Quality and performance prediction method and prediction system of hydrogenated and dehydrogenated titanium-based raw material for MIM (Metal Injection Molding)

Abstract

The invention relates to the technical field of analyzing and predicting the performance of a molded product based on the property of a raw material, and particularly provides a quality and performance prediction method and prediction system of a hydrogenated and dehydrogenated titanium-based raw material for MIM (Metal Injection Molding). The prediction method is based on parameters, such as particle size D90, average roundness, angle of repose, Karl index, oxygen content, Ti < 4 + > content proportion and powder critical loading capacity, which are obtained by a specific detection method and have representativeness of an HDH-Ti raw material, performance data, such as tensile strength, of an MIM titanium-based molded product, and correlation coefficients of the parameters and the performance data; the model is trained through a machine learning method and a reverse neural network to obtain a data prediction model, a grading evaluation scheme and a result evaluation scheme generated by the model obtained by the former can more accurately judge the quality grade of the raw material, and the latter can estimate the performance data of the MIM molded product; the method and the system are not limited by the source and the component of the HDH-Ti raw material, and the quality of the raw material and the performance prediction of the MIM molding product can be quickly realized.

Description

Technical Field

[0001] The present application relates to the technical field of machine learning and neural network models, and specifically provides a method and a system for predicting the quality and performance of a hydrogenated dehydrogenated titanium-based raw material for MIM. Background Art

[0002] Metal injection molding (MIM) is a near-net shaping technology that combines traditional powder metallurgy technology and plastic injection molding technology. It is suitable for mass production of small and complex-shaped parts. The main processes include: mixing, granulation, injection molding, debinding, sintering, and post-processing. MIM combines the advantages of traditional powder metallurgy and plastic injection molding technologies, and has characteristics such as good forming ability, excellent product performance, high surface finish of products, high dimensional accuracy, wide application range, and low production cost, breaking through the limitations of traditional metal powder die pressing forming processes in product shape.

[0003] Metal powder is the basis for ensuring the MIM process. Ideal powder has the following characteristics: (1) an appropriate particle size distribution to ensure a high powder loading; (2) no particle agglomeration and no internal pores; for rapid sintering, the particle size is preferably less than 20 μm; (3) a spherical or equiaxed shape; (4) sufficient inter-particle friction to maintain the shape of the debound blank; (5) non-toxic, clean surface, and no harmful reaction with the binder.

[0004] At present, the stainless steel MIM process is quite mature, while the MIM forming of titanium alloys with higher added value, higher specific strength, better corrosion resistance, lighter material, and better biocompatibility is restricted by lower apparent density, higher impurity content, larger shrinkage rate, and higher-cost spherical raw materials, and the development is relatively slow. Compared with expensive gas atomized spherical titanium powder, the HDH-Ti powder (i.e., hydrogenated dehydrogenated titanium powder) produced by the hydrogenation dehydrogenation process has become the most potential raw material in the MIM process due to its simple preparation, fine particle size, and low price. However, HDH-Ti has defects such as irregular shape, high impurity content, poor fluidity, uneven shrinkage deformation, and excessive oxygen concentration, resulting in poor mechanical elongation of the workpieces prepared therefrom. M.S. Moghadam et al. used inexpensive high-oxygen HDH-Ti powder as the raw material and prepared, through MIM, a material with a hardness of 490 HV and a density of 4.1 g / cm 3For products, due to the excessive oxygen content, the alloy has excellent hardness but poor ductility; when the oxygen concentration increases from 0.32 wt% to 0.45 wt%, the elongation rate of TC4 alloy drops sharply from >10% to <5% (Fabrication of titanium components by low-pressure powder injection moulding using hydride-dehydride titanium powder, Powde Technology, 377, 70-79, doi:10.1016 / j.powtec.2020.08.075.). In addition, the mixing and granulation of MIM are processes of multiple kneading and forming of metal powder and polymer systems, and the performance of the mixture directly determines the quality of the subsequent injection-molded workpieces; therefore, for HDH-Ti powder, controlling its solid content, viscosity, plasticity, density, fluidity, etc. are all necessary steps in the production process.

