Material mechanical property testing method based on displacement and load response of in-situ indentation
Through deep learning and integrated learning technology, combining load, pressure depth and displacement cloud maps, a mapping model of material mechanical properties is established, which solves the problem of high demand for traditional in situ indentation test data, and achieves efficient acquisition of material mechanical properties and improves model accuracy .
Patent Information
- Application Number
- CN202411004490.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-25
- Publication Date
- 2024-11-22
AI Technical Summary
Traditional in situ indentation testing methods require a large amount of expensive experimental data, and model training relies on a large amount of displacement cloud map data, resulting in high data requirements and it is difficult to efficiently obtain material mechanical properties.
Combining deep learning models, data augmentation, transfer learning, integrated learning and reinforcement learning, by establishing a mapping from load, pressure depth and displacement cloud map to material mechanical properties, the model's dependence on experimental data is reduced, and the finite element simulation data is used for pre-pre- Training and fine-tuning, integrating multiple sub-models for improved accuracy.
It realizes the acquisition of material mechanical properties while acquiring in-situ evolution information, reduces the demand for experimental data of the same precision model and improves the overall accuracy and efficiency of the model.
Smart Images

Figure CN119004889A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of material mechanical properties testing, and in particular to a material mechanical properties testing method based on displacement response and load response of in-situ indentation. The instrument of the present invention can be applied to the research and development of metal materials in the fields of aerospace, military equipment, electromagnetic emission, etc. Background Art
[0002] Artificial intelligence has boosted the development of the fourth industrial revolution, empowered human production and life, and promoted the emergence of many new technologies, such as intelligent assisted driving, mechanical system fault diagnosis, industrial defect detection, reverse design of metamaterial structures, and intelligent robots. The core issue of artificial intelligence is how to accurately realize the mapping from variable to variable. Compared with traditional numerical fitting methods (such as least squares method), its advantage lies in the ability to establish the correlation between variables such as vectors, numbers, images, and signals that represent certain information. At the same time, ensemble learning can further improve the regression accuracy of the model and integrate multiple weak learners into a strong learner. In addition, the reinforcement learning framework promotes the optimization of strategies, so that the optimization is not limited to the optimization of mapping accuracy.
[0003] The traditional indenter uses a conical indenter to press a pit on the surface of the workpiece, while the in-situ indenter (see the patented micro-nano-level indentation test device and test method for in-situ characterization of the indentation area) uses the small side of the plate sample to press, while the large side of the plate sample is used for in-situ observation. The so-called in-situ test is that the sample integrates imaging and other detection modules during the loading process to observe the surface evolution information of the material at all times. This evolution information includes EBSD information, surface SEM evolution morphology information, and DIC strain information. Compared with traditional mechanical properties testing instruments, in-situ testing can obtain more information about the evolution of materials at all times, so it is more conducive to exploring the failure mechanism and evolution deformation of materials under displacement loading conditions. This in-situ information can not only provide material evolution information, but also contain information such as material mechanical properties. For example, the grain curvature map and grain boundary map measured by EBSD contain information about the mechanical properties of the material (Acta Materialia, DOI: 10.1016 / j.actamat.2023.118954). At the same time, the indentation curve also contains information such as the mechanical properties of the material (PANS, DOI: 10.1073 / pnas.1922210117). Artificial intelligence can associate image information and vectors with vectors, which makes it possible to map the displacement cloud map and load and mechanical properties measured from DIC. This allows the in-situ indentation test to not only obtain in-situ information but also the mechanical properties of the material. However, training artificial intelligence models usually requires a large amount of data, which increases the difficulty of applying artificial intelligence to in-situ indenters to obtain material mechanical properties. Because a large number of in-situ indentation tests are expensive. At the same time, the DIC displacement cloud map may have a similar distribution to the displacement cloud map obtained by finite element static simulation. Therefore, the displacement cloud map obtained by simulation can be used in the fine-tuning stage of the model, and the displacement cloud map obtained by DIC can be used for fine-tuning the model, thereby reducing the demand for experimental data for training. At the same time, if a set of experimental data only provides data for one picture, it is also expensive. If multiple groups of compression depth, load and displacement cloud maps of a specimen are divided into multiple groups of data, the amount of available data will increase a lot, while the complexity of the model will only increase a little. The ensemble learning algorithm that integrates multiple trained models will further improve the accuracy of the model. The weight corresponding to each sub-model in the ensemble learning can be optimized through reinforcement learning to accelerate the establishment of a model with better performance.
