Method for detecting quality of foundation steel bar welded joint

Abstract

The invention provides a method for detecting the quality of a foundation steel bar welded joint, and belongs to the technical field of steel bar welded joint quality detection.The method comprises the steps that a stress distribution matrix is obtained through ultrasonic detection equipment, and a fatigue distribution matrix is established in combination with a fatigue analysis algorithm; a stress fatigue correlation analysis method and a bipartite graph maximum matching algorithm are used to construct a stress fatigue correlation matrix, a damage evaluation deep learning model based on a Transform architecture is used to generate a damage distribution matrix, and a time sequence prediction algorithm is used to consider a fatigue center change effect to establish a damage expansion trend matrix. A damage prediction matrix is deduced through a damage prediction deep learning model of a bidirectional long-short-term memory network, and finally, a quality detection result matrix is comprehensively evaluated and output by utilizing a game model, a multi-objective optimization algorithm and a gating weighting function, so that the technical problem that the quality detection accuracy of the steel bar welded joint is not enough is solved.

Description

Technical Field

[0001] This invention belongs to the field of quality inspection technology for welded joints of reinforcing bars, and more specifically, it relates to a method for quality inspection of welded joints of basic reinforcing bars. Background Technology

[0002] Quality inspection of welded rebar joints is a crucial technical aspect of structural safety assessment in building engineering. Traditional inspection methods primarily employ ultrasonic non-destructive testing (NDT) to obtain information on internal defects in the joints, judging weld quality by analyzing sound wave propagation characteristics and reflected signal intensity. Fatigue performance testing is also widely used, assessing the fatigue life and damage accumulation characteristics of joints through cyclic load tests. These methods play a vital role in the quality control of important engineering structures such as bridges, high-rise buildings, and industrial plants. However, traditional inspection techniques have significant limitations. Ultrasonic testing, based primarily on single physical quantity analysis, struggles to comprehensively reflect the complex stress state and damage mechanisms of the joints. While fatigue performance testing can assess material properties, it cannot accurately predict the damage evolution process under actual working conditions. Inspection results are often limited to static defect identification, lacking in-depth analysis of the coupling effects of stress concentration and fatigue damage under dynamic load conditions. In current quality assessments of welded rebar joints, the inability of inspection methods to establish a comprehensive correlation mechanism between stress distribution, fatigue damage, and material properties leads to insufficient inspection accuracy. This makes it difficult to accurately identify potential damage risks and predict the actual service performance of the joints, impacting the reliability of structural safety assessments. In other words, existing technologies suffer from insufficient accuracy in the quality inspection of welded rebar joints. Summary of the Invention

[0003] In view of this, the present invention provides a method for quality inspection of welded joints of basic reinforcing bars, which can solve the technical problem of insufficient accuracy in quality inspection of welded joints of reinforcing bars in the prior art.

[0004] This invention is implemented as follows: A method for quality inspection of welded joints of basic reinforcing bars includes: using ultrasonic testing equipment to perform non-destructive testing on the welded joints, acquiring internal stress wave propagation time data and reflection intensity data, and establishing a stress distribution matrix; performing fatigue performance testing on the welded joints using a fatigue analysis algorithm, and establishing a fatigue distribution matrix by combining the historical cyclic load data of the welded joints and material fatigue performance parameters; using a stress-fatigue correlation analysis method to couple the stress distribution matrix and the fatigue distribution matrix to establish a stress-fatigue correlation matrix; and optimizing the stress distribution matrix and fatigue distribution matrix using a bipartite graph maximum matching algorithm. The matching relationship of the matrix is ​​analyzed; the stress-fatigue correlation matrix is ​​analyzed based on the deep learning model for damage assessment, and a damage distribution matrix is ​​generated by combining the material damage accumulation theoretical parameters and fracture mechanics parameters; the damage distribution matrix is ​​analyzed for trend using a time series prediction algorithm, and the impact of fatigue center change effect on damage evolution path is considered to establish a damage expansion trend matrix; the damage expansion trend matrix is ​​used to extrapolate the future damage state of the damage expansion trend matrix through a deep learning model for damage prediction, and a damage prediction matrix is ​​generated; the damage prediction matrix is ​​comprehensively evaluated based on a game model, the game equilibrium solution is solved through a multi-objective optimization algorithm, the evaluation weight parameters are adjusted using a gating weight function, and the quality inspection result matrix is ​​output.

[0005] Specifically, the stress distribution matrix is ​​used to characterize the stress distribution state of the welded joint of the steel bar in three-dimensional space. Each element in the matrix represents the stress value at the corresponding spatial location, which is obtained by inversion calculation using stress wave propagation time data and reflection intensity data acquired by ultrasonic detection equipment.

[0006] The fatigue distribution matrix is ​​specifically used to quantify the degree of fatigue accumulation in various parts of the welded joint. It is established based on material fatigue performance parameters and historical cyclic load data, and the matrix element values ​​represent the relative degree of fatigue damage.

[0007] Specifically, the stress-fatigue correlation matrix is ​​used to describe the relationship between the stress distribution matrix and the fatigue degree distribution matrix. Through mathematical transformation, the stress distribution matrix and the fatigue degree distribution matrix are coupled into a unified correlation analysis result, reflecting the correspondence between the stress concentration area and the fatigue damage area.

[0008] Specifically, the damage distribution matrix is ​​used to represent the current damage state distribution of the welded joint. It is calculated by combining the stress fatigue correlation matrix, material damage accumulation theoretical parameters, and fracture mechanics parameters. The matrix element values ​​represent the damage degree values ​​at the corresponding locations.

[0009] Specifically, the damage expansion trend matrix is ​​used to predict the development trend of damage over time. It is established based on the time series data of the damage distribution matrix through a time series prediction algorithm, reflecting the speed and direction characteristics of damage expansion.

[0010] Specifically, the damage prediction matrix is ​​used to represent the predicted damage state of the welded joint within a future time period. It is obtained by extrapolating the damage expansion trend matrix over time using a deep learning model for damage prediction.

[0011] Specifically, the quality inspection result matrix is ​​used to represent the final quality evaluation result of the steel bar welded joint. It is calculated based on the damage prediction matrix through a game model, multi-objective optimization algorithm and gated weight function. The matrix element values ​​represent the quality level evaluation of the corresponding position.

[0012] The fatigue center shift effect specifically refers to the phenomenon that the main fatigue damage area shifts due to changes in material properties or load conditions during cyclic loading of a welded joint, which affects the damage evolution path and the final failure mode.

[0013] Specifically, the gating weight function step is used to adjust the gating mechanism parameters of the neural network. The equilibrium value is calculated based on the stress distribution entropy value, fatigue distribution mean value, damage gradient maximum value and material strength coefficient. When the equilibrium value is less than 0.3, a linear weight adjustment function is used; when the equilibrium value is between 0.3 and 0.7, a nonlinear sigmoid weight adjustment function is used; and when the equilibrium value is greater than 0.7, an exponential decay weight adjustment function is used to adjust the gating parameters.

[0014] Specifically, the bipartite graph maximum matching algorithm optimizes the matching relationship between the stress distribution matrix and the fatigue distribution matrix. It uses the high-stress region in the stress distribution matrix as one set of vertices and the high-fatigue region in the fatigue distribution matrix as another set of vertices. The Hungarian algorithm is used to find the optimal matching scheme between the two sets, ensuring the accuracy and computational efficiency of the stress-fatigue correlation matrix analysis.

[0015] The damage assessment deep learning model is specifically a multi-scale feature fusion network based on the Transformer architecture. It contains three parallel encoder branches that process stress features, fatigue features, and material features respectively. Each encoder contains six layers of self-attention mechanism and feedforward neural network layer. The number of attention heads is set to eight, and the hidden state dimension is set to 512. The deep fusion of the three features is achieved through the cross-attention mechanism, and the damage assessment result is finally output through a fully connected layer.

[0016] Specifically, the damage prediction deep learning model is a time series prediction network based on a long short-term memory network architecture. It contains three layers of bidirectional long short-term memory units, with the hidden state dimension of each layer set to 256. It uses an attention mechanism to weightedly fuse features at different time steps, and sets the sequence length to 50 to capture sufficient temporal dependencies. Finally, it outputs the prediction result of the future damage state through a fully connected layer.

[0017] Specifically, the steps of the multi-objective optimization algorithm are used to solve the equilibrium solution of the game model. Based on the genetic algorithm framework, it simultaneously optimizes two objectives: detection accuracy and computational efficiency. The population size is set to 100, the number of generations is set to 200, the crossover probability is set to 0.8, and the mutation probability is set to 0.1. The optimal solution set is selected through the Pareto front as the basis for calculating the quality detection result matrix.

