness 70 Shore A, steel plate yield strength 235 MPa) and stress conditions (e.g., transverse shear force 50 kN, longitudinal compressive force 200 kN), generating 5000 frames of simulation images (each frame labeled with the force direction and damage type), and merging them with static images to form an image dataset containing 6000 frames.

[0041] The gradient orientation distribution of lesion edges in the image dataset was statistically analyzed using the Histogram of Oriented Gradients (HOG) method, extracting eight principal orientations (0°, 45°, 90°, 135°, 180°, 225°, 270°, and 315°). For each orientation, a 5×5 convolutional kernel was designed. The pixel values ​​within each kernel were weighted according to the gradient orientation distribution of the lesion edge. For example, the kernel for the 0° orientation emphasized horizontal edges, with a weight of 0.5 in the center row, while the kernel for the 90° orientation emphasized vertical edges, with a weight of 0.7 in the center column. Finally, the eight optimized kernels were combined into an orientational convolutional kernel group to ensure high responsiveness to lesions in specific orientations.

[0042] S3.2. Damage direction identification and analysis are performed using each frame of the separated structural deformation image sequence and directional convolutional kernel groups. Specifically, each frame of the separated structural deformation image sequence is input into 8 directional convolutional kernel groups, and parallel computation is performed with a stride of 1 unit to generate 8 response maps. Each response map represents the damage intensity in the corresponding direction (and a label is given to the damage intensity in each direction). For example, the 0° response map highlights the transverse cracks of the bridge bearing, and the 90° response map highlights the longitudinal deformation of the bridge bearing. Simultaneously, the average pixel intensity of the 8 response images is calculated (by traversing each row and column of the response image and summing the intensity values ​​of all pixels; for example, the sum of the 1920×1080 pixels in the 0° response image is 1,244,160, which, divided by the total number of pixels 2,073,600, yields an average pixel intensity of 0.6) and combined into an 8-dimensional directional response vector (using feature concatenation, the average pixel intensity of the 8 response images is arranged according to the direction of the convolution kernel, i.e., 0°, 45°, 90°, 135°, 180°, 225°, 270°, 315°, forming an ordered vector). The numerical sequence, i.e., the response vector. For example, if the calculated average pixel intensities are 0.6 (0° response map), 0.3 (45° response map), 0.2 (90° response map), 0.1 (135° response map), 0.05 (180° response map), 0.03 (225° response map), 0.02 (270° response map), and 0.01 (315° response map), then the sequence is [0.6, 0.3, 0.2, 0.1, 0.05, 0.03, 0.02, 0.01], which is used to initially determine the dominant direction of damage.

[0043] S3.3 Next, a direction discriminator needs to be constructed and trained (the structure is a 3-layer fully connected neural network, with an 8-dimensional response vector in the input layer, 128 neurons in the hidden layer using the ReLU activation function to enhance nonlinear expression, and 4 neurons in the output layer corresponding to transverse cracks, longitudinal deformation, oblique compound damage, and no obvious damage, respectively). Specifically, the 6000-frame image dataset is divided into training, validation, and test sets in an 8:1:1 ratio. A stochastic gradient descent (SGD) combined with momentum optimization strategy is used, processing only one sample per iteration (batch size set to 1). The momentum factor (set to 0.9) is used to accelerate gradient convergence and avoid getting trapped in local maxima. The initial learning rate is set to 0.01, and it is dynamically adjusted every 10 epochs (training batches) based on the validation set loss. If the loss decreases by less than 0.001, the learning rate is halved (minimum value to 0.0001).

[0044] Each iteration of training consists of three core steps: forward propagation, loss calculation, and backpropagation. The forward propagation process is as follows: An 8-dimensional response vector from a randomly selected frame of image in the training set is used, for example, a numerical combination representing the intensity of eight directions [0.6, 0.2, 0.1, 0.05, 0.03, 0.01, 0.0, 0.01], as the input values ​​for the eight neurons in the input layer. The input values ​​of these eight neurons are weighted by the weights between the input layer and the hidden layer to calculate the inputs for the 128 neurons in the hidden layer. The input to each hidden neuron is a weighted sum of the input values, which is then activated using the ReLU activation function. If the activation result is less than 0, it is set to 0; otherwise, it remains unchanged, generating the output vector of the hidden layer (containing 128 values). Next, the hidden layer output is weighted again by the weights between the hidden layer and the output layer to calculate the inputs for the four neurons in the output layer. Each neuron corresponds to a damage direction (transverse crack, longitudinal deformation, oblique compound damage, no obvious damage). The output layer values ​​are processed by Softmax and converted into probability form, such as [0.85, 0.10, 0.04, 0.01], representing the probability of each direction.

