A steel structure welding quality evaluation system and method based on distributed computing

Abstract

The application discloses a steel structure welding quality evaluation system and method based on distributed computing. The system includes a distributed gradient sensing unit, a physical field reconstruction unit, an entropy calculation unit, a chaotic feature extraction unit, a resource scheduling unit, a physical constraint verification unit, a chaotic synchronization analysis unit, a data aggregation processing unit, a quality comprehensive evaluation unit and a knowledge base management unit. By distributed acquisition of multi-physical field signals in the welding process and calculation of spatial gradient, the data transmission amount is significantly reduced; the complete physical field is reconstructed from the gradient information by using the multi-grid method; the potential defect area is identified and the computing resources are dynamically optimized based on information entropy analysis; the nonlinear characteristics of the sound signal are analyzed by using the chaos theory; multi-dimensional cross verification is carried out through physical constraint verification and chaotic synchronization analysis; and the federal learning is used to realize multi-site knowledge sharing. The application realizes real-time, accurate and comprehensive evaluation of the welding quality, and provides reliable technical support for steel structure welding quality control.

Description

Technical Field

[0001] This invention belongs to the field of welding quality inspection technology, and in particular relates to a steel structure welding quality assessment system and method based on distributed computing. Background Technology

[0002] Steel structure welding is widely used in the construction of major infrastructure projects such as building construction, bridge construction, and marine engineering. Welding quality directly affects structural safety and service life. Traditional welding quality inspection mainly relies on non-destructive testing technologies such as ultrasonic testing, radiographic testing, and magnetic particle testing, which have the following shortcomings: 1. Existing detection methods are mostly post-construction inspections, which cannot achieve real-time quality monitoring of the welding process; 2. The accuracy of a single detection method is limited, and it is easy to miss detections or false alarms; 3. Centralized data processing mode suffers from problems such as high data transmission latency and concentrated computational load when dealing with multi-point parallel welding in large-scale projects; 4. Finally, traditional methods are unable to capture the dynamic physical field changes and nonlinear characteristics during the welding process, and cannot deeply analyze the defect formation mechanism.

[0003] Therefore, there is an urgent need for a new technical solution that can evaluate welding quality in real time, accurately, and efficiently. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a steel structure welding quality assessment system and method based on distributed computing. Through innovative technologies such as distributed physical field gradient perception and reconstruction, information entropy-driven computing resource scheduling, and chaotic feature analysis, it achieves real-time and accurate assessment of welding quality.

[0005] The technical solution of the present invention is as follows: A steel structure welding quality assessment system based on distributed computing includes a distributed gradient sensing unit, a physical field reconstruction unit, an entropy calculation unit, a chaotic feature extraction unit, a resource scheduling unit, a physical constraint verification unit, a chaotic synchronization analysis unit, a data aggregation and processing unit, a comprehensive quality assessment unit, and a knowledge base management unit. The distributed gradient sensing unit consists of multiple sensing nodes non-uniformly deployed in the welding area. It collects multi-physics field signals of the welding process through temperature sensors, stress sensors, magnetic field sensors and sound sensors. Each node calculates the spatial gradient of the physical field using the finite difference method. The distributed gradient sensing unit transmits the temperature gradient and stress gradient to the physical field reconstruction unit, the magnetic field gradient to both the physical field reconstruction unit and the physical constraint verification unit, and the sound signal to the chaotic feature extraction unit. The physical field reconstruction unit is distributed across multiple edge computing locations. After receiving gradient information, it performs inverse reconstruction of the physical field based on the Poisson equation using the geometric multigrid method. The physical field reconstruction unit transmits the reconstructed temperature field to the entropy calculation unit and the physical constraint verification unit, the stress field to the physical constraint verification unit and the data aggregation and processing unit, and the magnetic field to the physical constraint verification unit. The entropy calculation unit receives temperature field data, calculates Shannon entropy, sample entropy and permutation entropy and weights them together to generate an entropy distribution map which is transmitted to the resource scheduling unit. The coordinates of high-entropy areas are transmitted to the data aggregation and processing unit. The chaotic feature extraction unit reconstructs the phase space of the sound signal, calculates the Lyapunov exponent, correlation dimension and multifractal spectrum, transmits the chaotic feature vector to the chaotic synchronization analysis unit, and transmits the defect identification result to the data aggregation and processing unit. The resource scheduling unit constructs an entropy gradient field based on the entropy value distribution, implements a three-level scheduling strategy, and sends resource adjustment instructions to each unit. The physical constraint verification unit substitutes the physical field into the corresponding physical equation to verify consistency, and transmits the residual exceeding the standard area to the data aggregation and processing unit and the quality comprehensive evaluation unit. The chaotic synchronization analysis unit evaluates the synchronicity of chaotic features based on the Kuramoto model, transmits the synchronicity index to the data aggregation and processing unit, and transmits the synchronization failure position to the quality comprehensive evaluation unit. The data aggregation and processing unit integrates the outputs of each unit, performs spatiotemporal registration to generate a fused dataset, and transmits it to the quality comprehensive evaluation unit. The comprehensive quality assessment unit uses a multi-attribute decision-making method to output the quality level, defect type, and location, and transmits the results to the knowledge base management unit. The knowledge base management unit stores the detection data and uses federated learning to achieve multi-site sharing, providing updated templates and parameters to each unit; Furthermore, the system also includes an adaptive mesh refinement controller and a deep reinforcement learning agent for dynamically optimizing detection accuracy and strategy.

[0006] On the other hand, the present invention provides a method for evaluating the welding quality of steel structures based on distributed computing. This method utilizes the collaborative work of the various components of the aforementioned system, and specifically includes the following steps: Step S1: Distributed multiphysics signal acquisition and gradient conversion. The distributed gradient sensing unit acquires temperature, stress, magnetic field, and sound signals through various sensors, calculates the spatial gradient locally, and transmits the gradient data and sound signals to the corresponding processing units.

