Diamond wire metallographic intelligent analysis standard method based on deep learning

Abstract

The invention relates to a diamond wire metallographic intelligent analysis standard method based on deep learning, which is mainly realized through a structure intelligent identification module, a sorbite rate determination module and an original austenite grain size automatic rating module, and specifically comprises the following steps: firstly, selecting a rolled diamond wire through a standard metallographic sample preparation method to manufacture a standard metallographic sample; after 4% nitric acid alcohol erosion, a microstructure image is shot under a metallographic microscope, and a microstructure intelligent identification module is used for carrying out metallographic structure intelligent determination; secondly, automatically calculating the sorbite rate through a sorbite rate calculation module; and finally, after quenching, finishing the grading of the original austenite grain size of the diamond wire rod by utilizing an automatic grain size grading module. According to the method, a traditional metallographic inspection mode is changed, the rating result is high in reproducibility and precision, the influence of human subjectivity can be eliminated, and the accuracy and efficiency of metallographic analysis work are improved.

Description

Technical Field

[0001] This patent application belongs to the field of metallographic analysis technology, and more specifically, it relates to a standard method for intelligent metallographic analysis of diamond wire based on deep learning. Background Art

[0002] Diamond wire is the abbreviation of diamond cutting wire, also known as diamond cutting wire or diamond wire. In industry, many hard materials are cut with cutting steel wire or higher quality diamond wire, such as polysilicon slices, single crystal silicon, and crystal rods in the photovoltaic field. At present, the preparation technology of diamond wire is relatively developed mainly in the West and Japan. my country has also invested considerable research and development in this area. Metallographic analysis is the most important part of the research and development of diamond wire, mainly involving organization analysis, grain size rating, and troostite rate determination. At present, metallographic analysis mainly adopts the method of optical microscope observation and recording + manual analysis. This method has strict requirements on professional talents, and the rating personnel must have a high level of professionalism and rich experience. At the same time, different inspectors have different evaluation standards for similar samples, which are greatly affected by subjective consciousness. And because the rating calculation is large and time-consuming, high-throughput detection and analysis cannot be completed at present.

[0003] With the rapid iteration and update of computer hardware in recent years, related technologies such as artificial intelligence and big data have also developed rapidly. Various algorithms and new theories are widely used in various fields. The use of deep learning algorithms combined with image processing can largely replace some manual rating tasks and reduce workload.

[0004] Therefore, it is necessary to establish a standard module for intelligent metallographic analysis of diamond wire based on deep learning, which can realize intelligent metallographic analysis of diamond wire, greatly improve R&D efficiency, and effectively save personnel costs and time costs. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a standard method for intelligent metallographic analysis of diamond wire based on deep learning, which can realize intelligent metallographic analysis of diamond wire, greatly improve research and development efficiency, and effectively save personnel costs and time costs.

[0006] In order to solve the above problems, the technical solution adopted by the present invention is: A standard method for intelligent metallographic analysis of diamond wire based on deep learning, comprising the following steps: S1. Preparation of standard metallographic specimens: Select representative diamond wires for sample preparation and complete image acquisition, standardize the sample preparation process, and use a fully automatic grinding and polishing machine for sample grinding and polishing. Polishing must be performed when manually preparing samples. Samples are corroded using a 4% nitric acid alcohol solution by volume. In order to maintain stable performance of the corrosive solution, the corrosive solution should be sealed and not exposed to the air for a long time to prevent evaporation and oxidation of the corrosive solution.

[0007] S2. Metallographic image acquisition. The samples were imaged using a Zeiss Imager M2m microscope. The magnification of the acquired images was 200X or 500X.

[0008] S3. Metallographic structure determination: The deep learning model Res2net50 is used to integrate multi-scale features to establish a metallographic structure intelligent recognition module, and the metallographic structure of diamond wire is effectively determined through the metallographic structure intelligent recognition module.

[0009] S4. Determination of troostite rate: The troostite rate can ensure that the strength and plasticity of the wire are greatly improved, especially in the drawing process to maintain uniform deformation and not prone to breakage. Therefore, the troostite rate is an important means of metallographic analysis of wires. A wire troostite rate determination module based on the U-Net deep learning network is established, and the troostite rate of diamond wire is automatically determined by the module to ensure the quality of finished steel.

[0010] S5. Original austenite grain size rating: U-Net network is built based on TensorFlow framework, and the automatic grain size rating module is established according to the straight line intercept method in the national standard "GB / T6394-2017 Metal Average Grain Size Determination Method". After quenching, the original austenite grain size rating of the diamond wire is completed using the automatic grain size rating module.

