Multi-axis random load fatigue damage prediction method with fusion cross-attention mechanism
By introducing a Transformer model with a cross-attention mechanism, the problem of insufficient accuracy in multiaxial load prediction of traditional fatigue analysis methods is solved, and high-precision fatigue damage prediction is achieved, which is applicable to structural life prediction under multiaxial random load conditions.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- DONGFANG ELECTRIC MACHINERY
- Filing Date
- 2026-04-10
- Publication Date
- 2026-07-10
AI Technical Summary
Traditional fatigue analysis methods cannot fully consider the coupling effect between multiaxial loads and the dynamic characteristics of load history, resulting in insufficient accuracy in fatigue life prediction.
The Transformer model employs a cross-attention mechanism to achieve the fusion of multi-axis load information and high-precision prediction by preprocessing and extracting features from multi-axis random load signals and introducing a cross-attention mechanism and a fully connected layer.
It significantly improves the accuracy and generalization ability of fatigue damage prediction under multi-axis random loading conditions, can automatically learn the contribution weight of load components, capture long-sequence long-distance dependence features, and improve the accuracy and robustness of fatigue life prediction under complex working conditions.
Smart Images
Figure CN122365353A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a multi-axis random load fatigue damage prediction method that incorporates a cross-attention mechanism, belonging to the field of fatigue life prediction technology. Background Technology
[0002] In complex service environments, multiaxial random loads lead to the accumulation of fatigue damage in structural components, posing a significant engineering challenge to fatigue life prediction. Current traditional fatigue analysis methods, such as those based on Miner's linear accumulation rule and stress transformation techniques, struggle to fully consider the coupling effects between multiaxial loads and the dynamic characteristics of load history. In recent years, deep learning technology has seen increasing application in fatigue prediction, with the Transformer demonstrating unique advantages in time series processing due to its powerful sequence modeling capabilities. However, standard Transformer models struggle to capture the interaction characteristics between multiaxial loads, and their prediction accuracy still needs improvement. Therefore, there is an urgent need for a fatigue life prediction method that can integrate multiaxial load information and achieve higher prediction accuracy. Summary of the Invention
[0003] The purpose of this invention is to provide a multi-axis random load fatigue damage prediction method that integrates a cross-attention mechanism to address the aforementioned problems.
[0004] The technical solution adopted in this invention is as follows: A multi-axis stochastic load fatigue damage prediction method incorporating a cross-attention mechanism includes a data processing stage: S1. Acquire multi-axis random load signals and preprocess the signals to align the load signals of different axes according to time steps to obtain multi-axis time-series data segments. S2. Convert the preprocessed multi-axis time series data segments into high-dimensional feature representations that the model can process; S3. Input the high-dimensional features into the Transformer encoder and introduce a cross-attention mechanism. The Transformer encoder finally outputs multi-axis fused temporal features. S4. Input the multi-axis fused temporal features from S3 into the fully connected layer to obtain fatigue damage.
[0005] The optional steps include: Step 1: Using the known random load as the input feature and fatigue damage as the output feature, the features are input into the neural network for training through the data processing stages S1-S4 to form an evaluation model for subsequent prediction of fatigue damage. Step 2: Input the random load to be predicted as the input feature into the evaluation model through S1-S4 of the data processing stage, output fatigue damage, and determine that the structure has failed due to fatigue when the cumulative damage D reaches 1; calculate and output the predicted fatigue damage value.
[0006] Among the options available in S1 are: S11. Obtain the multi-axis random load signal of the target structure under actual working conditions; S12. Normalize, denoise, and normalize the signal obtained in S11; S13. Align the load signals of different axes obtained in S12 according to the time step to form a multidimensional input matrix X with dimensions T×M, where T is the time step size and M is the number of axial channels (M≥2).
[0007] Among the options available in S2 are: S21. Map the input multi-axis sequence data to a unified high-dimensional feature space through independent linear mapping layers.
[0008] in, It is the i-th axial signal sequence. It is a weight matrix. It is the bias vector; S22. Using a positional coding method, for Inject absolute position information This enables the model to perceive the order of the sequence; ultimately, the input sequence for each axis is transformed into a single dimension. The embedding matrix,
[0009]
[0010] in, Axial load The result after mapping to a higher dimension It is the result after position encoding. It is the input sequence for each axis.
[0011] Among the options available in S3 are: S31, transfer each axial signal Each sequence is input as an independent Transformer encoder to extract the temporal dependencies within the axial direction. Each Transformer encoder consists of a multi-head self-attention layer and a feedforward neural network; S32. To capture the interaction between signals along different axes, a cross-attention mechanism is introduced: Selecting the characteristics of a certain axis signal , As the query vector Q, the other axial signals As key K and value V; computed via cross-attention mechanism and The weighted fusion between them yields cross-attention:
[0012] S33, Integrate cross-attention output with features Perform residual connections and layer normalization to output local interactive feature sequences. ; .
