A fatigue performance analysis system for friction stir welding joints

Abstract

The present invention provides a fatigue performance analysis system for stir friction welding joints, which includes specimen data acquisition, fatigue loading, feature extraction, life prediction, path evaluation and visualization display modules; the system adopts a generalized regularized Transformer model to predict fatigue life, introduces an optimal convex potential function regularization term and a spectral space generalization contribution analysis mechanism to improve the prediction stability and accuracy of the model under heterogeneous data; combines a sparse optimization strategy driven by differential inclusion to reduce the overfitting risk caused by redundant channels; in addition, constructs an asymmetric pseudo-distance function and a fatigue state transfer strategy diagram, cooperates with a deep Q network to perform strategy search, and realizes high-precision estimation of remaining life; the system enhances the interpretability of fatigue path analysis results, and has good intelligence, generalization ability and engineering adaptability.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, in particular to a fatigue performance analysis system for a friction stir welding joint. Background Art

[0002] With the development of new material technology and green manufacturing process, Friction Stir Welding (FSW), as a new solid-phase connection method, is widely used in high-reliability fields such as aerospace, rail transportation, and automobile manufacturing. It is particularly suitable for the efficient connection of refractory materials such as high-strength aluminum alloys. Since FSW joints often bear complex cyclic loads during service, their fatigue performance is directly related to the safety and life of the structure. Therefore, fatigue performance evaluation and life prediction of friction stir welded joints have become a research hotspot in the current material service safety assessment. However, existing fatigue performance analysis methods mostly rely on traditional stress-strain constitutive relationship modeling and main SN curve fitting, and lack the ability to systematically integrate crack propagation behavior, microstructure and multi-source strain response data, making it difficult to fully reflect the performance of FSW joints under actual service conditions. Fatigue evolution process; at the same time, traditional methods have limited ability to express nonlinear fatigue characteristics in the process of data modeling, with large prediction errors and lack of intelligent data-driven analysis mechanisms, making it difficult to adapt to the needs of high-precision fatigue life assessment under complex working conditions; in recent years, with the rise of artificial intelligence and deep learning technologies, fatigue life prediction and path assessment methods driven by big data have gradually attracted attention; in particular, the Transformer model has shown superior global expression capabilities in time series feature modeling, and the application potential of deep reinforcement learning in strategy optimization has provided a new technical path for intelligent analysis of fatigue performance; but in actual applications, there is still a lack of a systematic analysis platform that can integrate multi-source data, has high generalization capabilities, and can dynamically track fatigue evolution paths. Summary of the Invention

[0003] The present invention provides a fatigue performance analysis system for friction stir welding joints, aiming to improve the intelligent evaluation capability of fatigue performance of complex welded structures; unlike existing methods that rely only on a single stress-strain feature or SN curve fitting, the present invention introduces a generalized regularized Transformer model, integrates the optimal convex potential function regularization term and the spectral space generalization contribution analysis mechanism, and significantly improves the prediction accuracy and robustness of the system under heterogeneous fatigue data; at the same time, combined with the sparse optimization strategy driven by differential inclusion, dynamic structural compression is performed on low-contribution channels, effectively alleviating the system overfitting problem; further, the present invention innovatively constructs an asymmetric pseudo-distance function and a state transition strategy graph, combined with a deep Q network strategy search algorithm, to achieve modeling of fatigue evolution paths and high-precision prediction of remaining life; the system realizes full-process analysis starting from the original microstructure and strain response data, covering fatigue feature extraction, life prediction, path evaluation to visual output, breaking through the bottlenecks of traditional fatigue analysis methods in modeling capability, path interpretability and life prediction generalization.

[0004] The present invention provides a fatigue performance analysis system for a friction stir welding joint, which includes a specimen data acquisition module, a fatigue loading module, a fatigue feature extraction module, a fatigue life prediction module, a fatigue path evaluation module, and a visualization module;

[0005] The specimen data acquisition module uses a 3D laser scanner to perform non-contact scanning of the weld area to obtain weld contour geometry data and extract weld profile parameters. A high-power microscope is used to observe the cross-section of the joint and obtain microstructure images. Strain gauges are placed in the specimen area to perform multi-frequency strain response testing and generate initial strain curve data.

