Machine vision-based real-time monitoring system for fatigue cracking of welding seam of steel structure

Abstract

The invention relates to the technical field of machine vision structure health monitoring, and discloses a steel structure weld fatigue cracking real-time monitoring system based on machine vision. The system comprises a space-time registration and fusion module, a multi-scale feature analysis module, a health monitoring module, a crack deduction calculation module and a regulation and control strategy generation module. Performing space-time registration and pixel-level fusion through the visual data of the plurality of image sensors to generate a synchronous multi-source image stream; a multi-level welding seam characteristic spectrum is constructed through multi-scale characteristic analysis, and a welding seam structure knowledge base is dynamically updated; the knowledge base and the real-time characteristic spectrum are used for monitoring the welding seam health state, and abnormity is recognized; deducing a crack initiation position and an evolution path in combination with historical damage data; and real-time load information is fused to pre-estimate the remaining service life, and a structural integrity regulation and control strategy is generated online. According to the invention, high-precision fusion of the multi-source visual data and active prediction of the crack trend are realized, and the monitoring accuracy and the early warning capability are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of machine vision structural health monitoring, in particular to a steel structure weld fatigue cracking real-time monitoring system based on machine vision. BACKGROUND

[0002] The steel structure weld fatigue cracking monitoring has long relied on manual inspection or fixed sensors. In the prior art, a visual monitoring system is usually used, and a single camera or a plurality of independent cameras arranged simply are used. These methods only perform time synchronization or rough splicing on images, lack strict spatio-temporal alignment and data fusion, and thus result in limited monitoring range and inaccurate feature extraction. The data sources are scattered and inconsistent, and it is difficult to form a comprehensive and reliable weld state representation.

[0003] The existing monitoring scheme usually only identifies and alarms the existing damage, and cannot predict the initiation and evolution trend of the crack. The historical damage data and the real-time monitoring information are disconnected, and the dynamic knowledge update is insufficient, so that the evaluation stays in the passive reaction stage, and the early warning capability is limited. The structure safety maintenance relies on post-processing, which increases the risk of sudden failure and maintenance cost.

[0004] The present application needs to solve the problem of information fragmentation caused by physical location difference of multi-source visual data, realize high-precision fusion to improve the monitoring accuracy. At the same time, the accumulated data resources are used to realize the active deduction of the crack initiation position and evolution path, and promote the health monitoring from state recognition to trend prediction. SUMMARY

[0005] The present application aims to provide a steel structure weld fatigue cracking real-time monitoring system based on machine vision to solve the problems proposed in the background.

[0006] To achieve the above-mentioned purpose, the present application provides a steel structure weld fatigue cracking real-time monitoring system based on machine vision, which comprises: a spatio-temporal registration and fusion module, which performs spatio-temporal registration and pixel-level fusion on visual data from a plurality of image sensors at different physical locations to generate a synchronous multi-source image stream; a multi-scale feature analysis module, which performs multi-scale feature analysis and enhancement on the synchronous multi-source image stream to construct a multi-level weld feature atlas; based on the multi-level weld feature atlas, a weld structure dynamic knowledge base is updated and maintained through a dynamic learning mechanism; a health monitoring module, which utilizes the weld structure dynamic knowledge base and the multi-level weld feature atlas acquired in real time to perform multi-dimensional monitoring and abnormality identification on the real-time health status of the steel structure weld; a crack deduction calculation module, which performs deduction calculation of the crack initiation position and evolution path according to the real-time health status monitoring result combined with the historical damage data; The control strategy generation module dynamically estimates the remaining service life of the weld based on the inference calculation results of the crack initiation position and evolution path and the fusion of real-time load information; and generates an online structural integrity control strategy according to the remaining service life estimation results of the weld and the real-time health state monitoring results.

[0007] Preferably, the specific steps of spatiotemporal registration and pixel-level fusion of visual data from multiple image sensors at different physical positions to generate a synchronous multi-source image stream include: Obtaining a sequence of original images captured by an image sensor array composed of a visible light camera, a thermal imager, and a three-dimensional profilometer; Using a pre-calibrated spatial position relationship and time stamp of the sensors, performing spatial geometric correction and time axis alignment on the sequence of original images, so that the data of different sensors for the same weld point are consistent in the spatiotemporal dimension; Mapping the aligned multi-modal image data, including intensity information, temperature field information, and three-dimensional point cloud information, to a unified reference coordinate system; In the reference coordinate system, using a weighted fusion method based on pixel confidence, the intensity information, temperature field information, and three-dimensional point cloud information are combined into a synchronous multi-source image stream with multi-dimensional attributes.

[0008] Preferably, the specific steps of performing multi-scale feature analysis and enhancement on the synchronous multi-source image stream to construct a multi-level weld feature map include: Receiving the synchronous multi-source image stream and performing multi-resolution pyramid decomposition to obtain image subbands covering multiple scales from macrostructure to microtexture; For each scale of the image subband, respectively applying a directional filter bank and a local binary pattern operator to extract edge orientation features and surface texture roughness features of the weld area; Cross-scale correlation and aggregation of the edge orientation features and the surface texture roughness features extracted at different scales to form a primary feature set with spatial hierarchical relationship; Introducing an attention weight mechanism to adaptively enhance feature channels with high correlation to fatigue damage in the primary feature set and suppress irrelevant background features; Reorganizing the enhanced features according to spatial position and scale level and encoding them into a structured multi-level weld feature map.

[0009] Preferably, the specific steps of updating and maintaining a weld structure dynamic knowledge base based on the multi-level weld feature map through a dynamic learning mechanism include: Initialize a weld structure dynamic knowledge base containing weld topology connection relationship, material attribute baseline and historical typical defect mode; Decode the geometric feature vector and surface state feature vector of the current weld from the real-time input multi-level weld feature map; Differentially compare the geometric feature vector and the surface state feature vector with the baseline features stored in the weld structure dynamic knowledge base point by point to generate a feature change vector; Use the incremental learning method to fuse the feature change vector into the weld structure dynamic knowledge base to update the state description and historical trajectory of the corresponding weld entry; According to the long time series of the feature change vector, learn the implicit law of weld state evolution through a probabilistic graph model, and correct the state transition probability in the weld structure dynamic knowledge base accordingly.