[0045] Loss calculation and backpropagation are performed simultaneously. Specifically, the loss is calculated based on the difference between the predicted probability value and the true label. For example, if the true label of a frame is a horizontal crack, the expected output is [1,0,0,0], while the actual output is [0.85,0.10,0.04,0.01], then the loss is based on the difference between the predicted probability value and the expected output, which amplifies the error in the direction of the horizontal crack. Backpropagation starts from the output layer, calculating the error of each output neuron (e.g., the error of the horizontal crack neuron is the difference between the predicted probability value and the expected output, i.e., the difference between 0.85 and 1), and propagating it back to the hidden layer through weights. This adjusts the contribution of each neuron in the hidden layer. Specifically, the error of the output layer is distributed according to the weight ratio, and the direction of the hidden layer output adjustment is determined according to the ReLU activation characteristic (i.e., the error is propagated when the input is greater than 0, otherwise it is blocked). The error of the hidden layer continues to propagate back to the input layer, calculating the change in the weights of each layer. For example, the change in the output layer weights is based on the product of the output error and the hidden layer output, and the change in the hidden layer weights is based on the product of the hidden layer error and the input value. The weights are updated using the momentum method until one epoch of 4800 iterations is completed.

[0046] To optimize the performance of the orientation discriminator, classification accuracy and loss values ​​are evaluated on the validation set every 5 epochs, and the current performance of the discriminator is recorded. If the classification accuracy on the validation set does not improve for three consecutive epochs, the current parameters are saved and training is terminated early. To improve the recognition ability of rare categories (such as oblique compound damage), the loss calculation for rare categories is weighted by 1.5 times. Specifically, the error impact of this category is amplified during loss calculation. For example, if the true label is oblique compound damage, the penalty for the prediction probability error is increased by 50%. After training, the orientation discriminator is validated on the test set (requiring a classification accuracy of over 94% and a misclassification rate of less than 5% for each direction), and confusion matrix analysis is used to ensure there is no significant bias.

[0047] S3.4. The 8-dimensional response vector is input to the orientation discriminator to determine the final direction of damage propagation. Specifically, the orientation discriminator generates an orientation probability output for each frame of the 8-dimensional response vector. For example, the output for a certain frame is [0.85, 0.10, 0.04, 0.01]. Then, 16 frames are selected, and the 4-dimensional probability vector of each frame is recorded to form a 16×4 probability matrix. For example, the first frame is [0.85, 0.10, 0.04, 0.01], the second frame is [0.80, 0.12, 0.06, 0.02], and so on. Next, a temporal weight is assigned to each frame of the image. The most recent frame (the 16th frame) has the highest weight, set to 0.5, the previous frame (the 15th frame) has a slightly lower weight, set to 0.48, and so on down to 0.1 for the first frame.

[0048] Then, the predicted probability values ​​for each direction are weighted and averaged according to time weights. Specifically, for the transverse crack direction, the probability values ​​of 16 frames of images (e.g., 0.85, 0.80, ...) are taken, multiplied by their corresponding weights (0.1, 0.12, ..., 0.5), and the products are summed to obtain the weighted average. For example, if the result is 0.82, the overall probability value for the transverse crack is 0.82. Similarly, the weighted average values ​​for longitudinal deformation, oblique composite damage, and no obvious damage are calculated to generate the final 4-dimensional probability vector, for example, [0.82, 0.13, 0.03, 0.02]. Simultaneously, the direction with the highest overall probability value is selected as the final damage propagation direction. For example, 0.82 corresponds to a transverse crack; since the overall probability value is the highest, the damage direction of this data stream segment is determined to be a transverse crack.