[0011] S6. Intelligent analysis standard module: The tissue intelligent identification module, the troostite rate determination module, and the original austenite grain size rating module are uniformly deployed to form a diamond wire metallographic intelligent analysis standard module based on deep learning, thus realizing an integrated solution for the intelligent metallographic analysis of diamond wire.

[0012] Due to the adoption of the above technical solution, the beneficial effects achieved by the present invention are: The present invention changes the traditional metallographic inspection method, and the grading results have strong reproducibility and high precision, and can eliminate the influence of human subjectivity, improve the accuracy and efficiency of metallographic analysis, realize the intelligent analysis of diamond wire metallography, and have a higher promotion and application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 is a flow chart of an embodiment of the present invention.

[0014] Figure 2 Schematic diagram of the tissue intelligent identification module.

[0015] Figure 3 It is a metallographic structure image of an embodiment of the present invention.

[0016] Figure 4 This is a troostite ratio measurement image of an example of the present invention.

[0017] Figure 5 This is an image of automatic grading of original austenite grain size according to an embodiment of the present invention. DETAILED DESCRIPTION

[0018] The present invention is further described in detail below with reference to the embodiments.

[0019] A standard method for intelligent metallographic analysis of diamond wire based on deep learning, such as Figure 1 As shown, the specific steps are: S1. Preparation of standard metallographic specimens: Select representative diamond wires for sample preparation, and then complete the following metallographic image acquisition, and standardize the sample preparation process. For example, it is recommended to use a fully automatic grinder and polisher for sample preparation, and polishing must be performed when manually preparing samples; after grinding and polishing, the sample is corroded using a 4% nitric acid alcohol solution by volume. In order to maintain the stable performance of the corrosive liquid, the corrosive liquid should be sealed and not exposed to the air for a long time to prevent the corrosive liquid from evaporating.

[0020] S2. Metallographic image acquisition. The samples were imaged using a Zeiss Imager M2m microscope, with the image magnification being 200X or 500X, 500X in this embodiment. Figure 3 As shown, the acquisition unit is 20 μm.

[0021] S3. Metallographic structure determination: Use deep learning models and integrate multi-scale features to establish a metallographic structure intelligent recognition module, such as Figure 2 As shown, the specific operations are: (1) A deep learning model Res2net50 was built based on the TensorFlow framework. An adaptive multi-scale attention enhancement module was proposed to address the problem of multiple magnifications of metallographic images, so as to enhance the correlation between images of different magnifications. Starting from the image input, it goes through convolution, maximum pooling, and four bottleneck layers, and an adaptive multi-scale attention enhancement module is introduced in the third layer.

[0022] (2) To enhance the extraction of texture features, the TripletAttention module is introduced into the residual structure of the last layer of the network in Res2net50, thereby enhancing the attention to details.

[0023] (3) A multi-scale dynamic adjustment feature fusion module is used to perform multi-scale feature fusion to increase the number of effective features captured by the network. The multi-scale dynamic adjustment feature fusion module can accurately classify a total of twenty types of metallographic structures, including single-phase, dual-phase, three-phase, and four-phase structures. After average pooling, a fully connected layer, and softmax output results, the accuracy can reach more than 95%, for example, up to 95.53% in this embodiment.

[0024] Use the intelligent identification module to effectively determine the metallographic structure of diamond wire, such as Figure 4 Shown are troostite and pearlite, and the collection unit is also 20μm, see the table below.

[0025]

[0026] S4. Determination of the sorbite rate: The sorbite rate can ensure that the strength and plasticity of the wire are greatly improved, especially in the drawing process to maintain uniform deformation and not prone to breakage. Therefore, the sorbite rate is an important means of metallographic analysis of wires. Therefore, a sorbite rate determination module based on the U-Net deep learning network is established, and the sorbite rate of diamond wire is automatically determined by the sorbite rate determination module.

[0027] S5. Rating of original austenite grain size: After quenching, the diamond wire is ground and polished, and then etched with a new original austenite etchant. Then, a U-Net network is built based on the TensorFlow framework. According to the straight line intercept method in the national standard "GB / T 6394-2017 Metal Average Grain Size Determination Method", the original austenite grain size rating of the diamond wire is completed using the grain size automatic rating module. Figure 5 shown.

[0028] The U-Net network structure here includes 4 groups of upsampling and 4 groups of downsampling, and the upsampling and downsampling are connected through the connection layer Concatenate. The downsampling method uses the maximum pooling, the pooling window sliding step is 2, the convolution kernel is 3*3, and the activation function is RELU; the upsampling method uses the deconvolution method, and the convolution kernel is 3*3.