[0004] In summary, by combining artificial intelligence with in-situ indentation testing, the in-situ indenter can obtain the mechanical properties of the material while obtaining the in-situ evolution information. In terms of implementation methods, the combined use of deep learning models, deep learning-based image data augmentation models, transfer learning, ensemble learning models, and reinforcement learning models can reduce the model's dependence on data, making it possible to obtain the mechanical properties of the material based on the in-situ information of the in-situ indenter. This method can be used in many fields such as metal material research and development and testing. Summary of the invention
[0005] The purpose of the present invention is to provide a material mechanical property testing method based on displacement response and load response of in-situ indentation, by establishing a mapping from load information and displacement cloud map information to material mechanical properties to help the in-situ indentation instrument obtain the indentation information on the side of the sample while also obtaining the mechanical properties of the material. At the same time, the method combines the use of deep learning models, data augmentation models based on deep learning models, transfer learning ideas, integrated learning models, and reinforcement learning models to reduce the demand for experimental data required for models with the same accuracy.
[0006] To achieve the above object, the present invention adopts the following technical solutions:
[0007] 1) Select several typical cheap copper alloys and aluminum alloys, cut them into thin slices and polish them, use the in-situ indenter to obtain the load at different indentation depths and the strain cloud map near the in-situ indentation area, and establish an experimental data set. Each set of data in the data set includes the indentation depth, load, modulus, yield stress, power hardening index, and the displacement cloud map of the indentation side under the corresponding conditions.
[0008] 2) Perform static simulation on the working conditions corresponding to the in-situ indentation test, use the power hardening law as the constitutive model, select several groups of modulus (10-210Gpa), yield stress (30-3000Mpa) and power hardening coefficient (0.1-0.5) to generate the plastic strain-stress curve of the plastic stage of the material through programming methods, and use several groups of modulus, yield stress and plastic stress-strain curve as material properties to perform static simulation of the corresponding working conditions. Establish a simulation data set, which includes the corresponding indentation depth, load, modulus, yield stress, power hardening index in the simulation calculation, and the displacement cloud map of the indentation side under the corresponding conditions.
[0009] 3) Data augmentation is performed on the experimental data set and the simulation data set. The modified diffusion model is used for data augmentation. The indentation depth, load, modulus, yield stress, and power hardening index are embedded into the model through the fully connected projection method (like position encoding information), and the mapping from indentation depth, load, modulus, yield stress, and power hardening index to the corresponding displacement cloud map is realized, thereby realizing data augmentation. The corresponding augmented data sets are the experimental augmented data set and the simulation augmented data set.
[0010] 4) Then, a deep learning model from indentation depth, load and corresponding displacement cloud map to mechanical properties is established. The model is first trained on the simulation augmented dataset, and then fine-tuned on the simulation dataset and the experimental augmented dataset respectively. Finally, the model is finally fine-tuned on the experimental dataset to obtain the final sub-model.
[0011] 5) Repeat step 4 n times to obtain n different sub-models. For each model, m are selected, and the m×n sub-models are integrated into an integrated learning model through an ensemble learning algorithm.
[0012] 6) Assign a weight to each sub-model in the integrated learning model, and optimize the weight corresponding to each sub-model through the reinforcement learning algorithm DQN to obtain the final model. During the reinforcement learning training process, the model is trained using the training set of the experimental data set, and each sub-model randomly selects data. The input data of the sub-model of an integrated learning model can come from different groups of experiments.
[0013] 7) Use new experimental data to test the integrated model. The test data used by the sub-models corresponding to the model needs to come from the same set of experimental data.
[0014] 8) Determine the model accuracy of the proposed modeling method, conduct ablation experiments, and compare the effectiveness of each part of the proposed modeling method. And compare the accuracy of the model corresponding to the proposed modeling method with the model based only on the mapping of a single image to the mechanical properties of the material.
[0015] The test method is suitable for in-situ indentation testing of metal materials, the indentation depth ranges from 0 to 200 μm, the use range of the test material is limited to metal solid materials, the modulus range of the material is 10 to 300 Gpa, the yield strength range of the material is 30 to 3000 Mpa, and the power hardening coefficient range of the material is 0.1 to 0.5.
[0016] The beneficial effects of the present invention are:
[0017] The present invention realizes the mapping from load, indentation depth and sample side displacement cloud map to material mechanical properties (modulus, yield stress and power hardening coefficient) through this method, so that the in-situ indentation test can obtain the mechanical properties of the material while obtaining the in-situ evolution information, and expands the function of the in-situ indenter. The sub-model used is multimodal data, the load comes from the force sensor, the indentation depth comes from the displacement sensor, and the displacement cloud map comes from the digital speckle test system. The load can be considered as the independent variable, and the indentation depth and displacement cloud map are dependent variables determined by the load and the mechanical properties of the material. The model reversely obtains the mechanical properties of the independent variable material through the dependent variable displacement cloud map and the indentation depth and the independent variable load.