[0007] Step S2: Inverse reconstruction of the physical field based on gradient information. The physical field reconstruction unit uses the multigrid method to solve the Poisson equation, recovers the complete physical field from the gradient information, and transmits the reconstruction result to the subsequent units.

[0008] Step S3: Temperature field information entropy feature extraction. The entropy calculation unit calculates the three types of information entropy and fuses them with weights to generate an entropy distribution map to guide resource scheduling. High-entropy regions participate in data fusion.

[0009] Step S4: Nonlinear dynamics analysis of the sound signal. The chaotic feature extraction unit extracts chaotic features through phase space reconstruction, matches the defect template to identify the type, and outputs feature vectors and identification results.

[0010] Step S5: Entropy-driven dynamic optimization of computing resources. The resource scheduling unit implements three-level scheduling based on the entropy gradient field, sending parameter adjustment instructions to each unit to optimize resource allocation.

[0011] Step S6: Physical Law Consistency Verification. The physical constraint verification unit verifies whether the physical field satisfies the basic physical laws and outputs the residual abnormal region.

[0012] Step S7: Multi-point chaotic feature synchronization assessment. The chaotic synchronization analysis unit assesses the phase synchronization at different locations and identifies synchronization failure locations.

[0013] Step S8: Spatiotemporal fusion of multi-source heterogeneous data. The data aggregation and processing unit integrates various detection results and performs spatiotemporal registration to generate a fused dataset.

[0014] Step S9: Comprehensive evaluation of welding quality and defect location. The comprehensive quality assessment unit calculates the quality score and determines the welding grade and defect information.

[0015] Step S10: Detect knowledge accumulation and model evolution. The knowledge base management unit updates system knowledge, achieves experience sharing through federated learning, and feeds back optimization parameters to each unit.

[0016] During execution, the adaptive mesh refinement controller and the deep reinforcement learning agent dynamically optimize the detection accuracy and strategy.

[0017] Beneficial effects:

[0018] 1. Data transmission optimization: Gradient transmission mechanism significantly reduces network bandwidth requirements; 2. Improved detection accuracy: Multi-dimensional feature fusion and cross-validation significantly improve reliability; 3. Intelligent resource utilization: Entropy-driven scheduling enables efficient allocation of computing resources; 4. Real-time processing guarantee: The distributed parallel architecture meets real-time monitoring requirements; 5. Accuracy adaptive adjustment: Dynamic mesh refinement optimizes detection resolution; 6. Secure knowledge sharing: Federated learning protects privacy while accumulating experience. Attached Figure Description

[0019] Figure 1 A schematic diagram of the system architecture described in this invention is shown; Figure 2 A flowchart illustrating the steps of the method described in this invention is shown. Detailed Implementation

[0020] Exemplary embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0021] On the one hand, combined with Figure 1 This invention provides a steel structure welding quality assessment system based on distributed computing, including a distributed gradient perception unit, a physical field reconstruction unit, an entropy calculation unit, a chaotic feature extraction unit, a resource scheduling unit, a physical constraint verification unit, a chaotic synchronization analysis unit, a data aggregation and processing unit, a comprehensive quality assessment unit, and a knowledge base management unit.

[0022] The distributed gradient sensing unit consists of multiple sensing nodes, responsible for collecting multi-physics information during the welding process. Considering the physical field distribution characteristics of the weld heat-affected zone, a non-uniform deployment strategy is adopted. A coordinate system is established with the weld centerline as the reference, and dense monitoring strips with a width three times the weld width are set on both sides of the weld, covering the main heat-affected area. The spacing between sensing nodes within the monitoring strips is set to 0.5 times the weld width to ensure that subtle changes in the physical field can be captured; the node spacing outside the monitoring strips increases exponentially. ; in: Let be the spacing of the i-th node; The baseline spacing is the distance between nodes within a dense monitoring zone. The attenuation coefficient is taken as 0.1 to 0.3 based on the thermal diffusion characteristics of the material; This is the vertical distance from the node to the centerline of the weld. is the base of the natural logarithm.

[0023] Each sensing node is equipped with multiple sensors to acquire comprehensive physical field information. The temperature sensor uses a type K thermocouple with a temperature measurement range of -200℃ to 1370℃, covering the temperature changes throughout the entire steel structure welding process. Its response time of less than 100ms ensures the ability to capture rapid temperature changes. The stress sensor uses a fiber Bragg grating sensor, based on the principle that the wavelength of the grating reflection changes with strain. Its strain measurement range is ±5000με, with a resolution of 1με, enabling accurate detection of the welding stress field. The magnetic field sensor uses an anisotropic magnetoresistive sensor with a measurement range of ±8 Gauss and a resolution of 0.01 milligauss. The sound sensor uses a wideband piezoelectric microphone with a frequency response range of 20Hz to 100kHz and a sensitivity of 50mV / Pa, comprehensively acquiring the welding arc sound signal.

[0024] The core innovation of the distributed gradient sensing unit lies in local gradient calculation. Traditional methods that directly transmit raw physical field data incur a huge network load. This invention calculates the spatial gradient of the physical field at each sensing node, transmitting only gradient information. A fourth-order precision central difference scheme is used for gradient calculation. ; For internal nodes, the temperature gradient component in the x-direction is calculated using a five-point difference scheme: in: This is the temperature gradient vector; The partial derivative in the x-direction; Spatial grid points Temperature value at that location; and These are the temperature values ​​of adjacent and spaced one grid point in the x-direction, respectively; Let x be the spatial step size in the x-direction.

[0025] The gradient components in the y and z directions are calculated using the same method, employing node data for each direction. For boundary nodes, one-sided difference or virtual node techniques are used to ensure calculation accuracy. Stress and magnetic field gradients are calculated using the same numerical format.

