Composite material connecting structure fatigue life prediction method and device and storage medium

Abstract

The application discloses a composite material connection structure fatigue life prediction method and device and a storage medium. Through the physical cause chain of "substrate degradation-glue layer degradation-mixed connection degradation" objectively existing in the composite material multi-configuration connection system, a cascaded neural network structure corresponding to the physical cause chain is constructed. The previous level network outputs the hidden feature vector representing the degradation state of the level while outputting the life prediction value of the level configuration, and the hidden feature vector is directionally transmitted to the next level network as part of the input, so that the high-level configuration can explicitly use the degradation information of the low-level configuration, and cross-configuration knowledge sharing is realized.

Description

Technical Field

[0001] This application relates to the field of material fatigue life prediction technology, specifically to a method, device, and storage medium for predicting the fatigue life of composite material connection structures. Background Technology

[0002] Carbon fiber reinforced plastics (CFRP) have been widely used in aerospace, rail transportation, and high-end equipment due to their high specific strength, high specific stiffness, and excellent lightweight potential. As a critical weak point in composite material components, the reliability of the connection structure directly affects the safety and durability of the overall structure. Common composite material connection methods include mechanical connections (load transfer via fasteners such as bolts), adhesive connections (load transfer via structural adhesives), and hybrid adhesive-riveting connections (load transfer achieved through both adhesive bonding and riveting).

[0003] In actual service, connection structures are typically subjected to both cyclic loading and humid heat environments. On the one hand, cyclic loading leads to the continuous accumulation of fatigue damage; on the other hand, humid heat environments cause performance degradation in the substrate, adhesive layer, and interface areas, thereby altering the local damage evolution process of the structure. Due to the differences in the load-bearing mechanisms, failure modes, and degradation paths of different connection configurations, fatigue life prediction is significantly complex.

[0004] Existing fatigue life prediction methods mainly fall into the following three categories:

[0005] (1) Empirical statistical methods. This type of method is represented by SN curve fitting, which establishes an empirical relationship between load level and fatigue life through statistical regression. However, this type of method requires the establishment of independent models for each configuration and each environmental condition, resulting in low sample utilization efficiency and weak model transferability.

[0006] (2) Numerical simulation methods based on damage mechanics. These methods simulate the fatigue damage evolution process by establishing a damage constitutive model and finite element simulation. Although they can reflect the damage mechanism to a certain extent, the modeling and calculation costs are high, making it difficult to achieve large-scale and rapid prediction under complex working conditions.

[0007] (3) Conventional data-driven methods. These methods use machine learning models such as neural networks to establish a direct mapping relationship between input operating parameters and fatigue life. However, conventional data-driven methods usually model different connection configurations independently or use parallel multi-task learning to make joint predictions, which fails to effectively represent the physical hierarchy dependencies between different configurations. Therefore, they are prone to insufficient generalization ability in small sample scenarios.

[0008] The aforementioned existing technologies overlook a crucial physical fact: for multi-configuration composite bonding systems, different configurations are not independent of each other, but rather exhibit a clear hierarchical dependency. Specifically, the fatigue response of CFRP is primarily controlled by the substrate degradation state; adhesive-bonded structures are further affected by adhesive layer degradation on top of substrate degradation; and hybrid adhesive-riveted structures are simultaneously affected by substrate degradation, adhesive layer degradation, and their coupling effect. This hierarchical causal relationship is not effectively utilized in existing prediction methods, resulting in the waste of shared degradation information among different configurations.

[0009] Therefore, there is an urgent need for a technical method that can explicitly characterize the hierarchical degradation information transmission relationship for composite multi-configuration connection systems and achieve unified lifetime prediction under small sample conditions. Summary of the Invention

[0010] The purpose of this application is to provide a method, apparatus, and storage medium for predicting the fatigue life of composite material connection structures.

[0011] To achieve the above objectives, the first aspect of this application provides a method for predicting the fatigue life of composite material connection structures, the method comprising:

[0012] Obtain the basic parameters of the composite material substrate under the predicted working condition;

[0013] The basic parameters are input into a trained hierarchical degradation state propagation neural network to obtain the fatigue life prediction value of the composite material;

[0014] The hierarchical degradation state propagation neural network is a multi-level cascaded neural network. The number of levels in the hierarchical degradation state propagation neural network is determined according to the physical hierarchical structure of the composite material connection configuration. The lowest level network in the hierarchical degradation state propagation neural network corresponds to the composite matrix material configuration to be predicted, and the higher level networks correspond to the composite material connection configuration to be predicted. They are connected sequentially from the lower level to the higher level in a physical causal order.

[0015] The lowest-level network is configured to generate a first degraded state vector and a fatigue life prediction value for that level based on the basic parameters.

[0016] Each layer of the network, except for the lowest layer, is configured to: receive an input formed by concatenating the basic parameters and the degradation state vector output from the previous layer network, generate the degradation state vector of the current layer network and the fatigue life prediction value corresponding to the current layer; the degradation state vector of the previous layer is used to encode the degradation state of the corresponding layer.

[0017] In this embodiment, the basic parameters include load level parameters and environmental aging parameters. The hierarchical degradation state transmission neural network is a three-layer neural network, including a first-layer network, a second-layer network, and a third-layer network. The first-layer network corresponds to the carbon fiber reinforced resin composite matrix structure and outputs a first fatigue life prediction value and a first degradation state vector based on the load level parameters and environmental aging parameters.

[0018] The second-level network corresponds to the adhesive bonding structure, receives the input formed by splicing the load level parameter, environmental aging parameter and the first degradation state vector, and outputs the second fatigue life prediction value and the second degradation state vector;

[0019] The third-level network corresponds to the glue-riveting hybrid connection structure, receives input formed by splicing the load level parameter, environmental aging parameter, first degradation state vector, and second degradation state vector, and outputs a third fatigue life prediction value.

[0020] In this embodiment of the application, the first-level network includes a first encoder network and two output branches. The load level parameter and the environmental aging parameter are extracted by the first encoder network. The first fatigue life prediction value is generated by the first output branch of the first-level network, and a multi-dimensional first degradation state vector is generated by the second output branch of the first-level network.

[0021] The second-level network includes a second encoder network and two output branches. The multidimensional vector formed by concatenating the load level parameter, the environmental aging parameter, and the first degradation state vector is processed by the second encoder network for feature extraction. The second fatigue life prediction value is generated by the first output branch of the second-level network, and the multidimensional second degradation state vector is generated by the second output branch of the second-level network.

[0022] The third-level network includes a third encoder network. The multi-dimensional vector obtained by concatenating the load level parameter, the environmental aging parameter, the first degradation state vector, and the second degradation state vector is processed by the second encoder network for feature extraction to generate the third fatigue life prediction value.

[0023] In the embodiments of this application, the first encoder network, the second encoder network and the third encoder network each contain 1 to 3 fully connected layers, each layer contains 8 to 64 neurons, and a nonlinear activation function is used as the output transformation of the layer.

