Method for predicting vickers hardness by nanoindentation hardness using deep learning neural network
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HARBIN INST OF TECH
- Filing Date
- 2023-10-17
- Publication Date
- 2026-06-26
AI Technical Summary
Existing Vickers hardness testing methods are destructive, time-consuming, and unsuitable for thin films and coatings, making it difficult to meet the needs of modern engineering for rapid, accurate, and non-destructive hardness measurement.
A deep learning neural network model was established to predict Vickers hardness through nanoindentation hardness. A multi-layer neural network model was constructed to handle the nonlinear relationship between nanoindentation hardness and Vickers hardness. The model was trained using M50NiL steel data that had undergone nitriding and carburizing treatments to predict its hardness in the range of 400-1000 HV.
It enables rapid, accurate, and non-destructive hardness measurement, applicable to M50NiL steel and other steels, improving the efficiency and accuracy of hardness measurement and reducing production costs.
Smart Images
Figure CN117409897B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of materials science, specifically relating to a method based on deep neural networks (DNN) for predicting the Vickers hardness (HV) of M50NiL steel after nitriding and carburizing treatment. Background Technology
[0002] Hardness is an important indicator of a material's mechanical properties, relating to its resistance to deformation or fracture. This property has significant applications in materials science, mechanical engineering, manufacturing, and many other fields. While various hardness measurement methods exist, such as Vickers hardness (HV), Brinell hardness (HB), and Rockwell hardness (HR), Vickers hardness (HV) is a method that applies a known, constant load to the material using a pyramid-shaped hard diamond indenter, holds it for a period of time, and then measures the diagonal length of the indentation to determine the hardness value. This value is typically expressed as HV, and its calculation formula is:
[0003]
[0004] Where F is the applied load (in N), and d is the diagonal length of the indentation (in mm).
[0005] Vickers hardness testing has several advantages:
[0006] 1. Wide applicability: Suitable for various types of materials such as metals, ceramics, and plastics.
[0007] 2. High repeatability: Due to the use of a pyramid-shaped indenter and precise pressure measurement, the test results are usually highly repeatable.
[0008] 3. Microscale: This hardness testing method can be used for material samples and tissues at smaller scales, including microstructures.
[0009] However, the Vickers hardness test also has some limitations:
[0010] 1. Destructive: This is a destructive testing method that leaves permanent indentations on the surface of the material.
[0011] 2. Time consumption: The testing process requires a period of time to stabilize the pressure head and perform accurate measurements.
[0012] 3. Limited applicability to thin films and coatings: Due to the need to apply a certain load, it is difficult to use for microscale materials such as thin films and coatings.
[0013] Therefore, overall, the development of rapid, accurate, and non-destructive hardness measurement methods has enormous potential value in meeting modern engineering needs, reducing production costs, improving production efficiency, and promoting progress in related scientific and technological fields. Summary of the Invention
[0014] This invention addresses the problems of existing Vickers hardness testing methods being destructive, time-consuming, and inefficient by providing a deep neural network-based model for processing and analyzing the hardness data of M50NiL steel (after nitriding and carburizing treatment), and accurately predicting its Vickers hardness in the range of 400-1000 HV.
[0015] This invention utilizes a deep learning neural network to predict Vickers hardness through nanoindentation hardness, and is implemented according to the following steps:
[0016] I. M50NiL steel that has undergone carburizing and plasma nitriding treatment was selected as the sample.
[0017] 2. The nanoindentation hardness and Vickers hardness of the samples were tested separately to construct a penetration depth-nanoindentation hardness-Vickers hardness dataset. The penetration depth-nanoindentation hardness-Vickers hardness dataset contains both the nanoindentation hardness and Vickers hardness measured at the penetration depth.
[0018] 3. Establish a depth-hardness DNN model, which includes one input layer, three hidden layers and one output layer. The first hidden layer contains 32 neurons, the second hidden layer contains 64 neurons and the third hidden layer contains 128 neurons. The sigmoid is used as the activation function and MSE is used as the evaluation function.
[0019] Using the infiltration depth and nanoindentation hardness data from the infiltration depth-nanoindentation hardness-Vickers hardness dataset, with the infiltration depth as input, a depth-hardness DNN model is used to extend and train the nanoindentation hardness data, resulting in extended nanoindentation hardness data. Then, using the infiltration depth and Vickers hardness data from the same dataset, with the infiltration depth as input, a depth-hardness DNN model is used to extend and train the Vickers hardness data, resulting in extended Vickers hardness data. Finally, the extended nanoindentation hardness data and the extended Vickers hardness data are merged to obtain an extended dataset.