[0018] Specifically, the game model includes an upper-level model that aims to minimize the uncertainty of quality assessment and a lower-level model that aims to maximize the sensitivity of damage identification. The upper-level model takes the damage prediction matrix as input, the objective function is to minimize the variance between different prediction methods, and the output is the quality assessment weight allocation scheme. The lower-level model takes the damage prediction matrix and the material strength coefficient as input, the objective function is to maximize the sensitivity index of damage detection, and the output is the damage identification threshold parameter.

[0019] Prior to the step of establishing the deep learning model for damage assessment, the step also includes establishing a training dataset for the deep learning model for damage assessment. Specifically, this involves collecting a total of 10,000 samples of steel bar welded joints under different material grades and welding process conditions, performing standardized non-destructive testing on each sample to obtain stress distribution data, obtaining real fatigue performance parameters through laboratory fatigue tests, and using industrial CT scanning technology to obtain the real distribution state of internal damage as label data.

[0020] This invention establishes a stress-fatigue correlation analysis method, coupling the stress distribution matrix obtained from ultrasonic testing with the fatigue degree distribution matrix. A bipartite graph maximum matching algorithm is used to optimize the matching relationship, constructing a multi-dimensional information fusion detection and analysis framework. A deep learning model for damage assessment is used to analyze the stress-fatigue correlation matrix, generating an accurate damage distribution matrix by combining material damage accumulation theory and fracture mechanics parameters. A damage expansion trend prediction mechanism is established by considering the fatigue center change effect through a time series prediction algorithm. A deep learning model for damage prediction is used to accurately predict future damage states. Finally, a game theory model is used to comprehensively evaluate the damage prediction results, and a multi-objective optimization algorithm is used to solve the game equilibrium solution. Evaluation parameters are dynamically adjusted using a gating weight function, establishing a comprehensive correlation mechanism between stress distribution, fatigue damage, and material properties. This achieves a comprehensive and accurate analysis of the internal stress state, fatigue accumulation degree, and damage evolution law of welded steel joints, significantly improving the accuracy of quality inspection and solving the technical problem of insufficient accuracy in quality inspection of welded steel joints. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention.

[0022] Figure 2 This is a schematic diagram of the structure of the deep learning model for damage prediction involved in the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0024] like Figure 1 The diagram shown is a flowchart of a method for quality inspection of welded joints of basic reinforcing bars provided by the present invention. This method includes the following steps:

[0025] S01. Use ultrasonic testing equipment to perform non-destructive testing on the welded joints of steel bars, obtain the propagation time data and reflection intensity data of the internal stress wave of the welded joint, and establish a stress distribution matrix.

[0026] S02. The fatigue performance of the welded joint of the steel bar is tested by the fatigue analysis algorithm. The fatigue distribution matrix is ​​established by combining the historical data of the cyclic load of the welded joint and the fatigue performance parameters of the material.

[0027] S03. Using the stress-fatigue correlation analysis method, the stress distribution matrix and the fatigue degree distribution matrix are coupled and calculated to establish a stress-fatigue correlation matrix. The matching relationship between the stress distribution matrix and the fatigue degree distribution matrix is ​​optimized by the bipartite graph maximum matching algorithm.

[0028] S04. Based on the deep learning model for damage assessment, the stress-fatigue correlation matrix is ​​analyzed, and the damage distribution matrix is ​​generated by combining the material damage accumulation theoretical parameters and fracture mechanics parameters.

[0029] S05. Use time series prediction algorithm to perform trend analysis on damage distribution matrix, consider the effect of fatigue center change on damage evolution path, and establish damage expansion trend matrix.

[0030] S06. Using a deep learning model for damage prediction, the damage expansion trend matrix is ​​used to extrapolate the future damage state and generate a damage prediction matrix.

[0031] S07. Based on the game model, a comprehensive evaluation of the damage prediction matrix is ​​performed. The game equilibrium solution is solved by a multi-objective optimization algorithm. The evaluation weight parameters are adjusted by a gating weight function, and the quality detection result matrix is ​​output.

[0032] The stress distribution matrix is ​​used to characterize the stress distribution state of the welded joint of the steel bar in three-dimensional space. Each element in the matrix represents the stress value at the corresponding spatial location, which is obtained by inversion calculation using stress wave propagation time data and reflection intensity data acquired by ultrasonic detection equipment.

[0033] The fatigue distribution matrix is ​​used to quantify the degree of fatigue accumulation in various parts of the welded joint. It is established based on material fatigue performance parameters and historical cyclic load data, and the matrix element values ​​represent the relative degree of fatigue damage.

[0034] The stress-fatigue correlation matrix is ​​used to describe the relationship between the stress distribution matrix and the fatigue degree distribution matrix. Through mathematical transformation, the stress distribution matrix and the fatigue degree distribution matrix are coupled into a unified correlation analysis result, reflecting the correspondence between the stress concentration area and the fatigue damage area.

[0035] The damage distribution matrix is ​​used to represent the current damage state distribution of the welded joint. It is calculated by combining the stress fatigue correlation matrix, material damage accumulation theoretical parameters, and fracture mechanics parameters. The matrix element values ​​represent the damage degree values ​​at the corresponding locations.

[0036] The damage expansion trend matrix is ​​used to predict the development trend of damage over time. It is established based on the time series data of the damage distribution matrix through a time series prediction algorithm, reflecting the speed and direction characteristics of damage expansion.

[0037] The damage prediction matrix is ​​used to represent the predicted damage state of the welded joint within a future time period. It is obtained by extrapolating the damage expansion trend matrix over time using a deep learning model for damage prediction.

[0038] The quality inspection result matrix is ​​used to represent the final quality evaluation result of the steel bar welded joint. It is calculated based on the damage prediction matrix through a game model, multi-objective optimization algorithm and gated weight function. The matrix element values ​​represent the quality level evaluation of the corresponding position.

[0039] The fatigue center shift effect refers to the phenomenon that the main fatigue damage area shifts during the cyclic loading of a welded joint due to changes in material properties or load conditions, which affects the damage evolution path and the final failure mode.

[0040] The gating weight function is used to adjust the gating mechanism parameters of the neural network. It calculates the equilibrium value based on the stress distribution entropy, fatigue distribution mean, damage gradient maximum value and material strength coefficient. When the equilibrium value is less than 0.3, a linear weight adjustment function is used. When the equilibrium value is between 0.3 and 0.7, a nonlinear sigmoid weight adjustment function is used. When the equilibrium value is greater than 0.7, an exponential decay weight adjustment function is used to adjust the gating parameters.

[0041] The bipartite graph maximum matching algorithm is used to optimize the matching relationship between the stress distribution matrix and the fatigue distribution matrix. The high-stress region in the stress distribution matrix is ​​used as one set of vertices, and the high-fatigue region in the fatigue distribution matrix is ​​used as another set of vertices. The optimal matching scheme between the two sets is found by using the Hungarian algorithm, which ensures the accuracy and computational efficiency of the stress-fatigue correlation matrix analysis.

[0042] The specific structure of the damage assessment deep learning model is a multi-scale feature fusion network based on the Transformer architecture, containing three parallel encoder branches that process stress features, fatigue features, and material features respectively. Each encoder contains six layers of self-attention mechanisms and feedforward neural network layers, with eight attention heads and a hidden state dimension of 512. Deep fusion of the three features is achieved through a cross-attention mechanism, and the damage assessment result is finally output through a fully connected layer. The model's attention mask parameter is used to control the intensity of information interaction between different features, and the inference batch size is set to 32 to balance computational efficiency and memory usage. The steps for establishing the training dataset for the damage assessment deep learning model specifically include collecting a total of 10,000 samples of steel bar welded joints under different material grades and welding process conditions, and performing standardized non-destructive testing on each sample to obtain the stress distribution. The data is obtained through laboratory fatigue tests to obtain real fatigue performance parameters. Industrial CT scanning technology is used to obtain the real distribution of internal damage as label data. The data is divided into training set, validation set and test set in an 8:1:1 ratio. The input features are normalized to ensure numerical stability. The training steps of the deep learning model for damage assessment specifically include using the AdamW optimizer to update parameters with a learning rate of 0.0001, using a cosine annealing scheduling strategy to dynamically adjust the learning rate, using a combination of weighted mean square error and structural similarity loss as the loss function, setting the batch size to 32, and a total of 200 training rounds. Every 10 rounds, the model performance is evaluated on the validation set and the optimal weights are saved. An early stopping strategy is used to prevent overfitting. Gradient pruning is used to stabilize the training process. Data augmentation techniques, including noise injection and feature perturbation, are used to improve the model's generalization ability.