[0010] Preferably, the specific steps of using the weld structure dynamic knowledge base and the multi-level weld feature map obtained in real time to monitor the real-time health status of the steel structure weld and identify abnormalities include: Real-time receive the multi-level weld feature map, and call the baseline features and state evolution probability of the corresponding weld from the weld structure dynamic knowledge base; Parallelly compare the real-time features and baseline features in multiple preset damage sensitive dimensions, including geometric deformation, texture anomaly and local thermal gradient; Accumulate the comparison differences in multiple dimensions using a sliding time window to form a state difference sequence evolving over time; Model the state difference sequence through a trained deep state evolution network, which outputs a comprehensive health score and multiple abnormal confidence levels in sub-dimensions; When the health score is lower than the dynamic threshold or any of the abnormal confidence levels exceeds its corresponding threshold, it is determined that the weld area is abnormal, and a monitoring event containing the abnormal type and location is generated.

[0011] Preferably, the specific steps of performing crack initiation location and evolution path deduction calculation according to the real-time health status monitoring result combined with historical damage data include: Receive the monitoring event generated from the real-time health status monitoring step and extract the local high-resolution feature area related to the event; Retrieve the historical damage data and material fatigue performance parameters of the weld from the weld structure dynamic knowledge base; Taking the monitoring event position as an initial node, a time-space deduction graph is constructed with the weld microstructure as a guide; on the time-space deduction graph, a crack propagation potential function is defined based on strain energy density and stress intensity factor, and the crack propagation potential function is used to calculate the instantaneous probability of crack propagation in different directions; A set of possible evolution path sets is generated by using a Monte Carlo simulation method, combining the material fatigue performance parameters with the real-time stress spectrum to simulate the crack initiation and propagation process multiple times; The evolution path sets are clustered and analyzed for convergence to extract the high-probability crack initiation area and the dominant evolution direction, forming a crack deduction report.

[0012] Preferably, the specific steps of dynamically estimating the remaining service life of the weld based on the crack initiation position and evolution path deduction calculation results include: The crack deduction report is obtained, from which the high-probability crack initial size, position and dominant propagation direction are analyzed; the real-time load information provided by the structural health monitoring system is simultaneously accessed, including dynamic stress spectrum, cycle number and environmental temperature; A micro-mechanical constitutive equation is established to describe the damage accumulation process of the weld material under cyclic loading; the crack size expansion process with load cycles is iteratively calculated by numerically solving the micro-mechanical constitutive equation, taking the crack initial size as the starting point and the real-time load information as the input; The crack propagation process is coupled with the component residual carrying capacity model, and when the residual carrying capacity decreases to the safety threshold, the corresponding load cycle number is the predicted remaining service life; According to the real-time newly added monitoring data, the parameters of the micro-mechanical constitutive equation are calibrated online, and the predicted remaining service life value is dynamically updated accordingly.

[0013] Preferably, the specific steps of generating an online structural integrity control strategy according to the remaining service life estimation results of the weld and the real-time health state monitoring results include: The updated remaining service life value from the remaining service life dynamic estimation step and the continuous health score and abnormal confidence from the real-time health state monitoring step are received; An optimization objective function is constructed to extend the overall structure safe operation time, and the constraint conditions of the optimization objective function include the differences in the remaining service life of the current welds, real-time load distribution and adjustable structure boundary conditions; Taking the health score and the abnormal confidence as weights, a set of recommended load distribution adjustment amounts or local stiffness modification suggestions is solved in the solution space of the optimization objective function; The load distribution adjustment amount or local rigidity modification suggestion is encoded as structural integrity regulation instructions that can be understood by an execution mechanism or management process.

[0014] Preferably, the system further comprises a step of deep feature learning on the synchronous multi-source image stream, specifically: A multi-level cascaded deep feature learning network is established, and the input of the deep feature learning network is the synchronous multi-source image stream; The deep feature learning network is trained in an end-to-end manner, and the output is a deep weld seam descriptor containing rich semantic information; the deep weld seam descriptor will be supplemented as high-level features to the multi-level weld seam feature atlas.

[0015] Preferably, the first sub-network of the deep feature learning network focuses on segmenting the weld body area and the background to generate a pixel-level segmentation mask; the second sub-network takes the region of interest constrained by the segmentation mask as input to perform semantic segmentation of subtle texture abnormalities and micro-defects; the third sub-network integrates the outputs of the first two sub-networks to construct a complete feature pyramid from pixels to semantics.

[0016] Compared with the prior art, the beneficial effects of the present application are: Through spatio-temporal registration and pixel-level fusion technology, the data from multiple image sensors at different physical locations are aligned in time and space dimensions and fused into a synchronous multi-source image stream. The errors introduced by the perspective deviation and the acquisition time delay are eliminated, and an image sequence with high spatial consistency and information density is generated. The subsequent feature analysis is based on unified and accurate input, and the subtle features of the weld surface are enhanced and extracted, and the sensitivity and reliability of the monitoring system for early fatigue signs are improved.

[0017] Based on the dynamically updated weld structure knowledge base, the real-time multi-level feature atlas and the historical damage record are integrated, and the deduction calculation of the crack initiation position and evolution path is performed. The fatigue damage accumulation law is simulated, and the crack development trend is quantified. The monitoring function is expanded from abnormal alarm to potential risk prediction, and the structure integrity evaluation has foresight. The maintenance decision can be based on the deduction result to deploy intervention measures in advance, optimize resource allocation, and prolong the safe service period of the weld. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 The timing diagram of the steel structure weld fatigue cracking real-time monitoring system based on machine vision according to the present application; Figure 2 The flowchart of the spatio-temporal registration and pixel-level fusion step; Figure 3 The flowchart of the weld structure dynamic knowledge base updating and maintenance; Figure 4characteristics diagram of the remaining service life estimation accuracy changing with the amount of monitoring data; Figure 5 performance analysis diagram of the feature pyramid level. DETAILED DESCRIPTION

[0019] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work are within the protection scope of the present application.

[0020] Please refer to Figure 1 The present application provides a steel structure weld fatigue cracking real-time monitoring system based on machine vision, which comprises a space-time registration and fusion module, a multi-scale feature analysis module, a health monitoring module, a crack deduction calculation module and a regulation strategy generation module. The system performs space-time registration and pixel-level fusion on visual data from multiple image sensors in different physical locations to generate a synchronous multi-source image stream. Multi-scale feature analysis and enhancement are performed on the synchronous multi-source image stream to construct a multi-level weld feature atlas, and a weld structure dynamic knowledge base is updated and maintained based on the multi-level weld feature atlas through a dynamic learning mechanism. The real-time health status of the steel structure weld is monitored and abnormally identified using the weld structure dynamic knowledge base and the multi-level weld feature atlas obtained in real time. According to the real-time health status monitoring result combined with historical damage data, the deduction calculation of the crack initiation position and evolution path is performed. Based on the deduction calculation result of the crack initiation position and evolution path and the real-time load information, the remaining service life of the weld is dynamically estimated, and the structure integrity regulation strategy is generated online according to the remaining service life estimation result and the real-time health status monitoring result of the weld.