[0006] The fatigue loading module collects standard fatigue loading parameters, including load range, stress ratio, and loading frequency. Combining weld profile parameters, microstructure images, and initial strain curve data, fatigue loading tests are conducted on a servo fatigue testing machine. During the loading process, stress-strain response sequences are generated, and crack growth image sequences are acquired, providing data support for fatigue performance analysis and life modeling.

[0007] The fatigue feature extraction module performs time-domain and frequency-domain analysis based on the stress-strain response sequence obtained by the fatigue loading module to extract characteristic parameters, including stress amplitude, plastic strain amplitude, and cyclic stability index. It also uses a convolutional neural network to extract features from the crack growth image sequence, identify the crack initiation location and growth path, and calculate the crack growth rate. Ultimately, it generates a fatigue feature parameter set.

[0008] The fatigue life prediction module introduces a generalized regularized Transformer model. The generalized regularized Transformer model is constructed by introducing an optimal convex potential function regularization term and a generalized error limit control mechanism, combining spectral space generalization contribution analysis with a sparse optimization machine driven by differential inclusion. The fatigue feature parameter set is processed by the generalized regularized Transformer model to generate fatigue life prediction results. The fatigue life prediction results include the main SN curve expression, the core regression parameter set, the error distribution map and the confidence interval. The generalized regularized Transformer model includes the Transformer model.

[0009] The fatigue path assessment module introduces a fatigue evolution path assessment model, which is constructed by combining an asymmetric pseudo-distance modeling mechanism based on a pseudo-metric loss function with the DQN strategy search algorithm of the deep Q network model. The stress-strain response sequence and crack growth image sequence are used as inputs to the fatigue evolution path assessment model to generate remaining life prediction results, including remaining fatigue life, fatigue evolution path diagram, and remaining life interval estimation.

[0010] The visualization module combines the fatigue life prediction results and the remaining life prediction results to construct a multi-dimensional life distribution visualization map, which is used to intuitively display the life indicators, state change paths and prediction error characteristics in the entire fatigue evolution cycle, and realize intelligent analysis and auxiliary decision-making of the fatigue performance of materials and components from a full life perspective.

[0011] Furthermore, by generalizing the regularized Transformer model, the process of generating fatigue life prediction results includes the following steps:

[0012] Step S1: normalize and label the fatigue feature parameter set to obtain a standardized feature set;

[0013] Step S2: Input the standardized feature set into the Transformer model for training. During the training process, the optimal convex potential function regularization term is introduced to regulate the implicit bias behavior of the Transformer model, thereby enhancing the model's expression stability for fatigue life under heterogeneous feature distributions. A nonlinear system of equations is introduced to dynamically obtain the generalization error limit, and the regularization term coefficient of the optimal convex potential function regularization term is regulated to obtain the trained Transformer model.

[0014] Step S3: Combined with the trained Transformer model, fatigue life is predicted, the main SN curve is fitted, residual errors are statistically analyzed, and confidence intervals are constructed to obtain the initial fatigue life prediction results; and the covariance spectrum matrix and optimal potential distribution of the trained Transformer model are extracted;

[0015] Step S4: Using the covariance spectrum matrix and the optimal latent distribution, the generalization contribution of each layer parameter channel in the trained Transformer model in the spectral space is calculated. Parameter channels with low generalization contribution are marked as low-contribution channels, and a differential inclusion-driven sparse optimization mechanism is introduced to perform structural sparsification on the low-contribution channels to remove redundancy, suppress overfitting, optimize the initial fatigue life prediction results, and generate fatigue life prediction results.