[0049] S3.5. Based on the determined crack and damage propagation directions, generate a damage distribution map with vector attributes. Specifically, extract the corresponding direction maps (e.g., the 0° response map for a transverse crack) from the eight response maps. Mark the damage areas using an adaptive threshold segmentation method (the adaptive threshold can be 1.3 times the average pixel intensity of the response map). Pixels with values ​​below the adaptive threshold are classified as background, while those above the adaptive threshold are retained as damage points. For each pixel of a damage point, attach vector attributes. Specifically, the direction of the damage point pixel is determined by a direction discriminator; for example, 0° is transverse. The direction length is linearly mapped according to the average pixel intensity of the response map; for example, a pixel intensity of 0.5 in the response map is mapped to a long arrow of 5 pixels.

[0050] To enhance the coherence of the mapping, a 3×3 structuring element is used for two dilation operations to connect adjacent damaged pixel points to form continuous regions. Boundary smoothing (i.e., 3×3 median filtering) is then used to identify the boundaries between these continuous regions. Since the boundaries between continuous regions may exhibit small irregular protrusions or isolated points (i.e., burrs) due to the dilation operation or adaptive threshold segmentation, a 3×3 neighborhood window is set centered on each damaged pixel, containing damaged pixel A and its eight surrounding neighboring pixels. Next, the average intensity values ​​of the nine pixels within the neighborhood window are calculated and sorted by size. The median value (i.e., the fifth value) is selected as the new intensity value for the central damaged pixel. For example, if the average intensity of 9 pixels within a certain neighborhood window is [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9], then the median value is 0.5, and this median value is updated to the new intensity value of the pixel at the center of the damage point (traversing the entire damage area, the updated intensity value replaces the original intensity value, making abrupt changes at the boundary smoother). For example, if there is an isolated high-intensity point (e.g., 0.9) surrounded by low-intensity points (e.g., 0.2) at a damage edge, median filtering will adjust the isolated high-intensity point to a lower value (i.e., close to 0.2), thereby eliminating glitch. After glitch elimination, the connectivity of the damage area is checked. If there are still tiny isolated points (i.e., connected components less than 3 pixels), they are directly removed. The final output is a vector damage distribution map.

[0051] S4. Compare and correlate the vector damage distribution map with the design parameters of the bridge bearing to assess the impact of damage on the bearing capacity and stability of the bridge bearing and generate a bridge bearing structural inspection report.

[0052] Includes the following steps, S4.1. Compare the vector damage distribution map with the design parameters of the bridge bearing to achieve accurate classification and real-time output of the bridge bearing condition. The design parameters of the bridge bearing are derived from the technical specifications when designing the bearing, specifically including the maximum allowable displacement, material fatigue limit (such as rubber shear strength, steel plate tensile strength), and tolerance of the force direction. The comparison process is as follows: First, extract the vector attributes (i.e., direction and length) of each damaged area from the vector damage distribution map. The vector length is recorded in pixels; for example, the vector length of a transverse crack area is 10 pixels. Then, convert the pixel length into actual physical displacement based on the spatial resolution of the image (i.e., the actual distance of each pixel to the bearing surface). The spatial resolution of the image needs to cover the actual size of the bearing (assuming 500mm × 300mm). Through proportional calculation, if each pixel corresponds to 0.52mm laterally and 0.56mm vertically, then take the average value of 0.5mm as a unified conversion factor. Therefore, the displacement amplitude of a damage point with a vector length of 10 pixels is 10 × 0.5mm = 5mm. If the vector direction is horizontal, a displacement of 5mm represents a horizontal displacement; if it is vertical, it represents a vertical displacement. The conversion is repeated for each damaged area, and the displacement amplitude of all damaged points is recorded. For example, if the vector length of another area is 60 pixels, the displacement is 60 × 0.5mm = 30mm. S4.2. Based on the comparison results, a support stress model is constructed. Specifically, firstly, the geometric and material data of the support are collected (e.g., support diameter is 500mm, thickness is 100mm, rubber hardness is 70 Shore A, and steel plate yield strength is 235MPa). A digital model of the support is generated using SolidWorks 3D modeling software. The digital model of the support is then imported into ANSYS analysis software, and boundary conditions (i.e., the bottom of the support is fixed, and a load is applied to the top) and mesh generation (mesh element size is 2mm, with a total of approximately 500,000 elements) are set. After the settings are completed, typical stress scenarios are simulated, such as applying a 50kN shear force laterally and a 200kN compressive force longitudinally. The stress distribution of the support under different damage levels (e.g., crack length 0~200mm) is calculated, generating 10,000 sets of stress distribution simulation data. Each set includes the damage area, damage location, damage direction, damage length, and corresponding stress value.