[0024] In the embodiments of this application, the first encoder network, the second encoder network and the third encoder network each contain two fully connected layers, each containing 16 neurons, and the Tanh activation function is used as the output transformation of the layer.

[0025] In this embodiment of the application, the training process of the hierarchical degenerate state propagation neural network includes:

[0026] Obtain the hierarchical degradation state propagation neural network to be trained and the multi-configuration fatigue life dataset of various composite materials;

[0027] The load level parameters and environmental aging parameters in the fatigue life dataset are input into the lowest level network of the hierarchical degradation state propagation neural network to obtain the first degradation state vector and the first fatigue life prediction value.

[0028] For each intermediate-level network in the multi-level state transfer module, except for the lowest-level network and the highest-level network, an input vector formed by splicing the load level parameter, the environmental aging parameter, and the degradation state vector output from the previous-level network is input into the intermediate-level network to obtain the degradation state vector and fatigue life prediction value corresponding to each level.

[0029] The load level parameter, the environmental aging parameter, and the degradation state vector output from the previous level network of the highest level network are concatenated to form an input vector, which is then input into the highest level network of the multi-level state transfer module to obtain the highest level fatigue life prediction value.

[0030] Based on the first fatigue life prediction value, the fatigue life prediction values ​​corresponding to each intermediate level network, and the fatigue life prediction value of the highest level, the network parameters of each level network are jointly updated in the same optimization process, so that each level network can simultaneously learn the life mapping relationship of the corresponding configuration and the degradation state representation shared across configurations, and obtain the trained target level degradation state propagation neural network.

[0031] In this embodiment of the application, the method for jointly updating the network parameters of each level network in the same optimization process based on the first fatigue life prediction value, the fatigue life prediction values ​​corresponding to each intermediate level network, and the highest level fatigue life prediction value includes:

[0032] Based on the first predicted fatigue life value, the predicted fatigue life values ​​corresponding to each intermediate level network, and the predicted fatigue life value of the highest level, and the logarithmic values ​​of the experimental life values ​​Y1, Y2, Y3...Y... of the corresponding levels... n The total loss value L is calculated using the following joint loss function:

[0033] L=λ1·MSE(Life1,Y1)+λ2·MSE(Life2,Y2)+λ3·MSE(Life3,Y3)....+λ n ·MSE (Life) n ,Yn );

[0034] Backpropagation is performed on the total loss value L, and the gradient is cascaded along the direction of the highest-level network, the third-level network, the second-level network to the first-level network. In the same optimization step, the Adam optimizer is used in conjunction with the cosine annealing learning rate scheduling strategy to synchronously update all network parameters from the first-level network to the highest-level network.

[0035] Repeat the above steps until the joint loss function converges to obtain the trained target-level degenerate state propagation neural network.

[0036] In the embodiments of this application, λ1=λ2=λ3...=λ n =1.

[0037] A second aspect of this application provides a fatigue life prediction device for composite material connection structures, comprising: a memory configured to store instructions; a processor configured to retrieve instructions from the memory and, when executing the instructions, to implement a fatigue life prediction method for composite material connection structures.

[0038] A third aspect of this application provides a machine-readable storage medium storing instructions for causing a machine to perform a method for predicting the fatigue life of composite material connection structures.

[0039] This application's method for predicting the fatigue life of composite material bonding structures constructs a corresponding cascaded neural network structure based on the objectively existing physical causal chain of "substrate degradation → adhesive layer degradation → hybrid bonding degradation" in multi-configuration composite material bonding systems. Each layer of the network outputs the predicted fatigue life value for its current configuration, along with a hidden feature vector representing the degradation state at that level. This vector is then directionally passed to the next layer as part of the input, thereby achieving the sharing and transmission of cross-configuration degradation information.

[0040] Compared with existing technologies, the composite material multi-configuration fatigue life prediction method provided by this invention has the following significant advantages:

[0041] (1) Based on the objectively existing hierarchical dependencies in composite multi-configuration connection systems, this invention constructs a cascaded network structure with a degradation state directional transmission mechanism. The degradation state vector output by the previous level network is used as part of the input of the next level network, enabling higher-level configurations to explicitly utilize the degradation information of lower-level configurations and achieve cross-configuration knowledge sharing. This design differs from the existing independent modeling methods where configurations are not related to each other, and also from the parallel multi-task methods where configurations are of equal status.

[0042] (2) This invention encodes physical causal relationships into the network topology itself, and the information transmission direction is strictly consistent with the physical direction of "substrate degradation → adhesive layer degradation → hybrid connection degradation". Compared with the method of adding physical constraint terms to the loss function, this invention does not require additional design and adjustment of constraint term weights, thus reducing the difficulty of model parameter tuning.

[0043] (3) The present invention can simultaneously output fatigue life prediction results of multiple connection configurations under different load levels and damp heat aging conditions under a unified model framework, avoiding the problem of maintaining multiple models separately for independent modeling, improving data utilization efficiency and reducing the cost of repeated modeling.

[0044] Other features and advantages of the embodiments of this application will be described in detail in the following detailed description section. Attached Figure Description

[0045] The accompanying drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the following detailed description to explain the embodiments of this application, but do not constitute a limitation on the embodiments of this application. In the drawings:

[0046] Figure 1 A schematic flowchart of a method for predicting the fatigue life of a composite material connection structure according to an embodiment of this application is shown.

[0047] Figure 2 This illustration schematically shows the overall architecture of the hierarchical degradation state propagation neural network in the fatigue life prediction method for composite material connection structures according to an embodiment of this application.

[0048] Figure 3 This illustration schematically shows the distribution of fatigue life test data for multi-configuration joint structures in the fatigue life prediction method for composite material joint structures according to embodiments of this application.

[0049] Figure 4 The schematic diagram illustrates the training convergence curve of the hierarchical degradation state transfer neural network model of the fatigue life prediction method for composite material connection structures according to an embodiment of this application.

[0050] Figure 5 This schematically illustrates a scatter plot comparing the predicted fatigue life values ​​with experimental values ​​for each configuration of the fatigue life prediction method for composite material connection structures according to embodiments of this application.

[0051] Figure 6 The diagram illustrates the leave-one-out cross-validation results of the fatigue life prediction method for composite material connection structures according to an embodiment of this application.

[0052] Figure 7The diagram schematically illustrates fatigue life prediction surface plots for various configurations of the fatigue life prediction method for composite material connection structures according to embodiments of this application.

[0053] Figure 8 The diagram schematically illustrates a histogram of the prediction error distribution of a fatigue life prediction method for composite material connection structures according to an embodiment of this application.

[0054] Figure 9 A schematic diagram of a computer device according to an embodiment of this application is shown. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only for illustration and explanation of the embodiments of this application and are not intended to limit the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0056] Figure 1 The illustration schematically shows a flowchart of a method for predicting the fatigue life of a composite material connection structure according to an embodiment of this application. Figure 1 and Figure 2 As shown in the embodiments of this application, a method for predicting the fatigue life of composite material connection structures is provided. This method may include the following steps:

[0057] Step 101: Obtain the basic parameters of the composite material substrate under the predicted working condition.