[0016] Furthermore, the stress-strain response sequence and crack growth image sequence are used as inputs to the fatigue evolution path assessment model to generate the remaining life prediction results. The process specifically includes the following steps:

[0017] Step B1: construct a state vector based on the stress-strain response sequence and the crack growth image sequence to generate a fatigue state sequence;

[0018] Step B2: Use the interval quasi-metric embedding encoder to map each state of the fatigue state sequence into a high-dimensional interval and construct an embedded tensor representation; based on the embedded tensor representation, define an asymmetric quasi-distance function;

[0019] Step B3: Construct state transition samples, expressed as triples, which specifically include the current fatigue state, the subsequent state, and the non-successor state; based on the state transition samples, construct a quasi-metric boundary control loss function, integrate local consistency and global separation constraints, update the parameters to be learned of the asymmetric quasi-distance function, complete asymmetric distance modeling, and obtain the fatigue state path strategy diagram;

[0020] Step B4: Using the deep Q-network model, perform an action strategy search based on the minimum pseudo-distance on the fatigue state path strategy graph, extract the approximate optimal fatigue evolution path, estimate the remaining fatigue life, and output the remaining life prediction result.

[0021] By adopting the above scheme, the beneficial effects achieved by the present invention are as follows:

[0022] By introducing a generalized regularized Transformer model, the present invention realizes deep modeling and global expression of multi-source heterogeneous fatigue characteristics, thereby improving the accuracy and stability of the system in fatigue life prediction of complex welded structures; by constructing an optimal convex latent function regularization term to constrain model parameters, and combining spectral space generalization contribution analysis, it solves the problems of weak generalization ability and sensitivity to data changes of traditional deep learning models in welding fatigue life prediction; enables the system to robustly adapt to multiple types of inputs such as weld profile parameters, stress-strain response characteristics and microstructure images, realizes reliable output of fatigue life prediction results, and provides more engineering feasible technical support for weld structure life management.

[0023] The present invention further combines the sparse optimization strategy driven by differential inclusion to dynamically compress channels with low generalization contribution in the Transformer model, achieving the dual goals of redundant parameter elimination and model structure optimization, effectively alleviating the overfitting risk caused by excessive model complexity, and enhancing the generalization ability and deployment efficiency of the fatigue life prediction model; this strategy is based on spectral space structure masking and dynamic sparse evolution mechanism, and adapts to the dynamic change requirements of the present invention for specimen data of different welding states, significantly improving the prediction stability of the model under finite sample conditions, and meeting the actual application requirements of industrial sites for high-precision and high-efficiency fatigue assessment tools.

[0024] In terms of path modeling, the present invention constructs an asymmetric pseudo-distance function and a fatigue state transfer strategy diagram, and combines it with a deep Q-network search algorithm to achieve dynamic modeling of the fatigue evolution process and accurate estimation of the remaining life; compared with the traditional path tracing method that can only analyze the hysteresis problem of the crack propagation trajectory a posteriori, the present invention can perform predictive reasoning at the early stage of fatigue state evolution, and effectively identify potential failure trends in advance; this mechanism enhances the interpretability of fatigue path analysis results, provides intelligent early warning and decision-making basis for welded structures during service, and enhances the practical value and technological advancement of the system of the present invention in the full life cycle management of welded structures. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 The fatigue life prediction distribution diagram provided in Example 6;

[0026] Figure 2 This is the fatigue state evolution path diagram provided in Example 6.

[0027] Figure 1 In the figure, horizontal axis: stress amplitude (MPa), vertical axis: predicted fatigue life (Cycles), features: fitted main SN curve and predicted life ± confidence interval;

[0028] Figure 2 Node colors: green represents long remaining life, red represents close to failure, and arrows represent the direction of the path of fatigue state evolution from initial to failure. DETAILED DESCRIPTION

[0029] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0030] In embodiment 1, the present invention provides a fatigue performance analysis system for a friction stir welding joint, the system comprising a specimen data acquisition module, a fatigue loading module, a fatigue feature extraction module, a fatigue life prediction module, a fatigue path evaluation module, and a visualization module;

[0031] The specimen data acquisition module uses a 3D laser scanner to perform non-contact scanning of the weld area to obtain weld contour geometry data and extract weld profile parameters. A high-power microscope is used to observe the cross-section of the joint and obtain microstructure images. Strain gauges are placed in the specimen area to perform multi-frequency strain response testing and generate initial strain curve data.