[0013] S3 also includes the following options: S34. Using a cyclic cross-attention mechanism, random loads along different axes are used sequentially as query vectors, while other axes are used as auxiliary inputs to obtain the updated feature sequence:
[0014] S35. After multiple rounds of interaction, an enhanced multi-axis temporal feature set is obtained: { }
[0015] S3 also includes the following options: S36. Introduce a multi-head mechanism in cross-attention computation, dividing the features into multiple subspaces for parallel computation.
[0016]
[0017] Where h is the number of heads, It is one of the heads of the multi-head attention mechanism.
[0018] S3 also includes the following options: S37. Input the enhanced features into the fully connected layer. Perform dimension mapping; S38. Obtain the final multi-axis fusion timing features. , .
[0019] Alternatively, in S4, weighted accumulation is performed using the Miner linear accumulation criterion.
[0020] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. This invention provides a multi-axis random load fatigue damage prediction method that integrates a cross-attention mechanism. By introducing a cross-attention mechanism to dynamically model the coupling relationship between multi-axis loads, it can automatically learn the contribution weights of different load components to fatigue damage, achieving high-precision prediction of fatigue damage under multi-axis random load conditions. Compared with traditional equivalent stress methods, this method avoids information loss caused by manual assumptions and can characterize temporal nonlinear coupling effects, thereby significantly improving prediction accuracy and generalization ability.
[0021] 2. The present invention provides a multi-axis random load fatigue damage prediction method that integrates cross-attention mechanism. It uses the multi-head self-attention mechanism of Transformer to mine long-distance dependence features in long sequences and comprehensively capture the long-term dependence and non-stationary characteristics of multi-axis random loads. The introduction of cross-attention mechanism realizes information interaction and dynamic fusion between load features of different axes. It improves the accuracy and robustness of fatigue life prediction under complex working conditions. Attached Figure Description
[0022] Figure 1 This is a flowchart of the present invention. Detailed Implementation
[0023] The present invention will now be described in detail with reference to the accompanying drawings.
[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0025] A multi-axis stochastic load fatigue damage prediction method incorporating cross-attention mechanism, such as Figure 1 As shown, the data processing stage: S1. Acquire multi-axis random load signals and preprocess the signals to align the load signals of different axes according to time steps to obtain multi-axis time-series data segments. S2. Convert the preprocessed multi-axis time series data segments into high-dimensional feature representations that the model can process; S3. Input the high-dimensional features into the Transformer encoder and introduce a cross-attention mechanism. The Transformer encoder finally outputs multi-axis fused temporal features. S4. Input the multi-axis fused temporal features from S3 into the fully connected layer to obtain fatigue damage.
[0026] Fatigue is the process of localized damage to a component caused by cyclic loading. It is a comprehensive result of a cumulative process including crack initiation, propagation, and eventual fracture. As time increases, the number of loading cycles a component endures increases, leading to greater damage. S1 aligns load signals from different axes according to time steps to obtain multi-axis time-series data segments, arranging them chronologically to better reflect actual operational patterns. S2 provides a high-quality, standardized data foundation for subsequent feature extraction and modeling, avoiding model training distortion or biased judgment results due to data quality issues. S3 introduces a cross-attention mechanism in the Transformer encoder, allowing elements in one sequence to focus on elements in another different sequence, thus achieving information alignment and fusion between the two sequences. This matches the external forces acting in multiple directions or multiple degrees of freedom between multi-axis loads, realizing the correlation between multi-axis loads. S4 In the Transformer architecture, the fully connected layer is a crucial component after the self-attention layer in the encoder and decoder. It further extracts complex features and enhances the model's expressive power through a combination of two linear transformations and activation functions, used to integrate features and output the final prediction.
[0027] As another specific implementation method, the following steps are included: Step 1: Using the known random load as the input feature and fatigue damage as the output feature, the features are input into the neural network for training through the data processing stages S1-S4 to form an evaluation model for subsequent prediction of fatigue damage. Step 2: Input the random load to be predicted as the input feature into the evaluation model through S1-S4 of the data processing stage, output fatigue damage, and determine that the structure has failed due to fatigue when the cumulative damage D reaches 1; calculate and output the predicted fatigue damage value.