[0005] Ideal injection molding powders have requirements in many aspects such as particle size distribution, particle morphology, particle size and shape, friction force, surface state, oxygen impurity content, reaction with binder, etc. In particular, a balance needs to be found between fluidity and shape retention. For example, a large specific surface area not only leads to poor fluidity but also improves shape retention. Therefore, it is obviously inaccurate to judge the performance of raw material powders from a single index, and the MIM process parameters have too much influence and are mutually coupled; there is still a lack of an evaluation system for reasonably predicting the quality of HDH-Ti raw materials by establishing the relationship between HDH-Ti raw material parameters and the performance of MIM titanium-based products prepared from them. Summary of the Invention

[0006] To solve the above technical problems, the present invention first provides a method for predicting the quality and performance of a hydrided-dehydrided titanium-based raw material for MIM (injection molding), including the following steps: S1. Detect the HDH-Ti raw material parameters to obtain a raw material parameter data set; S2. Detect the performance data of the MIM titanium-based molded products of the HDH-Ti raw materials to obtain an initial performance data set; S3. Use the random forest machine learning algorithm and the particle swarm optimization-based backpropagation neural network algorithm (PSO-BPNN) to process the data sets in S1 and S2, construct a model and perform training and testing; S4. Input the parameters of the raw material to be tested, and use the random forest prediction data model to obtain the predicted quality grade of the raw material to be tested; S5. Input the parameters of the raw materials to be tested, and use the trained particle swarm optimization-based backpropagation neural network (PSO-BPNN) model to predict the performance evaluation data of MIM titanium-based formed products, and determine whether they meet the requirements for MIM forming.

[0007] Further, the HDH-Ti raw material parameters include particle size D90, average roundness, angle of repose, Carr index, oxygen content, Ti 4+ content ratio, and powder critical loading.

[0008] Further, the performance data of the MIM titanium-based formed products is the tensile strength.

[0009] Further, the test of the HDH-Ti raw material particle size D90 needs to be carried out according to the national standard "GB / T 19077-2016 Particle Size Distribution - Laser Diffraction Method".

[0010] Further, the average roundness of the HDH-Ti raw material is tested by the method of combining scanning electron microscope photos with image analysis software, and the calculation is carried out according to the following formula (I): ………… Formula (I); In the formula: R - roundness; A - cross-sectional area of the particles on the observation surface, mm 2 ; d max - the longest diameter of the particles on the observation surface, mm.

[0011] Further, the average roundness should at least take the average sphericity of 200 powder particles; the roundness of an ideal sphere is 1.

[0012] Further, the angle of repose of the HDH-Ti raw material is tested according to the content specified in Section 4.5 of "GB / T 16913-2008 Test Methods for Physical Properties of Dust".

[0013] Further, the calculation of the Carr index is carried out according to the following formula (II): ………… Formula (II); In the formula, ρ t is the tapped density of the powder, ρ a is the loose bulk density of the powder, ρ t and ρ b are measured according to the provisions of Part 1 Funnel Method of "GB / T 1479.1-2011 Determination of Loose Bulk Density of Metal Powders" and "GB / T 5162-2006 Determination of Tapped Density of Metal Powders".

[0014] Further, the oxygen content of the HDH-Ti raw material was tested in accordance with "Methods for Chemical Analysis of Sponge Titanium, Titanium and Titanium Alloys - Part 7: 2011".

[0015] Further, the Ti content ratio of the HDH-Ti raw material was calculated based on the XPS test spectrum. 4+ Specifically: Using the monochromatic Al Ka spectral line of the x-ray source, the raw material powder was ion-etched at a voltage of 15 kV and a current of 20 mA, and the photoelectron spectrum of titanium under 5 nm on the surface of the powder was tested; the spectrum was calibrated according to the binding energy of C 1s; finally, the ratio of tetravalent titanium was calculated by XPS peak fitting and based on the integral area.