[0026] The distributed gradient sensing unit transmits the calculated temperature gradient and stress gradient to the physical field reconstruction unit. The magnetic field gradient is simultaneously transmitted to both the physical field reconstruction unit and the physical constraint verification unit. The sound signal, which requires time-series analysis, is transmitted to the chaotic feature extraction unit while maintaining its original format.

[0027] The physics field reconstruction units are distributed across multiple edge computing locations, receiving gradient data and then performing inverse physics field reconstruction. The mathematical essence of this process is solving the Poisson equation: ; in: The physical field to be reconstructed, including the temperature field. Stress field or magnetic field ; The Laplace operator is defined as follows: ; The source term is obtained by calculating the divergence of the gradient, i.e. .

[0028] To efficiently solve this equation, a geometric multigrid method is employed. This method rapidly eliminates error components of different frequencies by iterating between grids of different scales. A two-layer system of fine and coarse grids is constructed. The size of the fine grid is consistent with the minimum spacing between sensing nodes to fully utilize the acquired data, while the size of the coarse grid is four times that of the fine grid to quickly eliminate low-frequency errors.

[0029] The V-loop iterative process first performs pre-smoothing on the fine mesh, then updates the mesh point values ​​using the red-black Gauss-Seidel method: in: For nodes The value after the (n+1)th iteration; , , The values ​​of the six adjacent nodes in the nth iteration; For finer grid step size; This is the value of the source item at this node.

[0030] After 5 iterations of fine mesh calculation, the residual is calculated. And it is passed to the coarse mesh via the constraint operator: ;in, For coarse grid residuals; For fine-grid residuals; To constrain the operator matrix, a full-weight constraint is used to ensure accuracy.

[0031] Solving the error equation on a coarse mesh Obtaining error Then, the error is propagated back to the fine mesh for correction using interpolation operators: ;in, The corrected fine-grid physical field; The interpolation operator matrix is ​​obtained by using trilinear interpolation. This represents the error on the coarse grid.

[0032] To achieve parallel computing, the physical field reconstruction unit employs a domain decomposition method, dividing the computational domain into multiple subdomains, each assigned to a computational process for independent iteration. Subdomains exchange boundary data through an overlapping region of two grid cells, using asynchronous communication to reduce waiting time. After reconstruction, the temperature field distribution is transmitted to the entropy calculation unit and the physical constraint verification unit, the stress field distribution to the physical constraint verification unit and the data aggregation processing unit, and the magnetic field distribution to the physical constraint verification unit.

[0033] The entropy calculation unit receives temperature field data and evaluates the stability of the welding process through information entropy analysis. A sliding window mechanism is used for real-time entropy calculation, with the window length adjusted according to the welding speed. and sampling frequency Dynamic adjustment: ;in, The length of the window; The window coefficient is 2 to 5, depending on the welding process. The sampling frequency; This refers to the welding speed.

[0034] This adaptive window design ensures that each window contains enough data points for statistical analysis, while also ensuring that the entropy value reflects the current welding status.

[0035] The entropy calculation unit simultaneously calculates three types of information entropy to comprehensively evaluate data characteristics. Shannon entropy reflects the statistical characteristics of temperature distribution. First, the temperature range is divided into M equal-width intervals, and the probability distribution of each interval is statistically analyzed: ;in, Shannon entropy; Let be the probability that the temperature falls within the i-th interval; This is the number of discrete intervals, ranging from 20 to 50 depending on the data resolution; It represents the logarithm to the base 2.

[0036] Sample entropy assesses the regularity and predictability of time series. Its calculation process includes constructing an m-dimensional template vector and counting the number of template pairs whose distance is less than the tolerance r. in: The sample entropy; The embedding dimension is typically 2 or 3; The tolerance threshold is set to 0.1 to 0.25 times the standard deviation of the sequence. The sequence length; The number of template matches in the m+1 dimension; The number of m-dimensional template matches; It is the natural logarithm.

[0037] Permutation entropy assesses dynamic complexity by analyzing the distribution of sequential patterns. ;in, The permutation entropy; Let be the relative frequency of the j-th permutation pattern; Let m be the factorial of m, representing the number of all possible permutations of the m-dimensional vector.

[0038] The three entropy values ​​are weighted and fused to obtain a comprehensive entropy value: ;in, This is the overall entropy value; , , The weighting coefficients satisfy the normalization condition. The optimal weights are determined by principal component analysis.

[0039] Entropy value is higher than the threshold The region marked as a high-entropy region indicates high data complexity and potential welding defects. The mean of the global entropy value. The standard deviation of the entropy value is denoted as . The entropy calculation unit transmits the entropy distribution map to the resource scheduling unit for optimizing the allocation of computing resources, and the coordinates of high-entropy regions are transmitted to the data aggregation and processing unit for subsequent fusion analysis.

[0040] The chaotic feature extraction unit is specifically designed to handle the nonlinear dynamic characteristics of welding sound signals. During welding, the arc sound contains rich process information. Under normal welding conditions, the sound signal exhibits quasi-periodic characteristics, while defects cause changes in the chaotic characteristics of the sound signal. First, the one-dimensional sound time series is reconstructed into a high-dimensional phase space using the delay coordinate method, with a delay time... The embedding dimension is determined by calculating the autocorrelation function and finding the first zero point. Determined by the spurious nearest neighbor method: ;in, Let i be the i-th phase space vector; This represents the i-th data point in the original time series. For delay time; The embedding dimension is typically between 3 and 7.

[0041] In the reconstructed phase space, the maximum Lyapunov exponent is calculated to quantify the system's sensitivity to initial conditions: ;in, The maximum Lyapunov index; Let Euclidean distance be the distance between two neighboring trajectories at the initial moment. The distance after n time steps; The sampling time interval is denoted by . A positive Lyapunov exponent indicates that the system exhibits chaotic characteristics; the larger the value, the stronger the degree of chaos.