[0043] The specific structure of the damage prediction deep learning model is a time series prediction network based on a long short-term memory (LSTM) network architecture, containing three layers of bidirectional LSM units. The hidden state dimension of each layer is set to 256. An attention mechanism is used to weight and fuse features from different time steps. The sequence length is set to 50 to capture sufficient temporal dependencies. Finally, a fully connected layer outputs the predicted future damage state. The steps for establishing the training dataset for the damage prediction deep learning model specifically include constructing time series data based on the output of the damage assessment deep learning model. Each sequence contains records of damage distribution matrix changes over 50 consecutive time steps. Training samples are extracted from historical data using a sliding window method, with the first 80% of the time series data used as the training set and the last 20% as the test set. The time series data is standardized to eliminate the influence of units. The specific steps for training the damage prediction deep learning model include using the Adam optimizer to update parameters, setting the learning rate to 0.001, using a learning rate decay strategy to halve the learning rate every 20 rounds, using the root mean square error as the loss function, setting the batch size to 64, and a total of 150 training rounds. Cross-validation is used to evaluate the model's generalization performance, regularization is used to prevent overfitting, and gradient cutoff is used to stabilize the training process.

[0044] The multi-objective optimization algorithm is used to solve the equilibrium solution of the game model. Based on the genetic algorithm framework, it simultaneously optimizes the two objectives of detection accuracy and computational efficiency. The population size is set to 100, the number of generations is set to 200, the crossover probability is set to 0.8, and the mutation probability is set to 0.1. The optimal solution set is selected through the Pareto front as the basis for calculating the quality detection result matrix.

[0045] The game theory model includes an upper-level model that aims to minimize the uncertainty of quality assessment and a lower-level model that aims to maximize the sensitivity of damage identification. The upper-level model takes the damage prediction matrix as input, and its objective function is to minimize the variance between different prediction methods. The constraints include the accuracy limit of the ultrasonic detection equipment and the computational resource constraint. The output is the weight allocation scheme for quality assessment. The lower-level model takes the damage prediction matrix and the material strength coefficient as input, and its objective function is to maximize the sensitivity index of damage detection. The constraints include the algorithm complexity limit and the real-time requirement. The output is the damage identification threshold parameter. The two objective functions influence each other through the confidence of the detection result as a coupling term. The weight allocation of the upper-level model will affect the sensitivity calculation of the lower-level model, and the threshold setting of the lower-level model will improve the uncertainty assessment of the upper-level model.

[0046] The specific implementation methods of the above steps are described in detail below.

[0047] Step S01 involves establishing a stress distribution matrix through non-destructive testing of the welded joint of the reinforcing steel using ultrasonic testing equipment. First, ultrasonic transducers are placed on the surface of the welded joint in a grid pattern with a grid spacing of 2mm, ensuring complete coverage of the welded joint and its heat-affected zone. Next, the ultrasonic testing equipment is activated, transmitting an ultrasonic signal at a frequency of 5MHz. The propagation time of the stress wave within the material is recorded using time-domain analysis, while amplitude analysis is employed to obtain the reflection intensity data. Subsequently, an inversion calculation algorithm is used to convert the time and intensity data into stress values. This algorithm is based on acoustoelastic theory and utilizes the linear relationship between stress wave velocity and the material's stress state for calculation. Finally, the stress values ​​from all measuring points are arranged according to spatial coordinates to form a three-dimensional stress distribution matrix. The matrix dimensions are set to 100×100×50, corresponding to the spatial resolution in the length, width, and depth directions, respectively. The purpose of this step is to obtain the true stress distribution state within the welded joint, providing fundamental data support for subsequent analysis.

[0048] Step S02 is implemented by establishing a fatigue distribution matrix using a fatigue analysis algorithm. First, historical load data for the welded steel joint is collected, including load amplitude, load frequency, and load cycle count, with a data collection period of at least six months. Next, statistical analysis of the historical cyclic load data is performed using the rainflow counting method to identify the distribution of load cycle counts for different amplitudes. Then, combining the material's fatigue performance parameters, including the fatigue strength limit, fatigue life curve slope, and material constants, the linear cumulative damage theory is used to calculate the degree of fatigue accumulation at each location. During the calculation, the fatigue strength limit threshold is set to 60% of the material's yield strength, and the fatigue life curve slope parameter ranges from 3 to 12. Finally, the calculated fatigue accumulation values ​​are arranged according to spatial location to form a fatigue distribution matrix. Matrix element values ​​are normalized to 0 to 1, where 0 represents no fatigue damage and 1 represents fatigue failure. The purpose of this step is to quantify the fatigue damage accumulation state of each part of the welded joint, providing basic fatigue damage data for stress-fatigue correlation analysis.

[0049] Step S03 involves establishing a stress-fatigue correlation matrix using a stress-fatigue correlation analysis method. First, the stress distribution matrix and fatigue degree distribution matrix are preprocessed, including data standardization and noise filtering. Standardization uses the zero-mean, unit variance method, and noise filtering uses the median filtering algorithm with a 3×3 filtering window size. Next, a threshold segmentation method is used to extract high-stress regions from the stress distribution matrix, with the threshold set to the stress mean plus 1.5 times the standard deviation. Similarly, high-fatigue regions are extracted from the fatigue degree distribution matrix with a threshold set to 0.7. Then, the high-stress regions are treated as one vertex set in a bipartite graph, and the high-fatigue regions as another vertex set. The Hungarian algorithm is used to solve the maximum matching problem in the bipartite graph. This algorithm, based on the augmented path search principle, can find the optimal one-to-one matching relationship between the two sets. Finally, based on the matching results, a weighted fusion method is used to couple the stress distribution matrix and fatigue degree distribution matrix. The fusion weight is determined by the matching degree; regions with high matching degrees have a weight of 0.8, and regions with low matching degrees have a weight of 0.2. The purpose of this step is to establish a quantitative correlation between stress distribution and fatigue damage, revealing the spatial correspondence between stress concentration areas and fatigue damage areas.

[0050] Step S04 is implemented by generating a damage distribution matrix based on the stress-fatigue correlation matrix using a deep learning model for damage assessment. First, the stress-fatigue correlation matrix is ​​input into a pre-trained deep learning model for damage assessment, which employs a multi-scale feature fusion network structure based on the Transformer architecture. Then, the model processes stress features, fatigue features, and material features through three parallel encoder branches. Each encoder contains six layers of self-attention mechanisms and a feedforward neural network layer, with eight attention heads and a hidden state dimension of 512. Next, a cross-attention mechanism is used to achieve deep fusion of the three features. The attention mask parameter controls the intensity of information interaction between different features; the interaction intensity between stress and fatigue features is set to 0.9, and the interaction intensity with material features is set to 0.6. Simultaneously, material damage accumulation theory parameters, including damage threshold, damage growth rate, and critical damage value, as well as fracture mechanics parameters, including stress intensity factor and crack propagation rate, are incorporated. Finally, a damage distribution matrix is ​​output through a fully connected layer, where the matrix element values ​​represent the damage severity at the corresponding location, ranging from 0 to 1. The purpose of this step is to utilize the nonlinear mapping capability of the deep learning model to accurately assess the current damage state distribution of the welded joint.

[0051] Step S05 involves establishing a damage propagation trend matrix using a time series prediction algorithm. First, time series data of the historical damage distribution matrix is ​​collected, with a data collection frequency of once a week and a data window length of 24 weeks. Next, the time series data is preprocessed, including missing value interpolation and outlier detection. Linear interpolation is used for missing value interpolation, and a 3-standard-deviation criterion is used for outlier detection; data points exceeding the threshold are removed or corrected. Then, an autoregressive integral moving average model is used to analyze the trend of the damage distribution matrix. The model parameters are determined using the autocorrelation function and partial autocorrelation function, with the autoregression order set to 3 and the moving average order set to 2. During the analysis, the impact of fatigue center shift on the damage evolution path is considered. Fatigue center shifts are identified by monitoring changes in the centroid position of the main fatigue damage area; a shift exceeding 5 mm is considered a fatigue center shift. Finally, a damage propagation trend matrix is ​​established based on the trend analysis results. Matrix element values ​​represent the speed and direction of damage development over time, with the speed value represented by the weekly damage growth rate and the direction by the gradient vector. The purpose of this step is to predict the development trend of damage over time, providing evolutionary support for predicting future damage states.

[0052] Step S06 is implemented by using a damage prediction deep learning model to predict the future damage state of the damage propagation trend matrix. First, the damage propagation trend matrix is ​​input into a pre-trained damage prediction deep learning model, which employs a time-series prediction network structure based on a long short-term memory (LSTM) network architecture. Next, the model processes the time-series features through three layers of bidirectional LSM units, with each layer having a hidden state dimension of 256. An attention mechanism is used to weight and fuse features from different time steps, and the sequence length is set to 50 to capture sufficient temporal dependencies. Then, the model performs time extrapolation calculations based on historical damage evolution patterns, with a prediction time span of 12 weeks and a required prediction accuracy of a root mean square error (RMSE) of less than 0.05. Simultaneously, the model considers corrections for the prediction results based on factors such as load changes, environmental factors, and material aging. The correction coefficients for load changes are set to 1.1 to 1.3, for environmental factors to 0.9 to 1.2, and for material aging to 1.05 to 1.15. Finally, a damage prediction matrix is ​​output through a fully connected layer, where each element represents the predicted damage state at the corresponding location within a future time period. The purpose of this step is to utilize the time series prediction capabilities of deep learning models to accurately predict the future damage state of welded joints.