[0053] The training process of the support stress model is as follows: The Support Vector Machine (SVM) model is selected as the core architecture. 10,000 sets of stress distribution simulation data are divided into a training set (8,000 sets) and a test set (2,000 sets). The input features are damage length, damage direction, damage location, and the specific coordinates of the damage location. The output is the stress value. During training, the SVM maps the input features to a high-dimensional space through the radial basis kernel function. The hyperplane is adjusted according to each input feature to maximize the classification boundary. After every 10 iterations, the prediction error (i.e., mean square error) is evaluated on the test set. If the prediction error does not decrease for 5 consecutive iterations, the kernel parameters are adjusted (i.e., the bandwidth is increased or decreased). A total of 50 iterations are performed. The final prediction error needs to be controlled within ±0.5MPa. Otherwise, the training is repeated.

[0054] S4.3. Utilize the constructed support stress model to evaluate the impact of damage on the support's bearing capacity and stability, obtaining a comprehensive result of the support's condition. Specifically, read the displacement amplitude and damage direction of each damage point, for example, a 30mm transverse crack, and input them into the support stress model. The support stress model, based on the pre-training results, queries the corresponding stress value for the scenario and outputs it, for example, 18MPa. Iterate through all damage areas to record the stress distribution; for example, another 10mm longitudinal deformation corresponds to a 5MPa distribution. Calculate the percentage decrease in bearing capacity based on the stress distribution of all damage areas. If the stress distribution in multiple damage areas exceeds the limit (i.e., exceeds 15MPa), then the percentage decrease in bearing capacity is summed (e.g., a total decrease of 30% is considered a 30% decrease in the support's bearing capacity, placing it at risk of shear failure).

[0055] Next, the stability of the support needs to be evaluated. Specifically, the displacement amplitude and damage direction of each damage point are read from the vector damage distribution map. For example, a transverse crack is read with a displacement amplitude of 30 mm and a direction of 0° (transverse), and another longitudinal deformation has a displacement amplitude of 10 mm and a direction of 90° (longitudinal). The displacement amplitude and damage direction of each damage point are recorded as analysis units, for example, damage 1: 30 mm, 0°, damage 2: 10 mm, 90°.

[0056] Based on the real-time video stream of dynamic vehicle loads and the real-time displacement distance of the bridge bearing surfaces in the bridge bearing data, the current dominant force direction of the bridge bearings is determined. For example, if the bridge mainly bears the vertical load of vehicles, the force direction is 90° (longitudinal). If there are lateral wind loads or vehicle eccentric loads, the force direction is 0° (lateral) or a combined direction (e.g., 45°). Assuming that the current scenario is dominated by longitudinal loads, it is recorded as the dominant force direction: 90°, as a reference benchmark for stability analysis.

[0057] The damage direction and the main stress direction at each damage point are compared (comparing the included angles) to determine the mechanical correlation between the damage direction and the stress direction. Specifically, for damage 1: 30mm, 0°, the included angle with the main stress direction (90°) is 90°, indicating an orthogonal relationship. The transverse crack is orthogonal to the longitudinal load, which may lead to shear stress concentration and increase the risk of instability. For damage 2: 10mm, 90°, the included angle with the main stress direction (90°) is 0°, completely consistent (the longitudinal deformation is in the same direction as the longitudinal load, exhibiting compressive strengthening, and does not directly threaten the stability of the support). After the included angle calculation is completed, each damage point is labeled with a correlation tag, such as orthogonal, high risk, or same direction, low risk.