[0042] The correlation dimension describes the geometric complexity of the phase space attractor and is obtained by calculating the correlation integral at different scales: ;in, For the correlation dimension; The correlation integral is defined as the proportion of point pairs with a distance less than r out of the total number of point pairs; This is the distance threshold; in actual calculations, it is obtained by fitting the slope of a straight line segment in a double logarithmic coordinate system.

[0043] Multifractal spectral analysis is achieved using the wavelet leader method, which first involves performing a continuous wavelet transform on the sound signal: ;in, These are wavelet coefficients; The scaling parameter controls the scaling of the wavelet; The displacement parameter controls the translation of the wavelet; For the complex conjugate of the mother wavelet, Morlet wavelet or Daubechies wavelet is used.

[0044] The local Hölder exponent is calculated based on wavelet coefficients, and then the multifractal spectrum is obtained:

[0045] in: It is a multifractal spectrum; It is the singularity index; is the order of moments, with values ​​ranging from -5 to 5; It is a quality index function, determined by the scaling behavior of the partition function.

[0046] The chaotic feature extraction unit established a feature template library for typical defects. Porosity defects exhibit a Lyapunov exponent ranging from 0.05 to 0.15, indicating weak chaos; the correlation dimension is 2.3 to 2.7, with a relatively simple attractor structure; and the peak Hölder exponent is between 0.4 and 0.5. Crack defects exhibit a Lyapunov exponent ranging from 0.15 to 0.25, indicating a higher degree of chaos; the correlation dimension is 2.8 to 3.2, with a complex attractor structure; and the peak Hölder exponent is between 0.6 and 0.7. Unfused defects fall between these two extremes, with a Lyapunov exponent of 0.08 to 0.18, a correlation dimension of 2.5 to 2.9, and a peak Hölder exponent of 0.45 to 0.55. Defect identification is performed by calculating the Euclidean distance between the real-time features and the templates. The chaotic feature vector is transmitted to the chaotic synchronization analysis unit, and the defect type identification results are transmitted to the data aggregation and processing unit.

[0047] After receiving the entropy distribution map, the resource scheduling unit constructs an entropy gradient field covering the entire welding area to guide the optimal allocation of computing resources. The spatial gradient of the entropy is calculated using the Sobel operator. in: The entropy gradient is in the x-direction. The entropy gradient is in the y-direction. It is an entropy matrix; This represents the convolution operation.

[0048] The magnitude and direction of the entropy gradient are: ; in: This represents the magnitude of the entropy gradient, reflecting the degree of drastic change in entropy. The entropy gradient direction indicates the direction in which the entropy value increases the fastest.

[0049] Based on the entropy gradient field, the resource scheduling unit implements a three-level scheduling strategy. Intra-node scheduling dynamically adjusts the sampling frequency and calculation accuracy according to the local entropy value. in: The adjusted sampling frequency; The reference sampling frequency; This is the frequency adjustment coefficient, ranging from 0.5 to 2.0; The local entropy value of the node; The mean of the global entropy; is the standard deviation of the entropy value.

[0050] Inter-node scheduling is achieved through a load balancing mechanism, defining the node load level: in: Let be the load degree of the i-th node; This represents the current computational workload. For node computing power; The entropy influence factor is set between 0.3 and 0.5. Let be the entropy value at node i.

[0051] When the load of adjacent nodes differs When the threshold of 0.3 is exceeded, task migration is triggered, transferring some tasks from high-load nodes to low-load nodes.

[0052] Global scheduling employs the Lagrange relaxation method to optimize overall resource allocation, with the objective function being to minimize the total processing latency. ; in: The workload of node i; The computing resources allocated to node i; Total system computing resources; It is a Lagrange multiplier; This represents the total number of nodes.

[0053] The optimal resource allocation is obtained by solving the KKT conditions: ; in: The optimal resource allocation for node i.

[0054] The resource scheduling unit supports unified management of four types of heterogeneous computing resources: CPU, GPU, FPGA, and DSP, and maintains a task-resource adaptation matrix. , of which elements This indicates the relative execution efficiency of task type i on resource type j: ; in: This is a relative efficiency value, ranging from 0 to 1; CPU execution time was selected as the baseline execution time. This represents the actual execution time of task i on resource j.

[0055] Efficiency values ​​were determined through benchmark testing: matrix operations on GPUs achieve 5 to 10 times the speedup compared to CPUs, with efficiency values ​​set at 0.8 to 0.9; signal processing achieves 3 to 5 times the speedup on DSPs, with efficiency values ​​at 0.7 to 0.85; real-time control on FPGAs reduces latency by an order of magnitude, with efficiency values ​​at 0.85 to 0.95; general-purpose control logic is most flexible to execute on CPUs, with an efficiency value set at 1.0 as the baseline.

[0056] Resource allocation decisions take into account both efficiency and availability: ; in: Assigning task i to resource j as a suitability score; Let be the current utilization rate of resource j; The degree of matching between task characteristics and resource capabilities.

[0057] The physical constraint verification unit receives reconstructed physical field data and identifies anomalous regions by verifying whether the data satisfies fundamental physical laws. For the temperature field, it verifies the heat conduction equation: in: For material density, carbon steel is taken as 7850 kg / m³; The specific heat capacity is approximately 450 J / (kg·K) at room temperature; For temperature; For time; Thermal conductivity; This is the heat source for welding.

[0058] Considering the temperature dependence of material parameters, a piecewise linear model is used for thermal conductivity: in: The thermal conductivity at room temperature is approximately 50 W / (m·K); The temperature influence coefficient is taken as 0.3 to 0.5; Room temperature; This is the melting point temperature.