[0058] First, the composite material connection structures (including CFRP matrix structures, adhesive-bonded structures, and hybrid adhesive-riveted structures) requiring fatigue life prediction are identified, along with their corresponding service conditions to be predicted. These service conditions are defined by the fatigue load state experienced by the structure and the environmental aging state it undergoes. Therefore, in this embodiment, load parameters and environmental aging parameters are selected as inputs, with the load parameters primarily being load level parameters and the environmental aging parameters primarily being environmental damp-heat aging parameters.

[0059] The load level parameter q is defined as the ratio of the applied load to the static strength of the CFRP matrix structure, used to uniformly characterize the severity of the load on each layer configuration. Specifically, it is obtained as follows: the most representative load amplitude is extracted from the fatigue load spectrum of the predicted working condition, typically the maximum load F_max in the fatigue load cycle; the static strength F_static of the CFRP laminate with the same layup and molding process as the CFRP substrate in the connection structure is obtained through standard static tests or material property data sheets; the dimensionless load level parameter q is calculated using the formula q = F_max / F_static. When the actual working condition is a variable amplitude load spectrum, the equivalent stress amplitude can also be calculated using the rainflow counting method based on the fatigue equivalence principle, and then compared with the static strength to obtain the equivalent load level. The acquisition of this parameter does not depend on the specific form of the upper-layer connection configuration, allowing the same q value to be used as a direct input to the first-level network and, through degradation state transmission, to provide reference load information for subsequent layers.

[0060] The environmental aging parameter RH is defined as a damp heat aging characterization parameter, with a value range of [0,1]. 0 represents the initial state without aging, and 1 represents the critical state of complete aging. It can be obtained through historical calculations based on the actual service environment of the structure or through direct measurement. A typical method is to collect data on ambient temperature T(t), relative humidity H(t), and time t during the structure's service life, calculate the equivalent aging time using a damp heat aging acceleration model (e.g., an equivalent aging time model based on the Arrhenius relation), and then normalize it using a pre-built "equivalent aging time - aging degree" calibration curve to obtain RH. Another method is to obtain the damp heat aging characteristic indicators of the current material through on-site non-destructive testing or sampling analysis. For example, measure the saturated moisture absorption rate M_current of the composite material, and use the moisture absorption rate M_saturated corresponding to the fully saturated moisture absorption state of the material system at the same temperature as a benchmark, calculate RH according to the formula RH=M_current / M_saturated, and truncate it to the [0,1] interval. When aging history data is lacking, RH can be obtained by linear mapping based on key performance degradation indicators (such as glass transition temperature retention rate, interlaminar shear strength reduction rate, etc.). The RH parameter reflects the overall aging degree reached by the material and serves as a shared environmental input for the three-level network, ensuring that the lifetime prediction of each configuration is based on a consistent aging state.

[0061] After obtaining q and RH, a two-dimensional input vector [q, RH] can be formed, which can be used by the hierarchical degradation state propagation neural network for subsequent training to perform multi-configuration unified fatigue life prediction.

[0062] Step 102: Input the load level parameters and environmental aging parameters into the trained hierarchical degradation state transmission neural network to obtain the fatigue life prediction value of the composite material.

[0063] The hierarchical degradation state propagation neural network is a multi-level cascaded neural network. The number of levels in the hierarchical degradation state propagation neural network is determined according to the physical hierarchical structure of the composite material connection configuration. The lowest level network in the hierarchical degradation state propagation neural network corresponds to the composite matrix material configuration to be predicted, and the higher level networks correspond to the composite material connection configuration to be predicted. They are connected sequentially from the lower level to the higher level in a physical causal order.

[0064] The lowest-level network is configured to generate a first degradation state vector and a fatigue life prediction value for that level based on the load level parameters and environmental aging parameters.

[0065] Each network level, except the lowest level network, is configured to: receive an input formed by concatenating the load level parameter, environmental aging parameter, and degradation state vector output from the previous level network; generate a degradation state vector for the current level network and a fatigue life prediction value corresponding to the current level; the degradation state vector of the previous level network is used to encode the degradation state of the corresponding level; it should be noted that the fatigue life prediction value output by the highest level network is the fatigue life prediction value of the composite material.

[0066] After obtaining the load level parameter q and the environmental aging parameter RH for the operating condition to be predicted, these two scalar values ​​are used to construct a two-dimensional input vector [q, RH]. This vector is then fed into a pre-trained hierarchical degradation state propagation neural network.

[0067] The hierarchical degradation state propagation neural network in this application is a multi-layer cascaded neural network. The number of layers in the hierarchical degradation state propagation neural network is determined according to the physical hierarchical structure of the composite material connection configuration, and different connection configurations correspond to different layer networks. The lowest layer network in the hierarchical degradation state propagation neural network corresponds to the composite material matrix configuration to be predicted, and the highest layer network corresponds to the target connection configuration of the composite material to be predicted. Each layer network is connected sequentially from the lowest to the highest level according to physical causal order. For example, the hierarchical degradation state propagation neural network can be a three-layer cascaded neural network, which may include a first layer network, a second layer network, and a third layer network; the first layer network corresponds to the carbon fiber reinforced resin composite material matrix structure; the second layer network corresponds to the adhesive bonding structure; and the third layer network corresponds to the adhesive-riveting hybrid bonding structure. In other embodiments, the hierarchical degradation state propagation neural network can also be a four-layer cascaded neural network, adding a fourth layer to accommodate more complex hybrid connection forms, etc.

[0068] When the hierarchical degradation state propagation neural network is a multi-level cascaded neural network, the lowest level network is configured to generate a first degradation state vector based on the load level parameters and environmental aging parameters. This first degradation state vector is not directly used as the lifetime output, but rather as an intermediate representation for propagation to subsequent levels, used to encode the comprehensive degradation state of the substrate layer under the current load level and aging conditions.

[0069] Each intermediate-level network, except the lowest-level network, is configured to receive an input vector formed by concatenating the load level parameters, environmental aging parameters, and degradation state vectors output from its preceding-level network, and generate a degradation state vector for that level of network. The degradation state vector generated by each level of network encodes the comprehensive degradation information of the structural configuration represented by that level under the current operating conditions and the influence of the degradation state of the preceding level.

[0070] The highest-level network is configured to receive an input vector formed by concatenating the load level parameter, environmental aging parameter, and degradation state vectors output from its preceding network, and output a predicted fatigue life value for the composite material. Since the highest-level network incorporates the load level parameter, environmental aging parameter, and degradation state vectors from all preceding levels into its input, its output predicted fatigue life value comprehensively reflects the cumulative degradation effect across the entire chain from the substrate to the final target configuration.

[0071] The key design feature of the aforementioned hierarchical degenerate state propagation neural network is that the direction of information transmission is strictly set from lower-level configurations to higher-level configurations, i.e., unidirectional transmission along the physical causal chain, rather than reverse transmission or parallel sharing. This design allows the network topology itself to encode the physical hierarchical relationship between composite material connection configurations.

