Metal powder quality evaluation method based on machine learning driving

Abstract

The invention provides a metal powder quality evaluation method based on machine learning driving, and belongs to the field of material science and artificial intelligence. The method comprises the following steps: acquiring a plurality of metal powder samples and three-dimensional images thereof, and analyzing quality characterization parameters of each metal powder sample; setting an evaluation standard and evaluating each metal powder sample; constructing a neural network-based metal powder quality evaluation model, and training the model by taking the three-dimensional image of each metal powder sample as a training sample and the evaluation result of each metal powder sample as a sample label; and obtaining a three-dimensional image of the metal powder to be detected, inputting the three-dimensional image into the trained metal powder quality evaluation model, and obtaining an evaluation result of the model. According to the method, the quality of the metal powder can be evaluated more efficiently and accurately, the quality of the additive manufacturing metal powder can be checked, the method can also be used as an optimization basis of a metal powder manufacturing process, and therefore the service performance of additive manufacturing parts is improved.

Description

Technical Field

[0001] The present invention relates to the fields of materials science and artificial intelligence, and particularly to a method for evaluating the quality of metal powders driven by machine learning. Background Art

[0002] The quality of metal powders is a key factor affecting the product quality and production efficiency in technologies such as 3D printing and powder metallurgy. Traditional methods for evaluating the quality of metal powders mainly rely on manual inspection, microscopic observation, sieving methods, or chemical composition analysis. These methods are often cumbersome, time-consuming, and greatly affected by human factors. In addition, with the development of technology, the types and application scenarios of metal powders are becoming increasingly diverse, and traditional evaluation methods are difficult to meet the requirements of rapidity, high efficiency, and accuracy. Therefore, developing an automated method for evaluating the quality of metal powders based on machine learning has important practical significance. Summary of the Invention

[0003] Based on the problems in the prior art, the object of the present invention is to provide a method for evaluating the quality of metal powders driven by machine learning. The present invention collects physical and chemical characteristic data of various metal powders and uses machine learning algorithms to automatically evaluate the quality of metal powders, thereby achieving rapid and accurate evaluation of the quality.

[0004] To solve the above technical problems, the present invention provides the following technical solutions:

[0005] The present invention provides a method for evaluating the quality of metal powders driven by machine learning, wherein the test method includes the following steps:

[0006] S1: Obtain a plurality of metal powder specimens and their three-dimensional images, and use the three-dimensional images of each metal powder specimen as training samples; analyze the quality characterization parameters of each metal powder specimen, and the quality characterization parameters include fluidity, bulk density, purity, porosity, oxygen content, compressibility, and impurity content; among them, the fluidity, bulk density, and porosity of the metal powder specimen are obtained through three-dimensional image analysis;

[0007] S2: Set evaluation criteria for the quality characterization parameters; compare the quality characterization parameters of the metal powder specimens with the evaluation criteria, and evaluate each metal powder specimen as excellent, good, or unqualified, and the evaluation results are used as sample labels;

[0008] S3: Construct a neural network-based metal powder quality evaluation model, and the model is used to output an evaluation result according to the three-dimensional image of the metal powder. Use the training samples and sample labels to perform forward propagation training on the model until the output error meets the requirements or reaches the set maximum number of training times;

[0009] S4: Obtain the three-dimensional image of the metal powder to be measured, input the three-dimensional image into the trained metal powder quality evaluation model, and obtain the evaluation result of the model.

[0010] Compared with the prior art, the present invention has the following beneficial effects:

[0011] (1) By means of three-dimensional characterization, the present invention establishes a three-dimensional model of metal powder, extracts and analyzes the quality characteristics of metal powder, so as to more accurately evaluate the quality of metal powder, ensure the quality of additive manufacturing metal powder, improve the service performance of additive manufacturing parts, and at the same time obtain other quality characterizations through relevant experiments;

[0012] (2) By quantifying the quality characteristics of metal powder, the present invention can obtain metal powder quality parameters. Through more accurate detection of parameters such as particle size distribution, fluidity, and bulk density, it can better predict the powder forming and the performance of the final product. Improve production efficiency, reduce costs, improve powder consistency, and at the same time realize intelligent monitoring and real-time feedback, reducing human error. Description of the Drawings