[0053] Step S07 is implemented by using a game theory model to comprehensively evaluate the damage prediction matrix and output a quality detection result matrix. First, a two-layer game theory model is constructed. The upper-layer model aims to minimize the uncertainty of quality evaluation, while the lower-layer model aims to maximize the sensitivity of damage identification. The upper-layer model takes the damage prediction matrix as input, with the objective function being to minimize the variance between different prediction methods. Constraints include a 0.1mm accuracy limit for the ultrasonic detection equipment and an 8GB memory constraint. The output is the quality evaluation weight allocation scheme. The lower-layer model takes the damage prediction matrix and material strength coefficient as input, with the objective function being to maximize the sensitivity index of damage detection. Constraints include an algorithm complexity limit of O(n^2). 2 The system has a real-time requirement of 100ms, and the output is the damage identification threshold parameter. Next, a multi-objective optimization algorithm is used to solve the game equilibrium solution. Based on a genetic algorithm framework, it simultaneously optimizes both detection accuracy and computational efficiency. The population size is set to 100, the number of generations to 200, the crossover probability to 0.8, and the mutation probability to 0.1. Then, the evaluation weight parameters are adjusted through a gating weight function. This function calculates the equilibrium value based on the stress distribution entropy, the fatigue distribution mean, the maximum damage gradient, and the material strength coefficient. When the equilibrium value is less than 0.3, a linear weight adjustment function is used; when the equilibrium value is between 0.3 and 0.7, a nonlinear sigmoid weight adjustment function is used; and when the equilibrium value is greater than 0.7, an exponential decay weight adjustment function is used. Finally, a quality inspection result matrix is ​​output. The matrix element values ​​represent the quality level evaluation at the corresponding position, divided into three levels: excellent, qualified, and unqualified, with corresponding numerical ranges of 0.8 to 1.0, 0.6 to 0.8, and 0 to 0.6, respectively. The purpose of this step is to use game theory and multi-objective optimization methods to comprehensively consider detection accuracy and computational efficiency, and output reliable quality detection results.

[0054] The deep learning model for damage prediction is structured based on a long short-term memory (LSTM) network architecture, comprising five main components: an input layer, a feature extraction layer, a temporal modeling layer, an attention fusion layer, and an output layer. The input layer receives damage amplification trend matrix data, with a data dimension of batch size × sequence length × feature dimension. The batch size is set to 64, the sequence length to 50, and the feature dimension to 128. The feature extraction layer uses a one-dimensional convolutional neural network structure with three convolutional layers. The kernel sizes are 3, 5, and 7, and the number of kernels is set to 64, 128, and 256, respectively. The ReLU activation function is used, and max pooling is employed to reduce the feature dimension. The temporal modeling layer uses three layers of bidirectional LSM units, with each layer having a hidden state dimension of 256. The sigmoid function is used for the forget gate, input gate, and output gate, while the tanh function is used for the candidate values. Gradient clipping is used to prevent gradient explosion. The attention fusion layer employs a multi-head attention mechanism with 8 attention heads. Attention weights are calculated through linear transformations of the query matrix, key matrix, and value matrix. Weight normalization uses the softmax function, and residual connections and layer normalization techniques are employed to improve model stability. The output layer contains two fully connected layers. The first fully connected layer has 512 neurons and uses the ReLU activation function. The second fully connected layer has the same number of neurons as the dimension of the output matrix and uses the sigmoid activation function to ensure that output values ​​are between 0 and 1.

[0055] The process of establishing the training dataset for the deep learning model for damage prediction includes four main steps: data collection, data preprocessing, feature construction, and data partitioning. In the data collection phase, time-series data is constructed based on the output of the deep learning model for damage assessment. Records of damage distribution matrix changes obtained through continuous monitoring under different working conditions are collected. The data collection period spans two years, with a collection frequency of once a week, resulting in damage distribution data for 104 time points. In the data preprocessing phase, the raw data undergoes quality checks and cleaning, including missing value handling, outlier detection, and data format standardization. Missing values ​​are handled using forward imputation, and outliers are detected using the quartile interval method, with a detection threshold set to 1.5 times the quartile interval. The data format is standardized to 32-bit floating-point numbers. In the feature construction phase, training samples are extracted from historical data using the sliding window method. Each sequence contains records of damage distribution matrix changes over 50 consecutive time steps, with a sliding step size of 1, generating a total of 55 training sequences. The target value for each sequence is the damage distribution matrix for the next time step. Auxiliary features are also constructed, including load history, environmental parameters, and material properties. In the data partitioning phase, the first 80% of the time series data is used as the training set and the last 20% as the test set. The training set is further divided into a training subset and a validation subset in a ratio of 9:1. The time series data is standardized to eliminate the influence of units. The standardization method adopts min-max normalization to ensure that all feature values ​​are between 0 and 1.

[0056] It should be noted that the stress-fatigue correlation analysis method couples the stress distribution matrix and fatigue degree distribution matrix through mathematical transformation, establishing a quantitative correspondence between stress concentration regions and fatigue damage regions. This approach has significant advantages over traditional methods. Traditional detection methods typically treat stress analysis and fatigue assessment as independent processes. Stress analysis focuses on the stress distribution under static load conditions, while fatigue assessment emphasizes the damage accumulation characteristics of materials under cyclic loading. The lack of an effective correlation mechanism between the two leads to test results that cannot fully reflect the true mechanical behavior of the joint under complex working conditions. The stress-fatigue correlation analysis method of this invention establishes a mathematical mapping relationship between two different physical quantities and optimizes the matching relationship using a bipartite graph maximum matching algorithm. This achieves a deep fusion analysis of stress state and fatigue damage, accurately identifying the correspondence between stress concentration regions and fatigue-prone regions, revealing the damage mechanism of materials under combined load conditions, and fundamentally improving the scientific rigor and accuracy of the detection and analysis.

[0057] The deep learning-based damage evolution prediction method employs a dual architecture of a deep learning model for damage assessment and a deep learning model for damage prediction. This enables accurate assessment of the damage state of welded steel joints and precise prediction of their future evolution trends, offering significant advantages over traditional prediction methods. Traditional damage prediction methods primarily rely on empirical formulas and simplified mathematical models, calculating based on linear cumulative damage theory. However, these methods struggle to handle complex nonlinear damage evolution processes and cannot effectively consider the combined effects of multiple factors such as material property changes, load history effects, and fatigue center shifts, resulting in limited prediction accuracy. This invention utilizes a Transformer-based damage assessment model, simultaneously processing stress, fatigue, and material features through a multi-scale feature fusion network. It captures the complex relationships between different features using a self-attention mechanism, combined with a damage prediction model based on a long short-term memory network. This attention mechanism accurately models the long-term dependencies in damage evolution, effectively addressing the impact of fatigue center shifts on the damage evolution path and achieving high-precision prediction of future damage states.

[0058] The synergistic effect of stress-fatigue correlation analysis and deep learning-based damage evolution prediction constructs a complete technical chain from data fusion to intelligent prediction, exhibiting significant comprehensive advantages over existing technologies. The organic combination of these two approaches enables deep fusion and intelligent analysis of multi-dimensional information. The stress-fatigue correlation analysis method provides high-quality input data for the deep learning model, eliminating the problems of data isolation and information redundancy in traditional methods by establishing a precise correspondence between stress distribution and fatigue damage. The deep learning model then fully utilizes this correlation information for intelligent analysis and prediction, leveraging the powerful learning capabilities of neural networks to uncover damage evolution patterns hidden within complex data. This synergistic mechanism not only improves the effectiveness of individual techniques but, more importantly, generates a synergistic effect, enabling the entire detection system to accurately identify complex damage patterns that are difficult to detect using traditional methods. This achieves a comprehensive and accurate assessment and reliable prediction of the quality status of welded steel joints, fundamentally solving the problem of insufficient detection accuracy in existing technologies.