[0072] In the prediction phase, a two-dimensional input vector [q,RH] consisting of the load level parameter q and the environmental aging parameter RH of the working condition to be predicted is input into the trained hierarchical degradation state propagation neural network. After one complete forward propagation, the fatigue life prediction value of the composite material output by the highest-level network can be obtained. This fatigue life prediction value is the logarithm of the fatigue life cycle number N, with base 10. 10 The predicted value of (N).

[0073] Optionally, while the hierarchical degradation state propagation neural network outputs the highest-level fatigue life prediction value, each intermediate-level network can also synchronously output the fatigue life prediction value of the corresponding connection configuration, thereby achieving unified parallel prediction of fatigue life of multiple configurations through one forward propagation.

[0074] The fatigue life prediction method for composite material bonding structures in this embodiment constructs a corresponding cascaded neural network structure based on the objectively existing physical causal chain of "substrate degradation → adhesive layer degradation → hybrid bonding degradation" in the multi-configuration bonding system of composite materials. Each layer of the network outputs the predicted fatigue life value for its current configuration, along with a hidden feature vector representing the degradation state of that layer, and then directs this vector to the next layer as part of the input, thereby achieving the sharing and transmission of cross-configuration degradation information.

[0075] Compared with existing technologies, the composite material multi-configuration fatigue life prediction method provided by this invention has the following significant advantages:

[0076] (1) Based on the objectively existing hierarchical dependencies in composite multi-configuration connection systems, this invention constructs a cascaded network structure with a degradation state directional transmission mechanism. The degradation state vector output by the previous level network is used as part of the input of the next level network, enabling higher-level configurations to explicitly utilize the degradation information of lower-level configurations and achieve cross-configuration knowledge sharing. This design differs from the existing independent modeling methods where configurations are not related to each other, and also from the parallel multi-task methods where configurations are of equal status.

[0077] (2) This invention encodes physical causal relationships into the network topology itself, and the information transmission direction is strictly consistent with the physical direction of "substrate degradation → adhesive layer degradation → hybrid connection degradation". Compared with the method of adding physical constraint terms to the loss function, this invention does not require additional design and adjustment of constraint term weights, thus reducing the difficulty of model parameter tuning.

[0078] (3) The present invention can simultaneously output fatigue life prediction results of multiple connection configurations under different load levels and damp heat aging conditions under a unified model framework, avoiding the problem of maintaining multiple models separately for independent modeling, improving data utilization efficiency and reducing the cost of repeated modeling.

[0079] In another embodiment of this application, see Figure 1 and Figure 2 The hierarchical degradation state transmission neural network is a three-layer neural network, including a first-layer network, a second-layer network, and a third-layer network; the first-layer network corresponds to the carbon fiber reinforced resin composite matrix structure, and outputs a first fatigue life prediction value and a first degradation state vector based on the load level parameters and environmental aging parameters.

[0080] The second-level network corresponds to the adhesive bonding structure, receives the input formed by splicing the load level parameter, environmental aging parameter and the first degradation state vector, and outputs the second fatigue life prediction value and the second degradation state vector;

[0081] The third-level network corresponds to the glue-riveting hybrid connection structure, receives input formed by splicing the load level parameter, environmental aging parameter, first degradation state vector, and second degradation state vector, and outputs a third fatigue life prediction value.

[0082] The first-level network includes a first encoder network and two output branches. The load level parameter and environmental aging parameter are extracted by the first encoder network. The first fatigue life prediction value is generated by the first output branch of the first-level network, and the multi-dimensional first degradation state vector is generated by the second output branch of the first-level network.

[0083] The second-level network includes a second encoder network and two output branches. The multidimensional vector formed by concatenating the load level parameter, the environmental aging parameter, and the first degradation state vector is processed by the second encoder network for feature extraction. The second fatigue life prediction value is generated by the first output branch of the second-level network, and the multidimensional second degradation state vector is generated by the second output branch of the second-level network.

[0084] The third-level network includes a third encoder network. The multi-dimensional vector obtained by concatenating the load level parameter, the environmental aging parameter, the first degradation state vector, and the second degradation state vector is processed by the second encoder network for feature extraction to generate the third fatigue life prediction value. The third fatigue life prediction value is the final overall fatigue life prediction value of the composite material.

[0085] Specifically, the first, second, and third encoder networks each contain 1 to 3 fully connected layers, each containing 8 to 64 neurons, and employ a non-linear activation function as the output transformation of that layer. In this embodiment, the first, second, and third encoder networks preferably each contain 2 fully connected layers, each containing 16 neurons, and employ the Tanh activation function as the output transformation of that layer. It is understood that in other embodiments, the encoder network may also have a 3-layer fully connected layer structure, with each layer potentially containing more neurons, and the non-linear activation function may also be ReLU, Sigmoid, GELU, or other non-linear activation functions.

[0086] Specifically, based on the above embodiments, the hierarchical degradation state propagation neural network strictly follows the physical hierarchical progression of "CFRP substrate → adhesive bonding → hybrid adhesive-riveting bonding" in its structure, and is composed of a first-level network, a second-level network, and a third-level network connected sequentially. Its forward propagation and prediction process is as follows:

[0087] Step 1: Forward computation and state generation of the first-level network (CFRP matrix structure). The input vector [q, RH] first enters the encoder part of the first-level network. This encoder consists of two fully connected hidden layers, each containing 16 neurons, both using the hyperbolic tangent (Tanh) activation function for nonlinear transformation. The input vector is passed through the two hidden layers to progressively extract deep features related to substrate fatigue degradation. At the end of the encoder, the network splits into two parallel output branches:

[0088] (1) Life prediction branch: Based on the extracted features, a scalar output value is directly mapped to generate the first fatigue life prediction value Life1. This value represents the logarithm of the fatigue life cycle number N of the CFRP substrate under a given working condition. 10 The prediction results of (N).

[0089] (2) Degenerate state branch: A 4-dimensional vector is generated based on the extracted features, and the values ​​of each dimension are compressed to the interval [-1, 1] by the Tanh activation function to obtain the first degenerate state vector h1∈ℝ 4 Each element of h1 does not have direct physical dimensions, but rather is an implicit, distributed characterization encoding of the comprehensive degradation state of the CFRP substrate under a given load level and degree of damp heat aging. For example, it can implicitly reflect comprehensive information such as the fiber / matrix interface debonding tendency, microcrack density trend, and residual strength level of the substrate. This degradation state vector is not directly used as the final lifetime output, but rather as an intermediate information carrier transmitted across levels.

[0090] Step 2: Feature fusion and forward computation of the second-level network (adhesive joint structure). The degradation state vector h1 output by the first-level network is concatenated with the original working condition input parameters to form a 6-dimensional extended input vector [q,RH, h1]. This concatenation operation explicitly injects the degradation state information of the substrate into the adhesive joint layer. The physical meaning is that the fatigue life of the adhesive joint is not only directly affected by external loads and aging, but also closely related to the current state of the bonded substrate—the surface and stiffness degradation of the aged substrate will change the stress state of the adhesive layer.