[0020] IV. Establishing a nanoindentation-Vickers hardness DNN model. The structure (architecture) of the nanoindentation-Vickers hardness DNN model is the same as that of the depth-hardness DNN model. The nanoindentation-Vickers hardness DNN model includes one input layer, three hidden layers and one output layer. The first hidden layer contains 32 neurons, the second hidden layer contains 64 neurons and the third hidden layer contains 128 neurons. The sigmoid is used as the activation function and MSE is used as the evaluation function.
[0021] 5. The nanoindentation-Vickers hardness DNN model is trained using the extended dataset from step 3 to obtain the trained nanoindentation-Vickers hardness DNN model.
[0022] VI. Using the trained nanoindentation-Vickers hardness DNN model, with nanoindentation hardness as input data, Vickers hardness is predicted.
[0023] This invention utilizes M50NiL steel subjected to nitriding and carburizing treatments to obtain a hardness curve with a broad and smooth gradient. A correlation model between nanoindentation hardness and Vickers hardness (HV) is then constructed. This model can predict the Vickers hardness (HV) of G13Cr4Mo4Ni4V steel and other steels using nanoindentation hardness. This model is not only applicable to G13Cr4Mo4Ni4V steel but can also be extended to other types of steel.
[0024] This invention constructs a multi-layer deep neural network to handle the nonlinear relationship between nanoindentation hardness and Vickers hardness. The model can process and analyze hardness data of M50NiL steel (after nitriding and carburizing treatment) and accurately predict its Vickers hardness in the range of 400-1000 HV. Attached Figure Description
[0025] Figure 1 This is a flowchart of the method for predicting Vickers hardness using a deep learning neural network through nanoindentation hardness, as described in this invention.
[0026] Figure 2 This is a test graph showing the depth-Vickers hardness dataset and model fitting in step three of the embodiment.
[0027] Figure 3 This is a comparison chart of the actual and simulated training of the depth-Vickers hardness model in step three of the embodiment;
[0028] Figure 4 This is a test graph showing the depth-nanoindentation hardness dataset and model fitting in step three of the embodiment.
[0029] Figure 5 This is a comparison chart of the actual and simulated training of the depth-nanoindentation hardness model in step three of the embodiment;
[0030] Figure 6 This is a comparison chart of the actual Vickers hardness of M50NiL steel and the model prediction in the embodiments.
[0031] Figure 7 These are the actual curves and model-predicted curves of nanoindentation hardness-Vickers hardness for M50NiL steel in the examples. Detailed Implementation
[0032] Specific Implementation Method 1: This implementation method utilizes a deep learning neural network to predict Vickers hardness through nanoindentation hardness, and is carried out according to the following steps:
[0033] I. M50NiL steel that has undergone carburizing and plasma nitriding treatment was selected as the sample.
[0034] 2. The nanoindentation hardness and Vickers hardness of the samples were tested separately to construct a penetration depth-nanoindentation hardness-Vickers hardness dataset. The penetration depth-nanoindentation hardness-Vickers hardness dataset contains both the nanoindentation hardness and Vickers hardness measured at the penetration depth.
[0035] 3. Establish a depth-hardness DNN model, which includes one input layer, three hidden layers and one output layer. The first hidden layer contains 32 neurons, the second hidden layer contains 64 neurons and the third hidden layer contains 128 neurons. The sigmoid is used as the activation function and MSE is used as the evaluation function.
[0036] Using the infiltration depth and nanoindentation hardness data from the infiltration depth-nanoindentation hardness-Vickers hardness dataset, with the infiltration depth as input, a depth-hardness DNN model is used to extend and train the nanoindentation hardness data, resulting in extended nanoindentation hardness data. Then, using the infiltration depth and Vickers hardness data from the same dataset, with the infiltration depth as input, a depth-hardness DNN model is used to extend and train the Vickers hardness data, resulting in extended Vickers hardness data. Finally, the extended nanoindentation hardness data and the extended Vickers hardness data are merged to obtain an extended dataset.
[0037] IV. Establishing a nanoindentation-Vickers hardness DNN model. The structure (architecture) of the nanoindentation-Vickers hardness DNN model is the same as that of the depth-hardness DNN model. The nanoindentation-Vickers hardness DNN model includes one input layer, three hidden layers and one output layer. The first hidden layer contains 32 neurons, the second hidden layer contains 64 neurons and the third hidden layer contains 128 neurons. The sigmoid is used as the activation function and MSE is used as the evaluation function.