[0013] Figure 1 is a flowchart of a metal powder quality evaluation method driven by machine learning in an embodiment of the present invention;

[0014] Figure 2 is a flowchart of a metal powder quality evaluation model in an embodiment of the present invention. Detailed Embodiments

[0015] The following describes the preferred embodiments of the present invention with reference to the drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, but do not limit the present invention in any form. It should be noted that those of ordinary skill in the art can make several changes and improvements without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0016] Embodiment 1

[0017] As Figure 1 shown, the present invention proposes a metal powder quality evaluation method driven by machine learning, including the following steps:

[0018] Prepare a sufficient number of batches of metal powder with different process parameters, sample and make samples of each batch of metal powder to meet the sample size requirements of CT equipment or SXCT equipment;

[0019] Then, use the CT or SXCT method to observe the metal powder specimens. Based on the minimum size of the powder, set the observation parameters. Reconstruct the three-dimensional data obtained from the CT or SXCT observations to obtain the three-dimensional images of the metal powder specimens; the three-dimensional images of each metal powder specimen serve as the training samples for machine learning;

[0020] Extract the fluidity, bulk density, and porosity of each metal powder specimen through the analysis of the three-dimensional images. Obtain the purity of each metal powder specimen through the electrochemical method, obtain the oxygen content of each metal powder specimen through the flash and extinction method, obtain the compressibility of each metal powder specimen through the pressure test, and analyze and extract the impurity content of each metal powder specimen through the mass spectrometry method; the fluidity, bulk density, purity, porosity, oxygen content, compressibility, and impurity content are all quality characterization parameters of the metal powder;

[0021] According to the requirements, set the evaluation criteria for each quality characterization parameter respectively. Compare the evaluation criteria with the quality characterization parameters of each metal powder specimen, and evaluate the metal powder into three categories: excellent, good, and unqualified. This evaluation result is the sample label of each specimen;

[0022] Construct a metal powder quality evaluation model based on a neural network. The model is used to output the evaluation result (excellent, good, and unqualified) according to the three-dimensional image of the metal powder and the evaluation criteria. Use the three-dimensional image and sample label of the metal powder sample as the training data set for forward propagation training until the output error meets the requirements or reaches the set maximum number of training times;

[0023] Based on the probability sampling method, prepare the samples and scan the three-dimensional images of the metal powder to be measured. Input the three-dimensional image of the metal powder to be measured into the trained metal powder quality evaluation model to obtain the evaluation result of the model.

[0024] The metal powder mentioned above includes powders containing metal elements, such as ferroalloys, aluminum alloys, titanium alloys, copper alloys, and various metal matrix composites.

[0025] The CT or SXCT device mentioned above is all devices that can perform three-dimensional imaging analysis on specimens, including CT instruments using various light sources.

[0026] The sufficient number of batches of metal powder mentioned above should be able to support the completion of the training of the metal powder quality evaluation model.

[0027] The powder quality evaluation characterization parameters mentioned above include fluidity, bulk density, purity, oxygen content, compressibility, porosity, and impurity content.

[0028] Among them, the probability theory sampling method is mainly based on the central limit theorem. After determining the confidence interval and error according to the actual powder quantity, the corresponding required sample quantity can be obtained to reflect the overall powder making situation.

[0029] The present invention uses a machine learning model based on neural network to evaluate the quality of metal powders, which can be used to guide the metal powder preparation process, optimize the metal powder preparation equipment, thereby improving the powder quality and output, and providing high-quality and low-cost powders for additive manufacturing. The metal powder preparation process described above can be the current atomization powder making process, including water atomization, gas atomization, vacuum induction gas atomization, plasma atomization, rotating electrode atomization, etc. Guiding the metal powder preparation process means adjusting the process parameters that affect the powder quality and can be controlled, such as gas pressure, metal liquid temperature and flow rate, etc.; optimizing the metal powder preparation equipment includes adjusting the nozzle structure of the powder making equipment and additional equipment.