[0059] Specifically, the principle of this invention is as follows: The fundamental reason why this invention can solve the technical problem of insufficient accuracy in the quality inspection of welded steel rebar joints lies in the construction of a comprehensive inspection system that integrates multi-dimensional information fusion and intelligent analysis. The essential reason for the insufficient accuracy of traditional inspection methods is the single dimension of inspection information and the lack of an effective data fusion mechanism. In contrast, this invention uses stress wave propagation time and reflection intensity data obtained through ultrasonic detection to accurately reflect the three-dimensional stress distribution state inside the joint. The fatigue analysis algorithm, based on cyclic load history and material performance parameters, can quantify the degree of fatigue accumulation in each part. Both provide comprehensive information on the mechanical state of the joint from different physical dimensions. The stress-fatigue correlation analysis method couples physical quantities of different dimensions through mathematical transformation. The bipartite graph maximum matching algorithm can find the optimal correspondence between stress concentration areas and fatigue damage areas. This precise correspondence reveals the fatigue damage mechanism of materials under complex stress states, providing a reliable physical basis for damage assessment. The deep learning model for damage assessment adopts a multi-scale feature fusion network based on the Transformer architecture. It processes stress features, fatigue features, and material features through three parallel encoder branches, and uses self-attention and cross-attention mechanisms to achieve deep fusion between features, which can accurately identify complex damage patterns that are difficult to detect by traditional methods. The time-series prediction algorithm considers the fatigue center change effect and establishes an accurate damage evolution prediction model by dynamically tracking the transfer process of the main fatigue damage area. The deep learning model for damage prediction is based on a bidirectional long short-term memory network architecture and captures the long-term dependencies of damage evolution through an attention mechanism, achieving accurate prediction of future damage states. The game theory model, through a dual optimization mechanism of minimizing the uncertainty of quality evaluation in the upper-level model and maximizing the sensitivity of damage identification in the lower-level model, combined with the global optimization capability of the multi-objective optimization algorithm and the adaptive adjustment mechanism of the gating weight function, ensures high accuracy and reliability of the detection results, achieving a comprehensive and accurate assessment of the quality status of welded steel bar joints.

[0060] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0061] Step S 01The specific implementation method involves establishing a stress distribution matrix through non-destructive testing of welded steel bar joints using ultrasonic testing equipment. First, ultrasonic transducers are placed on the surface of the welded joint in a grid pattern, with a grid spacing of 2mm, ensuring complete coverage of the welded joint and its heat-affected zone. Next, the ultrasonic testing equipment is activated, transmitting an ultrasonic signal at a frequency of 5MHz. The propagation time of the stress wave within the material is recorded using time-domain analysis, while amplitude analysis is employed to obtain the reflection intensity data. Subsequently, an inversion calculation algorithm is used to convert the time and intensity data into stress values. This algorithm is based on the linear relationship between stress wave velocity and material stress state according to acoustoelastic theory, specifically expressed as follows:

[0062]

[0063] In the formula, σ ij K represents the stress value at position (i, j), in Pa; acoustic The acoustic elastic coefficient is expressed in Pa·s. -1 ;Δt ij ti represents the change in stress wave propagation time at position (i, j), in seconds; t0 represents the reference propagation time, in seconds; C ij Let A be the reflection intensity coefficient at position (i, j), which is dimensionless; r，ij denoted as (i, j), where is the reflection intensity at position (i, j), in dB; ∈1 represents the measurement error term, in Pa.

[0064] The parameter acquisition method is as follows: K acoustic The acoustoelastic coefficients were obtained through material acoustoelastic calibration experiments. The experimental steps included preparing a standard specimen, applying a known stress load, measuring the corresponding change in sound wave propagation time, and obtaining the acoustoelastic coefficients through least squares fitting. The values ​​ranged from 10. 8 Up to 10 10 Pa·s -1 t0 is obtained by measuring the sound wave propagation time under stress-free conditions, with a typical value of 10. -5 Up to 10 -4 s. C ij The values ​​were obtained through a reflection intensity calibration experiment, ranging from 0.1 to 0.9. A r，ij The stress values ​​are obtained directly through ultrasonic equipment. The range of ∈1 is 0.02 to 0.05 times the material's yield strength. Finally, the stress values ​​of all measuring points are arranged according to spatial coordinates to form a three-dimensional stress distribution matrix S, with the matrix dimensions set to 100×100×50, corresponding to the spatial resolution in the length, width, and depth directions, respectively.

[0065] Step S 02The specific implementation method involves establishing a fatigue distribution matrix using a fatigue analysis algorithm. First, historical load data for welded steel bar joints is collected, including load amplitude, load frequency, and load cycle count, with a data collection period of at least six months. Next, statistical analysis of the historical cyclic load data is performed based on the rainflow counting method to identify the distribution of load cycle counts for different amplitudes. Then, combined with the material's fatigue performance parameters, the linear cumulative damage theory is used to calculate the degree of fatigue accumulation at each location, as detailed below:

[0066]

[0067] In the formula, D mn Here is the cumulative fatigue damage value at position (m, n); n cycle n represents the total number of cyclic load types, determined based on historical load statistics. k N represents the number of cycles under the k-th stress amplitude. k (σ k ) represents the stress amplitude σ k Corresponding fatigue life; σ k σ represents the k-th stress amplitude, in Pa; α, β, and γ are material fatigue parameters; limit f is the fatigue strength limit; k ∈2 represents the frequency of the k-th load; ∈2 represents the cumulative error term.

[0068] The parameter acquisition method is as follows: N k (σ k The fatigue life function is obtained through fatigue testing. The testing steps include preparing fatigue specimens, applying cyclic loads with different stress amplitudes, recording the number of cycles at specimen failure, and fitting a SN curve to obtain the fatigue life function. Where A is the material fatigue constant, m is the slope of the fatigue life curve, and its value ranges from 3 to 12. α ranges from 0.8 to 1.2, β ranges from 0.5 to 0.9, and γ ranges from 1.5 to 2.5. σ limit Set to 60% of the material's yield strength. k The data was obtained statistically through rainflow counting. k The values ​​were obtained through analysis of historical load data, with units in Hz. The range of ∈2 is 0.01 to 0.03 times the fatigue strength limit. Finally, the calculated cumulative fatigue values ​​were arranged according to spatial location to form a fatigue distribution matrix F, with matrix element values ​​using a normalized representation from 0 to 1.

[0069] Step S 03The specific implementation method involves establishing a stress-fatigue correlation matrix using stress-fatigue correlation analysis. First, the stress distribution matrix S and fatigue degree distribution matrix F are preprocessed, including data standardization and noise filtering. Next, a threshold segmentation method is used to extract high-stress regions from the stress distribution matrix, with the threshold set to the stress mean plus 1.5 times the standard deviation. Similarly, high-fatigue regions are extracted from the fatigue degree distribution matrix. Finally, a weighted fusion method is used to couple the stress distribution matrix and fatigue degree distribution matrix for calculation, as detailed below:

[0070] R pq =w stress ·S pq ·M pq +w fatigue ·F pq ·(1-M pq )+λ·||S pq -F pq ||+∈3;

[0071] In the formula, R pq S represents the stress-fatigue correlation value at location (p, q); pq F represents the stress distribution matrix element at position (p, q); pq The elements of the fatigue distribution matrix at position (p, q); w stress and w fatigue For the combined weight of stress and fatigue; M pq The elements of the matching degree matrix are: ||Si, λ is the regularization parameter, and ||Si is the matching degree matrix element. pq -F pq || represents the L2 norm of the difference between stress and fatigue distribution; ∈3 represents the fusion error term.

[0072] The parameter acquisition method is as follows: M pq The region M with high matching degree is obtained by solving the maximum matching problem of a bipartite graph using the Hungarian algorithm. pq Region M with a low matching degree close to 1 pq Approaching 0. w stress The value is determined based on the matching degree; it is set to 0.8 when the matching degree is high and 0.2 when the matching degree is low. fatigue =1-w stress The value of λ ranges from 0.1 to 0.3. The value of ∈3 ranges from 0.005 to 0.015. The final stress-fatigue correlation matrix R is formed.

[0073] Step S 04 The specific implementation method is the same as the description of the damage assessment deep learning model mentioned above, and will not be repeated in detail here.

[0074] Step S 05The specific implementation method involves establishing a damage expansion trend matrix using a time series prediction algorithm. First, time series data of historical damage distribution matrices are collected, with a data collection frequency of once a week and a data window length of 24 weeks. Then, an autoregressive integral moving average model is used to perform trend analysis on the damage distribution matrix, while also considering the fatigue center change effect, as detailed below:

[0075]

[0076] In the formula, T rs (t) represents the damage expansion trend value at location (r, s) at time t; t is the current time, in weeks; i is the time step of the autoregressive term, i = 1, 2, ..., p; j is the time step of the moving average term, j = 1, 2, ..., q; D rs (ti) represents the damage distribution value at time (ti); φ i θ is the autoregressive coefficient. j ε is the moving average coefficient; p and q are the orders of the autoregression and moving average, respectively; rs (tj) represents the residual term at time (tj); δ represents the fatigue center change effect coefficient; ΔC center (t) represents the change in the fatigue center position; ζ is the gradient influence coefficient; ∈4 represents the damage distribution gradient; ∈4 represents the prediction error term.