[0016] Further, the calculation formula for the critical powder loading amount is Formula (III): …………Formula (III); In the formula, ρ b and ρ p are the binder density and the theoretical density of the powder, and w b and w p are the masses of the binder and the powder.

[0017] Among them, the calculation formula for the binder density is Formula (IV): …………Formula (IV); In the formula, w i and ρ i are the mass fraction and density of the i-th component in the binder.

[0018] Further, the binder is a five-component basic binder system, and the composition components and densities of the binder are shown in Table 1 below: Table 1

[0019] After calculation, the density of the five-component basic binder is = 1.33 g / cm 3 .

[0020] Further, after obtaining the raw material parameters in S1, it also includes the step of normalizing the raw material parameters.

[0021] Further, the raw material parameters in S1 include positive indicators and negative indicators. The positive indicators include average roundness and critical powder loading amount, and the negative indicators include particle size D90, angle of repose, Carr index, oxygen content, and Ti 4+ content ratio. The normalization method is as follows: Furthermore, the normalization method for the positive indicators is carried out according to Formula (V): ………… Formula (V); Furthermore, the normalization method for negative indicators is carried out according to Formula (VI): …………Formula (VI).

[0022] In the above normalization methods for positive and negative indicators, max a refers to the maximum threshold of any parameter a, and min a refers to the minimum threshold of any parameter a, and m n refers to the actual value of any parameter a, and x n refers to the value of the actual value of any parameter a after normalization.

[0023] Furthermore, the raw material parameter dataset consists of no less than 100 groups of HDH-Ti raw material parameters.

[0024] Furthermore, the HDH-Ti raw materials can be selected from HDH-Ti raw materials of different manufacturers, different models, different batches, HDH-Ti raw materials mixed in any proportion, or mixed Ti raw materials composed of HDH-Ti raw materials and gas atomized spherical titanium powder with the same composition in any proportion.

[0025] Furthermore, the tensile strength of the MIM titanium-based formed product in S2 is obtained by testing according to "GB / T 228.1-2021 Metallic materials - Tensile testing - Part 1: Method of test at room temperature".

[0026] Furthermore, the preparation method of the MIM titanium-based formed product in S2 is as follows: Mix and load the HDH-Ti raw materials and the binder according to the proportion of 97% powder critical loading amount, heat up to 180°C for mixing on an argon-protected mixer, granulate and then inject and mold the green body on an injection molding machine, the injection temperature is 180°C, and the injection pressure is 80 Mpa. After completion, degrease at 120°C, and sinter in a high-vacuum sintering furnace above 10 -3 pa, keep the temperature at 600°C for 2 h to remove the residual polymer, and then sinter at 1500°C for 2 h to obtain the finished product.

[0027] Furthermore, in step S3, use the random forest machine learning algorithm and the PSO-BPNN algorithm based on particle swarm optimization to process the normalized S1 and S2 datasets, construct a data model and conduct training and testing; Furthermore, in step S4, the model based on the machine learning random forest algorithm constructs a multi-decision tree model, conducts supervised training and prediction on the multi-dimensional feature data of the HDH-Ti raw material parameters, and realizes a quality grade classification evaluation with higher precision and recall rate; Furthermore, S4 specifically includes: S41. Collect multi-dimensional parameter data from the production process of HDH-Ti raw materials, including but not limited to chemical composition (such as oxygen content, Ti 4+ Content ratio), physical properties (such as particle size D90, average roundness, angle of repose, Carr index, critical powder loading). The collected data is cleaned, outliers are removed, and the quality level of the target variable (such as "A", "B", "C" and "unqualified") is labeled. The cleaned data is divided into training set, validation set and test set in a ratio of 8:1:1.

[0028] S42, initialize the data model, set hyperparameters; perform bootstrap sampling on the training set to generate multiple different subsets, each of which is used to train a decision tree. In addition, in the process of building the decision tree, randomly select some features at each node, and perform splitting operations based on the criteria of maximizing the Gini coefficient and information gain until the stopping condition is met.