[0030] Example 2

[0031] Taking nickel-based superalloy powders and titanium alloy powders as examples, the method described in Example 1 is used to evaluate the quality of metal powders, which specifically includes the following steps:

[0032] I. Sampling and sample preparation:

[0033] Multiple batches of nickel-based superalloy powders and titanium alloy powders are prepared by using different production processes. Considering the stability of the production process and the differences between batches, multi-point random sampling is carried out from different positions (such as the top, middle, bottom and different storage areas, etc.) of the finished powder of each batch. An appropriate amount of powder is taken out from each sampling point, and after full mixing, finally, it is ensured that the mass of each powder sample for detection is about 5 grams. Such a sampling method aims to ensure to the greatest extent that the sample can truly reflect the overall quality characteristics of the powder of this batch.

[0034] For example, for a batch of titanium alloy powders with a total amount of 500 kg, 10 sampling points are set at the four corners, the center position and different height levels of its packaging container, etc. About 10 grams of powder is taken out from each sampling point, and after mixing evenly, 5 grams is accurately weighed from it to cover the possible quality fluctuations.

[0035] A cylindrical sample container is selected, and the container material is polytetrafluoroethylene; the weighed metal powder is slowly poured into the container, and the powder is gently stirred to make it evenly distributed in the container, simulating the natural accumulation state of the powder in the actual application scenario (such as the powder spreading process in additive manufacturing); subsequently, a compaction device with a pressure sensor is used to compact the powder according to the preset pressure value (this pressure value is determined through multiple experimental comparisons, which can ensure that the powder is relatively stable and convenient for scanning, and will not overly change its original loose structure), and multiple metal powder samples meeting the CT scanning requirements are made.

[0036] II. CT Scanning:

[0037] In this embodiment, a CT scanner is used to perform CT scanning on each metal powder sample, and the CT scanning parameters are set as follows:

[0038] For nickel-based superalloy powders, since they contain multiple high-density alloying elements and have a relatively strong X-ray absorption ability, the scanning voltage is set to 140 kV and the scanning current is set to 100 μA. Such a parameter combination can ensure that the X-ray has sufficient energy to penetrate the powder particles, and at the same time, it can form a clear and well-contrasted image on the detector, avoiding the situation of blurred images and inability to accurately distinguish the internal structure due to insufficient voltage or current. For titanium alloy powders, considering their relatively low density and X-ray absorption characteristics, the scanning voltage is set to 120 kV and the scanning current is set to 80 μA. Through this targeted parameter adjustment, the scanning image can clearly present the details of the powder particles and complete the detection within a reasonable radiation dose range, reducing excessive wear on the equipment and potential safety hazards. In order to accurately observe extremely small undulations, edges and corners on the surface of the powder particles and possible micron-sized holes and inclusions inside, the scanning resolution is set to 5 μm; the scanning angle range is set to 360°, and the scanning step is set to 0.5°.

[0039] III. Data Analysis and Quality Assessment:

[0040] Based on the three-dimensional images obtained by CT scanner scanning, a supporting analysis software is used to reconstruct the three-dimensional model, and statistical analysis is carried out on the fluidity, bulk density, purity and porosity of the metal powder samples. After the analysis is completed, each metal powder sample is tested to obtain its oxygen content, compressibility and impurity content.

[0041] The fluidity of metal powder refers to its flow performance during the processing, the bulk density refers to the mass of metal powder per unit volume, and the porosity is the ratio of the volume occupied by pores in the metal powder sample to the total volume of the sample, which is an important influencing factor for the filling and compression properties of metal powder. The fluidity, bulk density and porosity can all be obtained by analyzing the three-dimensional images of the metal powder samples using the analysis software.

[0042] The purity, oxygen content, compressibility and impurity content can be obtained by performing performance characterization tests on each metal powder.

[0043] The purity is obtained by electrochemical analysis method. Specifically, the metal purity is determined by measuring the oxidation-reduction current of metal ions on the electrode, or by measuring the potential change of metal ions on the electrode through voltammetry. The oxygen content is obtained by reduction method. Specifically, in an inert atmosphere, the metal powder is heated to a high temperature, so that the oxygen in it reacts with carbon to generate gases such as carbon monoxide, and the oxygen content is determined by detecting the amount of the generated gas. The compactibility is evaluated by putting the metal powder into a specific mold and compressing it under a certain pressure, and measuring parameters such as the compacted density and height change of the powder under different pressures. The impurity content is obtained by inductively coupled plasma mass spectrometry (ICP-MS), which can realize rapid trace quantitative detection of multiple inorganic elements and isotopes simultaneously.