[0021] Embodiment 1: Please refer to Figure 2The space-time registration and fusion module acquires the original image sequence captured by the image sensor array composed of the visible light camera, the thermal imager and the three-dimensional profilometer, and performs spatial geometric correction and time axis alignment on the original image sequence by using the pre-calibrated spatial position relationship and time stamp of the sensors, so that the data of different sensors for the same weld point are consistent in space-time dimension. The aligned multi-modal image data including intensity information, temperature field information and three-dimensional point cloud information are mapped to a unified reference coordinate system, and in the reference coordinate system, a weighted fusion method based on pixel confidence is used to synthesize the intensity information, the temperature field information and the three-dimensional point cloud information into a synchronous multi-source image stream with multi-dimensional attributes. The multi-scale feature analysis module receives the synchronous multi-source image stream and performs multi-resolution pyramid decomposition thereon to obtain image subbands of multiple scales covering from macrostructure to microtexture, and applies a directional filter bank and a local binary pattern operator to each scale of image subband to extract edge orientation features and surface texture roughness features of the weld area. The edge orientation features and the surface texture roughness features extracted at different scales are cross-scale correlated and aggregated to form a primary feature set with spatial hierarchical relationship, an attention weight mechanism is introduced to adaptively enhance the feature channels with high correlation degree to fatigue damage in the primary feature set and suppress the interference of irrelevant background features. The enhanced features are reorganized according to the spatial position and scale level, and are encoded into a structured multi-level weld feature map.

[0022] In a specific implementation, inside the bridge steel box girder, along a key longitudinal weld, an image sensor array composed of a high-resolution visible light camera, an infrared thermal imager and a laser three-dimensional profilometer is deployed, the visible light camera, the thermal imager and the three-dimensional profilometer are fixedly installed at a specific angle and interval in space to cover the same weld area about 2 meters long. The image sensor array synchronously triggers to capture the original image sequence at a rate of 10 frames per second, and the original image sequence includes visible light intensity images, infrared temperature images and three-dimensional point cloud data.

[0023] In a specific implementation, the spatio-temporal registration and fusion module acquires the original image sequences captured by the visible light camera, the thermal imager and the three-dimensional profilometer, each frame of which is provided with a time stamp accurate to microseconds. The spatial geometric correction of the original image sequences is performed by using the pre-calibrated spatial position relationship of the sensors in the installation phase through a high-precision calibration board. The process of the spatial geometric correction is to project the image data captured by the thermal imager and the three-dimensional profilometer to the reference coordinate system with the light center of the visible light camera as the origin according to the rotation and translation matrix between the thermal imager and the three-dimensional profilometer. The time axis alignment is to synchronize the original image sequence data at different sampling times to a unified time reference point by using an interpolation algorithm according to the time stamp carried by the image data, so that the data of different sensors for the same physical point on the weld surface are synchronized in spatial position and time dimension. The aligned multi-modal image data, including the intensity information provided by the visible light camera, the temperature field information provided by the thermal imager and the three-dimensional point cloud information provided by the three-dimensional profilometer, are all mapped to the same two-dimensional reference coordinate system with the starting point of the weld as the origin. In this reference coordinate system, a weighted fusion method based on pixel confidence is used to synthesize the intensity information, the temperature field information and the three-dimensional point cloud information. The pixel confidence is dynamically calculated and determined by the signal-to-noise ratio of each sensor under the current environmental light and temperature conditions. The mathematical expression form of the weighted fusion is: ; wherein: represents the multi-dimensional attribute vector of the fused synchronous multi-source image stream at the coordinate , represents the number of sensor types participating in the fusion, represents the feature data provided by the th sensor after correction and alignment, represents the real-time calculated pixel confidence weight corresponding to , and satisfies . The synthesis result is to generate a synchronous multi-source image stream containing pixel position, gray scale, temperature and surface height multi-dimensional attributes, which is used to represent the comprehensive state of the weld area.

[0024] In some embodiments, the multi-scale feature analysis module receives the synchronized multi-source image stream generated by the spatio-temporal registration and fusion module, which is first sent into a multi-resolution pyramid decomposition process. The multi-resolution pyramid decomposition process performs Gaussian blur and down-sampling operations on each dimensional attribute in the synchronized multi-source image stream, respectively, to generate a series of image subbands with successively reduced resolutions, which collectively constitute a multi-scale representation from the original resolution to the lowest resolution. For each scale of image subbands in the pyramid, a set of directional filter banks and local binary pattern operators are applied for processing, respectively. The directional filter banks are composed of multiple Gabor filters with different directions, which are used to extract the edge orientation features of the weld region in various directions. The local binary pattern operators are used to calculate the texture coding in the local neighborhood of each pixel, which is used to quantify the texture roughness features of the weld surface. The edge orientation features and surface texture roughness features extracted from different scale image subbands are cross-scale associated and aggregated through up-sampling and stitching operations of the feature maps, which ensures that the macroscopic deformation features and microscopic texture change features can correspond in spatial position, thereby forming a primary feature set with a clear spatial hierarchical relationship. An attention weight mechanism is introduced to modulate the primary feature set, which calculates the contribution of each feature channel to fatigue damage through a small neural network to generate a corresponding attention weight map. The attention weight map is used to adaptively enhance the feature channels with high relevance to fatigue damage in the primary feature set, while suppressing the response intensity from irrelevant background feature channels such as weld background, spatter or oil stains. Finally, all features enhanced by the attention mechanism are reorganized and encoded according to their original spatial position coordinates and scale hierarchy, and encoded into a structured multi-level weld feature map that contains spatial and scale dimension information. The multi-level weld feature map serves as the input of subsequent modules.

[0025] In specific implementations, spatio-temporal registration and fusion ensure the spatio-temporal consistency of multi-source data, while multi-scale feature analysis constructs feature descriptions covering different levels of detail from consistent data. In some embodiments, the number of directions of the directional filter bank and the neighborhood radius of the local binary pattern operator can be preset according to the typical texture features of the weld. Optionally, the number of layers of the multi-resolution pyramid can be adjusted according to the available computing resources of the monitoring system and the requirements for microscopic defect detection accuracy. The structured format of the multi-level weld feature map is conducive to efficient storage and retrieval of the weld structure dynamic knowledge base.