[0018] The in-situ indenter can obtain multiple sets of displacement cloud maps in one set of experimental data, because the digital speckle test system needs to continuously collect image data to obtain the displacement cloud map. At the same time, in order to obtain higher-precision indentation data, the indenter needs to stop collecting each picture. Therefore, even if only one picture is taken to predict the mechanical properties of the material, all the pictures taken before this picture need to be obtained through experiments. Therefore, it is very wasteful to use only one picture to predict the mechanical properties of the material. Therefore, it is preferred to use a mapping model from load, indentation depth and specimen side displacement cloud map to material mechanical properties (modulus, yield stress and power hardening coefficient) to make full use of experimental data.
[0019] The present invention simultaneously uses finite element simulation, data augmentation model based on deep learning, transfer learning, integrated learning and reinforcement learning methods to reduce the dependence of the model with the same accuracy on experimental data. The displacement cloud map obtained by finite element analysis under the same working conditions has a similar distribution to the displacement cloud map obtained by the digital speckle system of the in-situ indenter, so the displacement cloud map obtained by finite element analysis can be used for pre-training, and the model can be fine-tuned using experimental data to reduce the dependence of the model with the same accuracy on experimental data. All the above variables include load, compression depth, modulus, yield stress, power hardening index and displacement cloud map of the sample. In addition to the displacement cloud map, other variables can be obtained through reasonable permutations and combinations to obtain a new set of data. Therefore, for this case, the key to data augmentation is to generate a displacement cloud map, so a data augmentation model based on mapping load, compression depth, modulus, yield stress and power hardening index to displacement cloud map is established. The original data and augmented data also have similar distributions. Therefore, the augmented data group can be used for pre-training, and the original data is used for fine-tuning of the model, thereby reducing the number of original data required to train the model with the same accuracy.
[0020] In addition, ensemble learning can integrate multiple weak learners into a strong learner. This method also uses ensemble learning to integrate the above trained sub-models to improve the overall accuracy of the model. Finally, in order to further improve the accuracy of the overall model, each sub-model is given a weight, and the reinforcement learning algorithm is used to optimize the weight to accelerate the performance of the overall model. BRIEF DESCRIPTION OF THE DRAWINGS
[0021] The drawings described herein are used to provide further understanding of the present invention and constitute a part of this application. The illustrative examples of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.
[0022] Figure 1 Schematic diagram of the relationship between the in-situ indenter measurement method and various variables;
[0023] Figure 2 Schematic diagram of data augmentation model based on deep learning;
[0024] Figure 3 Schematic diagram of sub-model architecture, input and output;
[0025] Figure 4 Schematic diagram of the integrated learning model;
[0026] Figure 5 Schematic diagram of the weight optimization method for the sub-model;
[0027] Figure 6 Modeling flow chart for the whole set;
[0028] Figure 1 The measurement diagram of the in-situ indenter is shown in the figure. The indentation depth h is obtained by the displacement sensor, the load F is obtained by the force sensor, and the displacement cloud map I is measured by the digital speckle measurement system. The figure also shows the range of the displacement cloud map I. The figure also shows the relationship between the variables, mechanical properties, load-determined displacement and displacement cloud map.
[0029] Figure 2 The data augmentation model f is shown z The input of the model is the mechanical properties (modulus E, yield stress σ y and power hardening coefficient n), load F and compression depth h, and the output is the displacement cloud map I under the corresponding conditions. s And the experimental dataset D e To train the model, and generate the simulated data augmentation dataset D through the trained model sz And experimental data augmented dataset D ez .
[0030] Figure 3 The ensemble learning model F is showne The input and output of the sub-model f are the load F and the compression depth h, the displacement cloud diagram I under the corresponding conditions, and the output of the sub-model f is the mechanical properties of the material (modulus E, yield stress σ y and the power refers to the hardening coefficient n).
[0031] Figure 4 For the ensemble learning model F e Schematic diagram of the ensemble learning sub-model, which includes n different sub-models, and each sub-model includes m models.
[0032] Figure 5 This is a schematic diagram of the optimization method for the sub-model weights. The model uses the DQN algorithm and the state is the integrated learning model F e The sub-model f u The action is to add 0.1 to the corresponding weight item, and the weight is normalized after each action is executed. The difference between the accuracy after the action is executed and the accuracy before the action is executed is the reward. Define the decay coefficient that changes over time to attenuate the size of the addition and subtraction of the weight item.