[0029] S43, using the training set to build multiple decision trees to form a random forest classifier. Then, the test set data is classified and predicted, and finally a voting mechanism is used to make a comprehensive decision on the classification results of multiple decision trees to determine the final quality level.

[0030] S44. Use the confusion matrix to evaluate the classification results and calculate the accuracy, precision, recall and F1 score. Based on the error analysis of the classification results, optimize the hyperparameters of the random forest model (including the number of decision trees, maximum tree depth, etc.).

[0031] S45. Use the feature importance evaluation function of the random forest model to quantify the influence of each input parameter on the classification result. Based on the results of feature importance analysis, optimize the production process and adjust the parameters that have a greater impact on quality, such as controlling oxygen content or adjusting tensile strength.

[0032] S46. Deploy the optimized random forest model in the online quality inspection system to classify the newly input HDH-Ti raw material parameters in real time and output the quality grade evaluation results.

[0033] Furthermore, in step S5, the trained PSO-BPNN neural network model is used to predict the performance evaluation data of the MIM titanium-based molded product to determine whether it meets the MIM molding requirements. The method for iterative training of the PSO-BPNN neural network model is as follows: S51. Collect multi-dimensional parameter data from the production process of HDH-Ti raw materials, including but not limited to chemical composition (such as oxygen content, Ti 4+Content ratio), physical properties (such as particle size D90, average roundness, angle of repose, Carr index, critical powder loading). The collected data was cleaned, outliers were removed, and numerical labels of tensile strength were annotated. The cleaned data was divided into training set, validation set and test set in a ratio of 8:1:1.

[0034] S52. Construct a back propagation neural network, which includes an input layer, a hidden layer, and an output layer. The number of neurons in the input layer is equal to the number of input features (number of parameters), and the number of neurons in the output layer is 1, representing the predicted tensile strength. Select an appropriate number of hidden layers and neurons according to the specific problem. The number of neurons in the hidden layer can be determined by methods such as cross-validation. The hidden layer uses nonlinear activation functions such as Sigmoid, ReLU, or Tanh to enhance the ability of the neural network to handle nonlinear problems.

[0035] S53, initialize the particle swarm algorithm parameters, including the initial position and speed of the particle swarm, and calculate the fitness function of the particles. The fitness of each particle is determined by the error function of the neural network. The commonly used error function is the mean square error (MSE), that is: ………… Formula (VII); Where n is the total number of samples, is the true value of the i-th experimental sample, is the predicted value of the ith experimental sample.

[0036] S54, in each iteration, calculate the fitness value of each particle. If the fitness value of the current particle is better than the fitness value of the previous iteration, then update the individual best position of the particle (i.e., the best position of the current particle). If the current fitness value is not better than the fitness value of the previous iteration, then keep the individual best position in the previous iteration unchanged.

[0037] S55, the particle swarm optimizes the position and speed through continuous iteration until the predetermined stopping condition is reached (such as the maximum number of iterations or the fitness value converges), and the optimal weight and threshold parameters of the BP neural network are obtained, thereby obtaining the PSO-BP neural network model.

[0038] Particle velocity update formula: ………… Formula (VIII); in, is the speed of particle (i) at time (t), which determines the direction and step size of the next movement of the particle; w is the inertia weight, which simulates the "inertia" in physics and controls the tendency of the particle to maintain its original speed; c 1 and c 2is the acceleration factor, which respectively adjusts the step sizes for the particle to learn from its individual historical best position and the swarm historical best position; r 1 and r 2 are random numbers uniformly distributed within [0, 1], with the physical meaning of introducing search randomness; is the historical best position of particle (i), whose physical meaning is the best solution found by the particle individual during the search process; is the swarm historical best position, whose physical meaning is the best solution found by the entire population so far; is the position of particle (i) at time (t), representing the coordinates of the current solution in the search space; is the velocity of particle (i) at time (t + 1).

[0039] Particle position update formula: ………… Formula (IX); wherein, is the position of particle (i) at time (t + 1).