[0026] Embodiment 2: see Figure 3, the weld structure dynamic knowledge base is initialized to include a weld topological connection relationship, a material attribute baseline, and a historical typical defect mode, geometric morphology feature vectors and surface state feature vectors of the current weld are decoded from the real-time input multi-level weld feature map. The geometric morphology feature vectors and the surface state feature vectors are compared with the baseline features stored in the weld structure dynamic knowledge base point by point, a feature change vector is generated, and the feature change vector is fused into the weld structure dynamic knowledge base by using an incremental learning method, and the state description and the historical trajectory of the corresponding weld item are updated. According to the long time sequence feature change vector, the implicit law of weld state evolution is learned through a probability graph model, and the state transition probability in the weld structure dynamic knowledge base is corrected accordingly. The health monitoring module receives the multi-level weld feature map in real time and calls the baseline features and state evolution probability of the corresponding weld from the weld structure dynamic knowledge base, and compares the real-time features and the baseline features in multiple preset damage sensitive dimensions, including geometric deformation, texture anomaly and local thermal gradient. A sliding time window is used to accumulate the differences in multiple dimensions to form a time-evolving state difference sequence, and a trained deep state evolution network is used to model the state difference sequence, and the deep state evolution network outputs a comprehensive health score and multiple abnormal confidence levels in sub-dimensions. When the health score is lower than the dynamic threshold or any abnormal confidence exceeds the corresponding threshold, it is determined that the weld area is abnormal, and a monitoring event including the abnormal type and location is generated.

[0027] In a specific implementation, the multi-scale feature analysis module continuously outputs a structured multi-level weld feature map, the weld structure dynamic knowledge base is initialized at system startup, and the initialization process loads the design drawings and manufacturing archives of the target steel box girder weld to construct a digital knowledge base including weld topological connection relationship, material attribute baseline and historical typical defect mode. The weld topological connection relationship describes the geometric connection and mechanical transmission path between the welds, the material attribute baseline stores the elastic modulus, Poisson's ratio and fatigue strength coefficient parameters of the steel, and the historical typical defect mode includes the feature templates of cracks, pores and slag inclusion defects found in previous detection of similar components. In a specific implementation, the geometric morphology feature vectors and surface state feature vectors of the current weld are decoded from the real-time input multi-level weld feature map through a fixed decoding network, the geometric morphology feature vectors include the quantitative values of the weld's excess height, width and undercut depth parameters, and the surface state feature vectors include the texture consistency and local curvature distribution statistics. In each monitoring period, the geometric morphology feature vectors and surface state feature vectors decoded in real time are compared with the baseline features stored in the weld structure dynamic knowledge base for the weld, the difference or ratio in the corresponding feature dimensions is calculated, and a feature change vector describing the feature change direction and amplitude is generated.

[0028] The feature change vector is fused into the weld structure dynamic knowledge base by using an incremental learning method, the incremental learning method updates the state description of the corresponding weld item in the form of a sliding window, and the feature change vector is sequentially attached to the historical trajectory record of the weld item. According to the feature change vector accumulated in a long time sequence, the implicit law of weld state evolution is learned through a probabilistic graph model, the probabilistic graph model uses a hidden Markov model, the observation state is the cluster category of the feature change vector, and the hidden state is the invisible health level of the weld. The state transition probability matrix and the observation probability matrix are learned from the historical data through the Baum-Welch algorithm, and the preset value of the state transition probability in the weld structure dynamic knowledge base is corrected accordingly. The specific implementation of the hidden Markov model involves systematic processing of the feature change vector accumulated in a long time sequence to learn the implicit law of weld state evolution. First, the feature change vector is derived from the point-by-point difference comparison between the real-time decoded geometric feature vector, the surface state feature vector and the baseline feature in the weld structure dynamic knowledge base. These vectors are classified into discrete observation state categories by a clustering algorithm, representing different surface abnormal patterns or change degrees. At the same time, the hidden state is defined as the invisible health level of the weld, such as multiple discrete levels from perfect, slight damage to severe damage, thereby constructing a hidden Markov model framework. Then, the feature change vector sequence in the historical data is trained. The algorithm estimates the state transition probability matrix and the observation probability matrix through an iterative optimization process. After training, the system corrects the preset state transition probability in the weld structure dynamic knowledge base according to the learned probability matrix, so that the knowledge base can more accurately reflect the dynamic evolution trend of the weld state, providing adaptive reference for real-time health monitoring.

[0029] In some embodiments, the health monitoring module receives the multi-level weld feature map from the multi-scale feature parsing module in real time, and calls the baseline feature of the corresponding weld item and the latest learned state evolution probability from the weld structure dynamic knowledge base. The real-time features decoded from the real-time multi-level weld feature map are compared with the baseline features called from the weld structure dynamic knowledge base in parallel in multiple preset damage sensitive dimensions, including geometric deformation, texture anomaly and local thermal gradient. The geometric deformation dimension compares the difference between the weld contour and the reference contour, the texture anomaly dimension compares the difference in statistical characteristics of the surface texture pattern, and the local thermal gradient dimension compares the temperature distribution gradient difference between the weld area and the base material area. A fixed time length sliding time window is used to accumulate the multi-dimensional comparison differences, and the difference value of each dimension in the time window constitutes a time sequence, forming a time-evolving state difference sequence. A pre-trained deep state evolution network is used to model the state difference sequence. The deep state evolution network adopts the architecture of one-dimensional convolutional neural network combined with long short-term memory network. The one-dimensional convolutional neural network extracts local patterns in the state difference sequence, and the long short-term memory network captures the time sequence dependence of the state difference. The output layer of the deep state evolution network outputs a comprehensive health score between 0 and 1 and multiple sub-dimension anomaly confidence. The health monitoring module is provided with a dynamic threshold generation mechanism. The dynamic threshold is calculated according to the statistical distribution of the historical health score and the current load condition. When the health score is lower than the calculated dynamic threshold or the anomaly confidence of any dimension exceeds the corresponding threshold set independently, the system determines that the weld area is abnormal, and immediately generates a structured monitoring event. The monitoring event includes the type of anomaly, the coordinates of the abnormal position, the timestamp and the related feature snapshot.

