Intelligent metal material fatigue evolution analysis system

Abstract

The application discloses an intelligent metal material fatigue evolution analysis system and belongs to the technical field of material analysis. The system comprises a metal material data acquisition module, a metal material data preprocessing module, a metal material fatigue feature extraction module, a metal material fatigue feature evolution analysis module and an intelligent fatigue evolution analysis module. According to the environmental influence factor and the time step attention weight, the application obtains a two-dimensional calibration feature, constructs a fatigue prior vector, samples a fatigue prior embedding feature, and then obtains a fatigue residual optimization feature through residual attention mask optimization. According to an adaptive adjustment coefficient, a multi-scale position coding matrix is generated to obtain a prior-position collaborative fusion feature matrix, an exclusive prior vector is extracted, and finally, a prediction label is generated through a core feedforward branch and an adaptive gating branch, so that the reliability and accuracy of metal material fatigue evolution analysis are effectively improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of material analysis, and particularly refers to an intelligent metal material fatigue evolution analysis system. BACKGROUND

[0002] The metal material fatigue evolution analysis system is a kind of system which utilizes artificial intelligence technology, analyzes fatigue time series data of metal materials comprehensively, accurately identifies key characteristic parameters in different damage stages, and outputs fatigue state evaluation results in real time, so as to help realize early warning and accurate intervention of metal component fatigue risk, and thus improve equipment operation safety and reliability.

[0003] However, the existing metal material fatigue evolution analysis system has the problems of ignoring the dynamic coupling effect of environment and time, and not integrating prior knowledge of fatigue evolution, so as to be unable to accurately extract features of metal materials in different damage stages, and thus low accuracy of metal material fatigue state judgment; the existing metal material fatigue evolution analysis system has the problems of not considering the differentiated influence of fatigue damage stages on time series dependence strength, and rigid fusion mode of multi-source information, so as to be difficult to dynamically adapt to evolution law analysis requirements in different damage states. SUMMARY

[0004] In view of the above, in order to overcome the defects of the prior art, the application provides an intelligent metal material fatigue evolution analysis system, in order to overcome the defects of the prior art, the application provides an intelligent metal material fatigue evolution analysis system, in view of the fact that the existing metal material fatigue evolution analysis system can easily ignore the dynamic coupling effect of environment and time, and does not integrate prior knowledge of fatigue evolution, and cannot accurately extract the characteristics of metal materials at different damage stages, resulting in low accuracy of metal material fatigue state judgment, the present application generates environment calibration features according to environmental influence factors, obtains double-dimensional calibration features by combining time step attention weights, constructs fatigue prior vectors, obtains fatigue prior embedding features, samples fatigue prior embedding features according to an adaptive sampling rate, extracts features in parallel using a multi-scale convolution kernel, and then optimizes the features through residual attention mask optimization to obtain fatigue residual optimization features, so as to avoid feature redundancy or loss of key information, and make the extracted features more accurately represent the fatigue damage state, thereby effectively improving the reliability and accuracy of metal material fatigue evolution analysis; in view of the fact that the existing metal material fatigue evolution analysis system does not consider the differentiated influence of fatigue damage stage on time sequence dependence strength, and the fusion mode of multi-source information is rigid, which makes it difficult to dynamically adapt to the evolution law analysis demand of different damage states, the present application generates a multi-scale position encoding matrix according to an adaptive adjustment coefficient, sets fatigue prior fusion weights and position encoding fusion weights, obtains a prior-position collaborative fusion feature matrix, extracts a dedicated prior vector according to a prior mixed weight, obtains a multi-head collaborative attention feature, obtains a final feature through a core feedforward branch and an adaptive gating branch, generates a prediction label, and provides accurate and reliable stage judgment basis for metal material fatigue evolution analysis, and effectively adapts to the evolution law analysis demand of different damage states.

[0005] The application provides an intelligent metal material fatigue evolution analysis system, which comprises a metal material data acquisition module, a metal material data preprocessing module, a metal material fatigue feature extraction module, a metal material fatigue feature evolution analysis module and an intelligent fatigue evolution analysis module.

[0006] The metal material data acquisition module acquires historical metal material fatigue time series data and manually labels fatigue damage stages.

[0007] The metal material data preprocessing module performs data cleaning, data normalization and data encoding processing.

[0008] The metal material fatigue feature extraction module generates environment calibration features according to environmental influence factors, obtains double-dimensional calibration features by combining time step attention weights, constructs fatigue prior vectors, obtains fatigue prior embedding features, samples fatigue prior embedding features according to an adaptive sampling rate, extracts features in parallel using a multi-scale convolution kernel, and then optimizes the features through residual attention mask optimization to obtain fatigue residual optimization features.

[0009] The metal material fatigue feature evolution analysis module generates a multi-scale position coding matrix based on adaptive adjustment coefficients, sets fatigue prior fusion weights and position coding fusion weights to obtain a prior-position collaborative fusion feature matrix, extracts exclusive prior vectors based on prior mixed weights to obtain multi-head collaborative attention features, obtains the final features through core feedforward branches and adaptive gating branches, and generates prediction labels.