[0091] The 6-dimensional vector is fed into the encoder of the second-layer network. The encoder of the second-layer network has the same structure as the first-layer network, consisting of two fully connected hidden layers with 16 neurons each and Tanh activation. During feature extraction, the encoder automatically learns how to jointly model using load, aging information, and substrate degradation characteristics to extract features suitable for predicting the lifespan of adhesive-bonded structures. Similarly, two branches are set at the end of the encoder:

[0092] (1) Output the second fatigue life prediction value Life2, which is the logarithmic life prediction value of the adhesive joint structure.

[0093] (2) Output the 4-dimensional second degenerate state vector h2∈ℝ activated by Tanh. 4 h2 encodes a comprehensive degradation characterization of the adhesive layer under current operating conditions and the current degradation state of the substrate it is bonded to. For example, it may implicitly include information such as the degree of chemical aging of the adhesive layer, cohesive damage accumulation, and interfacial adhesion weakening, providing key information for predicting the lifetime of more complex hybrid bonding configurations.

[0094] Step 3: Final prediction of the third-level network (adhesive-riveted hybrid connection structure). The degradation state vector h1 output from the first-level network and the degradation state vector h2 output from the second-level network are concatenated with the original working condition input parameters to form a 10-dimensional combined input vector [q, RH, h1, h2]. The physical basis of this design is that the adhesive-riveted hybrid connection simultaneously includes three load-bearing elements: substrate, adhesive layer, and rivet. Its fatigue behavior is essentially a complex process of substrate degradation (characterized by h1) and adhesive layer degradation (characterized by h2) working together and coupling with each other.

[0095] The 10-dimensional vector is fed into the encoder of the third-level network. After feature extraction through two fully connected hidden layers, the third fatigue life prediction value, Life3, is directly output, which is the logarithmic life prediction value of the glue-riveted hybrid connection structure. Since this level is located at the end of the network and has already received complete degradation state information from the first two levels, it is no longer necessary to pass the degradation state to the next level. Therefore, there is no need to output the degradation state vector.

[0096] Through the complete hierarchical forward propagation process described above, the network can output three scalar prediction values, Life1, Life2, and Life3, at once, corresponding to the logarithmic fatigue life of the three configurations—CFRP substrate, adhesive joint, and hybrid adhesive-riveted joint—under current service conditions. This hierarchical design ensures that the information transmission direction is strictly consistent with the physical causal chain, i.e., from basic components to complex components, making the prediction process physically interpretable, and enabling unified evaluation of multiple configurations with a single input.

[0097] The fatigue life prediction method for composite material connection structures adopted in this application firstly achieves hierarchical degradation information transmission from simple to complex configurations, significantly improving the prediction accuracy of fatigue life for complex hybrid connection structures. Most existing prediction methods treat CFRP substrates, adhesive connections, and adhesive-riveted hybrid connections as independent prediction objects, severing the physical causal relationship between them. This invention innovatively constructs a hierarchical degradation state transmission neural network. By explicitly transmitting degradation state vectors between adjacent layers, the network can perform progressive reasoning along a causal chain of "substrate degradation → adhesive layer degradation → comprehensive degradation of hybrid connections." This scheme enables the highest-level network to comprehensively utilize degradation characterization information from both the substrate and adhesive layers when predicting the fatigue life of adhesive-riveted hybrid connections, capturing the complex coupling effect between substrate damage and adhesive layer damage, thereby fundamentally improving the prediction accuracy for complex configurations.

[0098] Secondly, this approach achieves unified parallel prediction of fatigue life for multiple configurations using a single network and a single forward propagation, avoiding the tedious work of building and maintaining multiple independent models for different configurations. Through multi-level joint training, this scheme enables the same neural network to simultaneously learn the life mapping relationships for three configurations. In practical applications, only the load level parameters and environmental aging parameters of the composite material substrate under the predicted working condition need to be input. A single complete forward propagation yields fatigue life prediction results for CFRP substrates, adhesive-bonded connections, and hybrid adhesive-riveted connections. This unified prediction framework significantly reduces the complexity of model deployment and data management in engineering applications, thereby improving prediction efficiency.

[0099] Third, the network topology explicitly encodes physical hierarchical relationships, enhancing the model's interpretability. Information transmission is strictly designed to proceed from lower-level configurations to higher-level configurations, rather than in reverse or in parallel sharing. This design aligns closely with the inherent physical logic of composite material bonding structures—the substrate is the basis of adhesive bonding, and adhesive bonding is the basis of hybrid connections. Therefore, the meaning and transmission paths of the degenerate state vectors at each level are supported by clear engineering physics intuition, making this model more transparent and interpretable than traditional "black box" deep neural networks, and easier for engineers to understand and trust.

[0100] Fourth, the separation of degradation state vectors from lifetime prediction outputs allows the degradation representations at intermediate levels to be utilized as valuable intermediate results. While outputting lifetime predictions for each configuration, the network also generates low-dimensional, bounded degradation state vectors. Although these vectors do not directly define physical dimensions, they automatically learn implicit representations highly correlated with the degradation mechanisms at each level through joint training. In practical applications, these degradation state vectors can provide engineers with additional reference information. For example, when the predicted lifetime is abnormally low, analyzing the distribution of degradation state vector values ​​can help determine whether the anomaly originates from substrate degradation, adhesive layer degradation, or a combination of both, providing guidance for structural maintenance decisions.

[0101] In another embodiment of this application, the training process of the hierarchical degenerate state propagation neural network includes:

[0102] Step S210: Obtain the hierarchical degradation state transfer neural network to be trained and the multi-configuration fatigue life dataset of various composite materials.

[0103] Fatigue life data of various composite material connection configurations under different service conditions were collected. The service conditions included at least load parameters and environmental aging parameters. Among them, the load parameters included the load level parameter q, which is a dimensionless parameter defined as the ratio of applied load to static strength, and is used to reflect the severity of fatigue load; the environmental aging parameters included the damp heat aging characterization parameter RH, which is a dimensionless parameter used to characterize the degree of damp heat aging experienced by the material, and its value ranges from [0,1], where 0 represents the unaged state and 1 represents the fully aged state.

[0104] The model input is a two-dimensional input vector [q, RH] composed of the load level parameters and the damp heat aging characterization parameters, and the model input is the logarithm of the fatigue life cycle number N, which is base 10. 10 (N) is used as the model output to construct the multi-configuration fatigue life dataset. The reason for taking the logarithm of the fatigue life is that the number of fatigue life cycles usually spans multiple orders of magnitude, and taking the logarithm can improve the numerical distribution, which is beneficial to model training.

[0105] Step S220: Construct a multi-level cascaded neural network.