[0077] The parameter acquisition method is as follows: φ i and θ j Determined by the autocorrelation function and partial autocorrelation function, p is set to 3 and q is set to 2. ΔC center (t) is obtained by monitoring changes in the center of gravity position of fatigue damage. A shift in the center of gravity position exceeding 5 mm is considered a change in the fatigue center. δ ranges from 0.3 to 0.7. ζ ranges from 0.1 to 0.4. ∈4 ranges from 0.02 to 0.06. Finally, a damage expansion trend matrix T is established.

[0078] Step S 06 The specific implementation method is the same as the description above of using a deep learning model for damage prediction to extrapolate future damage states, and will not be repeated in detail here.

[0079] Step S 07 The specific implementation method is to comprehensively evaluate the damage prediction matrix based on a game theory model and output a quality detection result matrix. First, a two-layer game theory model is constructed, with the objective functions of the upper and lower layers expressed as follows:

[0080] Upper-level model objective function:

[0081] Lower-level model objective function:

[0082] In the formula, J1 is the objective function value of the upper-level model, representing the uncertainty of quality assessment; J2 is the objective function value of the lower-level model, representing the sensitivity of damage identification; K is the total number of prediction methods, with a value of 3; P k Let P be the damage prediction matrix for the k-th prediction method, where k = 1, 2, 3 correspond to stress-based prediction, fatigue-based prediction, and comprehensive prediction, respectively; uv represents the damage prediction matrix element at location (u, v); u and v are the location indices in the lower-level model; The mean matrix for all prediction methods; w ij C represents the dimensionless quality evaluation weight for position (i, j); i and j are the position indices in the upper-level model; μ1, v1, μ2, and v2 are the objective function tradeoff coefficients; C represents the dimensionless weights. accuracy To constrain detection accuracy, the value range is 0.85 to 0.95; θ threshold The damage identification threshold is 0.3 to 0.7; T compute The calculation time is in milliseconds (ms); R sensitivity ∈5 and ∈6 are the sensitivity index, with a value range of 0.7 to 0.95; ∈5 and ∈6 are the error terms of the objective function, which are dimensionless.

[0083] Next, the evaluation weight parameters are adjusted using a gating weight function, as shown below:

[0084]

[0085] In the formula, Balance is the equilibrium value used to determine the type of gating weight function, and it is dimensionless; H(S) is the stress distribution entropy value. This represents the mean of the fatigue degree distribution. ξ represents the maximum value of the damage gradient. strength η1, η2, η3, and η4 are the material strength coefficients; η7 is the weight adjustment error term.

[0086]

[0087] In the formula, w gate (Balance) is the output value of the gating weight function, used to adjust the parameters of the neural network gating mechanism; a and b are linear function parameters; c is the steepness parameter of the nonlinear function; d and e param is the parameter of the exponential decay function; e is the base of the natural logarithm, approximately equal to 2.718.

[0088] In the formula, H(S) is the stress distribution entropy value; This represents the mean of the fatigue degree distribution. ξ represents the maximum value of the damage gradient. strength η1, η2, η3, and η4 are the material strength coefficients; a, b, c, d, and e are the gating weight function parameters; and ∈7 is the weight adjustment error term.

[0089] The parameter acquisition method is as follows: H(S) is obtained through the information entropy formula H(S)=-∑ i，j p ij logp ij Calculate the uncertainty of stress distribution, where p ij is the normalized probability of the stress value at position (i, j). It is obtained by averaging all elements of the fatigue distribution matrix. ξ is obtained by calculating the gradient of the damage distribution matrix and taking its maximum value. strength ηa represents the material strength coefficient, obtained through material mechanics testing, with a value ranging from 200 to 500 MPa. η1, η2, η3, and η4 have values ​​ranging from 0.2 to 0.4, 0.1 to 0.3, 0.3 to 0.5, and 0.15, respectively. a and b are linear function parameters, with a ranging from 1.2 to 1.8 and b ranging from 0.05 to 0.15. c is the nonlinear function steepness parameter, ranging from 8 to 12. d, e param Here, d is the parameter of the exponential decay function, with values ​​ranging from 0.8 to 1.2, and e... param The value range is 2 to 4. The range for ∈7 is 0.01 to 0.04. The final output is a quality inspection result matrix Q, where the matrix element values ​​represent the quality level evaluation at the corresponding position. The levels are divided into three grades: excellent, qualified, and unqualified, with corresponding numerical ranges of 0.8 to 1.0, 0.6 to 0.8, and 0 to 0.6, respectively.

[0090] It should be noted that the stress distribution calculation formula A quantitative relationship between stress wave propagation time and the internal stress state of a material was established based on acoustoelastic theory. The formula utilizes the change in the propagation speed of ultrasound waves in the stressed material to invert the internal stress distribution, improving detection accuracy by fusing information from both time variation and reflection intensity. Compared to existing detection methods that rely solely on a single ultrasound parameter, this formula comprehensively considers information from both propagation time and reflection intensity, significantly improving the accuracy and spatial resolution of stress distribution reconstruction. It can accurately identify stress concentration areas within welded joints, providing reliable foundational data for subsequent fatigue analysis.

[0091] Fatigue cumulative damage formula This formula introduces a logarithmic correction term for stress level and a power-law correction term for load frequency, building upon the traditional linear cumulative damage theory. The principle behind this formula is to comprehensively consider the combined effects of cycle number, stress amplitude, and load frequency on fatigue damage. The logarithmic term reflects the nonlinear influence of stress level on damage rate, while the power-law term embodies the load frequency effect. Compared to existing simplified models that only consider cycle number and stress amplitude, this formula can more accurately quantify the degree of fatigue accumulation under different working conditions, particularly significantly improving damage assessment accuracy under variable amplitude and variable frequency load conditions, thus providing a more reliable theoretical basis for predicting the fatigue life of welded joints.

[0092] Stress fatigue correlation formula R pq =w stress ·S pq ·M pq +w fatigue ·F pq ·(1-M pq )+λ·||S pq -F pq The formula ||+∈3 establishes an intelligent correlation mechanism between stress distribution and fatigue damage. The principle behind this formula is to achieve optimal fusion of stress and fatigue information through adaptive weight allocation and regularization constraints. It dynamically adjusts the fusion weights using a matching degree matrix and constrains the consistency of the two distributions through a second-normative constraint. Compared to traditional methods that analyze stress and fatigue independently, this formula achieves deep coupling analysis of the two physical quantities, accurately identifying the spatial correspondence between stress concentration and fatigue damage. It avoids potential misjudgments that may occur with single analysis methods, significantly improving the accuracy and reliability of damage identification.

[0093] Time series forecasting formula This paper innovatively introduces a fatigue center change effect term and a damage gradient term based on the autoregressive integral moving average model. The principle of this formula is to predict future damage development trends through time-series analysis of historical damage data. The fatigue center change term reflects the dynamic migration of the damage-dominant region, while the gradient term reflects the spatial characteristics of damage expansion. Compared to existing static damage analysis methods, this formula can dynamically capture the spatiotemporal characteristics of damage evolution, especially the damage center shift phenomenon under changing load conditions or material performance degradation. This provides a forward-looking decision-making basis for preventative maintenance and significantly improves the predictability and practicality of quality inspection.

[0094] The game-theoretic model's dual-level objective function establishes a balance between detection accuracy and computational efficiency through the synergistic optimization of minimizing quality assessment uncertainty in the upper-level model and maximizing damage identification sensitivity in the lower-level model. The model transforms the quality inspection problem into a multi-objective game problem, achieving intelligent adjustment of detection parameters through a dual-level optimization structure. The upper-level model focuses on the stability of evaluation results, while the lower-level model emphasizes damage identification sensitivity. Compared to existing single-objective optimization detection methods, this game-theoretic model can balance computational efficiency and real-time requirements while ensuring detection accuracy, avoiding the traditional trade-off between accuracy and efficiency, and realizing the intelligence and adaptability of the quality inspection system.

[0095] Equilibrium value calculation formula A multi-dimensional feature fusion decision index system was constructed. The principle of this formula is to integrate key features such as stress distribution entropy, mean fatigue degree, damage gradient, and material strength, and then weight and fuse them to form a unified balanced evaluation index to guide the selection of subsequent gating weight functions. Compared with existing decision-making methods based on single indicators, this formula achieves collaborative decision-making based on multi-dimensional information, comprehensively reflects the overall state of welded joints, and improves the scientificity and accuracy of quality evaluation.

[0096] The gated weight function implements an adaptive weight adjustment strategy through piecewise function design. The principle of this function is to select the appropriate weight adjustment function based on different ranges of the balance value: linear functions are suitable for low complexity cases, nonlinear functions for medium complexity cases, and exponential decay functions for high complexity cases. Compared to existing fixed-weight analysis methods, this gated mechanism achieves dynamic optimization of detection parameters and can automatically adjust the analysis strategy according to the actual state of the weld joint, significantly improving the adaptability and robustness of quality inspection.