[0010] The intelligent fatigue evolution analysis module collects real-time fatigue time series data of metallic materials, obtains the corresponding fatigue damage stage based on the predicted label, and completes the fatigue evolution analysis of metallic materials.

[0011] Furthermore, the metal material data acquisition module collects historical metal material fatigue time series data and manually marks the fatigue damage stages;

[0012] The fatigue time series data of the metallic materials includes time steps, load data, response data, environmental data, and basic property data of the metallic materials;

[0013] The fatigue damage stages include no damage stage, slight damage stage, moderate damage stage, severe damage stage, and failure stage, and the fatigue damage stages are used as data labels.

[0014] Furthermore, the metal material data preprocessing module performs data cleaning, data normalization, and data encoding on the collected metal material fatigue time series data.

[0015] Furthermore, the metal material fatigue feature extraction module includes an environment-time dual-dimensional calibration unit, a fatigue prior embedding unit, and a stage-aware convolutional fusion unit; specifically, it includes the following:

[0016] Environment-time dual-dimensional calibration unit: For the fatigue time series data of metal materials after data preprocessing, the environmental impact factor is calculated based on the environmental data to obtain the environmental calibration features. The environmental calibration features are then subjected to global average pooling, and time step attention weights are generated through a fully connected layer. The environmental calibration features are then multiplied element-wise with the time step attention weights to obtain the dual-dimensional calibration features.

[0017] Fatigue prior embedding unit: The two-dimensional calibration features are mapped to a high-dimensional vector space through layer normalization to obtain the embedding layer features. The fatigue prior vector is constructed based on the load data and response data. The embedding layer features and the fatigue prior vector are fused to obtain the fatigue prior embedding features.

[0018] Stage-aware convolutional fusion unit: Adaptive sampling rates are set for each fatigue damage stage. The fatigue prior embedding features are sampled according to the adaptive sampling rates. Multi-scale convolutional kernels are used to extract the sampled fatigue prior embedding features in parallel to obtain multi-scale convolutional fusion features. The multi-scale convolutional fusion features are then optimized by residual attention mask to obtain fatigue residual optimized features.

[0019] Furthermore, the fatigue characteristic evolution analysis module for metallic materials includes a multi-scale position encoding unit, a priori-position collaborative fusion unit, a fatigue perception multi-head collaborative attention unit, a gated feedforward enhancement unit, and an analysis unit; specifically, it includes the following:

[0020] Multi-scale position coding unit; using sine-cosine coding method to generate basic position coding matrix, setting adaptive adjustment coefficients for each fatigue damage stage to obtain stage adjustment position coding matrix, and combining basic position coding matrix and stage adjustment position coding matrix to obtain multi-scale position coding matrix;

[0021] Prior-location collaborative fusion unit: Fatigue prior fusion weight and location encoding fusion weight are set respectively, and fatigue residual optimization features, fatigue prior vector and multi-scale location encoding matrix are weighted and fused to obtain prior-location collaborative fusion feature matrix;

[0022] Fatigue perception multi-head collaborative attention unit: Based on the prior mixing weights of the attention heads, the exclusive prior vector of each attention head is extracted from the fatigue prior vector, and then combined with the prior-position collaborative fusion feature matrix to generate the query matrix, key matrix and value matrix of each attention head. The fatigue perception attention weight of each attention head is calculated, and the outputs of all attention heads are concatenated and linearly mapped to obtain the multi-head collaborative attention features.

[0023] Gated feedforward enhancement unit: Extracts core feedforward features from multi-head collaborative attention features through core feedforward branches, generates gating adjustment coefficients based on adaptive gating branches, multiplies the core feedforward features with the gating adjustment coefficients element by element to obtain gating filtering features, connects the residuals of multi-head collaborative attention features and gating filtering features and performs layer normalization to obtain the final features;

[0024] Analysis unit: Based on fully connected layers and the Softmax function, the final features are mapped to the probability distribution of five fatigue damage stages, and the fatigue damage stage with the highest probability is selected as the predicted label output.

[0025] Furthermore, the intelligent fatigue evolution analysis module collects real-time fatigue time series data of metallic materials and inputs it sequentially into the metallic material data preprocessing module, the metallic material fatigue feature extraction module, and the metallic material fatigue feature evolution analysis module. Based on the output prediction labels, the fatigue damage stage corresponding to the real-time fatigue time series data of metallic materials is obtained, thus completing the fatigue evolution analysis of metallic materials.