[0032] The fatigue loading module collects standard fatigue loading parameters, including load range, stress ratio, and loading frequency. Combining weld profile parameters, microstructure images, and initial strain curve data, fatigue loading tests are conducted on a servo fatigue testing machine. During the loading process, stress-strain response sequences are generated, and crack growth image sequences are acquired, providing data support for fatigue performance analysis and life modeling.

[0033] The fatigue feature extraction module performs time-domain and frequency-domain analysis based on the stress-strain response sequence obtained by the fatigue loading module to extract characteristic parameters, including stress amplitude, plastic strain amplitude, and cyclic stability index. It also uses a convolutional neural network to extract features from the crack growth image sequence, identify the crack initiation location and growth path, and calculate the crack growth rate. Ultimately, it generates a fatigue feature parameter set.

[0034] The fatigue life prediction module introduces a generalized regularized Transformer model. The generalized regularized Transformer model is constructed by introducing an optimal convex potential function regularization term and a generalized error limit control mechanism, combining spectral space generalization contribution analysis with a sparse optimization machine driven by differential inclusion. The fatigue feature parameter set is processed by the generalized regularized Transformer model to generate fatigue life prediction results. The fatigue life prediction results include the main SN curve expression, the core regression parameter set, the error distribution map and the confidence interval. The generalized regularized Transformer model includes the Transformer model.

[0035] The fatigue path assessment module introduces a fatigue evolution path assessment model, which is constructed by combining an asymmetric pseudo-distance modeling mechanism based on a pseudo-metric loss function with the DQN strategy search algorithm of the deep Q network model. The stress-strain response sequence and crack growth image sequence are used as inputs to the fatigue evolution path assessment model to generate remaining life prediction results, including remaining fatigue life, fatigue evolution path diagram, and remaining life interval estimation.

[0036] The visualization module combines the fatigue life prediction results and the remaining life prediction results to construct a multi-dimensional life distribution visualization map, which is used to intuitively display the life indicators, state change paths and prediction error characteristics in the entire fatigue evolution cycle, and realize intelligent analysis and auxiliary decision-making of the fatigue performance of materials and components from a full life perspective.

[0037] Example 2: This example is based on Example 1. In this example, the process of generating fatigue life prediction results by generalizing the regularized Transformer model specifically includes the following steps:

[0038] Step S1: normalize and label the fatigue feature parameter set to obtain a standardized feature set;

[0039] Step S2: Input the standardized feature set into the Transformer model for training. During the training process, the optimal convex potential function regularization term is introduced to regulate the implicit bias behavior of the Transformer model, thereby enhancing the model's expression stability for fatigue life under heterogeneous feature distributions. A nonlinear system of equations is introduced to dynamically obtain the generalization error limit, and the regularization term coefficient of the optimal convex potential function regularization term is regulated to obtain the trained Transformer model. The formula used is as follows:

[0040] Optimal convex potential function formula:

[0041] ;

[0042] in, represents the optimal convex potential function, which is a general form; represents the model parameters to be regularized, represents the distribution parameter, represents the normalization constant, represents logarithmic operation, represents the generalization error limit, represents the optimal potential distribution, represents the conditional probability density of the optimal underlying distribution;

[0043] The total loss function formula of the optimal convex potential function regularization term is:

[0044] ;

[0045] in, represents the overall loss, represents the regression loss, represents the weight coefficient of the regularization term of the optimal potential function, Represents an index variable, represents the total number of regularized variables, Indicates the predictor variables, Indicates the Variance estimation corresponding to each variable; represents the sum of regularization terms, which imposes structural penalties on all variables to be learned;

[0046] Step S3: Combined with the trained Transformer model, fatigue life is predicted, the main SN curve is fitted, residual errors are statistically analyzed, and confidence intervals are constructed to obtain the initial fatigue life prediction results; and the covariance spectrum matrix and optimal potential distribution of the trained Transformer model are extracted;