[0058] The displacement amplitude and direction of each damage point are input into the support stress model, which outputs the instability risk value of the damage point. (During training, the support stress model generates a large number of instability risk values ​​for damage scenarios through finite element simulation analysis. For example, a 30mm transverse crack under orthogonal force shows shear stress concentration in the simulation, resulting in an instability probability of 0.4). Specifically, for example, a 30mm transverse crack under orthogonal force corresponds to a shear instability risk increment of 0.4 (range 0-1, 1 being complete instability). Specifically: inputting damage 1: 30mm, 0°, with an included angle of 90°, the support stress model outputs a shear instability risk of 0.4; inputting damage 2: 10mm, 90°, with an included angle of 0°, the model outputs a compressive instability risk of 0.1. The query results are recorded as damage 1: shear instability risk 0.4, damage 2: compressive instability risk 0.1.

[0059] The instability risk values ​​of all damage points are summed up. For example, 0.4 (damage 1) + 0.1 (damage 2) = 0.5, indicating that the overall stability risk value of the support is 0.5. If there are multiple high-risk damages (e.g., 3 orthogonal cracks, each with an instability risk value of 0.4), the total risk may reach 1.2, which needs to be normalized to 1. Considering the impact of damage location on stability, if the damage location is located in the main stress area of ​​the support (such as the center or edge support point), the instability risk value increases by about 20%. For example, if the damage location of damage 1 is in the center of the support, the instability risk value is increased from 0.4 to 0.48, and the overall stability risk value is adjusted to 0.58.

[0060] A stability risk threshold is set (based on the safety margin requirements of the support design, i.e., the support can still withstand the expected load without instability when damage exists; the specific range is 0.3~0.6, and can also be adjusted according to the specific design of the support). When the overall stability risk value is less than 0.3, it is judged as low stability risk; when the overall stability risk value is between 0.3 and 0.6, it is judged as medium stability risk; and when the overall stability risk value is greater than 0.6, it is judged as high stability risk. Finally, the conclusion of the stability impact is generated and combined with the proportion of load-bearing capacity reduction. For example, if the support load-bearing capacity decreases by 30% and the stability risk is 0.58, the output support load-bearing capacity decreases by 30%, which is at medium shear failure risk; if the overall stability risk value exceeds 0.6, such as 0.8, the output support load-bearing capacity decreases by 30%, which is at high shear failure risk.

[0061] S4.4 A random forest model is selected as the classifier for the comprehensive results of the bearing condition. The random forest model consists of 100 decision trees. The input is the comprehensive results of damage length, damage direction, damage location, stress distribution, and bearing capacity, as well as risk weights (0~1). The training data consists of 5000 sets of historical detection samples (i.e., damage distribution images, condition labels, environmental parameters such as temperature 25°C, humidity 60%, load conditions such as vehicle load 200kN, and comprehensive results of bearing condition). Each set is labeled with a condition (normal, minor damage requiring monitoring, severe damage requiring repair). During training, each tree randomly selects 70~80% of the detection samples and sets node splitting (based on maximizing information gain). If the displacement amplitude is greater than 5mm and the stress distribution exceeds 15MPa, the bridge bearing is judged as severely damaged. At the same time, each tree votes independently, and the category with the most votes for the comprehensive result of the bearing condition is output. For example, 60 trees selected severe damage.

[0062] During real-time classification, a comprehensive result of the support condition is extracted, such as 5 damaged areas, a displacement amplitude of 6 mm, lateral dominance, a stress distribution of 18 MPa, and a risk weight (0.8, set based on the damage direction and load), which is then input into a random forest model. The random forest model analyzes features for each tree; for example, the first tree determines severe damage based on displacement amplitude and stress distribution, the second tree determines severe damage based on the risk weight, and finally, the voting results are statistically analyzed.

[0063] S4.5 To improve classification results, multi-frame consistency verification is required. Specifically, extract the classification results of 30 consecutive frames, for example, 20 frames for minor damage and 10 frames for severe damage. Each frame is assigned a risk weight, for example, the risk weight for minor damage is 0.4, and the risk weight for severe damage is 0.9. Count the occurrence of each damage state in the 30 frames. When the proportion of severe damage exceeds 60% (i.e., more than 18 frames), the state of severe damage is directly confirmed. If the proportion of severe damage does not meet the standard, compare the risk weights of each damage state. For example, severe damage occurs in 10 frames with a risk weight of 0.9 (10 multiplied by 0.9 equals 9), which is higher than the risk weight of minor damage in 20 frames (0.4 multiplied by 0.4 equals 8), so it is judged as severe damage. If the risk weights of the damage states are very close (i.e., the difference is less than 1.0), the damage state of the previous window is taken as a reference, and the final output consistency state is, for example, severe damage, which may cause clipping failure. After verification, the system outputs a bridge bearing structure inspection report (including a damage distribution map, with highlighted areas, red markings for severe damage, and yellow markings for minor damage), a status classification (e.g., severe damage requiring repair), a risk description (e.g., potential for shear failure), and maintenance recommendations (e.g., immediate bearing replacement).