[0040] Finally, the present invention also provides a quality and performance prediction system for the hydrogenated dehydrogenated titanium-based raw material for MIM, including a memory 41 and a processor 42. The memory includes a data input module 411, a normalization module 412, a raw material parameter library 413, and a performance database 414; the processor 42 includes an output module 421, a grading determination module 422, a prediction data module 423, and a simulation training optimization module 424.

[0041] The data input module 411 is an input unit for the particle size D90, average roundness, angle of repose, Carr index, oxygen content, Ti 4+ content ratio, and powder critical loading amount of the HDH-Ti raw material, the mixture of the HDH-Ti raw material and the gas atomized titanium powder of the same component in any ratio, and the subsequent new samples, as well as an input unit for the performance data of the MIM titanium-based formed product.

[0042] The normalization module 412 is a normalization processing unit for all parameters of the HDH-Ti raw material.

[0043] The raw material parameter library 413 is used to store the parameter data of the HDH-Ti raw material, the mixture of the HDH-Ti raw material and the gas atomized titanium powder of the same component in any ratio, and the parameter data of the subsequent new samples.

[0044] The performance database 414 is used to store the performance data of the MIM formed product.

[0045] The prediction data module 423 predicts the quality grade of the raw material sample and the performance evaluation data of its MIM titanium-based formed product based on the data prediction model.

[0046] When the processor 42 executes the program, the following steps are implemented: (1) Input the HDH-Ti raw material parameters through the data input module 411 and store them in the raw material parameter library 413; (2) Normalize the raw material parameters through the normalization module 412 to obtain a raw material parameter data set; (3) Input the performance data of the MIM titanium-based formed product into the performance database 414 to obtain an initial performance data set; (4) Through the simulation training optimization module 424, perform machine learning and neural network training to optimize the basic prediction data model in the prediction data module 423; (5) Input new sample parameters, use the output module 421 to predict the quality grade of the raw material and the performance evaluation data of its MIM titanium-based formed product, and judge and output whether the raw material meets the requirements for MIM forming.

[0047] Furthermore, the basic prediction data model evaluates the raw material performance indicators as first-class, second-class, third-class, and defective products. Among them, the first-class raw materials can be directly put into MIM forming for use, the second-class raw materials need to be subjected to a final sintering test for determination, the third-class raw materials need to adjust the MIM process and then conduct a sintering test for determination, and the defective products cannot be used for MIM forming.

[0048] Furthermore, the grading and determination module of the basic prediction data model proceeds according to the following process: Normalize the original HDH-Ti raw material parameter data, input it into the hydrogenation and dehydrogenation titanium-based raw material quality detection system based on the random forest model, then perform multi-decision tree prediction, and finally output the quality grade through a voting mechanism. At the same time, input the tensile strength prediction system of the MIM titanium-based formed product based on the PSO-BPNN model, and output the tensile strength value through the neural network based on particle swarm optimization.

[0049] Compared with the prior art, the present invention has the following beneficial effects: The present invention creatively uses 7 parameter indicators that are representative and obtained through specific detection methods to represent dozens of influencing parameters involved in the particle size distribution, particle morphology, particle size and shape, friction force, surface state, reaction with the binder, oxygen content, and MIM forming process of the original powder, greatly reducing the detection workload of operators, greatly reducing the interference between various parameters, greatly reducing production costs, and improving production efficiency.

[0050] The present invention takes seven parameter indicators of HDH-Ti raw materials as independent variables, and establishes and optimizes a machine learning model through the random forest algorithm, which can accurately predict the MIM molding possibility of HDH-Ti raw materials, greatly improve the use efficiency of products, and can adjust the loading and kneading process, providing data guidance for low-cost and high-quality MIM molding. By training the backpropagation neural network PSO-BPNN model, the tensile strength of MIM titanium-based molded products can be quickly predicted, with high prediction accuracy and stable prediction results.