[0026] The beneficial effects achieved by the present invention using the above solution are as follows:

[0027] (1) In view of the problem that the existing fatigue evolution analysis system for metallic materials easily ignores the dynamic coupling effect of environment and time, and does not incorporate prior knowledge of fatigue evolution, it is impossible to accurately extract the characteristics of metallic materials under different damage stages, resulting in low accuracy of fatigue state judgment of metallic materials. This scheme generates environmental calibration features based on environmental influence factors, and obtains dual-dimensional calibration features by combining time step attention weights. This enables the features to dynamically match the environmental effects under actual working conditions, suppress secondary time step interference, and more accurately reflect the fatigue evolution law under environment-time coupling. The fatigue prior vector is constructed to obtain fatigue prior embedded features, and the prior knowledge of metallic materials is directly embedded into the features to enhance the correlation between features and fatigue state. The fatigue prior embedded features are sampled according to the adaptive sampling rate, and features are extracted in parallel using multi-scale convolution kernels. Then, fatigue residual optimized features are obtained through residual attention mask optimization to avoid feature redundancy or loss of key information. This enables the extracted features to more accurately represent the fatigue damage state and effectively improve the reliability and accuracy of fatigue evolution analysis of metallic materials.

[0028] (2) To address the problem that existing fatigue evolution analysis systems for metallic materials do not consider the differential impact of fatigue damage stages on temporal dependence intensity and have rigid multi-source information fusion methods, making it difficult to dynamically adapt to the needs of evolution law analysis under different damage states, this scheme generates a multi-scale position coding matrix based on adaptive adjustment coefficients to accurately match the differences in temporal dependence intensity at each stage, gain a clearer understanding of the dynamic temporal process of fatigue evolution, and improve the sensitivity to changes in damage stages; it sets fatigue prior fusion weights and position coding fusion weights to obtain a prior-position collaborative fusion feature matrix, making the fused features more consistent with the evolution laws of different damage states and enhancing the feature's ability to represent fatigue evolution; it extracts exclusive prior vectors based on prior mixed weights to obtain multi-head collaborative attention features, accurately captures the core features under each damage state, and improves the ability of attention to focus on key information of fatigue damage; it obtains the final features through core feedforward branches and adaptive gating branches, generates prediction labels, and provides accurate and reliable stage judgment basis for fatigue evolution analysis of metallic materials, effectively adapting to the needs of evolution law analysis under different damage states. Attached Figure Description

[0029] Figure 1 A schematic diagram of an intelligent fatigue evolution analysis system for metallic materials provided by the present invention;

[0030] Figure 2 This is a schematic diagram of the fatigue feature extraction module for metallic materials.

[0031] Figure 3 This is a schematic diagram of the fatigue characteristic evolution analysis module for metallic materials.

[0032] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0033] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0034] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0035] Example 1, see Figure 1 The present invention provides an intelligent fatigue evolution analysis system for metallic materials, comprising a metallic material data acquisition module, a metallic material data preprocessing module, a metallic material fatigue feature extraction module, a metallic material fatigue feature evolution analysis module, and an intelligent fatigue evolution analysis module;

[0036] The metal material data acquisition module collects historical metal material fatigue time series data, manually marks the fatigue damage stages, and sends the data to the metal material data preprocessing module.

[0037] The metal material data preprocessing module receives data sent by the metal material data acquisition module, performs data cleaning, data normalization and data encoding processing, and sends the data to the metal material fatigue feature extraction module.

[0038] The metal material fatigue feature extraction module receives data sent by the metal material data preprocessing module, generates environmental calibration features based on environmental impact factors, obtains dual-dimensional calibration features by combining time step attention weights, constructs fatigue prior vectors, obtains fatigue prior embedding features, samples the fatigue prior embedding features according to an adaptive sampling rate, extracts features in parallel using multi-scale convolution kernels, and then optimizes them through residual attention masks to obtain fatigue residual optimized features, and sends the data to the metal material fatigue feature evolution analysis module.

[0039] The metal material fatigue feature evolution analysis module receives data sent by the metal material fatigue feature extraction module, generates a multi-scale position coding matrix according to the adaptive adjustment coefficient, sets fatigue prior fusion weights and position coding fusion weights to obtain a prior-position collaborative fusion feature matrix, extracts a specific prior vector according to the prior hybrid weights to obtain multi-head collaborative attention features, obtains the final features through the core feedforward branch and adaptive gating branch, generates prediction labels, and sends the data to the intelligent fatigue evolution analysis module.

[0040] The intelligent fatigue evolution analysis module receives data sent by the metal material fatigue characteristic evolution analysis module, collects real-time fatigue time series data of metal materials, obtains the corresponding fatigue damage stage based on the predicted label, and completes the fatigue evolution analysis of metal materials.

[0041] Example 2, see Figure 1 This embodiment is based on the above embodiment. In the metal material data acquisition module, historical metal material fatigue time series data is collected and fatigue damage stages are manually marked.

[0042] The fatigue time series data of the metallic materials includes time steps, load data, response data, environmental data, and basic property data of the metallic materials;

[0043] The load data includes load type, load amplitude, maximum load, minimum load, average load, load ratio, load waveform, and load frequency;

[0044] The response data includes axial / bending displacement, elastic strain amplitude, stiffness change rate, plastic strain amplitude, load-displacement hysteresis loop area change rate, and vibration natural frequency.