[0047] Step S4: Using the covariance spectrum matrix and the optimal potential distribution, calculate the generalization contribution of each layer parameter channel in the trained Transformer model in the spectral space. For parameter channels with low generalization contribution, mark them as low-contribution channels, and introduce a differential inclusion-driven sparse optimization mechanism to perform structural sparsification on the low-contribution channels, remove redundancy, suppress overfitting, optimize the initial fatigue life prediction results, and generate fatigue life prediction results. The formula used is as follows:

[0048] Generalization contribution calculation formula:

[0049] ;

[0050] in, Indicates the The generalization contribution under the variable dimension, represents the mathematical expectation, represents the derivative of the probability density;

[0051] Differential inclusion driven sparse optimization mechanism:

[0052] Step 1: Set the contribution threshold and generate a low-contribution channel set based on the generalized contribution obtained in the spectral space;

[0053] Step 2: Construct an initial structure mask based on the set of low-contribution channels; define a sparse optimization loss function with mask parameters, and iteratively optimize the initial structure mask using a dynamic sparse evolution strategy to obtain an updated structure mask, which is used to perform structural sparsification on low-contribution channels to remove redundant channels, suppress overfitting, and improve model generalization capabilities;

[0054] Sparse optimization loss function with mask parameter:

[0055] ;

[0056] in, represents the structure mask, represents the sparse optimization loss, Represents the weight parameter matrix in the Transformer model after training, represents the prediction loss with structure mask, represents the sparse regularization coefficient, express The L1 norm of Indicates that the difference includes the control coefficient, Indicates the current time The structure mask of Indicates the previous moment The structure mask of The difference penalty term before and after the mask is used to control the smoothness and physical consistency of the mask sparse evolution;

[0057] Dynamic sparse evolution strategy:

[0058] ;

[0059] in,, Indicates the time after update The structure mask of represents the interval projection operator, represents the learning rate, Indicates the current structure mask Partial derivative of the sparse optimization loss function.

[0060] Embodiment 3: This embodiment is based on embodiment 1. In this embodiment, the process of generating fatigue life prediction results specifically includes the following steps:

[0061] Step R1: normalize and label the fatigue feature parameter set to obtain a standardized feature set;

[0062] Step R2: Input the standardized feature set into the Transformer model for training to obtain the trained Transformer model;

[0063] Step R3: Combine the trained Transformer model to predict fatigue life, fit the main SN curve, calculate the residual error, and construct the confidence interval to obtain the initial fatigue life prediction result; and extract the covariance spectrum matrix and optimal potential distribution of the trained Transformer model;

[0064] Step R4: Use the covariance spectrum matrix and the optimal potential distribution to calculate the generalization contribution of each layer parameter channel in the spectral space in the trained Transformer model. For parameter channels with low generalization contribution, mark them as low-contribution channels, and introduce a differential inclusion-driven sparse optimization mechanism to perform structural sparsification on the low-contribution channels, remove redundancy, suppress overfitting, optimize the initial fatigue life prediction results, and generate fatigue life prediction results.

[0065] Example 4: This example is based on Example 2. In this example, the stress-strain response sequence and the crack growth image sequence are used as inputs of the fatigue evolution path assessment model to generate the remaining life prediction result. The process specifically includes the following steps:

[0066] Step B1: construct a state vector based on the stress-strain response sequence and the crack growth image sequence to generate a fatigue state sequence;

[0067] Step B2: Use the interval quasi-metric embedding encoder to map each state of the fatigue state sequence to a high-dimensional interval and construct an embedded tensor representation. Based on the embedded tensor representation, define an asymmetric quasi-distance function. The formula used is as follows:

[0068] Asymmetric quasi-distance function formula:

[0069] ;

[0070] in, Indicates that the index is Current fatigue status, Indicates that the index is target fatigue state; represents the parameters to be learned, express arrive Asymmetric pseudo-distance of In the embedding space Interval pseudo-distance under channel dimensions; represents the weighting factor, represents the maximum distance component among all channels, represents the average distance component of all channels;

[0071] The interval quasi-metric embedding encoder is a feature transformation structure that maps state data into a high-dimensional interval representation space. While maintaining the structural relationship between sample states, it introduces upper and lower bounds or interval uncertainty expressions to adapt to the state representation requirements of uncertain and dynamically evolving systems. It has two core functions: interval embedding and quasi-metric modeling.