[0051] The test method provided by the present invention has strong compatibility with HDH-Ti raw materials, is not restricted by the sources and components of HDH-Ti raw materials, and can uniformly determine the qualification of HDH-Ti raw materials from different manufacturers, different batches, and different specifications, as well as mixed raw materials configured with spherical titanium powder in any proportion.

[0052] The quality and performance prediction system provided by the present invention can quickly determine the performance of raw materials. Especially for HDH-Ti raw materials with great irregularity and high oxygen characteristics, they can be accurately predicted by this system and will not be directly determined as unqualified. Description of the Drawings

[0053] Appendix Figure 1 : Flowchart for establishing the test method of the hydrogenated dehydrogenated titanium-based raw material of the present invention.

[0054] Appendix Figure 2 : Ti 4+ Schematic diagram of the test analysis of the content ratio.

[0055] Appendix Figure 3 : Flowchart for the operation of the random forest quality classification determination module of the machine learning algorithm of the present invention.

[0056] Appendix Figure 4 : Block diagram of the quality and performance prediction system of the present invention; Appendix Figure 4 Among them, memory - 41, processor - 42, data input module - 411, normalization module - 412, raw material parameter library - 413, performance database - 414, output module - 421, classification determination module - 422, prediction data module - 423, simulation training optimization module - 424.

[0057] Appendix Figure 5 : Flowchart for establishing the particle swarm optimization backpropagation neural network model of the present invention. Detailed Embodiments

[0058] Embodiment This embodiment first provides a method for predicting the quality and performance of HDH-Ti powder for MIM, including the following steps: S1. Detect the parameters of HDH-Ti raw materials to obtain a raw material parameter data set; S2. Detect the performance data of the MIM titanium-based formed products of the HDH-Ti raw materials to obtain an initial performance data set; S3. Use the random forest machine learning algorithm and the particle swarm optimization-based backpropagation neural network algorithm (PSO-BPNN) to process the data sets of S1 and S2, construct a model and conduct training and testing; S4. Input the parameters of the raw material to be tested, use the prediction data model to obtain the predicted quality grade of the raw material to be tested and the performance evaluation data of its MIM titanium-based formed products, and judge whether it meets the requirements for MIM forming. See the attached Figure 1 .

[0059] The raw material parameters include the following 7 types: (1) Particle size D90: According to the characteristics of injection molding process, the particle size of the raw material cannot be too large, otherwise it is difficult to sinter effectively. If the particle size is too small, the viscosity of the bonding system will increase, the pores in the formed product will increase, and the density will decrease. The particle size is tested by a laser particle size analyzer according to the national standard "GB / T 19077-2016 Particle Size Distribution - Laser Diffraction Method".

[0060] (2) Average roundness: On the observation surface, the ratio of the actual cross-sectional area of the particle to the area corresponding to the longest diameter of the particle is the roundness. Taking the average value of a certain number of particles within a certain range is the average roundness. The higher the roundness of the particles, the greater the fluidity, the lower the corresponding friction force, and the worse the shape retention. Therefore, particles with a certain range of roundness are also required to meet the requirements.

[0061] The roundness of the powder is measured by the method of combining scanning electron microscope photos with image analysis software. The roundness is calculated according to the following formula (1): …………Formula (1) In the formula: R - roundness; A - the cross-sectional area of the particle on the observation surface, mm 2 ; d max - the longest diameter of the particle on the observation surface, mm.

[0062] The average roundness should be calculated by checking the roundness of at least 200 powder particles and taking the average value. The roundness of an ideal spherical body is 1.

[0063] (3) Angle of repose: It is obtained by calculating the plane angle of the conical knot formed when the powder flows vertically under the action of gravity. The smaller the angle of repose, the smaller the friction force and the better the fluidity. The angle of repose is related to parameters such as the particle size and its distribution, shape, friction coefficient, adhesion, electrostatic voltage, porosity, compressibility, moisture, temperature, and humidity in the air of the relevant material particles.

[0064] The angle of repose data is tested according to the content specified in Section 4.5 of "GB / T 16913-2008 Test Methods for Physical Properties of Dust".