[0059] For the stress field, verify the equilibrium equations and compatibility equations, while also considering the effects of thermal stress: in: The total stress; Mechanical stress; Thermal stress; The elastic modulus is approximately 200 GPa. The coefficient of thermal expansion is approximately 12 × 10⁻ 6 / ℃; For reference temperature; It is Poisson's ratio, approximately 0.3.

[0060] An adaptive threshold is introduced when calculating the physical field residuals: in: For dynamic thresholds; The baseline threshold is set to 0.1. This is the time adjustment factor; For welding time; It is the stable time constant.

[0061] When the actual residual exceeds the dynamic threshold, it is marked as a physical anomaly region. The physical constraint verification unit transmits the coordinates of the anomaly region and the residual value to the data aggregation and processing unit and the quality comprehensive evaluation unit.

[0062] The chaotic synchronization analysis unit receives chaotic feature vectors from multiple spatial locations and evaluates the synchronicity of the welding process at different locations based on the Kuramoto coupled oscillator model. During normal welding, the chaotic features at each location should evolve synchronously; when a defect occurs, the chaotic features at the defect location will deviate from the overall synchronous state.

[0063] Assuming there are N observation points, the phase evolution at each point follows the equation: in: The instantaneous phase of the i-th observation point is extracted from the chaotic feature time series using the Hilbert transform. The natural frequency is determined based on the Lyapunov index; The coupling strength reflects the degree of mutual influence between points, with a typical value of 2 to 5. The phase difference is a sine wave, which drives phase synchronization.

[0064] Define an order parameter to quantify the overall synchronization level: ; in: This is a synchronization indicator, with a value ranging from 0 to 1; The imaginary unit; It is a complex number representation on the unit circle.

[0065] Synchronicity assessment adopts a graded standard: when The timing indicates strong synchronization, and the welding process is stable and normal. This indicates weak synchronization, which may contain minor defects or process fluctuations. The timing indicates a loss of synchronization, which is a serious flaw.

[0066] To identify local synchronization failures, a synchronization matrix is ​​constructed: ; in: This measures the instantaneous synchronization between nodes i and j, with a value ranging from -1 to 1. , This represents the instantaneous phase of the corresponding node.

[0067] Synchronization groups are identified by analyzing the synchronicity matrix using spectral clustering algorithms. ; in: It is a Laplace matrix; For a degree matrix, the diagonal elements ; This is the synchronization matrix.

[0068] The eigenvalues ​​of the Laplacian matrix are calculated, and the second smallest eigenvalue (Fiedler value) reflects the network connectivity. A Fiedler value close to 0 indicates a significant cluster split, with the split location corresponding to the defect location. The chaotic synchronization analysis unit transmits the synchronization index and cluster analysis results to the data aggregation and processing unit, while the specific coordinates of the synchronization failure location are transmitted to the quality comprehensive evaluation unit.

[0069] The data aggregation and processing unit, serving as the system's data integration center, receives multi-source heterogeneous data from various processing units, including stress field data from the physical field reconstruction unit, high-entropy region coordinates from the entropy calculation unit, defect identification results from the chaotic feature extraction unit, residual exceeding-standard regions from the physical constraint verification unit, and synchronization indicators from the chaotic synchronization analysis unit. First, time alignment is performed. Due to differences in processing latency and sampling rates among the units, a dynamic time warping algorithm is employed. in: Let X be the dynamic time-warped distance between sequences X and Y; Align the path; This is the path length; It is a distance metric function; , The index of the path on the two sequences.

[0070] Find the optimal alignment path using dynamic programming: in: This is the cumulative distance matrix; , Given two sequences of data points; the optimal path is obtained by backtracking the minimum cumulative distance.

[0071] Spatial registration takes into account the thermal deformation during the welding process, and non-rigid registration is performed using thin-plate spline interpolation: ; in: It is a transformation function; For constant terms; This is the affine transformation coefficient vector; These are the weights of the radial basis functions; For thin plate spline kernel functions; For control points.

[0072] Data fusion employs an improved DS evidence theory, introducing a credibility factor to address evidence conflicts: in: To assign a value to the basic probability of hypothesis A after fusion; For the i-th source of evidence, the hypothesis is... The basic probability assignment; Let be the credibility factor of the i-th source of evidence; The normalization constant is .

[0073] The confidence factor is dynamically adjusted based on historical accuracy: ; in: Let be the credibility factor of the i-th source of evidence; This represents the historical accuracy of the evidence source.

[0074] The data aggregation and processing unit generates a unified format multimodal fusion dataset, which includes multidimensional information such as spatiotemporal coordinates, physical field values, entropy values, chaotic features, synchronization indicators, and physical residuals, and transmits it to the quality comprehensive evaluation unit.

[0075] After receiving the fused dataset, the comprehensive quality assessment unit uses a multi-criteria decision-making method combining the Analytic Hierarchy Process (AHP) and Top-Level Strategy (TOPSIS) to conduct a comprehensive quality assessment. First, a hierarchical structure of the assessment index system is constructed. The top layer represents the overall welding quality objective; the middle layer includes three criteria: physical consistency, process stability, and defect characteristics; and the bottom layer includes various specific indicators.

[0076] Construct a judgment matrix for pairwise comparisons: in: The importance ratio of criterion i to criterion j is expressed using a 1-9 scale.

[0077] Calculate the largest eigenvalue and eigenvector of the judgment matrix: ; in: It is the largest eigenvalue; The corresponding feature vector is normalized to obtain the weight vector.

[0078] Consistency check: ; in: As a consistency indicator; The order of the matrix; Consistency ratio; This is a random consistency indicator. When... It is assumed that the judgment matrix has satisfactory consistency.

[0079] Based on defined weights, the TOPSIS method is used for multi-attribute decision-making. A weighted normalized decision matrix is ​​constructed as follows: ; in: The weighted normalized value; Let be the weight of the j-th attribute; This represents the original value of the j-th attribute at the i-th evaluation point; To assess the number of points.