[0072] Step B3: Construct state transition samples, expressed as triples, which specifically include the current fatigue state, the successor state, and the non-successor state. Based on the state transition samples, construct a quasi-metric boundary control loss function, integrate local consistency and global separation constraints, update the parameters to be learned of the asymmetric quasi-distance function, complete asymmetric distance modeling, and obtain the fatigue state path strategy diagram. The formula used is as follows:

[0073] Quasi-metric boundary control loss function:

[0074] ;

[0075] in, represents the quasi-metric reinforcement learning loss function, Represents the penalty factor Maximize the constraints. represents a state transition sample, Indicates the current fatigue status. Indicates the subsequent state, Indicates a non-successor state, Indicates a penalty factor The state transition sample distribution is also called the probability distribution of the state transition sample; Indicates from Sampling triples , calculate the mathematical expectation of the loss term under the triple; express arrive The asymmetric pseudo-distance of express arrive Asymmetric pseudo-distance of represents the linear rectification function, represents a local consistency constraint, Represents the boundary buffer parameters, represents a monotonically increasing convex function, Represents a global separation term;

[0076] Step B4: Using the deep Q-network model, perform an action strategy search based on the minimum pseudo-distance on the fatigue state path strategy graph, extract the approximate optimal fatigue evolution path, estimate the remaining fatigue life, and output the remaining life prediction result.

[0077] Example 5: This example is based on Example 2. In this example, the process of generating the remaining life prediction result specifically includes the following steps:

[0078] Step E1: construct a state vector based on the stress-strain response sequence and the crack growth image sequence to generate a fatigue state sequence;

[0079] Step E2: Using an interval quasi-metric embedding encoder, each state of the fatigue state sequence is mapped to a high-dimensional interval to construct an embedded tensor representation; based on the embedded tensor representation, an asymmetric quasi-distance function is defined;

[0080] Step E3: Construct a state transition sample, expressed as a triplet, which specifically includes the current fatigue state, the subsequent state, and the non-successor state; based on the state transition sample, construct a distance function between states, quantify the relative evolution relationship between states, complete asymmetric distance modeling, and obtain a state relationship diagram;

[0081] Step E4: Based on the state relationship diagram, use the strategy search method to extract the fatigue evolution path, estimate the remaining fatigue life, and output the remaining life prediction result.

[0082] Example 6, according to Figure 1 、 Figure 2 This embodiment is based on the fifth embodiment. In this embodiment, the visualization module combines the fatigue life prediction results and the remaining life prediction results to construct a multi-dimensional life distribution visualization map, which is used to intuitively display the life indicators, state change paths and prediction error characteristics in the entire fatigue evolution cycle, and realize intelligent analysis and auxiliary decision-making of the fatigue performance of materials and components from a full life perspective;

[0083] In this embodiment,

[0084] The fatigue life prediction results are shown in Table 1:

[0085] Table 1

[0086] ;

[0087] The remaining life prediction results are shown in Table 2:

[0088] Table 2

[0089] ;

[0090] The multi-dimensional life distribution visualization map includes fatigue life prediction distribution map and fatigue state evolution path map.

[0091] The present invention and its embodiments are described above. Such description is not restrictive. What is shown in the accompanying drawings is only one of the embodiments of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in this field are inspired by it and do not depart from the purpose of the invention, they can creatively design structural methods and embodiments similar to the technical solution, which should all fall within the scope of protection of the present invention.