[0045] The environmental data includes ambient temperature, ambient humidity, type of corrosive medium, and concentration of corrosive medium;

[0046] The basic property data of the metallic materials includes the material grade and heat treatment status;

[0047] The fatigue damage stages include no damage stage, slight damage stage, moderate damage stage, severe damage stage and failure stage, and the fatigue damage stages are used as data labels.

[0048] The criteria for manually marking fatigue damage stages are as follows:

[0049] When a metallic material simultaneously meets the following conditions, such as stiffness change rate less than 2%, plastic strain amplitude less than 1.5%, and load-displacement hysteresis loop area change rate less than 3%, it is marked as a no-damage stage.

[0050] When a metallic material meets any one of the following conditions: stiffness change rate greater than or equal to 2% and less than 5%, plastic strain amplitude greater than or equal to 1.5% and less than 3%, load-displacement hysteresis loop area change rate greater than or equal to 3% and less than 8%, and all parameters have not reached the threshold of moderate damage stage, it is marked as slightly damaged stage.

[0051] When a metallic material meets any one of the following conditions: stiffness change rate greater than or equal to 5% and less than 15%, plastic strain amplitude greater than or equal to 3% and less than 10%, load-displacement hysteresis loop area change rate greater than or equal to 8% and less than 20%, and all parameters have not reached the threshold of severe damage stage, it is marked as moderate damage stage.

[0052] When a metallic material meets any one of the following conditions: stiffness change rate greater than or equal to 15% and less than 30%, plastic strain amplitude greater than or equal to 10% and less than 15%, load-displacement hysteresis loop area change rate greater than or equal to 20% and less than 50%, and all parameters have not reached the failure stage threshold, it is marked as a severely damaged stage.

[0053] When a metallic material meets any one of the following conditions, it is marked as being in the failure stage: stiffness change rate is greater than or equal to 30%, plastic strain amplitude is greater than or equal to 15%, or load-displacement hysteresis loop area change rate is greater than or equal to 50%.

[0054] Example 3, see Figure 1 This embodiment is based on the above embodiment. In the metal material data preprocessing module, the collected metal material fatigue time series data is cleaned, normalized and encoded.

[0055] The data cleaning process involves removing erroneous, missing, and outlier values ​​from the data.

[0056] The data normalization mentioned above uses the max-min normalization method to unify the data to the same scale;

[0057] The data encoding uses one-hot encoding to convert categorical data and data labels into numerical data.

[0058] Example 4, see Figure 1 and Figure 2This embodiment, based on the above embodiment, includes an environment-time dual-dimensional calibration unit, a fatigue prior embedding unit, and a stage-aware convolutional fusion unit in the metal material fatigue feature extraction module; specifically, it includes the following:

[0059] A dual-dimensional calibration unit based on environment and time is used. The fatigue evolution of metallic materials is influenced by both environmental and temporal factors. Traditional feature extraction often ignores the dynamic changes in environmental factors and fails to differentiate the importance of different time steps, resulting in extracted features that cannot accurately reflect the fatigue evolution pattern under the coupled effect of environment and time. For preprocessed metallic material fatigue time series data, environmental impact factors are calculated based on environmental data to obtain environmental calibration features. These features are then subjected to global average pooling, and time step attention weights are generated through a fully connected layer. The environmental calibration features are then multiplied element-wise with these time step attention weights to obtain dual-dimensional calibration features. This fully considers the coupling effect of environment and time in the fatigue evolution of metallic materials, enabling the extracted features to dynamically reflect fatigue damage changes under different environmental conditions and time steps, better reflecting the complex characteristics of actual fatigue evolution. Furthermore, the introduction of time step attention weights automatically identifies time step data that is more important for judging fatigue damage stages, avoiding feature redundancy or the submersion of key information caused by equal weights for all time steps. The formulas used are as follows:

[0060] ;

[0061] ;

[0062] ;

[0063] ;

[0064] In the formula, and X env These are environmental impact factors and environmental calibration characteristics, X. pre These are characteristics of fatigue time series data of metallic materials after data preprocessing, s time It is the attention weight of the time step, X calib It is a two-dimensional calibration feature, where T, H, and C are the ambient temperature, ambient humidity, and corrosive medium concentration collected at the current time step, respectively. ref H ref and C ref These are the reference temperature, reference humidity, and reference corrosive medium concentration, respectively. T ref =0.2, H ref =0.3, C ref =0.1, T max H max and C maxThese represent the maximum values ​​of ambient temperature, ambient humidity, and corrosive medium concentration, respectively, across all data points. W1 and W2 are the trainable weight matrices of the fully connected layer, and b1 and b2 are the trainable bias terms of the fully connected layer. It is a global average pooling operation. It is a linear rectified function. It is the Sigmoid activation function. It is the element-wise multiplication operator;