[0106] Based on the hierarchical dependencies between different connection configurations, a cascaded neural network with multiple layers is constructed, with information passed between layers through hidden degenerate state vectors. Taking a three-layer neural network as an example, the specific structure of the network is as follows:

[0107] The first-level network, corresponding to the CFRP matrix structure, is configured as follows:

[0108] The system receives an input [q, RH] consisting of load level parameters and environmental aging parameters. After feature extraction via an encoder network, it outputs a first fatigue life prediction value Life1 and a first degradation state vector h1 through two independent output branches. The encoder network comprises two fully connected hidden layers, each with 16 neurons, and uses the hyperbolic tangent (Tanh) activation function. The first degradation state vector h1 is a 4-dimensional vector. The Tanh activation function restricts the values ​​of each element to the interval [-1, 1], representing the comprehensive degradation state of the CFRP substrate under the current load level and aging conditions. This first degradation state vector is not directly used as the life output but serves as an intermediate representation passed to subsequent layers.

[0109] The second-level network, corresponding to the adhesive bonding structure, is configured as follows:

[0110] The system receives a 6-dimensional input vector [q, RH, h1] formed by concatenating the load level parameters, environmental damp heat aging characterization parameters, and the first degradation state vector. After extracting features through an encoder network with the same structure as the first-level network, it outputs a second fatigue life prediction value Life2 and a second degradation state vector h2 through two output branches. The second degradation state vector h2 is also a 4-dimensional vector activated by Tanh, used to encode the comprehensive degradation state of the adhesive layer under the current working condition and the substrate degradation state. By introducing the first degradation state vector h1, the second-level network can utilize degradation information from the substrate layer when predicting the fatigue life of the adhesive bond.

[0111] The third-level network, corresponding to the adhesive-riveted hybrid connection structure, is configured as follows:

[0112] The system receives a 10-dimensional input vector [q, RH, h1, h2] formed by concatenating the load level parameters, environmental damp heat aging characterization parameters, the first degradation state vector, and the second degradation state vector. After extracting features through an encoder network with the same hierarchical structure as described above, it outputs a third fatigue life prediction value, Life3. By simultaneously introducing the first degradation state vector h1 and the second degradation state vector h2, this third-level network can comprehensively utilize substrate degradation information and adhesive layer degradation information when predicting the life of adhesive-riveted hybrid connections, thereby capturing more complex coupling degradation effects.

[0113] The key design feature of the aforementioned network structure is that information transmission is strictly configured to proceed from lower-level configurations to higher-level configurations, i.e., from the substrate layer to the adhesive layer, and then from the adhesive layer to the hybrid connection layer, rather than in reverse or through parallel sharing. This design allows the network topology itself to encode the physical hierarchical relationships between composite material connection configurations, enhancing the interpretability of the model.

[0114] The load level parameters and environmental damp heat aging parameters in the fatigue life dataset are input into the first layer of the hierarchical degradation state propagation neural network to obtain the first degradation state vector and the first fatigue life prediction value.

[0115] Step S3: Multi-level joint training.

[0116] Using the multi-configuration fatigue life dataset constructed in step S210, the hierarchical degradation state transit neural network constructed in step S220 is jointly trained across multiple layers. During training, the network parameters of the first, second, and third layers are updated simultaneously in the same optimization process, enabling each layer to synchronously learn the lifetime mapping relationship for its corresponding configuration and the degradation state representation shared across configurations. The joint training refers to the simultaneous updating of network parameters of each layer in the hierarchical degradation state transit neural network in the same optimization process, rather than training them independently and then combining them, so that each layer can synchronously learn the lifetime mapping relationship for its corresponding configuration and the degradation state representation shared across configurations.

[0117] The specific training process is as follows:

[0118] Take a batch of training samples from the multi-configuration fatigue life dataset. This batch of samples simultaneously includes the first configuration sample corresponding to the lowest-level network, intermediate configuration samples corresponding to each intermediate-level network, and the target configuration sample corresponding to the highest-level network. Input the load level parameters and environmental aging parameters from the batch of samples into the hierarchical degradation state propagation neural network, and perform forward propagation step by step from the lowest level to the highest level in physical hierarchy:

[0119] First, the load level parameters and environmental aging parameters are input into the lowest-level network. After the encoder of the lowest-level network extracts features, the first degradation state vector and the first fatigue life prediction value are obtained.

[0120] For each intermediate layer network in the hierarchical degradation state propagation neural network, excluding the lowest and highest layer networks, the input vector formed by concatenating the load level parameter, the environmental aging parameter, and the degradation state vector output from the previous layer network is input into the encoder of the intermediate layer network for feature extraction, thereby obtaining the corresponding degradation state vector and fatigue life prediction value of the intermediate layer network.

[0121] Finally, the input vector, formed by concatenating the load level parameters, the environmental aging parameters, and the degradation state vector output from the previous layer of the highest-level network, is input into the encoder of the highest-level network for feature extraction to obtain the highest-level fatigue life prediction value.

[0122] After forward propagation, the joint loss function is calculated based on the fatigue life predictions of each network level and the experimental life values ​​of the corresponding configuration. In one embodiment, the joint loss function is a weighted sum of the life prediction errors of each level:

[0123] L=λ1·MSE(Life1,Y1)+λ2·MSE(Life2,Y2)+λ3·MSE(Life3,Y3)....+λ n ·MSE (Life) n ,Y n )

[0124] Wherein, Life1, Life2, and Life3 are the fatigue life prediction values ​​output by the three hierarchical networks, respectively, and Y1, Y2, Y3...Y... n The values ​​represent the logarithmic lifetimes measured experimentally for the corresponding configurations, MSE represents the mean squared error, and λ1, λ2, and λ3 are the weighting coefficients of the loss function at each level. In a preferred embodiment, each level is set with equal weights, i.e., λ1 = λ2 = λ3 ... = λ n =1.

[0125] The joint loss function value is backpropagated. Since the layers in the hierarchical degenerate state propagation neural network are cascaded and connected via degenerate state vectors, the gradient will propagate back along the propagation path of the degenerate state vectors, starting from the highest-level network and proceeding step-by-step to the lowest-level network, forming an end-to-end continuous gradient backpropagation path. Specifically, the gradient of the highest-level network is backpropagated via the degenerate state vector connection path between it and the preceding intermediate-level network; the gradients of each intermediate-level network are sequentially backpropagated via the degenerate state vector connection path between it and its respective preceding-level network, until the lowest-level network.

[0126] In the same optimization step, the optimizer uses the gradients of the network parameters at each level calculated by backpropagation to synchronously update all network parameters of the lowest-level network, each intermediate-level network, and the highest-level network. In one embodiment, the optimizer employs the Adam optimizer and uses a cosine annealing learning rate scheduling strategy for learning rate decay. Through this synchronous update method, the update of the lowest-level network parameters is driven not only by its own lifetime prediction error but also indirectly by the prediction errors from subsequent levels of the network. This allows the degradation state vector learned by the lowest-level network to better serve the prediction tasks of subsequent levels, achieving cross-configuration shared degradation state representation.