[0080] Determine the ideal solution and negative ideal solution : in: The optimal value for the j-th attribute; Let j be the worst value of the j-th attribute.

[0081] Calculate the distance from each evaluation point to the ideal solution: in: Let be the Euclidean distance from the i-th evaluation point to the positive ideal solution; This is the distance to the negative ideal solution.

[0082] Calculate the relative similarity as a comprehensive quality score: ; in: The quality score for the i-th evaluation point ranges from 0 to 1; a larger value indicates better quality.

[0083] Classification based on quality scores: The weld quality is rated as excellent and meets the design requirements. It is rated as acceptable; minor defects exist but do not affect its use. This is a substandard grade and requires repair or re-welding.

[0084] The probability distribution of defect types is determined through Bayesian inference: in: For the defect type under given evidence E The posterior probability; It is the likelihood function, determined based on the defect feature template; This is a priori probability, based on historical statistical data; This represents the total number of defect types.

[0085] Defect localization uses the weighted centroid method: ; in: Location of the defect center; Let i be the coordinates of the i-th anomaly point; Rate the quality of this point; This is the set of outliers.

[0086] The comprehensive quality assessment unit generates a detailed assessment report, including information such as welding quality grade, probability distribution of defect types, spatial location of defects, and confidence interval, and transmits the assessment results to the knowledge base management unit.

[0087] The knowledge base management unit is responsible for the accumulation, updating, and sharing of system knowledge, and adopts a hierarchical storage architecture to manage different types of knowledge. The historical data storage module saves the original detection data and evaluation results, and uses a time-series database to optimize storage and query efficiency; the feature template library stores standard feature patterns of various defects and supports dynamic updates; the model parameter library maintains the parameter configurations of all algorithms in the system and implements version control and rollback mechanisms.

[0088] Knowledge updates employ an incremental learning strategy, where new detection cases update the model through an online learning algorithm. in: These are the updated model parameters; For the current parameter; The learning rate is adaptively adjusted. The gradient of the loss function; This is for new sample data.

[0089] To prevent catastrophic forgetting, the Elastic Weight Consolidation (EWC) method is employed: in: This is the total loss function; Losses due to new missions; The regularization coefficient is used. These are the diagonal elements of the Fisher information matrix, used to measure parameter importance; These are the historically optimal parameters.

[0090] The knowledge base management unit employs a federated learning framework to achieve collaborative learning across multiple welding sites. Each site retains its original data and only shares model updates. in: This represents the parameter update amount for the i-th site. These are the parameters after local training; These are global model parameters.

[0091] The central server aggregates updates from all sites and uses a federated averaging algorithm. in: The number of participants on site; Let be the number of samples at the i-th site; This represents the total number of samples.

[0092] To protect privacy, differential privacy noise is added during parameter transmission: in: Update the parameters after adding noise; It is Gaussian noise; For noise scale; Sensitivity refers to the maximum impact of a single sample on a parameter.

[0093] Privacy budgets are calculated using the combinatorial theorem: in: Total privacy budget; This represents the number of iteration rounds. For a single round of privacy budget; This represents the probability of failure.

[0094] During system operation, all units work collaboratively to form a closed-loop control. The adaptive mesh refinement controller continuously monitors the system status and dynamically adjusts the detection strategy when an anomaly is detected. Refinement trigger conditions include: ; in: To refine the trigger flags; This is the local entropy value; The entropy threshold; For physical field residuals; The residual threshold; As a synchronicity indicator; This is the synchronization threshold.

[0095] After refinement is triggered, the mesh size is adjusted according to the multi-mesh strategy: in: This refers to the refined grid size; Original dimensions; To refine the levels, a range of 1 to 3 is used based on the severity of the abnormality.

[0096] Simultaneously adjust the relevant calculation parameters: ; in: For the new time step; The original time step is used; the square relation ensures numerical stability.

[0097] The system's configured deep reinforcement learning agent continuously optimizes the detection strategy through interaction with the environment. The state space is defined as follows: ; in: This is the system state vector; Characteristics of entropy distribution; This is a vector representing resource utilization rates. To improve detection accuracy; This is the average delay; To calculate the cost.

[0098] Action space includes discrete and continuous actions: in: For action vectors; This is the sampling frequency adjustment factor; For the calculation accuracy level; Choose an index for the algorithm; This refers to the proportion of resources allocated.

[0099] The agent is trained using the Proximal Policy Optimization (PPO) algorithm, with the objective function being: in: This is an alternative objective function for pruning; It represents the probability ratio; For the estimation of the advantage function; The trimming parameter is set to 0.2.

[0100] The advantage function employs generalized advantage estimation (GAE): ; in: Discount factor; For GAE parameters; This refers to timing difference error.

[0101] The system outputs an evaluation report presented in a visual format. The 3D cloud map of the physical field is displayed using volume rendering technology.

[0102] in: For pixels Color intensity; Opacity function; For color functions; The depth of vision.

[0103] The spatiotemporal evolution of entropy is represented by a heatmap, and the color mapping function is: in: The color corresponding to the entropy value; , , This is the segmentation threshold.

[0104] The chaotic feature spectrum shows the changes of the Lyapunov exponent over time and space, generated using wavelet time-frequency analysis: ; in: The time-frequency spectral density; For Lyapunov exponent time series; These are wavelet basis functions.

[0105] The defect probability distribution map is displayed using contour lines, and the defect probability of each spatial point is obtained through kernel density estimation. ; in: Let x be the defect probability density. The number of samples; For bandwidth parameters; The dimension of the space; The kernel function is a Gaussian kernel.

[0106] The confidence level of each evaluation result was estimated using the Bootstrap method: ; in: The confidence interval is 95%. and These are the 2.5 and 97.5 percentiles for the Bootstrap sample.