[0065] Fatigue Prior Embedding Unit: Traditional feature extraction relies solely on the statistical or structural features of the data itself, failing to incorporate prior knowledge of fatigue in metallic materials. This results in a weak correlation between the extracted features and the physical mechanisms of fatigue evolution, making it difficult for the model to understand the fatigue damage significance behind the features, thus affecting the interpretability and accuracy of the analysis. By mapping the two-dimensional calibration features to a high-dimensional vector space through layer normalization, embedding layer features are obtained. Fatigue prior vectors are constructed based on load and response data. The embedding layer features are then fused with the fatigue prior vectors to obtain fatigue prior embedded features. By integrating key prior parameters from the fatigue domain, the extracted features possess a clear physical meaning of fatigue damage. The high-dimensional mapping enhances the expressive power of the features, while the integration of prior knowledge guides the features towards optimization that reflects the essence of fatigue damage, making the features more discriminative. The formulas used are as follows:

[0066] ;

[0067] ;

[0068] ;

[0069] Among them, X embed These are embedding layer features. It is the layer normalization function, W emb and b emb These are the trainable weight matrix and bias term of the embedding layer, respectively. It is the fatigue prior value at the t-th time step and the f-th dimension, where t and f are the time step index and the dimension index, respectively, and d model It is the largest dimension, R t and These are the load ratio and stiffness change rate at time step t, respectively, R min and These are the minimum values ​​of the load ratio and the rate of change of stiffness among all the data, R. max and These are the maximum values ​​of the load ratio and the rate of change of stiffness among all the data, respectively. prior It is the fatigue prior embedding feature, where P is the fatigue prior vector;

[0070] Stage-aware convolutional fusion unit: The fatigue characteristics of metallic materials vary significantly across different fatigue damage stages. Traditional convolutional extraction often uses a fixed sampling rate and a single-scale convolutional kernel, which cannot adapt to the feature changes at different damage stages. This results in excessive redundancy in feature extraction for stages with gradual changes and insufficient feature extraction for stages with drastic changes. Furthermore, the multi-scale feature fusion effect is poor, failing to comprehensively capture fatigue damage information at different scales. Adaptive sampling rates are set for each fatigue damage stage: 0.25, 0.5, 1, 2, and 4 for no damage, slight damage, moderate damage, severe damage, and failure stages, respectively. Based on the adaptive sampling rate, fatigue priors are... Embedded features are sampled, and multi-scale convolutional kernels of sizes 1×3, 1×5, and 1×7 are used to extract fatigue prior embedded features in parallel, resulting in multi-scale convolutional fusion features. These multi-scale convolutional fusion features are then optimized using residual attention masks to obtain fatigue residual optimized features. An adaptive sampling rate is set to consider the feature change characteristics at different fatigue damage stages, avoiding feature redundancy or loss of key information caused by a fixed sampling rate. This ensures that feature extraction accurately adapts to the characteristics of fatigue evolution at each stage, improving the feature's ability to distinguish between different damage stages. Parallel extraction using multi-scale convolutional kernels comprehensively captures feature information at different scales during fatigue evolution, enhancing the representativeness and discriminative power of the features. The formulas used are as follows:

[0071] ;

[0072] ;

[0073] ;

[0074] Among them, X sample It is the fatigue prior embedding feature after sampling, r adapt It is an adaptive sampling rate. This is an average pooling operation; kernel size is the pooling kernel size. This is a bilinear interpolation upsampling operation, where scale is the upsampling scaling factor, stride is the stride, and F... conv It is a multi-scale convolutional fusion feature. , and These are one-dimensional convolution operations of sizes 1×3, 1×5, and 1×7, respectively. It is the residual attention mask branch, X multi These are fatigue residual optimization features; S1, S2, and S3 are convolution operation sequences, each containing a 1×1 convolution and a ReLU activation function. It rounds up.

[0075] By performing the above operations, this solution addresses the problems in existing fatigue evolution analysis systems for metallic materials, such as the tendency to overlook the dynamic coupling effect of environment and time, the lack of incorporation of prior knowledge of fatigue evolution, and the inability to accurately extract features of metallic materials at different damage stages, leading to low accuracy in judging the fatigue state of metallic materials. This solution generates environmental calibration features based on environmental influence factors and combines them with time-step attention weights to obtain dual-dimensional calibration features. This allows the features to dynamically match the environmental effects under actual working conditions, suppressing interference from secondary time steps and more accurately reflecting the fatigue evolution law under environment-time coupling. A fatigue prior vector is constructed to obtain fatigue prior embedded features, directly embedding prior knowledge of metallic materials into the features and enhancing the correlation between features and fatigue state. The fatigue prior embedded features are sampled according to an adaptive sampling rate, and features are extracted in parallel using multi-scale convolutional kernels. Then, residual attention masks are used for optimization to obtain fatigue residual optimized features, avoiding feature redundancy or loss of key information. This allows the extracted features to more accurately characterize the fatigue damage state, effectively improving the reliability and accuracy of fatigue evolution analysis of metallic materials.