[0033] Figure 6 For the entire modeling process, we first conduct experiments and simulations, then augment the data based on the experimental and simulated data, use the simulated data and augmented data to pre-train the model, and then use the experimental data to fine-tune the model to obtain the fine-tuned sub-model f u , each sub-model selects m, forms an integrated learning algorithm, and finally uses the reinforcement learning method to train the sub-model f u The weights are optimized. DETAILED DESCRIPTION
[0034] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0035] The purpose of the present invention is to provide a material mechanical property testing method based on displacement response and load response of in-situ indentation, by establishing a mapping from load information and displacement cloud map information to material mechanical properties to help the in-situ indentation instrument to obtain the in-situ information of the side indentation of the sample while also obtaining the mechanical properties of the material. At the same time, the method combines the use of deep learning models, data augmentation models based on deep learning models, transfer learning ideas, integrated learning models, and reinforcement learning models to reduce the demand for experimental data required for models with the same accuracy.
[0036] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.
[0037] 1) Figure 1 The test diagram of the in-situ indenter and the relationship between the parameters involved are shown. Several typical cheap copper alloys (chromium-zirconium-copper alloy, copper-beryllium alloy, copper-chromium alloy and copper-tin alloy) and aluminum alloys (7075 aluminum alloy, 6061 aluminum alloy and 2024 aluminum alloy) were selected, cut into 10mm×10mm×1mm slices and polished. The load F at different indentation depths h and the strain cloud diagram I near the in-situ indentation area were obtained using the in-situ indenter to establish the experimental data set D e Dataset D e Each set of data includes indentation depth h, load F, modulus E, yield stress σ y , power hardening index n and displacement cloud diagram I of the indentation side under corresponding conditions. Each set of data corresponds to a set of digital speckle images I taken experimentally and its corresponding other parameters. Among them, the mechanical properties of the sample (load F, modulus E and yield stress σ y ) is obtained by consulting the Handbook of Mechanical Properties of Metal Materials.
[0038] 2) Static simulation of the working conditions corresponding to the in-situ indentation test was performed based on Abaqus2023, using the power hardening law as the constitutive model, and selecting the modulus E (10 Gpa, 60 Gpa, 110 Gpa, 160 Gpa and 210 Gpa), yield stress σ y The plastic strain-stress curve of the plastic stage of the material is generated by programming method, and several groups of modulus E, yield stress σ y The plastic stress-strain curve is used as the material property to perform static simulation of the corresponding working condition. s , simulation data set D s Each set of data in the simulation includes the corresponding indentation depth h, load F, modulus E, yield stress σ y , power hardening exponent n and the displacement cloud diagram of the indentation side under the corresponding conditions. Each set of data corresponds to a convergence step of a set of simulation results.
[0039] 3) Experimental dataset D e And the simulation data set D s For data augmentation, the modified diffusion model f is used z (Main home structure such as Figure 2As shown in the figure, the five-dimensional vector (compression depth h, load F, modulus E, yield stress σ y , the power refers to the hardening coefficient n) is embedded into the model through the fully connected projection method (like position encoding information), realizing the indentation depth h, load F, modulus E, yield stress σ y , the mapping of the power hardening index n to the corresponding displacement cloud map I, thereby realizing data augmentation. The corresponding augmented data set is the experimental data augmented data set D ez And the simulated data augmented dataset D sz , D ez The number of array groups is D e 5 times of D sz The number of data sets is D s 5 times of that.
[0040] 4) Then establish the indentation depth h, load F and the corresponding displacement cloud map I to the mechanical properties (load F, modulus E and yield stress σ y )’s deep learning model f, such as Figure 3 As shown. Model f first simulates the augmented dataset D sz Then, we train on the simulation dataset D s And experimental augmented dataset D ez Fine-tuning is performed on the experimental dataset D e The final fine-tuning is completed on the training set (the training set and the test set are divided into 5:2, the training set comes from 5 sets of experimental data in 7 groups, and the test set comes from the other 2 sets of experimental data in 7 groups) to obtain the model f u .
[0041] 5) Repeat step 4 6 times to obtain 6 models f1 with different structures u ~f n u . For each model, 6 are selected and the m×n models are integrated into one model F through the ensemble learning algorithm. e ,like Figure 4 The CNN encoders used are based on LeNet, AlexNet, VGGNet, GoogLeNet, ResNet and DenseNet. The alignment and fusion modules are fully connected layers. The integrated learning F e It is a weighted summation.