[0127] Repeat the above forward propagation, loss calculation, back propagation and parameter synchronization update steps until the joint loss function converges or reaches the preset training rounds to obtain the target-level degenerate state propagation neural network that has been trained.

[0128] Optionally, a physical consistency constraint term may be added to the joint loss function. This physical consistency constraint term is used to ensure that the prediction results meet preset engineering rules. For example, the physical consistency constraint term can be set as follows: under the same environmental aging conditions, as the load level parameter q increases, the predicted fatigue life logarithm should show a monotonically decreasing trend. By introducing such constraints, the consistency between the model's prediction results and physical understanding can be further improved.

[0129] By designing multi-level joint training and a weighted loss function, the correlation between data from multiple configurations is effectively utilized, alleviating the limitation on model generalization ability caused by insufficient data from a single configuration. Fatigue life testing is costly, especially since test data for complex connection configurations is often scarce. This invention, through shared degradation representations and joint optimization, allows test data from relatively abundant simple configurations (such as substrates) to indirectly assist the predictive learning of complex configurations (such as glued connections) with limited data, thereby achieving more robust generalization performance under small sample conditions and reducing data acquisition costs.

[0130] In this embodiment, the method described above is used to study three types of carbon fiber reinforced resin matrix composite connection configurations to verify the feasibility and prediction accuracy of the proposed method. The three connection configurations are: CFRP substrate, adhesive bonding structure, and adhesive-riveting hybrid connection structure.

[0131] Step S1: Construct a multi-configuration fatigue life dataset

[0132] Fatigue life test data for three types of connection configurations under different load levels and different damp heat aging conditions were collected, totaling 74 sets of data. The specific composition of the dataset is shown in Table 1.

[0133] The distribution of experimental data for each configuration is as follows: Figure 3 As shown in the figure: (1) For the three configurations, under the same aging conditions, the higher the load level, the shorter the fatigue life; (2) The different configurations have different lifespans under the same working conditions, reflecting the influence of the connection form on fatigue performance.

[0134] Step S2: Construct a hierarchical degenerate state propagation neural network

[0135] A three-layer cascaded network is constructed according to the method described in this invention, with the specific parameter settings as follows:

[0136] Degenerate state vector dimension: d=4;

[0137] Encoder structure at each level: two fully connected hidden layers, each with 16 neurons;

[0138] Activation function: Tanh;

[0139] First-level network input dimensions: 2 (q and RH);

[0140] The second-level network input dimension is 6 (4 components: q, RH, and h1).

[0141] The third-level network input dimension is 10 (4 components of q, RH, and h1, and 4 components of h2).

[0142] Total number of trainable parameters in the network: 1339.

[0143] Unlike traditional independent modeling methods, in this embodiment, the three-level network forms a cascade structure through the hierarchical transmission of degenerate state vectors h1 and h2, and the information flow direction is consistent with the physical causal chain of "CFRP substrate → adhesive bonding → adhesive riveting".

[0144] Step S3: Model Training

[0145] The constructed dataset is then input into the network for joint training. The training hyperparameters are set as follows:

[0146] Optimizer: Adam (Adaptive Learning Rate Optimization Algorithm);

[0147] Initial learning rate: 0.005;

[0148] Weight decay coefficient: 1×10⁻³;

[0149] Learning rate scheduling strategy: Cosine Annealing LR, total training epochs T=3000;

[0150] Loss function: L = MSE(Life1, Y1) + MSE(Life2, Y2) + MSE(Life3, Y3), meaning the three levels are equally weighted.

[0151] The training convergence process is as follows Figure 4 As shown, the loss function steadily decreased during training, converging to 0.0372 by the 3000th round, indicating that the model training had fully converged.

[0152] Step S4: Prediction Accuracy Assessment

[0153] After training, the prediction accuracy of the training set was evaluated. The prediction accuracy for each configuration is shown in Table 2, and the scatter plot comparing the predicted and experimental values ​​for each configuration is shown below. Figure 5 As shown, ,

[0154] Where MAE is the mean absolute error (unit: log). 10 (cycles), RMSE is the root mean square error, and R² is the coefficient of determination. The overall R² of the training set is 0.9920, indicating that the model has excellent fitting accuracy. From Figure 5 It can be seen that the predicted values ​​of the three configurations are all closely distributed near the ideal prediction line and all fall within the ±0.5 error band.

[0155] Figure 7 The continuous predictive surfaces for three configurations in the two-dimensional space of load level-humid heat aging are shown. The surfaces are smooth and conform to the basic engineering principle that "the higher the load, the shorter the lifespan", indicating that the model has learned a reasonable input-output mapping relationship.

[0156] Figure 8 The prediction error distribution for each configuration is shown. The errors approximately follow a normal distribution with a mean of zero, and there is no obvious systematic bias, indicating that the model does not have structural underfitting or overfitting problems.

[0157] Step S5: Leave-one-out cross-validation

[0158] To verify the model's generalization prediction ability, leave-one-out cross-validation (LOO-CV) was performed on all 74 working conditions. Each time, one working condition was used as the test sample, and the remaining 73 working conditions were used as the training samples. The model was retrained, and the prediction results for the test samples were recorded. The LOO-CV results are shown in Table 3, with comparisons between predicted and experimental values. Figure 6 As shown, ,

[0159] The overall coefficient of determination for LOO-CV is R² = 0.9732, and the mean absolute error (MAE) is 0.1582 log. 10 (cycles), meaning the average prediction error is less than 0.2 orders of magnitude. Among the three configurations, the CFRP substrate has the highest accuracy (R²=0.9830), while the adhesive bonding accuracy is relatively lower (R²=0.9674), but all meet the requirements of engineering applications for fatigue life prediction accuracy.

[0160] In one embodiment, a fatigue life prediction device for composite material connection structures is provided. The device includes a processor and a memory, and the processor executes the aforementioned program modules stored in the memory to implement the corresponding functions.

[0161] The processor contains a kernel, which retrieves the corresponding program units from memory. One or more kernels can be configured, and adjusting kernel parameters allows for methods to predict the fatigue life of composite material join structures.

[0162] The memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0163] This application provides a storage medium storing a program that, when executed by a processor, implements the above-described method for predicting the fatigue life of composite material connection structures.

[0164] This application provides a processor for running a program, wherein the program executes the above-described method for predicting the fatigue life of composite material connection structures.

[0165] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 9 As shown. The computer device includes a processor A01, a network interface A02, memory (not shown), and a database (not shown) connected via a system bus. The processor A01 provides computing and control capabilities. The memory includes internal memory A03 and a non-volatile storage medium A04. The non-volatile storage medium A04 stores an operating system B01, a computer program B02, and a database (not shown). The internal memory A03 provides an environment for the operation of the operating system B01 and the computer program B02 stored in the non-volatile storage medium A04. The database is used to store data. The network interface A02 is used for communication with external terminals via a network connection. When executed by the processor A01, the computer program B02 implements a method for predicting the fatigue life of composite material connection structures.