[0028] Neural networks can learn complex nonlinear mappings from data without requiring manual derivation of precise mathematical formulas or physical models, significantly reducing reliance on prior knowledge and modeling experience. By continuously optimizing parameters through a large number of samples, they gradually improve prediction accuracy and generalization ability, adapting to different data distributions and complex patterns. After model training, only a new random load needs to be input to quickly obtain the corresponding fatigue damage, achieving efficient prediction. Each stress cycle generates a certain amount of fatigue damage; all cycle damage can be accumulated; when the accumulated damage reaches a critical value of 1, the material or structure is considered to have reached its fatigue limit. At this point, the fatigue damage value is the maximum stress level at which the material can withstand an infinite number of cycles without fatigue failure.
[0029] As another specific implementation, S1 includes: S11. Obtain the multi-axis random load signal of the target structure under actual working conditions; S12. Normalize, denoise, and normalize the signal obtained in S11; S13. Align the load signals of different axes obtained in S12 according to the time step to form a multidimensional input matrix X with dimensions T×M, where T is the time step size and M is the number of axial channels (M≥2).
[0030] S11 Acquires raw load data that closely matches actual working conditions to ensure the authenticity of subsequent analysis; S12 Cleans the data and unifies the scale through noise reduction and normalization, eliminating invalid interference and dimensional differences; S13 Achieves time alignment and structured integration of multi-axis signals, preserving the temporal and coupling characteristics of the load.
[0031] As another specific implementation, S2 includes: S21. Map the input multi-axis sequence data to a unified high-dimensional feature space through independent linear mapping layers.
[0032] in, It is the i-th axial signal sequence. It is a weight matrix. It is the bias vector; S22. Using a positional coding method, for Inject absolute position information This enables the model to perceive the order of the sequence; ultimately, the input sequence for each axis is transformed into a single dimension. The embedding matrix,
[0033]
[0034] in, Axial load The result after mapping to a higher dimension It is the result after position encoding. It is the input sequence for each axis.
[0035] S21 Mapping data to a high-dimensional space solves the problem of data being linearly inseparable in low-dimensional space, while the high-dimensional feature space can effectively represent complex data. S22 Positional encoding is a key component in the Transformer architecture, used to provide the model with information about the order of words in the sequence. Since the Transformer's self-attention mechanism is inherently permutation-invariant, the model cannot perceive the order of the input sequence; therefore, positional encoding is needed to explicitly inject positional information.
[0036] As another specific implementation, S3 includes: S31, transfer each axial signal Each sequence is input as an independent Transformer encoder to extract the temporal dependencies within the axial direction. Each Transformer encoder consists of a multi-head self-attention layer and a feedforward neural network; S32. To capture the interaction between signals along different axes, a cross-attention mechanism is introduced: Selecting the characteristics of a certain axis signal , As the query vector Q, the other axial signals As key K and value V; computed via cross-attention mechanism and The weighted fusion between them yields cross-attention:
[0037] S33, Integrate cross-attention output with features Perform residual connections and layer normalization to output local interaction features. ; .
[0038] As another specific implementation, S3 also includes: S34. Using a cyclic cross-attention mechanism, random loads along different axes are used sequentially as query vectors, while other axes are used as auxiliary inputs to obtain the updated feature sequence:
[0039] S35. After multiple rounds of interaction, an enhanced multi-axis temporal feature set is obtained: { }
[0040] The recurrent cross-attention mechanism introduces a recurrent iteration and step-by-step feature aggregation strategy on top of cross-attention, enabling global dependency modeling with lower computational complexity and memory usage. Traditional cross-attention directly performs pairwise interactions at all positions, resulting in a quadratic increase in computational and memory consumption with sequence length. In contrast, recurrent cross-attention performs attention weighting in different directions and steps, and gradually integrates contextual information through recurrent propagation, significantly reducing complexity while approximately maintaining global modeling capabilities. It can enhance the model's understanding of context by strengthening long-distance associations between features through multiple lightweight iterations without significantly increasing the number of parameters.
[0041] As another specific implementation, S3 also includes: S36. Introduce a multi-head mechanism in cross-attention computation, dividing the features into multiple subspaces for parallel computation.
[0042]
[0043] Where h is the number of heads, It is one of the heads of the multi-head attention mechanism.
[0044] The multi-head stacking mechanism allows cyclic cross-attention to extract complementary information from multiple independent feature subspaces while maintaining high computational efficiency. Different attention heads can focus on different types of relationships between features, avoiding the one-sidedness of a single attention head in modeling dependencies. Multi-head parallel processing can enhance the model's ability to capture complex feature patterns, improve the discrimination of long-distance dependencies and local details, and without significantly increasing the computational burden per step.
[0045] As another specific implementation, S3 also includes: S37. Input the cross-attention enhanced features into the fully connected layer. Perform dimension mapping; S38. Obtain the final multi-axis fusion timing features. , .