[0107] On the other hand, combining Figure 2 This invention provides a method for evaluating the welding quality of steel structures based on distributed computing. This method utilizes the collaborative work of the various components of the aforementioned system, and specifically includes the following steps: Step S1: Distributed Multiphysics Signal Acquisition and Gradient Conversion. The distributed gradient sensing unit is activated, and each sensing node synchronously acquires multiphysics signals from the welding area using temperature, stress, magnetic field, and sound sensors. Each node performs a fourth-order central difference operation locally, converting the raw physical field data into gradient information. The temperature and stress gradients are transmitted to the physical field reconstruction unit, while the magnetic field gradient is simultaneously transmitted to both the physical field reconstruction unit and the physical constraint verification unit. The sound signal is transmitted to the chaotic feature extraction unit while maintaining its original format.

[0108] Step S2: Inverse reconstruction of the physical field based on gradient information. The physical field reconstruction unit receives gradient data and recovers the complete physical field distribution by solving the Poisson equation. A geometric multigrid method is used, iterating between coarse and fine grids, and transmitting error and correction information through constraint operators and interpolation operators. The reconstructed temperature field is transmitted to the entropy calculation unit and the physical constraint verification unit; the stress field is transmitted to the physical constraint verification unit and the data aggregation processing unit; and the magnetic field is transmitted to the physical constraint verification unit.

[0109] Step S3: Temperature Field Information Entropy Feature Extraction. The entropy calculation unit performs complexity analysis on the temperature field data and calculates information entropy indices in three dimensions—Shannon entropy, sample entropy, and permutation entropy—using a sliding window mechanism. After weighted fusion, an entropy distribution map is generated. The entropy distribution map is transmitted to the resource scheduling unit to guide the allocation of computing resources, and the coordinates of high-entropy regions are transmitted to the data aggregation and processing unit for subsequent fusion.

[0110] Step S4: Nonlinear dynamics analysis of the sound signal. The chaotic feature extraction unit performs phase space reconstruction on the sound signal, extracting chaotic features such as the Lyapunov exponent, correlation dimension, and multifractal spectrum. Defect type identification is achieved by matching with the defect feature template library. The chaotic feature vector is transmitted to the chaotic synchronization analysis unit, and the defect identification results are transmitted to the data aggregation and processing unit.

[0111] Step S5: Entropy-driven dynamic optimization of computing resources. The resource scheduling unit constructs an entropy gradient field based on the entropy value distribution and identifies computation hotspots. A three-level scheduling strategy is implemented to optimize resource allocation, sending sampling adjustment instructions to the distributed gradient sensing unit, mesh adjustment instructions to the physics field reconstruction unit, and precision adjustment instructions to the computing unit.

[0112] Step S6: Physical Law Consistency Verification. The physical constraint verification unit substitutes the reconstructed physical field and the received magnetic field gradient into the corresponding physical equations to verify whether the constraints of heat conduction, mechanical equilibrium, and electromagnetic field are satisfied. It calculates the residual between the actual field and the theoretical prediction, and transmits the information on anomalous regions to the data aggregation and processing unit and the quality comprehensive evaluation unit.

[0113] Step S7: Multi-point Chaotic Feature Synchronization Assessment. The chaotic synchronization analysis unit analyzes the phase synchronization of chaotic features at different spatial locations based on the Kuramoto model, and determines the stability of the welding process through order parameters. The synchronization index is transmitted to the data aggregation and processing unit, and the synchronization failure location is transmitted to the comprehensive quality assessment unit.

[0114] Step S8: Spatiotemporal fusion of multi-source heterogeneous data. The data aggregation and processing unit integrates the detection results from various processing units, including stress field data, high-entropy regions, defect identification, residual anomalies, and synchronicity indicators. Temporal alignment and spatial registration are performed to generate a unified fused dataset, which is then transmitted to the quality comprehensive evaluation unit.

[0115] Step S9: Comprehensive Welding Quality Assessment and Defect Location. Based on fused data and directly received anomaly information, the comprehensive quality assessment unit uses a multi-attribute decision-making method to calculate a quality score, determine the welding grade, identify defect types, and locate their spatial positions. The assessment results are transmitted to the knowledge base management unit.

[0116] Step S10: Detection of knowledge accumulation and model evolution. The knowledge base management unit stores evaluation results and updates defect feature templates and detection parameters. Federated learning enables multi-site experience sharing, feeding back new templates to the chaotic feature extraction unit, providing optimization weights to the quality comprehensive evaluation unit, and providing load patterns to the resource scheduling unit.

[0117] During the execution of the above steps, when a local anomaly is detected, the adaptive mesh refinement controller automatically triggers an improvement in the accuracy of the corresponding area; the deep reinforcement learning agent continuously optimizes the system strategy to ensure dynamic optimization of detection performance.

[0118] In summary, the steel structure welding quality assessment system and method based on distributed computing provided by this invention solves the data transmission bottleneck through distributed gradient computation, achieves intelligent resource scheduling driven by information entropy, enhances defect identification capabilities through chaotic synchronization analysis, and realizes privacy-preserving knowledge sharing through a federated learning framework. This invention overcomes the limitations of traditional detection methods, achieving real-time, accurate, and comprehensive assessment of welding quality, and providing reliable technical support for steel structure welding quality control.