[0065] (4)Carr index: It characterizes the flowability of powder through its compressibility. The greater the difference between the bulk density and the tapped density, the greater the Carr coefficient, indicating a long compression stroke and good compressibility. A large Carr coefficient also means that the material is relatively loose, the voids between powder particles are larger, the powder particles are lighter, and the flowability is poorer. On the contrary, the flowability is better.

[0066] Measure ρ according to Part 1 Funnel Method of "GB / T 1479.1-2011 Determination of Apparent Density of Metal Powders" and "GB / T 5162-2006 Determination of Tapped Density of Metal Powders". t (Apparent density of powder), ρ a (Bulk density of powder); The calculation formula (II) of the Carr index is as follows: ………… Formula (II)

[0067] (5)Oxygen content: The oxygen content in the powder directly affects the mechanical properties of the final product; Determine the oxygen and nitrogen content according to "GB / T 4698.7-2011 Methods for Chemical Analysis of Sponge Titanium, Titanium and Titanium Alloys".

[0068] (6)Ti 4+ Content ratio: Depending on the different processes and raw materials of hydrogenated dehydrogenated titanium powder, the existence form of oxygen in the final titanium raw material is also different. If oxygen exists in the form of TiO2, it is difficult to remove and affects the mechanical properties of the final product. If it is dissolved in the material lattice or exists in other compound forms, the impact on the final material properties is relatively small. Therefore, the ratio of tetravalent titanium can be used to assist in judging the influence of oxygen on the material properties.

[0069] Calculate the Ti content ratio of the HDH-Ti raw material according to the XPS test spectrum. Specifically: Use the monochromatic AlKa spectrum of the x-ray source to perform ion etching on the raw material powder at a voltage of 15 kV and a current of 20 mA, and test the photoelectron spectrum of titanium under 5 nm on the surface of the powder; Calibrate the spectrum according to the binding energy of C 1s; Finally, perform XPS peak fitting and calculate the ratio of tetravalent titanium according to the integral area. Attached 4+ is the XPS energy spectrum and peak fitting results of typical HDH titanium-based raw material powder. It can be seen from the figure that Ti Figure 2 mainly exists in the form of TiO2, with a binding energy of about 530 eV, which is relatively stable and difficult to remove. 4+

[0070] ​(7) Critical powder loading: A high powder loading can reduce the shrinkage of injection-molded parts, which is beneficial to the control of dimensional accuracy. However, an excessive powder loading leads to an increase in the viscosity of the feedstock, making forming difficult. Therefore, for a certain powder, there is a critical binder content that enables the powder to tightly wrap the powder surface while filling the particle gaps. When the loading exceeds this critical powder loading, there is not enough binder to fill the inter-particle pores, resulting in a decrease in the feedstock density.

[0071] The calculation formula is formula (III): ………… Formula (III); In the formula, ρ b , ρ p are the binder density and the theoretical density of the powder, w b , w p are the mass of the binder and the powder.

[0072] Among them, the calculation formula of the binder density (IV) is: ………… Formula (IV); In the formula, w i , ρ i are the mass fraction and density of the i-th component in the binder.

[0073] If measured using a five-component basic binder system, the binder component composition and density are as follows in the table:

[0074] After calculation, the density of the binder is = 1.33 g / cm 3 .

[0075] The raw material parameters are divided into positive indicators and negative indicators. The positive indicators include roundness and critical powder loading, and the negative indicators include particle size D90, angle of repose, Carr index, oxygen content, and Ti 4+ content ratio.

[0076] Normalizing the raw material parameters in step S1 can greatly improve the performance and accuracy of the system. In particular, it can reduce the difficulty of numerical calculation, accelerate model training, and help accelerate the convergence of the learning algorithm.

[0077] The normalization method for positive indicators is: ………… Formula (V); The normalization method for negative indicators is: ………… Formula (VI); In the above normalization methods for positive and negative indicators, max arefers to the maximum threshold value of any parameter a, min a refers to the minimum threshold value of any parameter a, m n refers to the actual value of any parameter a, x n refers to the value of the actual value of any parameter a after normalization.