Claims

1. A fatigue performance analysis system for a friction stir welded joint, comprising a fatigue feature extraction module, which acquires a stress-strain response sequence and a crack growth image sequence, extracts features, and generates a fatigue feature parameter set; characterized in that: The system also includes a fatigue life prediction module and a fatigue path assessment module; The fatigue life prediction module introduces a generalized regularized Transformer model, which processes the fatigue feature parameter set and generates fatigue life prediction results. The fatigue path assessment module introduces a fatigue evolution path assessment model; the fatigue evolution path assessment model processes stress-strain response sequences and crack growth image sequences to generate remaining life prediction results; The process of generating fatigue life prediction results by generalizing the regularized Transformer model includes the following steps: Step S1: Process the fatigue feature parameter set to obtain a standardized feature set; Step S2: Input the standardized feature set into the Transformer model for training. During the training process, the optimal convex potential function regularization term is introduced to regulate the implicit bias behavior of the Transformer model. A nonlinear equation system is introduced to obtain the generalization error limit, and the regularization term coefficient of the optimal convex potential function regularization term is regulated to obtain the trained Transformer model. The formula used is as follows: Optimal convex potential function formula: ； in, represents the optimal convex potential function, which is a general form; represents the model parameters to be regularized, represents the distribution parameter, represents the normalization constant, represents logarithmic operation, represents the generalization error limit, represents the optimal potential distribution, represents the conditional probability density of the optimal underlying distribution; Step S3: Combine the trained Transformer model to obtain the initial fatigue life prediction results, and extract the covariance spectrum matrix and optimal potential distribution of the trained Transformer model; Step S4: Using the covariance spectrum matrix and the optimal potential distribution, the generalization contribution of each layer parameter channel in the trained Transformer model in the spectral space is calculated. Parameter channels with low generalization contribution are marked as low-contribution channels, and a differential inclusion-driven sparse optimization mechanism is introduced to perform structural sparsification on the low-contribution channels, optimize the initial fatigue life prediction results, and generate fatigue life prediction results. Differential inclusion-driven sparse optimization mechanism: Based on low-contribution channels, a differential penalty term before and after the mask is introduced into the sparse optimization loss function, driving the update of the structure mask through a dynamic sparse evolution strategy. The dynamic sparse evolution strategy iteratively updates the structure mask and combines the differential penalty term before and after the mask with the interval projection operator during the update process to drive the stable evolution of the mask in time. The fatigue evolution path assessment model includes a state construction unit, an embedding coding unit, a distance modeling unit and a path prediction unit; the state construction unit generates a fatigue state sequence; the embedding coding unit maps the fatigue state sequence to a high-dimensional interval and defines an asymmetric quasi-distance function; the distance modeling unit updates the parameters to be learned of the asymmetric quasi-distance function to obtain a fatigue state path strategy graph; the path prediction unit performs action strategy search on the fatigue state path strategy graph and outputs the remaining life prediction result.

2. The fatigue performance analysis system of a friction stir welding joint according to claim 1, characterized in that: The generalized regularized Transformer model includes the Transformer model.

3. The fatigue performance analysis system of a friction stir welding joint according to claim 1, characterized in that: The state construction unit generates a fatigue state sequence based on the stress-strain response sequence and the crack growth image sequence.

4. The fatigue performance analysis system of a friction stir welding joint according to claim 3, characterized in that: The embedding coding unit uses the interval quasi-metric embedding encoder to map the fatigue state sequence into high-dimensional intervals and construct an embedded tensor representation; based on the embedded tensor representation, an asymmetric quasi-distance function is defined.

5. The fatigue performance analysis system of a friction stir welding joint according to claim 4, characterized in that: The distance modeling unit constructs state transition samples, and based on the state transition samples, constructs a quasi-metric boundary control loss function, integrates local consistency and global separability constraints, updates the parameters to be learned of the asymmetric quasi-distance function, and obtains the fatigue state path strategy diagram.

6. The fatigue performance analysis system of a friction stir welding joint according to claim 5, characterized in that: The path prediction unit uses a deep Q-network model to perform action strategy search on the fatigue state path strategy graph, extract the optimal fatigue evolution path, and output the remaining life prediction result.

Images