[0046] The fully connected layer performs dimensional scaling and feature reconstruction on the joint features output by the attention module through point-by-point independent nonlinear mapping, thereby achieving information filtering and enhancement, and at the same time completing the adaptation from the high-dimensional feature space to the output dimension required by the task.
[0047] In another specific implementation, in S4, weighted accumulation is performed using the Miner linear accumulation criterion.
[0048] In linear equal damage models, the Miner criterion has the closest critical damage value to 1 under random load spectrum and is easy to calculate, with better lifetime estimation accuracy than other models.
[0049] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. The invention extends to any new features or combinations disclosed in this specification, and any modifications, equivalent substitutions, and improvements made within the spirit and principles of the invention should be included within the scope of protection of the invention. It is obvious to those skilled in the art that the invention is not limited to the details of the above exemplary embodiments, and that detailed technical features not disclosed in this embodiment, such as specific structures, are all prior art and can be obtained by those skilled in the art from the prior art. The connection method can be a fixed connection, a detachable connection, or an integral part; it can be a fixed connection, a movable connection, or a hinged connection; it can be a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific manner of the above terms in the embodiments of the present invention according to the specific circumstances, and this disclosure does not specifically limit this aspect.
Claims
1. A multi-axis stochastic load fatigue damage prediction method incorporating a cross-attention mechanism, characterized in that: Including the data processing stage: S1. Acquire multi-axis random load signals and preprocess the signals to align the load signals of different axes according to time steps to obtain multi-axis time-series data segments. S2. Convert the preprocessed multi-axis time series data segments into high-dimensional feature representations that the model can process; S3. Input the high-dimensional features into the Transformer encoder and introduce a cross-attention mechanism. The Transformer encoder finally outputs multi-axis fused temporal features. S4. Input the multi-axis fused temporal features from S3 into the fully connected layer to obtain fatigue damage.
2. The method as described in claim 1, characterized in that: Includes the following steps: Step 1: Using the known random load as the input feature and fatigue damage as the output feature, the features are input into the neural network for training through the data processing stages S1-S4 to form an evaluation model for subsequent prediction of fatigue damage. Step 2: Input the random load to be predicted as the input feature into the evaluation model through S1-S4 of the data processing stage, output fatigue damage, and determine that the structure has failed due to fatigue when the cumulative damage D reaches 1; calculate and output the predicted fatigue damage value.
3. The method as described in claim 1, characterized in that: S1 includes: S11. Obtain the multi-axis random load signal of the target structure under actual working conditions; S12. Normalize, denoise, and normalize the signal obtained in S11; S13. Align the load signals of different axes obtained in S12 according to the time step to form a multidimensional input matrix X with dimensions T×M, where T is the time step size and M is the number of axial channels (M≥2).
4. The method as described in claim 1, characterized in that: S2 includes: S21. Map the input multi-axis sequence data to a unified high-dimensional feature space through independent linear mapping layers. in, It is the i-th axial signal sequence. It is a weight matrix. It is the bias vector; S22. Using a positional coding method, for Inject absolute position information This enables the model to perceive the order of the sequence; ultimately, the input sequence for each axis is transformed into a single dimension. The embedding matrix, in, Axial load The result after mapping to a higher dimension It is the result after position encoding. It is the input sequence for each axis.
5. The method as described in claim 1, characterized in that: S3 includes: S31, transfer each axial signal Each sequence is input as an independent Transformer encoder to extract the temporal dependencies within the axial direction. Each Transformer encoder consists of a multi-head self-attention layer and a feedforward neural network; S32. To capture the interaction between signals along different axes, a cross-attention mechanism is introduced: Selecting the characteristics of a certain axis signal , As the query vector Q, the other axial signals As key K and value V; computed via cross-attention mechanism and The weighted fusion between them yields cross-attention: S33, Integrate cross-attention output with features Perform residual connections and layer normalization to output local interactive feature sequences. ; 。 6. The method as described in claim 5, characterized in that: S3 also includes: S34. Using a cyclic cross-attention mechanism, random loads along different axes are used sequentially as query vectors, while other axes are used as auxiliary inputs to obtain the updated feature sequence: S35. After multiple rounds of interaction, an enhanced multi-axis temporal feature set is obtained: { } 7. The method as described in claim 6, characterized in that: S3 also includes: S36. Introduce a multi-head mechanism in cross-attention computation, dividing the features into multiple subspaces for parallel computation. Where h is the number of heads, It is one of the heads of the multi-head attention mechanism.
8. The method as described in claim 7, characterized in that: S3 also includes: S37. Input the enhanced features into the fully connected layer. Perform dimension mapping; S38. Obtain the final multi-axis fusion timing features. , .
9. The method as described in claim 1, characterized in that: In S4, weighted accumulation is performed using the Miner linear accumulation criterion.