[0038] 5. The nanoindentation-Vickers hardness DNN model is trained using the extended dataset from step 3 to obtain the trained nanoindentation-Vickers hardness DNN model.
[0039] VI. Using the trained nanoindentation-Vickers hardness DNN model, with nanoindentation hardness as input data, Vickers hardness is predicted.
[0040] This implementation first employs nitriding and carburizing treatments to significantly alter the microstructure and hardness of M50NiL steel, thereby providing a more comprehensive and continuous dataset for model training. Then, by utilizing a deep neural network to learn from a large amount of data, it is able to capture the complex nonlinear relationship between the hardness of M50NiL steel and the nanoindentation hardness.
[0041] Specific Implementation Method Two: This implementation method differs from Specific Implementation Method One in that the carburizing process in step one is as follows:
[0042] Initial carburizing is first performed in an acetylene atmosphere at a temperature of 950–1000℃ for 15–30 hours. Then, a quenching process is carried out in a nitrogen atmosphere of 1–5 Bar at a temperature of 1050–1150℃ for 40–60 minutes to complete the carburizing treatment.
[0043] In this embodiment, the carburizing process uses an ECM vacuum carburizing furnace, while the quenching stage uses a BMI micro high-temperature vacuum furnace.
[0044] Specific Implementation Method Three: This implementation method differs from Specific Implementation Method One or Two in that the process of plasma nitriding treatment in step one is as follows:
[0045] The active barrier nitriding process is adopted, and the bias voltage of the pulse power supply is controlled at 500-1000V and the frequency is 40-60kHz. The furnace cavity is evacuated to 1-10Pa, and then N2:H2 volume flow rate ratio is introduced at 1:2-1:50, temperature is 450-500℃, duration is 40-50 hours, and then cooled to room temperature to complete the plasma nitriding treatment.
[0046] In this embodiment of the active screen nitriding process, the active screen is a porous cylinder with a diameter of 300 mm and a height of 200 mm. The sample frustum, with a diameter of 260 mm, is placed inside the active screen, 150 mm from the top. The active screen is directly connected to the pulse electrode. The sample stage is supported by an insulating ceramic column, keeping the sample at a floating potential. A K-series thermocouple is placed in the central hole of the sample for real-time temperature measurement and control. Two layers of metal heat shields are installed inside the furnace hood, and circulating water is connected outside the furnace for cooling. The vacuum pressure inside the furnace is measured using a mercury differential pressure gauge.
[0047] Specific Implementation Method Four: This implementation method differs from Specific Implementation Methods One to Three in that the continuous stiffness measurement method is used to detect the nano-indentation hardness of the sample in step two.
[0048] Specific Implementation Method 5: This implementation method differs from Specific Implementation Methods 1 to 4 in that in step 2, a micro Vickers hardness tester is used to test the Vickers hardness of the sample under a 500g load.
[0049] Specific Implementation Method Six: This implementation method differs from Specific Implementation Methods One to Five in that the amount of data for detecting the nanoindentation hardness and Vickers hardness of the sample in step two is 200-400.
[0050] Specific Implementation Method Seven: This implementation method differs from Specific Implementation Methods One through Six in that the number of rounds of extended training in step three is controlled to be 5000 to 10000, the learning rate is 0.1 to 0.001, and the Adam optimizer is used for optimization.
[0051] Specific Implementation Method Eight: This implementation method differs from Specific Implementation Methods One to Seven in that the extended dataset obtained in step three contains 20,000 to 30,000 nano-indentation hardness-Vickers hardness data points.
[0052] Specific Implementation Method Nine: This implementation method differs from Specific Implementation Methods One to Eight in that in step five, the number of training rounds is controlled to be 5000 to 10000, the learning rate is 0.1 to 0.001, and the Adam optimizer is used for optimization.
[0053] Specific Implementation Method 10: This implementation method differs from Specific Implementation Method 9 in that the number of training rounds is controlled to be 10,000 in step 5.
[0054] Example: This example utilizes a deep learning neural network to predict Vickers hardness through nanoindentation hardness, and is implemented according to the following steps:
[0055] 1. M50NiL steel is made into Φ35×10mm annealed cylindrical specimens. Initial carburizing is performed in an acetylene atmosphere at 960℃, followed by quenching in a nitrogen atmosphere of 2 Bar at 1100℃ to complete the carburizing process.