[0076] Example 5, see Figure 1 and Figure 3 This embodiment, based on the above embodiment, includes a multi-scale position encoding unit, a priori-position collaborative fusion unit, a fatigue perception multi-head collaborative attention unit, a gated feedforward enhancement unit, and an analysis unit in the fatigue feature evolution analysis module for metallic materials; specifically, it includes the following:

[0077] Multi-scale positional coding unit: Fatigue time series data for metallic materials exhibits temporal sequence characteristics, and the temporal relationships between different time steps are crucial for fatigue evolution analysis. Traditional positional coding often employs a fixed method, failing to consider the varying importance of temporal relationships across different fatigue damage stages. This results in the encoded temporal information not accurately reflecting the temporal characteristics of fatigue evolution at different stages, impacting the model's ability to capture temporal dependencies. This study adopts a sine-cosine coding method to capture the fundamental temporal relationships between time steps in the time series, generating a basic positional coding matrix. Adaptive adjustment coefficients are set for each fatigue damage stage: no damage, slight damage, moderate damage, severe damage, and failure stage. The adaptive adjustment coefficients for each segment are 0.2, 0.4, 0.6, 0.8, and 1, respectively, yielding the stage-adjusted position encoding matrix. Combining the basic position encoding matrix and the stage-adjusted position encoding matrix results in a multi-scale position encoding matrix. By combining basic temporal encoding and stage adaptive adjustment, the position encoding accurately reflects the chronological order of time steps and adjusts the weights of temporal information according to the characteristics of different fatigue damage stages. This better reflects the actual situation where the temporal dependence intensity varies at different stages of fatigue evolution in metallic materials. The multi-scale position encoding matrix provides the model with temporal information at different levels, meeting the model's requirement to capture the temporal relationships of fatigue evolution at different stages. The formulas used are as follows:

[0078] ;

[0079] ;

[0080] ;

[0081] In the formula, PE base PE adj and PE multi These are the base location encoding matrix, the stage-adjusted location encoding matrix, and the multi-scale location encoding matrix, respectively, where i is the location encoding dimension index, and t... curr This is the index of the current time step of interest. This is the stage adaptive adjustment coefficient, where L is the total number of time steps. and These are the base position encoding values ​​for time step t in even-numbered dimensions 2i and odd-numbered dimensions 2i+1, respectively. It is the stage-adjusted position encoding value at time step t in the i-th dimension;

[0082] Prior-location co-fusion unit; In fatigue feature evolution analysis, fatigue residual optimization features, fatigue prior vectors, and multi-scale location encoding matrices represent data features, domain prior knowledge, and temporal information, respectively. Traditional fusion methods often use fixed weights or simple concatenation, failing to consider the impact of key parameters in the fatigue evolution of metallic materials on the importance of different information sources. This results in the fused features being unable to dynamically adapt to the information needs under different damage states, affecting the comprehensive expressive ability of the features. Based on the stiffness change rate, fatigue prior fusion weights and location encoding fusion weights are set separately. The fatigue residual optimization features, fatigue prior vectors, and multi-scale location encoding matrices are then weighted and fused to obtain the prior-location co-fusion feature matrix. The fusion weights are dynamically adjusted based on the stiffness change rate, enabling the fused features to adaptively highlight more important information sources according to the degree of fatigue damage in metallic materials, improving the feature's adaptability to different damage states. The formulas used are as follows:

[0083] ;

[0084] ;

[0085] ;

[0086] in, and These are the fatigue prior fusion weights and position encoding fusion weights at time step t, respectively, β prior and β pos These are the fatigue prior fusion weight vector and the position encoding fusion weight vector, respectively, X prep It is a priori-location co-fusion feature matrix;

[0087] Fatigue perception multi-head collaborative attention unit: Traditional multi-head attention mechanisms, when processing fatigue features of metallic materials, use the same information source and weight allocation method for all attention heads, failing to consider the different needs of different attention heads to focus on different aspects of fatigue evolution, and not incorporating prior fatigue knowledge. This results in attention weights failing to accurately capture key information about fatigue damage, and the outputs of different attention heads lacking synergy, affecting the overall feature extraction effect. Based on the prior mixed weights of the attention heads, a unique prior vector for each attention head is extracted from the fatigue prior vector. Then, combined with the prior-position collaborative fusion feature matrix, a query matrix, key matrix, and value matrix are generated for each attention head. The fatigue perception attention weight of each attention head is calculated, and the outputs of all attention heads are concatenated and linearly mapped to obtain multi-head collaborative attention features. Each attention head has a unique prior vector, enabling it to focus on different key dimensions of fatigue evolution, avoiding the one-size-fits-all information processing method of traditional multi-head attention, and more comprehensively capturing multi-dimensional key information in the fatigue evolution of metallic materials, thus improving the discriminativeness of features. The formulas used are as follows:

[0088] ;

[0089] ;

[0090] ; ; ;

[0091] ;

[0092] ;

[0093] Where A is the total number of attention heads. It is the prior mixture weight of the a-th attention head, where a is the attention head index, and d is the weight of the attention head. h It is the feature dimension of a single attention head. , It is the specific prior value of the a-th attention head at the t-th time step and the f-th feature dimension. It is the specific prior vector of the a-th attention head. , and These are the query matrix, key matrix, and value matrix of the a-th attention head. , and It is the trainable weight matrix of the a-th attention head. yes transpose, It is a normalization function. , and These are the fatigue perception attention weights for the 1st, ath, and Ath attention heads, V1 and V2 respectively. A These are the value matrices of the 1st and Ath attention heads, respectively, W mh It is a multi-head attention fusion weight matrix. It is a splicing operation; MultiHead is a multi-head collaborative attention feature.