[0044] IV. Establish sample labels:

[0045] According to the requirements, evaluation criteria are respectively set for each quality characterization parameter. By comparing the evaluation criteria with the quality characterization parameters of each metal powder sample, the metal powder is evaluated into three categories: excellent, good and unqualified. This evaluation result is the sample label of each sample.

[0046] In this embodiment, the evaluation criteria set for nickel-based superalloy powder and titanium alloy powder are based on national standards, including excellent criteria and good criteria. When all the quality characterization parameters of the sample meet the excellent criteria, it is evaluated as excellent; when one or more of the quality characterization parameters of the sample do not meet the excellent criteria but all meet the good criteria, it is evaluated as good; if one or more do not meet, it is evaluated as unqualified.

[0047] V. Model construction and training:

[0048] Construct a metal powder quality evaluation model based on neural network. The model is used to output the evaluation result (excellent, good and unqualified) according to the three-dimensional image of the metal powder. The three-dimensional image and sample label of the metal powder sample are used as the training data set for forward propagation training until the output error meets the requirements or reaches the set maximum number of training times. The metal powder quality evaluation model based on neural network uses a deep learning classification model, such as a multi-layer perceptron, including an input layer, a hidden layer and an output layer, relying on its powerful non-linear fitting ability to learn the mapping relationship between parameters and classification results, and then giving feedback to optimize the process and parameters based on the feedback data.

[0049] Specifically, it includes the following steps:

[0050] Construct a metal powder quality evaluation model based on neural network, and set the network parameters of the metal powder quality evaluation method model. The network parameters include the maximum number of training times, learning accuracy, learning rate, initial weight and iteration stop condition.

[0051] After normalizing the training data, it is input into the model. Cross-entropy is used as the loss function for forward propagation training. During each training, the training error is used to adjust and update the weights using the backpropagation formula until the adjusted weights can make the output error meet the set learning accuracy requirement, or reach the maximum number of training times, then the model training is completed.

[0052] The processing flow of the input data by the metal powder quality evaluation model based on the neural network is as Figure 2 shown. The specific network layers included in the model are as follows:

[0053] Input layer: Receives the three-dimensional image of the metal powder and the evaluation criteria.

[0054] Hidden layer: Includes convolutional layer, pooling layer, fully connected layer, batch normalization layer, and Dropout layer. The three-dimensional image is first normalized through the batch normalization layer to accelerate model convergence, reduce the problem of gradient vanishing or explosion, and improve the stability and generalization ability of the model. The convolutional layer extracts local features in the normalized three-dimensional image, such as information about the edges and textures of particles, which helps to evaluate the shape and surface quality of the powder. The pooling layer downsamples the feature map extracted by the convolutional layer, reduces the data dimension, retains the main features, reduces the model calculation amount, and can improve the robustness of the model to a certain extent. The fully connected layer further performs non-linear transformation and combination on the features after convolution and pooling, maps the features to a higher-dimensional space, and obtains high-dimensional features. The Dropout layer is used to randomly set the outputs of some neurons in the high-dimensional features to 0 with a certain probability, thereby preventing the model from overfitting, enabling the model to learn more robust features, and improving the generalization performance of the model.

[0055] Output layer: Receives the image features finally output by the hidden layer and the evaluation criteria transmitted from the input layer, combines the image features with the evaluation criteria to perform the quality evaluation task, and classifies the quality of the metal powder into different grades, such as excellent, good, and unqualified.

[0056] Six: Final evaluation:

[0057] Sample and scan the three-dimensional image of the metal powder to be tested based on the probability sampling method, input the three-dimensional image of the metal powder to be tested into the trained metal powder quality evaluation model, and obtain the evaluation result of the model.

[0058] The above probability sampling method is specifically as follows:

[0059] Based on the central limit theorem, determine the number N of metal powders to be evaluated. By adopting the method of random sampling, a certain number of metal powder particles are extracted to ensure that the probability of each sample being extracted is equal.

[0060] Set the confidence level at 95%, and the corresponding Z value is 1.96. The confidence level represents the probability that the population parameter falls within the confidence interval.

[0061] Set the population standard deviation σ according to the historical data of the powder-making process and the standard metal powder quality parameters; calculate the error E of the sample mean according to the formula E = Z×σ / √n.