[0166] Those skilled in the art will understand that Figure 9 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0167] This application provides a computer (electronic) device, which includes a processor, a memory, and a program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of any of the above-mentioned composite material connection structure fatigue life prediction methods.

[0168] This application also provides a computer program product that, when executed on a data processing device, is suitable for executing a program that initializes the steps of a method for predicting the fatigue life of a composite material connection structure.

[0169] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0170] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0171] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0172] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0173] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0174] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0175] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0176] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0177] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for predicting the fatigue life of composite material connection structures, characterized in that, The method includes: Obtain the basic parameters of the composite material substrate under the predicted working condition; The basic parameters are input into a trained hierarchical degradation state propagation neural network to obtain the fatigue life prediction value of the composite material; The hierarchical degradation state propagation neural network is a multi-level cascaded neural network. The number of levels in the hierarchical degradation state propagation neural network is determined according to the physical hierarchical structure of the composite material connection configuration. The lowest level network in the hierarchical degradation state propagation neural network corresponds to the composite matrix material configuration to be predicted, and the higher level networks correspond to the composite material connection configuration to be predicted. They are connected sequentially from the lower level to the higher level in a physical causal order. The lowest-level network is configured to generate a first degraded state vector and a fatigue life prediction value for that level based on the basic parameters. Each layer of the network, except for the lowest layer, is configured to: receive an input formed by concatenating the basic parameters and the degradation state vector output from the previous layer network, generate the degradation state vector of the current layer network and the fatigue life prediction value corresponding to the current layer; the degradation state vector of the previous layer is used to encode the degradation state of the corresponding layer.

2. The method for predicting the fatigue life of composite material connection structures according to claim 1, characterized in that, The basic parameters include load level parameters and environmental aging parameters. The hierarchical degradation state transmission neural network is a three-layer neural network, including a first-layer network, a second-layer network, and a third-layer network. The first-layer network corresponds to the carbon fiber reinforced resin composite matrix structure and outputs a first fatigue life prediction value and a first degradation state vector based on the load level parameters and environmental aging parameters. The second-level network corresponds to the adhesive bonding structure of carbon fiber reinforced resin composite material, receives the input formed by splicing the load level parameter, environmental aging parameter and the first degradation state vector, and outputs the second fatigue life prediction value and the second degradation state vector. The third-level network corresponds to the glued and riveted hybrid connection structure of carbon fiber reinforced resin composite material. It receives inputs formed by splicing the load level parameters, environmental aging parameters, the first degradation state vector and the second degradation state vector, and outputs a third fatigue life prediction value.

3. The fatigue life prediction method for composite material connection structures according to claim 2, characterized in that, The first-level network includes a first encoder network and two output branches. The load level parameter and environmental aging parameter are extracted by the first encoder network. The first fatigue life prediction value is generated by the first output branch of the first-level network, and the multi-dimensional first degradation state vector is generated by the second output branch of the first-level network. The second-level network includes a second encoder network and two output branches. The multidimensional vector formed by concatenating the load level parameter, the environmental aging parameter, and the first degradation state vector is processed by the second encoder network for feature extraction. The second fatigue life prediction value is generated by the first output branch of the second-level network, and the multidimensional second degradation state vector is generated by the second output branch of the second-level network. The third-level network includes a third encoder network. The multi-dimensional vector obtained by concatenating the load level parameter, the environmental aging parameter, the first degradation state vector, and the second degradation state vector is processed by the second encoder network for feature extraction to generate the third fatigue life prediction value.

4. The fatigue life prediction method for composite material connection structures according to claim 3, characterized in that, The first encoder network, the second encoder network, and the third encoder network all contain 1 to 3 fully connected layers, each containing 8 to 64 neurons, and use a non-linear activation function as the output transformation of the layer.

5. The method for predicting the fatigue life of composite material connection structures according to claim 4, characterized in that, The first encoder network, the second encoder network, and the third encoder network each contain two fully connected layers, each containing 16 neurons, and the Tanh activation function is used as the output transformation of the layer.

6. The method for predicting the fatigue life of composite material connection structures according to any one of claims 1 to 5, characterized in that, The training process of the hierarchical degenerate state propagation neural network includes: Obtain the hierarchical degradation state propagation neural network to be trained and the multi-configuration fatigue life dataset of various composite materials; The basic parameters in the fatigue life dataset are input into the lowest level network of the hierarchical degradation state propagation neural network to obtain the first degradation state vector and the first fatigue life prediction value. For each intermediate level network in the multi-level degradation state transfer network except for the lowest level network and the highest level network, the input vector formed by splicing the basic parameters and the degradation state vector output from the previous level network is input into the next level network to obtain the degradation state vector and fatigue life prediction value corresponding to each level. The load level parameter, the environmental aging parameter, and the degradation state vector output from the previous level network of the highest level network are concatenated to form an input vector, which is then input into the highest level network of the multi-level state transfer module to obtain the highest level fatigue life prediction value. Based on the first fatigue life prediction value, the fatigue life prediction values ​​corresponding to each intermediate level network, and the fatigue life prediction value of the highest level, the network parameters of each level network are jointly updated in the same optimization process, so that each level network can simultaneously learn the life mapping relationship of the corresponding configuration and the degradation state representation shared across configurations, and obtain the trained target level degradation state propagation neural network.

7. The fatigue life prediction method for composite material connection structures according to claim 6, characterized in that, The method for jointly updating the network parameters of each level network in the same optimization process based on the first fatigue life prediction value, the fatigue life prediction values ​​corresponding to each intermediate level network, and the highest level fatigue life prediction value includes: Based on the first predicted fatigue life value, the predicted fatigue life values ​​corresponding to each intermediate level network, and the predicted fatigue life value of the highest level, and the logarithmic values ​​of the experimental life values ​​Y1, Y2, Y3...Y... of the corresponding levels... n The total loss value L is calculated using the following joint loss function: L=λ1·MSE(Life1,Y1)+λ2·MSE(Life2,Y2)+λ3·MSE(Life3,Y3)....+λ n ·MSE(Life n ,Y n ); Backpropagation is performed on the total loss value L, and the gradient is cascaded along the direction of the highest-level network, the third-level network, the second-level network to the first-level network. In the same optimization step, the Adam optimizer is used in conjunction with the cosine annealing learning rate scheduling strategy to synchronously update all network parameters from the first-level network to the highest-level network. Repeat the above steps until the joint loss function converges to obtain the trained target-level degenerate state propagation neural network.

8. The method for predicting the fatigue life of composite material connection structures according to claim 7, characterized in that, The above λ1=λ2=λ3...=λ n =1.

9. A device for predicting the fatigue life of composite material connection structures, characterized in that, include: The memory is configured to store instructions; A processor is configured to retrieve the instructions from the memory and, when executing the instructions, to implement the fatigue life prediction method for composite material connection structures according to any one of claims 1 to 8.

10. A machine-readable storage medium storing instructions thereon, characterized in that, When executed by a processor, this instruction causes the processor to be configured to perform the fatigue life prediction method for composite material connection structures according to any one of claims 1 to 8.

Images