[0097] It should be noted that the variables involved in this invention are explained in detail in Table 1.

[0098] Table 1. Variable Explanation Table

[0099]

[0100]

[0101] To better understand and implement this invention, a specific application scenario is provided below as Example 2: A power transmission tower maintenance team is responsible for structural health monitoring of a 500kV ultra-high voltage transmission line tower. The tower is 78 meters high and has been in service for 12 years. The welded joints of the reinforcing bars have shown varying degrees of fatigue damage under long-term wind loads and temperature cycles. The maintenance team uses the reinforcing bar welded joint quality inspection method of this invention to conduct a comprehensive inspection and analysis of the key welded joints at the connection points between the main tower structure and the crossarms.

[0102] First, execute step S. 01 The maintenance team selected a critical welded joint of reinforcing steel bars at the connection between the third-level crossarm and the main structure of the tower as the inspection target. This joint connects to HRB400 grade steel bars with a diameter of 25mm. Due to the exposed nature of the tower structure, it is subjected to long-term wind loads, temperature variations, and electromagnetic field influences, resulting in a complex and variable stress state at the welded joint. Ultrasonic transducers were arranged on the surface of the welded joint with a grid spacing of 2mm, covering an 80mm × 80mm rectangular area to ensure complete coverage of the welded joint and its heat-affected zone. The ultrasonic detection equipment was activated, transmitting ultrasonic signals at a frequency of 5MHz. Stress wave propagation time data was recorded using time-domain analysis. The ambient temperature during the inspection was 18℃, relative humidity was 78%, and wind speed was 12m / s. Amplitude analysis was used to obtain reflection intensity data, achieving a measurement accuracy of 0.1dB. An inversion calculation model was established based on acoustoelastic theory, and the acoustoelastic coefficient K was determined through material acoustoelastic calibration experiments. acoustic =7.8×10 9 Pa·s -1 Reference propagation time t0 = 3.9 × 10 -5 s. Reflection intensity coefficient C ij The values ​​were determined through calibration experiments and ranged from 0.2 to 0.7. The measurement error term ∈1 was controlled within 2.5% of the material's yield strength. The stress values ​​at each measuring point were obtained through inversion calculation, and a three-dimensional stress distribution matrix S was established with dimensions of 40×40×20. Table 2 shows some of the values ​​of the stress distribution matrix.

[0103] Table 2 Partial numerical values ​​of the stress distribution matrix

[0104] Position coordinates (i,j,k) Stress value (MPa) Position coordinates (i,j,k) Stress value (MPa) (8,8,4) 168.7 (16,12,6) 134.2 (12,10,5) 192.4 (20,14,8) 156.9 (14,11,5) 243.6 (24,16,10) 178.3 (18,13,6) 221.5 (28,18,12) 187.6 (20,15,7) 267.2 (32,20,14) 164.8

[0105] Next, proceed to step S. 02 The maintenance team collected historical wind load and temperature cycle data for the tower over the past 15 months. Data obtained from the tower's online monitoring system showed that the average annual wind speed in the area was 8.5 m / s, with a maximum wind speed of 35 m / s, and a temperature range of -15℃ to 42℃, with an average daily temperature difference of 18℃. Strong winds and temperature cycles in this area are the main causes of fatigue damage to the welded joints of the tower's steel reinforcement. Statistical analysis of the historical cyclic load data using the rainflow counting method identified six different wind load cycle types and four temperature cycle types. Combined with the fatigue performance parameters of HRB400 steel reinforcement, fatigue tests were conducted to determine the material constant A = 3.2 × 10⁻⁶ in the fatigue life function. 12 The fatigue life curve slope is m = 3.8. Fatigue parameters are α = 0.95, β = 0.68, and γ = 2.1. The fatigue strength limit σ... limitThe material yield strength was set to 60%, or 240 MPa. Considering the corrosive effects of the environment where the tower is located, the fatigue strength limit was reduced by 5% to 228 MPa. The linear cumulative damage theory was used to calculate the degree of fatigue accumulation at each location, and the cumulative error term ∈2 was controlled within 1.8% of the fatigue strength limit. The calculated fatigue accumulation values ​​were arranged according to spatial location to form a fatigue distribution matrix F. The matrix element values ​​were normalized to 0 and 1. Table 3 shows some of the values ​​of the fatigue distribution matrix.

[0106] Table 3. Partial numerical values ​​of the fatigue distribution matrix.

[0107] Position coordinates (m, n) Fatigue value Position coordinates (m, n) Fatigue value (10,6) 0.385 (22,18) 0.567 (13,9) 0.512 (25,20) 0.423 (16,12) 0.648 (28,22) 0.589 (19,15) 0.734 (31,24) 0.445 (22,17) 0.591 (34,26) 0.692

[0108] Then proceed with step S. 03 The operations and maintenance team preprocessed the stress distribution matrix S and fatigue distribution matrix F, using median filtering to remove noise interference caused by wind vibration. High-stress regions were extracted from the stress distribution matrix using a threshold segmentation method, with the threshold set to the mean stress plus 1.5 times the standard deviation, i.e., 201.3 MPa. Similarly, high-fatigue regions were extracted from the fatigue distribution matrix, with a threshold set to 0.62. The Hungarian algorithm was used to solve the bipartite graph maximum matching problem, treating high-stress regions as one set of vertices and high-fatigue regions as another set of vertices, to find the optimal matching scheme between the two sets. The matching results showed that 78% of the high-stress regions and high-fatigue regions had a good correspondence. This proportion is lower than that of indoor structures, mainly because the load types on the tower structure are more complex. A weighted fusion method was used to couple the two matrices for calculation, with the stress weight w for regions with high matching degrees. stress =0.75, fatigue weight w fatigue =0.25, with regions of low matching degree having opposite weights. Regularization parameter λ = 0.18, and fusion error term ∈3 controlled at 0.012. The final stress-fatigue correlation matrix R is formed, and Table 4 shows some values ​​of the stress-fatigue correlation matrix.

[0109] Table 4. Partial values ​​of the stress-fatigue correlation matrix

[0110] Position coordinates (p, q) Associated values Match Position coordinates (p, q) Associated values Match (12,8) 0.683 0.82 (24,18) 0.524 0.65 (15,11) 0.756 0.87 (27,20) 0.598 0.71 (18,14) 0.812 0.91 (30,22) 0.467 0.56 (21,16) 0.749 0.84 (33,24) 0.635 0.76 (24,18) 0.694 0.79 (36,26) 0.512 0.62

[0111] Next, proceed with step S. 04The operations and maintenance team analyzed the stress-fatigue correlation matrix based on a deep learning model for damage assessment. This model employs a multi-scale feature fusion network based on the Transformer architecture, containing three parallel encoder branches that process stress features, fatigue features, and material features respectively. Each encoder contains six layers of self-attention mechanisms and feedforward neural network layers, with eight attention heads and a hidden state dimension of 512. This analysis was combined with the theoretical parameters of cumulative material damage and fracture mechanics parameters of HRB400 steel reinforcement, including an elastic modulus of 200 GPa, Poisson's ratio of 0.3, and fracture toughness of 42 MPa·m. 0.5 Considering the corrosive effects of the environment, the material strength parameters were adjusted accordingly. A damage distribution matrix D was generated using a deep learning model, and Table 5 shows some of the values ​​for this matrix.

[0112] Table 5. Partial numerical values ​​of the damage distribution matrix.

[0113]

[0114]

[0115] Then proceed with step S. 05 The maintenance team collected historical damage distribution matrix time series data from the past 26 weeks, with data collection frequency once a week. Due to the influence of seasonal wind load changes on the tower structure, the damage evolution exhibits obvious periodic characteristics. An autoregressive integral moving average model was used for trend analysis, with model parameters set to p=3 and q=2. Monitoring revealed a 9.8mm shift in the fatigue damage centroid position at week 20, mainly due to the fatigue center change effect caused by significant changes in the direction and intensity of typhoon seasonal wind loads. The fatigue center change effect coefficient δ=0.58, and the gradient influence coefficient ζ=0.32. The autoregressive coefficients φ1=0.72, φ2=0.38, φ3=0.15, and the moving average coefficients θ1=0.45, θ2=0.29. The prediction error term ∈4 was controlled within 0.042. A damage expansion trend matrix T was established, and Table 6 shows some values ​​of the damage expansion trend matrix.