[0094] Gated feedforward enhancement unit: Multi-head collaborative attention features may contain redundant information or noise. Traditional feedforward networks process all feature dimensions equally, failing to filter effective information or suppress invalid information, leading to a decline in feature quality and affecting the accuracy of subsequent fatigue damage stage prediction. Furthermore, the impact of internal statistical characteristics on feature selection is not considered, further reducing the accuracy of feature processing. This unit extracts core feedforward features from the multi-head collaborative attention features through a core feedforward branch, generates gating coefficients based on an adaptive gating branch, and multiplies the core feedforward features element-wise with the gating coefficients to obtain the gating selection features. The residuals of the multi-head collaborative attention features and the gating selection features are concatenated and layer normalized to obtain the final features. The introduction of the gating coefficients adaptively filters effective information in the core feedforward features, eliminating redundancy and noise, avoiding interference from invalid information on fatigue damage stage prediction, and incorporating the standard deviation of the feature dimension direction to improve the rationality and accuracy of feature selection, providing high-quality feature support for the final fatigue damage stage prediction. The formulas used are as follows:

[0095] ;

[0096] ;

[0097] ;

[0098] Among them, FFN main and FFN gate These are the core feedforward features and the gating adjustment coefficient, and It is the trainable weight matrix of the core feedforward branch. and It is the trainable bias term of the core feedforward branch. It is the GELU activation function. and It is the trainable weight matrix of the adaptive gating branch. and It is a trainable bias term for the adaptive gating branch. X is the standard deviation of the feature dimension of multi-head collaborative attention. final It is the final characteristic;

[0099] Analysis unit: Based on fully connected layers and the Softmax function, the final features are mapped to the probability distribution of five fatigue damage stages, and the fatigue damage stage with the highest probability is selected as the predicted label output.

[0100] By performing the above operations, this solution addresses the problems in existing fatigue evolution analysis systems for metallic materials, such as failing to consider the differentiated impact of fatigue damage stages on temporal dependence intensity and having rigid multi-source information fusion methods, making it difficult to dynamically adapt to the needs of evolution law analysis under different damage states. This solution generates a multi-scale position encoding matrix based on adaptive adjustment coefficients, accurately matching the differences in temporal dependence intensity at each stage, providing a clearer understanding of the dynamic temporal process of fatigue evolution, and improving sensitivity to changes in damage stages. It sets fatigue prior fusion weights and position encoding fusion weights to obtain a prior-position co-fusion feature matrix, making the fused features more closely match the evolution laws of different damage states and enhancing the feature's representation ability for fatigue evolution. Based on the prior mixed weights, it extracts a dedicated prior vector to obtain multi-head collaborative attention features, accurately capturing the core features under each damage state and improving the ability of attention to focus on key fatigue damage information. Finally, it obtains the final features through core feedforward branches and adaptive gating branches, generating predictive labels to provide accurate and reliable stage judgment criteria for fatigue evolution analysis of metallic materials, effectively adapting to the needs of evolution law analysis under different damage states.

[0101] Example 6, see Figure 1 This embodiment is based on the above embodiment. In the intelligent fatigue evolution analysis module, real-time fatigue time series data of metal materials are collected and sequentially input into the metal material data preprocessing module, the metal material fatigue feature extraction module, and the metal material fatigue feature evolution analysis module. Based on the output prediction labels, the fatigue damage stage corresponding to the real-time fatigue time series data of metal materials is obtained, and the fatigue evolution analysis of metal materials is completed.

[0102] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0103] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. An intelligent fatigue evolution analysis system for metallic materials, characterized in that: It includes a metal material data acquisition module, a metal material data preprocessing module, a metal material fatigue feature extraction module, a metal material fatigue feature evolution analysis module, and an intelligent fatigue evolution analysis module; The metal material data acquisition module collects historical metal material fatigue time series data and manually marks the fatigue damage stages. The metal material data preprocessing module performs data cleaning, data normalization, and data encoding. The metal material fatigue feature extraction module generates environmental calibration features based on environmental impact factors, obtains two-dimensional calibration features by combining time step attention weights, constructs fatigue prior vectors, obtains fatigue prior embedding features, samples fatigue prior embedding features according to adaptive sampling rate, extracts features in parallel using multi-scale convolution kernels, and then optimizes them through residual attention mask to obtain fatigue residual optimized features. The metal material fatigue feature evolution analysis module generates a multi-scale position coding matrix based on adaptive adjustment coefficients, sets fatigue prior fusion weights and position coding fusion weights to obtain a prior-position collaborative fusion feature matrix, extracts exclusive prior vectors based on prior mixed weights to obtain multi-head collaborative attention features, obtains the final features through core feedforward branches and adaptive gating branches, and generates prediction labels. The intelligent fatigue evolution analysis module collects real-time fatigue time series data of metallic materials, obtains the corresponding fatigue damage stage based on the predicted label, and completes the fatigue evolution analysis of metallic materials.