[0030] It can be understood that the weld structure dynamic knowledge base serves as a dynamic reference system for real-time health state monitoring, providing personalized baseline. In some embodiments, the incremental learning update formula of the feature change vector can be expressed as: ; Wherein: represents the updated weld knowledge feature vector at time t, represents the knowledge feature vector at time t, represents the feature change vector calculated at time t, represents the feature change vector calculated at time t, represents the feature change vector calculated at time t, represents the feature change vector calculated at time t, is a forgetting factor between 0 and 1, which is used to control the weight proportion of historical information and current change. Optionally, the training data of the deep state evolution network is derived from historical monitoring data and labeled data of laboratory fatigue tests of the same type of structure. The process of multi-dimensional parallel comparison allows the system to focus on early signs of different types of damage at the same time, avoiding single feature false negatives.

[0031] In embodiment 3, the crack evolution calculation module receives the monitoring events generated from the real-time health state monitoring step and extracts the local high-resolution feature area related to the events, retrieves the historical damage data and material fatigue performance parameters of the weld from the weld structure dynamic knowledge base. Taking the monitoring event position as the initial node and the weld microstructure as the guide, a space-time evolution graph is constructed, a crack propagation potential function based on strain energy density and stress intensity factor is defined on the space-time evolution graph, and the instantaneous probability of crack propagation in different directions is calculated using the propagation potential function. The Monte Carlo simulation method is used to combine the material fatigue performance parameters and the real-time stress spectrum to simulate the crack initiation and propagation process for multiple rounds, generate a set of evolution path sets, and extract the high-probability crack initiation area and dominant evolution direction through clustering and convergence analysis of the evolution path sets, forming a crack evolution report. The regulation strategy generation module obtains the high-probability crack initial size, position, and dominant propagation direction from the crack evolution report, and synchronously accesses the real-time load information provided by the structure health monitoring system, including dynamic stress spectrum, cycle number, and environmental temperature. A micro-mechanical constitutive equation is established to describe the damage accumulation process of the weld material under cyclic loading, and the micro-mechanical constitutive equation is solved from the crack initial size as the starting point and the real-time load information as the input value, and the crack size expansion process with load cycles is iteratively calculated. The crack propagation process is coupled with the component residual carrying capacity model, and when the residual carrying capacity decreases to the safety threshold, the corresponding load cycle number is the predicted residual service life. According to the real-time monitoring data, the parameters of the micro-mechanical constitutive equation are calibrated online, and the predicted residual service life value is dynamically updated accordingly.

[0032] In a specific implementation, the inference calculation related to crack initiation and evolution and the dynamic prediction of the remaining service life are involved. The specific implementation process takes the scenario generated by the health monitoring module to generate monitoring events. The crack inference calculation module receives monitoring events generated from the real-time health state monitoring step. The monitoring events include the abnormal type "local micro-crack initiation" and the precise two-dimensional position coordinates and the corresponding abnormal confidence. The crack inference calculation module extracts the event-related local high-resolution feature area from the multi-level weld feature map according to the position coordinates. The local high-resolution feature area focuses on a range of about 10 mm x 10 mm around the abnormal position and includes complete multi-scale features from macro morphology to micro texture. The history damage data of the weld item at the specific position and the material fatigue performance parameters are retrieved from the weld structure dynamic knowledge base. The history damage data includes the evolution log of the past similar position recorded micro-defects, and the material fatigue performance parameters include the fracture toughness KIC and the fatigue crack propagation rate parameter.

[0033] In a specific implementation, the abnormal position marked by the monitoring event is taken as the initial node, and the microstructure orientation of the weld and the local principal stress direction obtained by finite element analysis are used as the guide to construct a space-time inference graph. The nodes of the space-time inference graph represent the positions of the material grain boundaries or potential weak points, and the edges represent the physical possibility of crack propagation in a certain direction within a time step. On the space-time inference graph, a crack propagation potential function based on strain energy density and stress intensity factor is defined. The crack propagation potential function is used to quantify the driving energy of crack propagation to different adjacent nodes under the current load state. The calculation of the crack propagation potential function integrates the strain energy density release rate at the node and the stress intensity factor amplitude calculated from the current crack morphology and the real-time stress field. The instantaneous probability of crack propagation in different directions is calculated using the crack propagation potential function, and the instantaneous probability is proportional to the value of the crack propagation potential function. The Monte Carlo simulation method is used to simulate the crack initiation and propagation process in multiple rounds, combined with the material fatigue performance parameters obtained from the weld structure dynamic knowledge base and the real-time stress spectrum accessed from the structure health monitoring system in real time. Each round of Monte Carlo simulation represents a random crack evolution process. The simulation process gradually advances the crack front according to the principles of fracture mechanics, generating an evolution path set containing thousands of evolution paths. Each evolution path records the change of crack length and direction with the number of load cycles. The evolution path set is clustered and analyzed for convergence based on spatial density and direction similarity. The high-probability crack initiation area and the most common dominant evolution direction in subsequent propagation are extracted from the main cluster. These information is summarized and formatted to form a crack inference report containing probabilistic prediction results.

[0034] In some embodiments, the regulation strategy generation module obtains a crack propagation report generated by the crack propagation calculation module, the crack propagation report is transmitted in a structured data format, and the regulation strategy generation module parses a high-probability crack initial size, position and dominant expansion direction therefrom. Real-time load information provided by the structural health monitoring system is accessed synchronously, the real-time load information including a dynamic stress spectrum, stress cycle number and environmental temperature, and the dynamic stress spectrum provides cycle information of a principal stress amplitude and mean value in a time series form. A microscopic mechanics constitutive equation describing a damage accumulation process of a weld material under a cyclic load is established, the microscopic mechanics constitutive equation characterizing a quantitative relationship between a crack propagation rate and a mechanics parameter. The crack initial size parsed from the crack propagation report is taken as a starting point, and the dynamic stress spectrum accessed in real time is taken as an input, and the microscopic mechanics constitutive equation is numerically solved, and an iterative loop algorithm is used in a numerical solving process, and a crack size increment is calculated in each load cycle step. The crack propagation process is coupled with a component residual carrying capacity model, the component residual carrying capacity model calculating a net section stress and fracture toughness of the component according to the crack size, position and direction, and when the residual carrying capacity of the component calculated by iteration reduces to a safety threshold, a corresponding load cycle number is the predicted residual service life, and the predicted residual service life is output in the form of a cycle number or an equivalent time. The microscopic mechanics constitutive equation of the crack propagation is described in a differential form: ; wherein: represents a crack propagation rate, i.e. a crack length change rate with a load cycle number is a stress intensity factor range under a current cycle, and are constants related to a material and an environment. According to real-time newly added monitoring data and crack propagation calculation results, material parameters in the microscopic mechanics constitutive equation are calibrated online, the online calibration being realized by comparing a predicted value of the crack size with an observed value based on a multi-level weld feature spectrum inversion, and a predicted residual service life value is dynamically updated accordingly.