[0116] Table 6 Partial values ​​of the damage expansion trend matrix

[0117]

[0118] Next, proceed to step S. 06The operations and maintenance team used a deep learning model for damage prediction to extrapolate the future damage state from the damage expansion trend matrix. This model is based on a long short-term memory (LSTM) network architecture, containing three layers of bidirectional LSM units, with each layer having a hidden state dimension of 256. An attention mechanism is used to weightedly fuse features from different time steps, and the sequence length is set to 50 to capture sufficient temporal dependencies. Considering that the region is about to enter typhoon season, the model predicts the evolution of the damage state over the next 10 weeks, generating a damage prediction matrix P. Table 7 shows some of the values ​​in the damage prediction matrix.

[0119] Table 7 Partial values ​​of the damage prediction matrix

[0120] Position coordinates (u,v) Current damage value Predicted value in 10 weeks growth rate(%) Risk level (13,9) 0.267 0.328 22.8 Low (16,12) 0.445 0.578 29.9 middle (19,15) 0.612 0.756 23.5 high (20,16) 0.687 0.823 19.8 high (22,18) 0.745 0.867 16.4 Extremely high

[0121] Finally, execute step S. 07 The operations and maintenance team comprehensively evaluated the damage prediction matrix based on a game theory model. A two-layer game theory model was constructed: the upper-layer model aimed to minimize the uncertainty of quality evaluation, while the lower-layer model aimed to maximize the sensitivity of damage identification. The upper-layer model considered three methods: stress-based prediction, fatigue-based prediction, and a comprehensive prediction, calculating the variance between different prediction methods as an uncertainty index. The lower-layer model set the damage identification threshold to 0.58 and calculated a sensitivity index of 0.79. The evaluation weight parameters were adjusted using a gating weight function, and the equilibrium value involved the stress distribution entropy H(S) = 2.89 and the fatigue distribution mean. Maximum damage gradient Material strength coefficient ξ strength =320MPa. The equilibrium value calculation coefficients are η1=0.35, η2=0.28, η3=0.42, η4=0.31, resulting in an equilibrium value of Balance=0.612, which falls between 0.3 and 0.7. A nonlinear Sigmoid weight adjustment function is used, with a steepness parameter c=10.2. A multi-objective optimization algorithm is used to solve the game equilibrium. The genetic algorithm population size is set to 100, the number of generations to 200, the crossover probability to 0.8, and the mutation probability to 0.1. The final output quality detection result matrix Q is shown in Table 8, which contains some values ​​of the quality detection result matrix.

[0122] Table 8. Partial numerical values ​​of the quality inspection results matrix.

[0123]

[0124]

[0125] Through inspection and analysis, the maintenance team discovered significant stress concentration and fatigue damage at the welded joints of the power tower's reinforcing bars, particularly at locations (18,15) and (20,16), where the damage was severe and deemed substandard. Immediate repair and reinforcement measures are required before the typhoon season. Considering the harsh environmental conditions, it is recommended that the substandard areas be reinforced with steel plates and protected with an additional anti-corrosion coating. The remaining areas are generally in good condition, but monitoring frequency needs to be increased during the typhoon season.

[0126] Traditional methods for quality inspection of welded joints in power transmission towers primarily rely on visual inspection, ultrasonic testing, or magnetic particle testing. These methods can only detect surface or shallow defects and cannot accurately assess fatigue damage under complex environmental loads, resulting in limited accuracy. The accuracy rate of traditional methods is typically around 72%, and they fail to account for the impact of wind load variations and temperature cycling on damage evolution. This invention, by establishing a stress-fatigue correlation analysis mechanism, fully considers the influence of complex environmental loads in the region. Combined with a deep learning model for intelligent analysis and prediction, the accuracy rate is increased to 85%, an 18% improvement compared to traditional methods. More importantly, this invention can predict damage evolution trends under extreme weather conditions such as typhoon seasons, providing a scientific basis for the safe operation and preventative maintenance of power systems. Traditional methods can only perform post-incident inspections, while this invention achieves a technological shift from passive detection to proactive prevention, and a functional upgrade from single defect identification to comprehensive risk assessment, bringing significant technological advancements to the quality inspection technology of welded joints in power transmission towers.

[0127] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for quality inspection of welded joints of foundation reinforcing bars, characterized in that, include: Ultrasonic testing equipment was used to perform non-destructive testing on the welded joints of steel bars to obtain the propagation time and reflection intensity data of the internal stress waves of the welded joints and to establish a stress distribution matrix. Fatigue performance testing of the welded joints of steel bars was carried out using fatigue analysis algorithms. Based on the historical data of cyclic loads of the welded joints and the fatigue performance parameters of the materials, a fatigue distribution matrix was established. The stress-fatigue correlation analysis method is adopted to couple the stress distribution matrix and the fatigue degree distribution matrix to establish a stress-fatigue correlation matrix. The matching relationship between the stress distribution matrix and the fatigue degree distribution matrix is ​​optimized by the bipartite graph maximum matching algorithm. A damage distribution matrix is ​​generated by analyzing the stress-fatigue correlation matrix using a deep learning model for damage assessment and combining material damage accumulation theoretical parameters and fracture mechanics parameters. A time series prediction algorithm is then used to analyze the trend of the damage distribution matrix, considering the impact of fatigue center change on the damage evolution path, and a damage expansion trend matrix is ​​established. A damage prediction matrix is ​​generated by extrapolating the future damage state from the damage expansion trend matrix using a deep learning model for damage prediction. Finally, a comprehensive evaluation of the damage prediction matrix is ​​performed based on a game theory model. A multi-objective optimization algorithm is used to solve for the game equilibrium solution, and a gating weight function is used to adjust the evaluation weight parameters, outputting a quality inspection result matrix.

2. The method for quality inspection of welded joints of foundation reinforcing bars according to claim 1, characterized in that, The stress distribution matrix is ​​specifically used to characterize the stress distribution state of the welded joint of the steel bar in three-dimensional space. Each element in the matrix represents the stress value at the corresponding spatial location, which is obtained by inversion calculation using stress wave propagation time data and reflection intensity data acquired by ultrasonic detection equipment.

3. The method for quality inspection of welded joints of foundation reinforcing bars according to claim 2, characterized in that, The fatigue distribution matrix is ​​specifically used to quantify the degree of fatigue accumulation in various parts of the welded joint. It is established based on material fatigue performance parameters and historical cyclic load data, and the matrix element values ​​represent the relative degree of fatigue damage.

4. The method for quality inspection of welded joints of foundation reinforcing bars according to claim 3, characterized in that, The stress-fatigue correlation matrix is ​​specifically used to describe the relationship between the stress distribution matrix and the fatigue degree distribution matrix. Through mathematical transformation, the stress distribution matrix and the fatigue degree distribution matrix are coupled into a unified correlation analysis result, reflecting the correspondence between stress concentration areas and fatigue damage areas.

5. The method for quality inspection of welded joints of foundation reinforcing bars according to claim 4, characterized in that, The damage distribution matrix is ​​specifically used to represent the current damage state distribution of the welded joint. It is calculated by combining the stress fatigue correlation matrix, material damage accumulation theoretical parameters, and fracture mechanics parameters. The matrix element values ​​represent the damage degree values ​​at the corresponding locations.

6. The method for quality inspection of welded joints of foundation reinforcing bars according to claim 5, characterized in that, The damage propagation trend matrix is ​​specifically used to predict the development trend of damage over time. It is established based on time series data of the damage distribution matrix through a time series prediction algorithm, reflecting the speed and direction characteristics of damage propagation.

7. The method for quality inspection of welded joints of foundation reinforcing bars according to claim 6, characterized in that, The damage prediction matrix is ​​specifically used to represent the predicted damage state of the welded joint within a future time period. It is obtained by extrapolating the damage expansion trend matrix over time using a deep learning model for damage prediction.

8. The method for quality inspection of welded joints of foundation reinforcing bars according to claim 7, characterized in that, The quality inspection result matrix is ​​specifically used to represent the final quality evaluation result of the steel bar welded joint. It is calculated based on the damage prediction matrix through a game model, multi-objective optimization algorithm and gated weight function. The matrix element values ​​represent the quality level evaluation of the corresponding position.

9. The method for quality inspection of welded joints of foundation reinforcing bars according to claim 8, characterized in that, The fatigue center shift effect refers specifically to the phenomenon that the main fatigue damage area shifts during the cyclic loading of a welded joint due to changes in material properties or load conditions, which affects the damage evolution path and the final failure mode.

10. The method for quality inspection of welded joints of foundation reinforcing bars according to claim 9, characterized in that, The steps of the gating weight function are specifically used to adjust the gating mechanism parameters of the neural network. The equilibrium value is calculated based on the stress distribution entropy value, fatigue distribution mean value, damage gradient maximum value and material strength coefficient. When the equilibrium value is less than 0.3, a linear weight adjustment function is used. When the equilibrium value is between 0.3 and 0.7, a nonlinear sigmoid weight adjustment function is used. When the equilibrium value is greater than 0.7, an exponential decay weight adjustment function is used to adjust the gating parameters.

Images