[0062] Under the condition that the number N of metal powders and the population standard deviation σ are determined, find the error range through the Z value corresponding to the selected confidence level.

[0063] The above-described embodiments only represent several implementation manners of the present invention and are not used to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, for those skilled in the art, they can still modify the technical solutions recorded in the foregoing embodiments or perform equivalent replacements for some of the technical features. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A metal powder quality assessment method driven by machine learning, characterized in that: include: S1: Acquire multiple metal powder samples and their three-dimensional images; Analyzing the quality characterization parameters of each metal powder sample, wherein the quality characterization parameters include fluidity, bulk density, purity, porosity, oxygen content, compactness and impurity content; wherein the fluidity, bulk density and porosity of the metal powder sample are obtained by three-dimensional image analysis; S2: setting evaluation criteria for quality characterization parameters; comparing the quality characterization parameters of the metal powder samples with the evaluation criteria, and evaluating each metal powder sample as excellent, good, or unqualified; S3: constructing a metal powder quality assessment model based on a neural network, wherein the model is used to output an assessment result according to a three-dimensional image of the metal powder, using the three-dimensional image of each metal powder sample as a training sample, and the assessment result of each metal powder sample in S2 as a sample label, and performing forward propagation training on the model until the output error meets the requirement or reaches the set maximum number of training times; S4: Obtain a three-dimensional image of the metal powder to be tested, input the three-dimensional image into the trained metal powder quality assessment model, and obtain an assessment result of the model.

2. The metal powder quality assessment method according to claim 1, characterized in that: In S1, the specific method of obtaining multiple metal powder samples is: Multi-point random sampling is performed from different positions of different batches of metal powders, the powders taken from different sampling points of the same batch are collected and fully mixed, the mixed powders are sampled again to make the weights of samples from different batches the same, and the obtained powders are compacted in a cylindrical container at the same pressure value to obtain multiple metal powder samples.

3. The metal powder quality assessment method according to claim 1, characterized in that: In S1, the specific method of obtaining the three-dimensional image of the metal powder sample is: performing CT scanning on each metal powder sample to obtain the three-dimensional image of each metal powder sample.

4. The metal powder quality assessment method according to claim 1, characterized in that: In S1, the fluidity, packing density and porosity of the metal powder sample are obtained by three-dimensional image analysis. The specific method is: input the three-dimensional image into the analysis software supporting the CT scanning equipment, reconstruct the three-dimensional model of the metal powder sample by the software, and obtain the fluidity, packing density and porosity of the metal powder sample based on the three-dimensional model analysis.

5. The metal powder quality assessment method according to claim 1, characterized in that: In S1, the analysis methods of purity, oxygen content, compactness and impurity content are specifically as follows: the purity is obtained by electrochemical analysis, the oxygen content is obtained by reduction analysis, the compactness is obtained by comprehensively evaluating the compaction density and height change of the metal powder under different pressures, and the impurity content is obtained by mass spectrometry analysis.

6. The metal powder quality assessment method according to claim 1, characterized in that: In S3, the neural network-based metal powder quality assessment model includes an input layer, a hidden layer and an output layer, the input layer receives a three-dimensional image of the metal powder and a set assessment standard, the hidden layer extracts features from the three-dimensional image, and the output layer combines the assessment standard and the features extracted by the hidden layer to classify the metal powder as excellent, good or unqualified.

7. The metal powder quality assessment method according to claim 6, characterized in that: The hidden layer includes a convolution layer, a pooling layer, a fully connected layer, a batch normalization layer and a Dropout layer; the three-dimensional image is first normalized by the batch normalization layer; the convolution layer extracts local features in the normalized three-dimensional image; the pooling layer downsamples the local features extracted by the convolution layer; the fully connected layer further nonlinearly transforms and combines the features after convolution and pooling, maps the features to a higher-dimensional space, and obtains high-dimensional features; The high-dimensional features are processed by the Dropout layer to obtain the output of the hidden layer.

8. The metal powder quality assessment method according to claim 1, characterized in that: The method of obtaining a three-dimensional image of the metal powder to be tested specifically includes: preparing a sample of the metal powder to be tested based on a probabilistic sampling method, and obtaining a three-dimensional image of the prepared metal powder sample to be tested by CT scanning.

Images