[0064] This embodiment also provides a bridge bearing visual inspection system based on artificial intelligence, including: a data acquisition module, which acquires bridge bearing data and performs preprocessing on the bridge bearing data to obtain image sequences and load event timelines. The bridge bearing data includes real-time video streams under dynamic vehicle loads, static images, real-time displacement distances of the bridge bearing surface, and temperature and humidity of the bridge bearing. The image separation module constructs a spatiotemporal dual-stream network based on image sequences and load event timelines, extracts and fuses temporal and spatial feature maps, performs temporal enhancement, decoding, and upsampling on the fused features, generates a dynamic mask matrix, and marks and removes dynamic interference regions based on the dynamic mask matrix to obtain the separated structural deformation image sequence. The distribution map generation module calculates the response intensity of each directional convolution kernel to the structural deformation image based on the separated structural deformation image sequence using a pre-constructed group of directional convolution kernels, generates a multi-channel response map, compares the directional responses, and obtains a vector damage distribution map. The report integration module correlates and compares the vector damage distribution map with the design parameters of the bridge bearing to assess the impact of damage on the bearing capacity and stability of the bridge bearing and generate a bridge bearing structural inspection report.

[0065] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0066] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the artificial intelligence-based visual inspection method for bridge supports as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0067] In summary, this invention achieves the following: First, by constructing a spatiotemporal dual-stream network that integrates the time axis of load events, it jointly analyzes image data collected from bridge bearings under dynamic vehicle loads. This distinguishes and separates dynamic interference regions from actual structural deformation regions, improving the accuracy and stability of structural deformation extraction and providing a more reliable data foundation for subsequent damage identification. Second, by utilizing directional convolution kernels corresponding to the stress characteristics of bridge bearings, it performs directional response analysis on the separated structural deformation image sequences, enabling directional identification of crack and damage propagation directions. Furthermore, it combines bearing design parameters to conduct load-bearing capacity and stability assessments, connecting visual inspection results with mechanical performance analysis. This enhances the engineering applicability, risk assessment capability, and maintenance decision-making reference value of bridge bearing inspection results.

[0068] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A visual inspection method for bridge bearings based on artificial intelligence, characterized in that: include, Bridge bearing data is collected and preprocessed to obtain image sequences and load event timelines. The bridge bearing data includes real-time video streams under dynamic vehicle loads, static images, real-time displacement distances of the bridge bearing surfaces, and temperature and humidity of the bridge bearings. A spatiotemporal dual-stream network is constructed based on image sequences and load event timelines to extract and fuse temporal and spatial feature maps. The fused features are temporally enhanced, decoded, and upsampled to generate a dynamic mask matrix. Based on the dynamic mask matrix, dynamic interference regions are marked and removed to obtain the separated structural deformation image sequence. Based on the separated structural deformation image sequence, the response intensity of each directional convolution kernel to the structural deformation image is calculated using a pre-constructed group of directional convolution kernels. A multi-channel response map is generated and the directional responses are compared to obtain a vector damage distribution map. By correlating and comparing the vector damage distribution map with the design parameters of the bridge bearing, the impact of damage on the bearing capacity and stability of the bridge bearing is assessed, and a bridge bearing structural inspection report is generated.

2. The bridge bearing visual inspection method based on artificial intelligence as described in claim 1, characterized in that: The preprocessing includes image denoising and illumination correction of real-time video streams and static images under dynamic vehicle loads. The real-time video frames corrected by illumination are time-stamped and synchronized with the real-time displacement distance of the bridge bearing surface and the temperature and humidity of the bridge bearing. Convert the aligned data to a multi-channel format.