[0056] Then, the active barrier nitriding process was adopted, the bias voltage of the pulse power supply was controlled at 800V and the frequency was 40kHz, the furnace cavity was evacuated to 5Pa, and then N2:H2 volume flow rate ratio of 1:20 was introduced, the temperature was 500℃, the duration was 50 hours, and then cooled to room temperature to complete the plasma nitriding treatment, and a sample with hardness gradient change was obtained.
[0057] 2. The samples were cleaned with anhydrous ethanol. The Vickers hardness of the samples was tested under a 500g load using a micro Vickers hardness tester. 300-400 hardness data points were collected. At the same time, the nanoindentation hardness was measured using the continuous stiffness measurement method. The test points were calibrated using optical and electron microscopes to construct a diffusion layer depth-nanoindentation hardness-Vickers hardness dataset. The diffusion layer depth-nanoindentation hardness-Vickers hardness dataset contains both the nanoindentation hardness and Vickers hardness measured at the diffusion layer depth.
[0058] 3. A deep-hardness DNN model is built based on TensorFlow. The deep-hardness DNN model includes one input layer, three hidden layers and one output layer. The first hidden layer contains 32 neurons, the second hidden layer contains 64 neurons and the third hidden layer contains 128 neurons. The sigmoid is used as the activation function and MSE is used as the evaluation function.
[0059] Using the infiltration depth-nanoplastin hardness-Vickers hardness dataset, the nanoindentation hardness data were expanded and trained using a depth-hardness DNN model with infiltration depth as input, for 10,000 iterations, resulting in expanded nanoindentation hardness data. Then, using the same dataset, the Vickers hardness data was expanded and trained using the same depth-nanoplastin hardness-Vickers hardness dataset with infiltration depth as input, resulting in expanded Vickers hardness data. Finally, the expanded nanoindentation hardness data and the expanded Vickers hardness data at the same infiltration depth were merged to obtain the expanded dataset.
[0060] IV. Establishing a nanoindentation-Vickers hardness DNN model. The structure (architecture) of the nanoindentation-Vickers hardness DNN model is the same as that of the depth-hardness DNN model. The nanoindentation-Vickers hardness DNN model includes one input layer, three hidden layers and one output layer. The first hidden layer contains 32 neurons, the second hidden layer contains 64 neurons and the third hidden layer contains 128 neurons. The sigmoid is used as the activation function and MSE is used as the evaluation function.
[0061] 5. The nanoindentation-Vickers hardness DNN model is trained using the extended dataset from step 3 to obtain the trained nanoindentation-Vickers hardness DNN model.
[0062] VI. Using the trained nanoindentation-Vickers hardness DNN model, with nanoindentation hardness as input data, Vickers hardness is predicted.
[0063] In step three of this embodiment, 25,000 nanoindentation and Vickers hardness data points were generated using a depth-hardness DNN model, covering a depth range of 40-2000 μm.
[0064] Table 1. Predicted, actual, and error values of Vickers hardness in the range of 400-1000 HV.
[0065]
[0066] Table 2. Predicted, actual, and error values of Vickers hardness outside the 400-1000 HV range.
[0067]
[0068] according to Figure 2-7 The data is used to illustrate that, in conventional understanding, the correlation between Vickers hardness and nanoindentation hardness is usually considered to be approximately linear. However, this approximately linear characterization cannot accurately predict the deviation between the two. Using deep neural network methods, this invention can delve into the complex relationship between Vickers hardness and nanoindentation hardness and achieve effective prediction of this relationship.
[0069] In numerous experiments and analyses, Vickers hardness and nanoindentation hardness show an approximately linear correlation. However, in both theoretical models and practical applications, the accuracy of predictions based on a single linear model is often limited. Deep neural networks, as an efficient tool capable of extracting and processing nonlinear relationships, offer a new approach to understanding and explaining these irregular phenomena. This invention, by learning from a large amount of experimental data, demonstrates how deep neural networks can capture the nonlinear relationship between Vickers hardness and nanoindentation hardness, thereby further revealing the unique response characteristics of material properties under different testing conditions.
Claims
1. A method for predicting Vickers hardness using nanoindentation hardness via deep learning neural networks, characterized in that... The method for predicting Vickers hardness is implemented according to the following steps: I. M50NiL steel that has undergone carburizing and plasma nitriding treatment was selected as the sample.