2. The intelligent fatigue evolution analysis system for metallic materials according to claim 1, characterized in that: The metal material fatigue feature extraction module includes an environment-time dual-dimensional calibration unit, a fatigue prior embedding unit, and a stage-aware convolutional fusion unit; specifically, it includes the following: Environment-time dual-dimensional calibration unit: For the fatigue time series data of metal materials after data preprocessing, the environmental impact factor is calculated based on the environmental data to obtain the environmental calibration features. The environmental calibration features are then subjected to global average pooling, and time step attention weights are generated through a fully connected layer. The environmental calibration features are then multiplied element-wise with the time step attention weights to obtain the dual-dimensional calibration features. Fatigue prior embedding unit: The two-dimensional calibration features are mapped to a high-dimensional vector space through layer normalization to obtain the embedding layer features. The fatigue prior vector is constructed based on the load data and response data. The embedding layer features and the fatigue prior vector are fused to obtain the fatigue prior embedding features. Stage-aware convolutional fusion unit.

3. The intelligent fatigue evolution analysis system for metallic materials according to claim 2, characterized in that: The stage-aware convolutional fusion unit sets an adaptive sampling rate for each fatigue damage stage, samples the fatigue prior embedding features according to the adaptive sampling rate, and uses multi-scale convolutional kernels to extract the sampled fatigue prior embedding features in parallel to obtain multi-scale convolutional fusion features. Finally, it optimizes the multi-scale convolutional fusion features through residual attention masking to obtain fatigue residual optimized features.

4. The intelligent fatigue evolution analysis system for metallic materials according to claim 3, characterized in that: The fatigue characteristic evolution analysis module for metallic materials includes a multi-scale position encoding unit, a priori-position collaborative fusion unit, a fatigue perception multi-head collaborative attention unit, a gated feedforward enhancement unit, and an analysis unit; specifically, it includes the following: Multi-scale position coding unit; Generate a basic position coding matrix, set adaptive adjustment coefficients for each fatigue damage stage to obtain a stage-adjusted position coding matrix, and combine the basic position coding matrix and the stage-adjusted position coding matrix to obtain a multi-scale position coding matrix; Prior-location collaborative fusion unit: Fatigue prior fusion weight and location encoding fusion weight are set respectively, and fatigue residual optimization features, fatigue prior vector and multi-scale location encoding matrix are weighted and fused to obtain prior-location collaborative fusion feature matrix; Fatigue perception multi-head collaborative attention unit; Gated feedforward enhancement unit; Analysis unit: Based on fully connected layers and the Softmax function, the final features are mapped to the probability distribution of five fatigue damage stages, and the fatigue damage stage with the highest probability is selected as the predicted label output.

5. The intelligent fatigue evolution analysis system for metallic materials according to claim 4, characterized in that: The fatigue-perceiving multi-head collaborative attention unit extracts the exclusive prior vector of each attention head from the fatigue prior vector based on the prior mixing weights of the attention heads, and then combines it with the prior-position collaborative fusion feature matrix to generate the query matrix, key matrix and value matrix of each attention head. It calculates the fatigue-perceiving attention weight of each attention head, and concatenates and linearly maps the outputs of all attention heads to obtain the multi-head collaborative attention features.

6. The intelligent fatigue evolution analysis system for metallic materials according to claim 5, characterized in that: The gated feedforward enhancement unit extracts core feedforward features from the multi-head collaborative attention features through the core feedforward branch, generates gated adjustment coefficients based on the adaptive gated branch, multiplies the core feedforward features and gated adjustment coefficients element by element to obtain gated screening features, and connects the residuals of the multi-head collaborative attention features and gated screening features and performs layer normalization to obtain the final features.

7. The intelligent fatigue evolution analysis system for metallic materials according to claim 6, characterized in that: The metal material data acquisition module collects historical metal material fatigue time series data and manually marks the fatigue damage stages. The fatigue time series data of the metallic materials includes time steps, load data, response data, environmental data, and basic property data of the metallic materials; The fatigue damage stages include no damage stage, slight damage stage, moderate damage stage, severe damage stage, and failure stage, and the fatigue damage stages are used as data labels.

8. The intelligent fatigue evolution analysis system for metallic materials according to claim 7, characterized in that: The intelligent fatigue evolution analysis module collects real-time fatigue time series data of metallic materials and inputs it sequentially into the metallic material data preprocessing module, the metallic material fatigue feature extraction module, and the metallic material fatigue feature evolution analysis module. Based on the output prediction labels, the fatigue damage stage corresponding to the real-time fatigue time series data of metallic materials is obtained, thus completing the fatigue evolution analysis of metallic materials.

Images