3. The bridge bearing visual inspection method based on artificial intelligence as described in claim 2, characterized in that: The construction of the spatiotemporal dual-stream network includes collecting spatiotemporal data of bridge bearings and generating a training sample library, and using 3D convolutional neural networks and 2D convolutional neural networks as the core networks. Multi-frame image data and single-frame image data are extracted from the training sample library. The multi-frame image data is input into a 3D convolutional neural network to construct a temporal flow network, and the single-frame image data is input into the 2D convolutional neural network to construct a spatial flow network.

4. The bridge bearing visual inspection method based on artificial intelligence as described in claim 3, characterized in that: The generation of the dynamic mask matrix includes extracting temporal and spatial feature maps using a spatiotemporal dual-stream network, and then concatenating the temporal and spatial feature maps to obtain a fused feature map. The fused feature maps corresponding to consecutive frames are input into a gated recurrent unit for temporal correlation enhancement to obtain temporally enhanced feature maps. Upsampling decoding is performed on the temporal enhanced feature map to obtain a dynamic mask matrix corresponding to the size of the input image.

5. The bridge bearing visual inspection method based on artificial intelligence as described in claim 4, characterized in that: The obtained separated structural deformation image sequence includes identifying periodic patterns of vibration noise using time feature maps, and determining pixel weight thresholds based on the correspondence statistics between the pixel prediction probability distribution in the validation set and the actual annotation results. Regions below the pixel weight threshold are marked as dynamic interference regions and removed to obtain a preliminary structural deformation image sequence. Morphological smoothing is performed on the initial structural deformation image sequence to output the separated structural deformation image sequence.

6. The artificial intelligence-based visual inspection method for bridge bearings as described in claim 5, characterized in that: The obtained vector damage distribution map includes constructing a group of directional convolution kernels based on the stress characteristics of the bridge bearings and using ANSYS finite element analysis software. The stress characteristics of the bridge bearings include transverse shear, longitudinal compression, and oblique composite stress. The separated structural deformation image sequence is input into a group of directional convolution kernels to generate a response map; Calculate the average pixel intensity of each response map and arrange them according to the direction of the convolution kernel to obtain the directional response vector; Construct a direction discriminator and input the direction response vector into the direction discriminator to obtain a vector damage distribution map.

7. The bridge bearing visual inspection method based on artificial intelligence as described in claim 6, characterized in that: The process of generating a bridge bearing structure inspection report includes determining the design parameters of the bridge bearing and comparing the design parameters with a vector damage distribution map to obtain the comparison results. Based on the comparison results, a support stress model is constructed. The support stress model is used to evaluate the impact of damage on the bearing capacity and stability of the support, and the comprehensive results of the support status are output. A random forest model was used as a classifier to classify the comprehensive results of the bearing status, and multi-frame consistency was used to verify the classification results to obtain a bridge bearing structure inspection report.

8. A bridge bearing visual inspection system based on artificial intelligence, based on the bridge bearing visual inspection method based on artificial intelligence as described in any one of claims 1 to 7, characterized in that: include, The data acquisition module collects bridge bearing data and performs preprocessing on the bridge bearing data to obtain image sequences and load event timelines. The bridge bearing data includes real-time video streams under dynamic vehicle loads, static images, real-time displacement distances of the bridge bearing surfaces, and temperature and humidity of the bridge bearings. The image separation module constructs a spatiotemporal dual-stream network based on image sequences and load event timelines, extracting and fusing temporal and spatial feature maps. The fused features are temporally enhanced, decoded, and upsampled to generate a dynamic mask matrix. Based on the dynamic mask matrix, dynamic interference regions are marked and removed to obtain the separated structural deformation image sequence. The distribution map generation module calculates the response intensity of each directional convolution kernel to the structural deformation image based on the separated structural deformation image sequence using a pre-constructed group of directional convolution kernels, generates a multi-channel response map, compares the directional responses, and obtains a vector damage distribution map. The report integration module correlates and compares the vector damage distribution map with the design parameters of the bridge bearing to assess the impact of damage on the bearing capacity and stability of the bridge bearing and generate a bridge bearing structural inspection report.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the artificial intelligence-based bridge bearing visual inspection method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the artificial intelligence-based visual inspection method for bridge bearings as described in any one of claims 1 to 7.

Images