[0119] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A steel structure welding quality assessment system based on distributed computing, characterized in that, It includes a distributed gradient perception unit, a physical field reconstruction unit, an entropy calculation unit, a chaotic feature extraction unit, a resource scheduling unit, a physical constraint verification unit, a chaotic synchronization analysis unit, a data aggregation and processing unit, a quality comprehensive evaluation unit, and a knowledge base management unit; The distributed gradient sensing unit collects temperature field, stress field, magnetic field and sound signal data of the welding process through temperature sensor, stress sensor, magnetic field sensor and sound sensor respectively. It calculates the spatial gradient of temperature field, stress field and magnetic field data, and transmits the temperature gradient and stress gradient to the physical field reconstruction unit. The magnetic field gradient is transmitted to the physical field reconstruction unit and the physical constraint verification unit at the same time. The sound signal is transmitted to the chaotic feature extraction unit. The physical field reconstruction unit, based on the received gradient information, uses the multigrid method to solve the Poisson equation and reconstructs the spatial distribution of the temperature field, stress field, and magnetic field. The reconstructed temperature field is transmitted to the entropy calculation unit and the physical constraint verification unit, the stress field is transmitted to the physical constraint verification unit and the data aggregation and processing unit, and the magnetic field is transmitted to the physical constraint verification unit. The entropy calculation unit calculates the information entropy of the temperature field data, generates an entropy distribution map and transmits it to the resource scheduling unit, and transmits the coordinates of the high-entropy area to the data aggregation and processing unit. The chaotic feature extraction unit performs nonlinear dynamic analysis on the sound signal, extracts chaotic feature vectors and transmits them to the chaotic synchronization analysis unit, and transmits the defect identification results to the data aggregation and processing unit. The resource scheduling unit constructs an entropy gradient field based on the entropy value distribution, optimizes the allocation of computational resources, and sends adjustment instructions to each unit. The physical constraint verification unit substitutes the reconstructed physical field into the physical equation to verify consistency, and transmits the residual exceeding the standard area to the data aggregation and processing unit and the quality comprehensive evaluation unit. The chaotic synchronization analysis unit evaluates the synchronicity of chaotic features at multiple spatial locations, transmits the synchronicity index to the data aggregation and processing unit, and transmits the synchronization failure location to the quality comprehensive evaluation unit. The data aggregation and processing unit integrates the output data of each unit, performs spatiotemporal registration to generate a fused dataset, and transmits it to the quality comprehensive evaluation unit. The comprehensive quality assessment unit outputs welding quality level, defect type, and location information based on fused data, and transmits the assessment results to the knowledge base management unit. The knowledge base management unit stores historical data and updates system parameters, providing optimized templates and parameters to each unit.

2. The system according to claim 1, characterized in that, The distributed gradient sensing unit adopts a non-uniform deployment strategy, setting up dense monitoring strips on both sides of the weld. The width of the monitoring strip is a set multiple of the weld width. The node spacing inside the monitoring strip is smaller than the node spacing outside the monitoring strip, and the node spacing outside the monitoring strip increases exponentially.

3. The system according to claim 1, characterized in that, The physical field reconstruction unit achieves inverse reconstruction from gradient to physical field by solving the Poisson equation. It adopts a cyclic iterative strategy of geometric multigrid method, performs local calculations in fine grid and global correction in coarse grid, and transmits information between grid levels through constraint operators and interpolation operators.

4. The system according to claim 1, characterized in that, The entropy calculation unit adopts a sliding window mechanism to simultaneously calculate three information entropy indices: Shannon entropy, sample entropy, and permutation entropy. The comprehensive entropy value is obtained through weighted fusion. When the entropy value exceeds the dynamic threshold, it is marked as a high-entropy region.

5. The system according to claim 1, characterized in that, The chaotic feature extraction unit uses the delayed coordinate method and embedding theorem to reconstruct the sound signal into a high-dimensional phase space, calculates the Lyapunov exponent, correlation dimension and multifractal spectrum features, and establishes a feature template library containing a variety of typical welding defects.

6. The system according to claim 1, characterized in that, The resource scheduling unit implements a three-level scheduling strategy: intra-node scheduling dynamically adjusts the sampling frequency and calculation accuracy based on the local entropy value; inter-node scheduling achieves load balancing through task migration. Global scheduling employs an optimization algorithm for overall resource allocation.

7. The system according to claim 1, characterized in that, The chaotic synchronization analysis unit calculates the sequence parameter based on the coupled oscillator model, and determines the strong synchronization, weak synchronization or desynchronization state according to the different intervals of the sequence parameter value, which correspond to normal welding, minor defects and serious defects, respectively.

8. The system according to claim 1, characterized in that, The comprehensive quality assessment unit uses the analytic hierarchy process (AHP) to determine the weights of each indicator, employs a multi-criteria decision-making method to calculate the quality score, determines the distribution of defect types through probabilistic reasoning, and uses a weighted location method to determine the spatial location of defects.

9. The system according to claim 1, characterized in that, The knowledge base management unit adopts a federated learning framework, where each welding site only shares the model parameter update amount and not the original data. Model optimization is achieved through parameter aggregation algorithms, and privacy protection technology and anti-forgetting mechanisms are used to ensure data security and knowledge accumulation.

10. A method for evaluating the welding quality of steel structures based on distributed computing, employing the system described in any one of claims 1 to 9, characterized in that, Includes the following steps: The temperature field, stress field, magnetic field and sound signal of the welding area are collected by a distributed gradient sensing unit, and the spatial gradient of temperature, stress and magnetic field are calculated. The complete temperature field, stress field, and magnetic field distribution are reconstructed based on gradient information using a physical field reconstruction unit. The temperature field data is analyzed by entropy calculation unit to generate an entropy distribution map. The sound signal is analyzed by nonlinear dynamics through a chaotic feature extraction unit to extract chaotic features and identify defect types. The resource scheduling unit optimizes the allocation of computing resources based on entropy distribution and sends adjustment instructions to each unit. The physical constraint verification unit verifies whether the reconstructed physical field satisfies the physical laws and identifies abnormal regions. The synchronicity of chaotic features at multiple spatial locations is evaluated using a chaotic synchronization analysis unit. The data aggregation and processing unit integrates the output results of each unit and performs spatiotemporal registration to generate a fused dataset. The comprehensive quality assessment unit evaluates welding quality levels and locates defects based on fused data. The system knowledge is updated through the knowledge base management unit, enabling multi-site experience sharing.

Images