[0029] The method provided by the present invention can realize intelligent metallographic analysis of diamond wire, which changes the traditional metallographic inspection method. The grading results are highly reproducible and accurate, and can eliminate the influence of human subjectivity, thereby improving the accuracy and efficiency of metallographic analysis. The accuracy rate reaches 95.53%, and the efficiency is increased by more than 3 times.

Claims

1. A standard method for intelligent metallographic analysis of diamond wire based on deep learning, characterized in that The following steps are involved: S1. Preparation of standard metallographic specimens: Select representative diamond wires for sample preparation and standardize the sample preparation process. It is recommended to use a fully automatic grinding and polishing machine for sample preparation. Polishing must be performed when manually preparing samples. Sample corrosion uses a 4% nitric acid alcohol solution by volume. In order to keep the corrosion liquid stable, the corrosion liquid should be sealed and not exposed to the air for a long time to prevent evaporation and oxidation of the corrosion liquid. S2, metallographic image acquisition: use a microscope to acquire images of the sample; S3. Metallographic structure determination: Use deep learning model and integrate multi-scale features to establish a tissue intelligent recognition module, and effectively determine the metallographic structure of diamond wire through the tissue intelligent recognition module; S4. Sorbite rate determination: Establish a Sorbite rate determination module based on the U-Net deep learning network, and use this module to automatically determine the Sorbite rate of diamond wire to ensure the quality of finished steel; S5. Original austenite grain size rating: Based on the TensorFlow framework, a U-Net network is built. According to the national standard, an original austenite grain size automatic rating module is established. After the diamond wire is quenched, the original austenite grain size of the diamond wire is automatically rated using this module. S6. Formulation of intelligent analysis standards: Unify the deployment of the tissue intelligent identification module, the troostite ratio determination module, and the original austenite grain size automatic rating module to form a diamond wire metallographic intelligent analysis standard module based on deep learning, and realize an integrated solution for diamond wire intelligent metallographic analysis.

2. According to claim 1, a standard method for intelligent metallographic analysis of diamond wire based on deep learning, characterized in that: In S1, "standardization of sample preparation process" means that the samples are ground and polished by fully automatic grinding and polishing machines, and polishing must be performed when the samples are prepared manually; the samples are corroded by using a 4% nitric acid alcohol solution by volume. In order to maintain the stable performance of the corrosive solution, the corrosive solution should be sealed and not exposed to the air for a long time to prevent evaporation and oxidation of the corrosive solution.

3. The standard method for intelligent metallographic analysis of diamond wire based on deep learning according to claim 1, characterized in that: In S2, the microscope used was the Zeiss Imager M2m, and the image acquisition magnification was 200X or 500X.

4. The standard method for intelligent metallographic analysis of diamond wire based on deep learning according to claim 1, characterized in that: In S3, "using deep learning models and integrating multi-scale features to establish an intelligent tissue recognition module" means: (1) Based on the TensorFlow framework, a deep learning model Res2net50 was built. Aiming at the problem of multiple magnifications of metallographic images, an adaptive multi-scale attention enhancement module was proposed to enhance the correlation between images of different magnifications. (2) To enhance the extraction of texture features, the TripletAttention module is introduced into the residual structure of the last layer of the Res2net50 network to enhance the attention to details; (3) A multi-scale dynamic adjustment feature fusion module is used to perform multi-scale feature fusion to increase the number of effective features captured by the network.

5. The standard method for intelligent metallographic analysis of diamond wire based on deep learning according to claim 4, characterized in that: The multi-scale dynamic adjustment feature fusion module can accurately classify a total of twenty types of metallographic structures, including single-phase, two-phase, three-phase, and four-phase structures, with an accuracy rate of over 95%.

6. The standard method for intelligent metallographic analysis of diamond wire based on deep learning according to claim 1, characterized in that: In S5, the U-Net network structure includes 4 groups of upsampling and 4 groups of downsampling. The downsampling method adopts maximum pooling, the pooling window sliding step is 2, the convolution kernel is 3*3, and the activation function is RELU; the upsampling method adopts the deconvolution method, the convolution kernel is 3*3; the upsampling and downsampling are connected through the connection layer Concatenate.

7. A standard method for intelligent metallographic analysis of diamond wire based on deep learning according to any one of claims 1 to 6, characterized in that: The original austenite grain size rating module adopts the straight line intercept method in the national standard "GB / T 6394-2017 Method for Determination of Average Grain Size of Metals" to perform grain size rating.

Images