2. The nanoindentation hardness and Vickers hardness of the samples were tested separately to construct a penetration depth-nanoindentation hardness-Vickers hardness dataset. The penetration depth-nanoindentation hardness-Vickers hardness dataset contains both the nanoindentation hardness and Vickers hardness measured at the penetration depth.
3. Establish a depth-hardness DNN model, which includes one input layer, three hidden layers and one output layer. The first hidden layer contains 32 neurons, the second hidden layer contains 64 neurons and the third hidden layer contains 128 neurons. The sigmoid is used as the activation function and MSE is used as the evaluation function. Using the penetration depth and nanoindentation hardness data from the penetration depth-nanoindentation hardness-Vickers hardness dataset, with penetration depth as input data, the nanoindentation hardness data is expanded and trained using a depth-hardness DNN model to obtain expanded nanoindentation hardness data; then, using the penetration depth and Vickers hardness data from the penetration depth-nanoindentation hardness-Vickers hardness dataset, the Vickers hardness data is expanded and trained using a depth-hardness DNN model to obtain expanded Vickers hardness data. Then, the expanded nanoindentation hardness data and the expanded Vickers hardness data are merged to obtain the expanded dataset; IV. Establishing a nanoindentation-Vickers hardness DNN model. The structure of the nanoindentation-Vickers hardness DNN model is the same as that of the depth-hardness DNN model. The nanoindentation-Vickers hardness DNN model includes one input layer, three hidden layers and one output layer. The first hidden layer contains 32 neurons, the second hidden layer contains 64 neurons and the third hidden layer contains 128 neurons. The sigmoid function is used as the activation function and the MSE function is used as the evaluation function.
5. The nanoindentation-Vickers hardness DNN model is trained using the extended dataset from step 3 to obtain the trained nanoindentation-Vickers hardness DNN model. VI. Using the trained nanoindentation-Vickers hardness DNN model, with nanoindentation hardness as input data, Vickers hardness is predicted.
2. The method for predicting Vickers hardness using nanoindentation hardness via deep learning neural networks according to claim 1, characterized in that... The carburizing process in step one is as follows: Initial carburizing is first performed in an acetylene atmosphere at a temperature of 950–1000℃ for 15–30 hours. Then, a quenching process is carried out in a nitrogen atmosphere of 1–5 Bar at a temperature of 1050–1150℃ for 40–60 minutes to complete the carburizing treatment.
3. The method for predicting Vickers hardness using nanoindentation hardness via deep learning neural networks according to claim 1, characterized in that... The process of plasma nitriding treatment in step one is as follows: The active barrier nitriding process is adopted, and the bias voltage of the pulse power supply is controlled at 500-1000V and the frequency is 40-60kHz. The furnace cavity is evacuated to 1-10Pa, and then N2:H2 volume flow rate ratio is introduced at 1:2-1:50, temperature is 450-500℃, duration is 40-50 hours, and then cooled to room temperature to complete the plasma nitriding treatment.
4. The method for predicting Vickers hardness using nanoindentation hardness via deep learning neural networks according to claim 1, characterized in that... In step two, the nanoindentation hardness of the sample is tested using the continuous stiffness measurement method.
5. The method for predicting Vickers hardness using nanoindentation hardness via deep learning neural networks according to claim 1, characterized in that... In step two, a micro Vickers hardness tester is used to test the Vickers hardness of the sample under a load of 500g.
6. The method for predicting Vickers hardness using nanoindentation hardness via deep learning neural networks according to claim 1, characterized in that... In step two, the number of data points for detecting the nanoindentation hardness and Vickers hardness of the sample is 200-400.
7. The method for predicting Vickers hardness using nanoindentation hardness via deep learning neural networks according to claim 1, characterized in that... In step three, the number of training rounds is controlled to be 5000-10000, the learning rate is 0.1-0.001, and the Adam optimizer is used for optimization.
8. The method for predicting Vickers hardness using nanoindentation hardness via deep learning neural networks according to claim 1, characterized in that... The extended dataset obtained in step three contains 20,000 to 30,000 nanoindentation hardness-Vickers hardness data points.
9. The method for predicting Vickers hardness using nanoindentation hardness via deep learning neural networks according to claim 1, characterized in that... In step five, the number of training rounds is controlled to be 5000-10000, the learning rate is 0.1-0.001, and the Adam optimizer is used for optimization.
10. The method for predicting Vickers hardness using a deep learning neural network via nanoindentation hardness according to claim 9, characterized in that... In step five, the number of training rounds is controlled to be 10,000.