[0035] It can be understood that the crack propagation calculation converts an abnormal monitoring event of a single point into a spatiotemporal probability prediction of a future development path of the crack. A Monte Carlo simulation method handles inherent material and load randomness in the crack propagation. The component residual carrying capacity model provides a bridge from the crack size to functional failure. An online calibration mechanism enables the life prediction model to adapt to changes in material performance. Optionally, a specific form of the crack propagation potential function can be adjusted according to anisotropic properties of the material.

[0036] ​Example 4: The control strategy generation module receives the updated remaining service life value from the dynamic remaining service life prediction step and the continuous health score and anomaly confidence level from the real-time health status monitoring step. It constructs an optimization objective function aimed at extending the overall structural safe operating time. The constraints of the objective function include the current differences in the remaining service life of each weld, real-time load distribution, and adjustable structural boundary conditions. Using the health score and anomaly confidence level as weights, it optimizes within the solution space of the objective function to obtain a set of recommended load distribution adjustments or local stiffness modification suggestions. These load distribution adjustments or local stiffness modification suggestions are encoded into structural integrity control instructions that can be understood by the executing mechanism or management process.

[0037] In practical implementation, a structural integrity control strategy is generated online based on the remaining service life prediction results and real-time health status monitoring results. The specific implementation process is described using the health management of a welded ball joint in a large stadium as an example. The control strategy generation module receives the updated remaining service life value from the dynamic remaining service life prediction step. The updated remaining service life value is given in the form of load cycle count or equivalent calendar days. Simultaneously, the control strategy generation module continuously receives health scores and anomaly confidence levels from multiple sub-dimensions output by the real-time health status monitoring step. In practical implementation, the control strategy generation module constructs an optimization objective function aimed at extending the overall structural safe operating time. The specific form of the optimization objective function is to minimize the overall structural risk index. ; in: This represents the overall structural risk indicator to be minimized. This indicates the total number of monitored critical welds. Indicates weld seam Weighting coefficients, weighting coefficients The importance factor of the weld in the structural system determines this. Indicates weld seam Comprehensive abnormal factors, comprehensive abnormal factors The health score output from the real-time health status monitoring steps and the anomaly confidence scores of each dimension are calculated through weighted fusion. Indicates weld seam Normalized remaining service life, normalized remaining service life is the ratio of the updated remaining service life prediction and the design life of the weld. The constraints of the optimization objective function include the difference of the remaining service life of each weld, the real-time load distribution described by the member axial force and bending moment data provided by the structural health monitoring system, and the adjustable boundary conditions of the structure, including the range of movement of the partially movable supports or the stiffness setting of the adjustable dampers. The comprehensive abnormal factor As a key input, a set of recommended load distribution adjustment or local stiffness modification suggestions is solved within the solution space of the optimization objective function that satisfies the structural mechanics balance and displacement constraints. The solving process uses a nonlinear programming algorithm with constraints.

[0038] In some embodiments, the recommended measures obtained by solving are organized and presented in a structured data table form, which summarizes the control suggestions and their expected effects for different welds. Subsequently, the control strategy generation module encodes the information in this table into structural integrity control instructions that can be understood by the execution mechanism or management process. The encoding process converts specific parameter adjustment amounts into device control instructions or maintenance work orders, such as converting "increase the prestress of member A-5 by 15 kN" into the corresponding hydraulic tensioning device control protocol instruction, or converting "add temporary support near node J" into a maintenance work order with accompanying drawings and construction points. The structural integrity control instructions are sent to the structural health management system or facility management platform of the stadium through the industrial network, used to guide manual intervention or trigger automatic control systems. Referring to Table 1, a simplified weld ball node network control strategy table is shown.

[0039] Table 1: Load distribution and maintenance strategy table for weld ball node network

[0040] It can be understood that the optimization objective function converts life prediction and health status into a quantifiable risk minimization problem. The comprehensive abnormal factor is the ratio of the normalized remaining life directly reflects the immediate risk level of the weld. The optimization solving process needs to find the most effective way to reduce the overall risk under complex structural constraints. Optionally, the solution of the optimization objective function can be set to run periodically or triggered by specific events. The structural integrity control instructions generated by encoding are a key step to realize the closed loop from state perception to maintenance decision. The setting of the weight coefficient can be pre-configured based on the analysis of the influence of weld failure on the overall structural stability.

[0041] Referring to Figure 4In the monitoring analysis of the remaining service life prediction accuracy, the correlation between the accuracy variation law and the monitoring data quantity is presented by combining the trend curve and the error band. In the specific operation, the monitoring data quantity (pieces) is taken as the horizontal axis, the remaining life prediction accuracy (0-1) is taken as the vertical axis, the core curve represents the prediction accuracy level under different data quantities, and the light blue area is the error band corresponding to the accuracy. In the actual analysis, when the monitoring data quantity increases from the initial value to 200, the prediction accuracy shows a rapid upward trend; when the data quantity exceeds 200, the accuracy increases gradually and tends to be stable, and finally maintains at a high level of about 0.95 when the data quantity reaches 1000. The dynamic change of the error band reflects the uncertainty of the prediction result: when the data quantity is low, the error band range is large, and as the data quantity accumulates, the error band width gradually narrows, indicating that the increase of the data quantity effectively reduces the dispersion of the prediction result. In terms of parameter dimension, the analysis corresponds to the remaining life prediction link of stage 8, and the quantitative benchmark of the accuracy index is the 0-1 interval (1 represents complete accuracy), and the boundary of the error band is determined by the accuracy extreme value of multiple parallel tests.

[0042] In embodiment 5, the system further comprises a step of deep feature learning of the synchronous multi-source image stream, and a multi-level cascaded deep feature learning network is established, and the input of the deep feature learning network is the synchronous multi-source image stream. The deep feature learning network is trained in an end-to-end manner, and the output of the deep feature learning network is a deep weld seam descriptor containing rich semantic information, and the deep weld seam descriptor is supplemented as high-level features into the multi-level weld seam feature atlas. The first sub-network of the deep feature learning network focuses on segmenting the weld body area and the background to generate a pixel-level segmentation mask, the second sub-network takes the region of interest constrained by the segmentation mask as the input to perform semantic segmentation of subtle texture anomalies and micro defects, and the third sub-network integrates the outputs of the first two sub-networks to construct a complete feature pyramid from pixels to semantics.