[0042] 6) Ensemble learning model F e Each model f u Assign a weight and optimize the weight corresponding to each sub-model through the reinforcement learning algorithm DQN. The definition and process of reinforcement learning training parameters are as follows Figure 5As shown, the state is the weight of the sub-model, the action is to adjust the weight of the sub-model, and the reward is the increased accuracy of the model after the weight adjustment compared to the model before the weight adjustment. The decay coefficient T in the figure is the negative exponent e -t Attenuation, t is the iteration step of reinforcement learning. During the reinforcement learning training process, the model uses the experimental data set D e The training set in is used for training, and each sub-model f u Randomly select data, the same ensemble learning model F e The corresponding sub-model f u The input can come from different sets of data.
[0043] 7) Using experimental dataset D e Test the model F on the test set e , each model F e The corresponding sub-model f u The test data used by (6×6) needs to come from the same set of experimental data. Finally, the overall modeling process of the model is as follows Figure 6 shown.
Claims
1. A material mechanical properties testing method based on displacement and load response of in-situ indentation, characterized by: The method described takes the following steps: 1) Select several typical cheap copper alloys and aluminum alloys, cut them into thin sheet specimens and polish them, use the in-situ indenter to obtain the load at different indentation depths and the displacement cloud map near the in-situ indentation area, and establish the experimental data set; each set of data in the data set includes the indentation depth, load, modulus, yield stress, power hardening index and the displacement cloud map of the indentation side under the corresponding conditions; 2) Perform static simulation on the working conditions corresponding to the in-situ indentation test, use the power hardening law as the constitutive model, select several groups of combinations of modulus, yield stress and power hardening index to generate the plastic strain-stress curve of the material in the plastic stage through programming methods, and use several groups of modulus, yield stress and plastic stress-strain curves as material properties to perform static simulation of the corresponding working conditions; establish a simulation data set, which includes the corresponding indentation depth, load, modulus, yield stress, power hardening index in the simulation calculation, and the displacement cloud map of the indentation side under the corresponding conditions; 3) Data augmentation is performed on the test data set and the simulation data set. The modified diffusion model is used for data augmentation. The indentation depth, load, modulus, yield stress, and power hardening coefficient are embedded into the model through the fully connected projection method to achieve mapping from indentation depth, load, modulus, yield stress, and power hardening index to the corresponding displacement cloud map, thereby achieving data augmentation; the corresponding augmented data sets are the test data augmentation data set and the simulation data augmentation data set; 4) Then, a deep learning model is established from responses such as indentation depth, load, and corresponding displacement cloud map to mechanical properties such as modulus, yield stress, and power hardening coefficient. The model is first trained on the simulation augmented dataset, then fine-tuned on the simulation augmented dataset and the experimental augmented dataset, and finally the final fine-tuning is completed on the experimental dataset to obtain the final sub-model; 5) Repeat step 4 n times to obtain n different models; select m models for each model, and integrate the m×n models into an integrated learning model through an ensemble learning algorithm; 6) Assign a weight to each sub-model in the integrated learning model, and optimize the weight corresponding to each sub-model through the reinforcement learning algorithm DQN to obtain the final model; During the reinforcement learning training process, the model is trained using the training set from all experimental data sets. Each sub-model randomly selects data. Different sub-models in an integrated learning model can select data from different groups of experiments. 7) Use new experimental data to test the integrated learning model. The test data used by the m×n groups of sub-models corresponding to each integrated learning model needs to come from the same set of experimental data.
2. The material mechanical properties testing method based on displacement and load response of in-situ indentation according to claim 1, characterized in that: This method realizes the mapping from load, indentation depth and specimen side displacement cloud map to material mechanical properties. The indentation depth h ranges from 0 to 200 μm. The use range of the test material is limited to metal solid materials. The material modulus ranges from 10 to 300 Gpa, the material yield strength ranges from 30 to 3000 Mpa, and the material power hardening coefficient ranges from 0.1 to 0.
5.
3. The material mechanical properties testing method based on displacement and load response of in-situ indentation according to claim 1, characterized in that: This method performs finite element statics simulation of the corresponding working conditions and realizes the mapping from modulus, yield stress and power hardening index to compression depth, load and displacement cloud map to obtain a simulation data set with a distribution similar to the experimental data set; this method proposes a mapping based on compression depth, load, modulus, yield stress, power hardening coefficient to displacement cloud map to achieve data augmentation; in addition, this method uses an integrated learning method to reduce the demand for expensive experimental data.