[0043] In a specific implementation, the system establishes a multi-level cascaded deep feature learning network, and the deep feature learning network is a specially designed convolutional neural network architecture. The input of the deep feature learning network is the synchronous multi-source image stream generated by the time and space registration and fusion module, and the synchronous multi-source image stream is a data cube containing pixel intensity, temperature and three-dimensional coordinate multi-dimensional attributes. The deep feature learning network is trained in an end-to-end manner, and the training data is composed of historical synchronous multi-source image streams and manually labeled defect position and class labels. The output of the forward propagation is a high-dimensional deep weld seam descriptor containing rich semantic information. The deep weld seam descriptor is supplemented as high-level features into the multi-level weld seam feature atlas, and is spliced and fused with the features extracted by the traditional method of directional filter bank and local binary pattern operator, thereby forming a more comprehensive weld seam state representation.

[0044] In some embodiments, the multi-layer cascaded deep feature learning network is composed of three sub-networks connected sequentially in structure and function. The first stage sub-network of the deep feature learning network focuses on segmenting the weld body region from the background. The first stage sub-network adopts an encoder-decoder structure, the encoder extracts abstract features through multi-layer convolution and down-sampling, and the decoder recovers spatial details through up-sampling and skip connection, finally outputs a pixel-level segmentation mask with the same spatial resolution as the input image, which classifies each pixel in the image as weld body or background. The second stage sub-network of the deep feature learning network takes the region of interest constrained by the pixel-level segmentation mask generated by the first stage sub-network as input, and only calculates features for the weld body region indicated by the segmentation mask in space. The structure of the second stage sub-network contains multiple parallel hollow convolution branches to capture multi-scale context information, which performs semantic segmentation of subtle texture anomalies and micro-defects, and outputs a probability map of each weld pixel belonging to the "normal", "micro-crack", "porosity", "inclusion" categories. The third stage sub-network of the deep feature learning network integrates the outputs of the first two stage sub-networks. The third stage sub-network receives the intermediate layer feature maps of the first stage sub-network, the semantic segmentation probability map of the second stage sub-network, and the original synchronous multi-source image stream, and performs feature fusion through a top-down and horizontal connection feature pyramid network structure to build a complete feature pyramid from pixel-level details to high-level semantics. The top layer features of the complete feature pyramid are processed by global pooling and fully connected layer, and then mapped into the final deep weld descriptor The calculation process can be formally represented as: ; Wherein: represents the final deep weld descriptor, represents the global pooling and fully connected transformation function, represents the feature pyramid fusion function, , , represent the feature maps from the low, middle and high layers of the network respectively.

[0045] It can be understood that the multi-layer cascaded design enables the deep feature learning network to abstract and focus step by step. The generation of the pixel-level segmentation mask effectively eliminates the interference of complex background on subsequent fine analysis. The operation of the second stage sub-network in the region of interest significantly improves the calculation efficiency and the detection accuracy of small targets. The construction of the feature pyramid ensures that the visual information of different scales is effectively preserved and fused. Optionally, the training of the deep feature learning network can adopt a phased pre-training and joint fine-tuning strategy. As a powerful complement to traditional handcrafted features, the deep weld descriptor significantly enhances the expression ability of the multi-layer weld feature atlas.

[0046] Referring toFigure 5 In the feature pyramid level performance analysis, the multi-dimensional performance of different resolution levels in the feature fusion evaluation stage is focused. Specifically, the horizontal axis divides P1 (1 / 8 resolution) to P4 (original resolution) four feature pyramid levels, and the vertical axis includes three types of core indicators, including feature channel number, effective receptive field (pixels), and defect detection rate (%). Among them, the feature channel number is the left vertical axis, and the effective receptive field and the defect detection rate are the right vertical axis. From the index change law: the feature channel number remains stable and high value at each level, which reflects the balanced allocation of channel resources in multi-scale feature analysis; the effective receptive field increases significantly linearly with the resolution, reflecting the enhanced capture ability of high-resolution levels to larger spatial range information; the defect detection rate maintains a high level at each level and rises slightly, indicating that after cross-scale feature fusion, the identification performance of different resolution levels to defects is effectively guaranteed.

[0047] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A machine vision based real-time monitoring system for fatigue cracking in welds of steel structures, characterized in that, The system comprises: a space-time registration and fusion module for performing space-time registration and pixel-level fusion on visual data from multiple image sensors at different physical locations to generate a synchronous multi-source image stream; a multi-scale feature analysis module for performing multi-scale feature analysis and enhancement on the synchronous multi-source image stream to construct a multi-level weld feature map; based on the multi-level weld feature map, a weld structure dynamic knowledge base is updated and maintained through a dynamic learning mechanism; a health monitoring module for utilizing the weld structure dynamic knowledge base and the multi-level weld feature map obtained in real time to perform multi-dimensional monitoring and abnormality identification on the real-time health status of a steel structure weld; a crack deduction calculation module for performing deduction calculation on the location and evolution path of a crack based on the real-time health status monitoring result and historical damage data; a regulation strategy generation module for dynamically estimating the remaining service life of a weld based on the deduction calculation result of the location and evolution path of a crack and fusion of real-time load information; and generating a structure integrity regulation strategy online according to the remaining service life estimation result of the weld and the real-time health status monitoring result.

2. The machine vision based real time monitoring system for fatigue cracking in welds of steel structures as claimed in claim 1 wherein, The specific steps of the space-time registration and fusion on visual data from multiple image sensors at different physical locations to generate a synchronous multi-source image stream include: obtaining original image sequences captured by an image sensor array composed of a visible light camera, a thermal imager and a three-dimensional profilometer; performing space geometry correction and time axis alignment on the original image sequences by using a pre-calibrated sensor space position relationship and a time stamp, so that the data of different sensors for the same weld point are consistent in space-time dimensions; mapping the aligned multi-modal image data, including intensity information, temperature field information and three-dimensional point cloud information, to a unified reference coordinate system; in the reference coordinate system, using a weighted fusion method based on pixel confidence to combine the intensity information, the temperature field information and the three-dimensional point cloud information into a synchronous multi-source image stream with multi-dimensional attributes.

3. The machine vision based real time monitoring system for fatigue cracking in welds of steel structures as claimed in claim 1 wherein, The specific steps of the multi-scale feature analysis and enhancement on the synchronous multi-source image stream to construct a multi-level weld feature map include: receiving the synchronous multi-source image stream and performing multi-resolution pyramid decomposition to obtain image subbands of multiple scales covering from macrostructure to microtexture; applying a directional filter bank and a local binary pattern operator to each scale of the image subbands to extract edge orientation features and surface texture roughness features of the weld area; cross-scale correlation and aggregation are performed on the edge orientation features and the surface texture roughness features extracted at different scales to form a primary feature set with spatial hierarchical relationship; an attention weight mechanism is introduced to adaptively enhance feature channels with high fatigue damage correlation in the primary feature set and suppress irrelevant background features; the enhanced features are reorganized according to spatial position and scale level and encoded into a structured multi-level weld feature map.

4. The machine vision-based real-time monitoring system for fatigue cracking in welds of steel structures as claimed in claim 1 wherein, The specific steps of updating and maintaining a weld structure dynamic knowledge base based on the multi-level weld feature map through a dynamic learning mechanism include: Initialize a weld structure dynamic knowledge base containing weld topology connection relationship, material attribute baseline and historical typical defect mode; Decode the geometric feature vector and surface state feature vector of the current weld from the real-time input multi-level weld feature map; Compare the geometric feature vector and surface state feature vector with the baseline features stored in the weld structure dynamic knowledge base point by point to generate a feature change vector; Use the incremental learning method to fuse the feature change vector into the weld structure dynamic knowledge base to update the state description and historical trajectory of the corresponding weld entry; According to the long time series of the feature change vector, learn the implicit law of weld state evolution through the probability graph model, and correct the state transition probability in the weld structure dynamic knowledge base accordingly.

5. The machine vision-based real-time monitoring system for fatigue cracking in welds of steel structures as claimed in claim 1 wherein, The specific steps of using the weld structure dynamic knowledge base and the multi-level weld feature map obtained in real time to monitor the real-time health status of the steel structure weld and identify abnormalities include: Real-time receive the multi-level weld feature map, and call the baseline features and state evolution probability of the corresponding weld from the weld structure dynamic knowledge base; Parallel comparison of real-time features and baseline features in multiple preset damage sensitive dimensions, including geometric deformation, texture anomaly and local thermal gradient; Use a sliding time window to accumulate the comparison differences in multiple dimensions to form a time-evolving state difference sequence; Model the state difference sequence through the trained deep state evolution network, which outputs a comprehensive health score and multiple abnormal confidence levels in sub-dimensions; When the health score is lower than the dynamic threshold or any of the abnormal confidence levels exceeds its corresponding threshold, it is determined that the weld area is abnormal, and a monitoring event containing the type and location of the abnormality is generated.

6. The machine vision-based real-time monitoring system for weld fatigue cracking in steel structures as claimed in claim 5 wherein, The specific steps of performing crack initiation location and evolution path deduction calculation according to the real-time health status monitoring result combined with historical damage data include: Receive the monitoring event generated from the real-time health status monitoring step and extract the local high-resolution feature area related to the event; Retrieve the historical damage data and material fatigue performance parameters of the weld from the weld structure dynamic knowledge base; Take the monitoring event location as the initial node and the weld microstructure as the guide to construct a space-time deduction graph; On the space-time deduction graph, define a crack propagation potential function based on strain energy density and stress intensity factor, and use the crack propagation potential function to calculate the instantaneous probability of crack propagation in different directions; Use the Monte Carlo simulation method to combine the material fatigue performance parameters and real-time stress spectrum to perform multiple rounds of probability simulation on the crack initiation and propagation process to generate a set of possible evolution path sets; Cluster and analyze the convergence of the evolution path set to extract the high-probability crack initiation area and dominant evolution direction to form a crack deduction report.

7. The machine vision-based real-time monitoring system for weld fatigue cracking in steel structures as claimed in claim 1 wherein, The specific steps of dynamically estimating the remaining service life of the weld based on the crack initiation location and evolution path deduction calculation result combined with real-time load information include: Obtaining the crack deduction report, and parsing high-probability crack initial size, position and dominant propagation direction from the crack deduction report; synchronously accessing real-time load information provided by a structure health monitoring system, including dynamic stress spectrum, cycle number and environmental temperature; Establishing a micro-mechanical constitutive equation describing damage accumulation process of weld material under cyclic loading; taking the crack initial size as a starting point and the real-time load information as input, numerically solving the micro-mechanical constitutive equation to iteratively calculate the crack size propagation process with load cycles; Coupling the crack propagation process with a component residual carrying capacity model, when the residual carrying capacity is calculated to drop to a safety threshold, the corresponding load cycle number is the predicted residual service life; According to the real-time newly added monitoring data, the parameters of the micro-mechanical constitutive equation are calibrated online, and the predicted residual service life value is dynamically updated accordingly.

8. The machine vision-based real-time monitoring system for weld fatigue cracking in steel structures as claimed in claim 1 wherein, The specific steps of generating the structure integrity control strategy online according to the residual service life estimation result of the weld and the real-time health state monitoring result include: Receiving the updated residual service life value from the residual service life dynamic estimation step and the continuous health degree score and abnormal confidence from the real-time health state monitoring step; Building an optimization objective function with the goal of prolonging the safe operation time of the overall structure, the constraint conditions of the optimization objective function including the residual service life difference of the current welds, real-time load distribution and adjustable structure boundary conditions; Taking the health degree score and the abnormal confidence as weights, in the solution space of the optimization objective function, a set of recommended load distribution adjustment amount or local stiffness modification suggestion is solved by optimization; Encoding the load distribution adjustment amount or local stiffness modification suggestion into structure integrity control instructions that can be understood by execution mechanisms or management processes.

9. The machine vision-based real-time monitoring system for weld fatigue cracking in steel structures as claimed in claim 2 wherein, The system further includes a step of deep feature learning of the synchronous multi-source image stream, specifically: Establishing a multi-level cascaded deep feature learning network, the input of the deep feature learning network being the synchronous multi-source image stream; The deep feature learning network is trained in an end-to-end manner, and its output is a deep weld descriptor containing rich semantic information; the deep weld descriptor will be supplemented as high-level features into the multi-level weld feature atlas.

10. The machine vision-based real-time monitoring system for fatigue cracking in welds of steel structures as claimed in claim 9 wherein, The first sub-network of the deep feature learning network focuses on segmenting the weld main body region and the background to generate a pixel-level segmentation mask; the second sub-network takes the region of interest constrained by the segmentation mask as input to perform semantic segmentation of subtle texture anomalies and micro-defects; the third sub-network integrates the outputs of the first two sub-networks to build a complete feature pyramid from pixels to semantics.

Images