[0078] In step S2, the performance of the MIM titanium-based product is the tensile strength.

[0079] The tensile strength of the MIM titanium-based product is the maximum stress value before the metal fractures under tension, also known as the ultimate strength. Its value is obtained by testing with a universal material testing machine in accordance with "GBT 228.1-2021 Metallic materials - Tensile testing - Part 1: Method of test at room temperature".

[0080] To eliminate the interference of factors such as binders, temperature, and equipment parameters during the MIM forming process, the MIM titanium-based products to be tested for tensile strength are all prepared using a unified process. Specifically: The HDH-Ti raw material and the binder are mixed and loaded in accordance with the ratio of 97% of the powder critical loading amount. First, they are mixed on an argon-protected mixer, and then heated to 180°C for mixing. After granulation, a green body is injection-molded on an injection molding machine. The injection temperature is 180°C, and the injection pressure is 80 Mpa. After completion, debinding is carried out at a debinding temperature of 120°C. After debinding, it is sintered in a high-vacuum sintering furnace above 10 -3 pa, held at 600°C for 2 h to remove residual polymers, and then sintered at 1500°C for 2 h to obtain the finished product.

[0081] In step S4, the above-mentioned data set is processed using a machine learning algorithm to construct and train a data model; specifically: Step S41, collect 200 groups of multi-dimensional parameter data from the production process of the HDH-Ti raw material. Clean the collected data. Divide the cleaned data into 160 groups of training sets, 20 groups of validation sets, and 20 groups of test sets, Step S42, initialize the data model and set the following hyperparameters: Number of decision trees (n_estimators): To ensure the stability of the algorithm, the set value is 200.

[0082] Maximum tree depth (max_depth): To prevent overfitting, the set value is 8.

[0083] Minimum number of samples for splitting (min_samples_split): Used to prevent overfitting, the set value is 5.

[0084] Maximum number of features (max_features): Limit the number of features available for node splitting, and the set value is "sqrt".

[0085] During the construction of the decision tree, at each node, a partial set of features is randomly selected, and the splitting operation is performed according to the criterion of minimizing the Gini coefficient until the number of training samples in the child nodes of the decision tree is less than 5, at which point the process stops.

[0086] Step S43: Use the 200 decision trees constructed from the training set to form a random forest classifier. Then, classify and predict the test set data. The final prediction result is obtained through voting, and the output category is output.

[0087] ………… Formula (X); Among them, is the final output result of the random forest model, and y1, y2....y 200 are the output results of each branch of the random forest.

[0088] Step S44: Calculate the accuracy rate A, precision rate P, and recall rate R based on the classification results of the test data. According to the error analysis of the classification results, optimize the hyperparameters of the random forest model (including the number of decision trees, maximum tree depth, etc.).

[0089] ………… Formula (XI); ………… Formula (XII); ………… Formula (XIII); Among them, TP is the true positive, that is, the number of samples correctly predicted as the positive class by the model; TN is the true negative, that is, the number of samples correctly predicted as the negative class by the model; FP is the false positive, the number of samples wrongly predicted as the positive class by the model; FN is the false negative, the number of samples wrongly predicted as the negative class by the model.

[0090] Step S45: Use the feature importance evaluation function of the random forest model to quantify the influence degree of each input parameter on the classification result. Based on the feature importance analysis result, optimize the production process and adjust the parameters that have a greater impact on quality, such as controlling the oxygen content or adjusting the tensile strength.

[0091] Step S46: Deploy the optimized random forest model to the online quality detection system, perform real-time classification on the newly input HDH-Ti raw material parameters, and output the evaluation result of the quality grade. The operation flow chart of the random forest quality grading determination module of the machine learning algorithm is shown in the appendix Figure 3 。

[0092] Step S5 specifically includes: constructing a BPNN neural network model, using the preprocessed data, and iteratively training the BPNN neural network in combination with the particle swarm optimization (PSO) algorithm to obtain an optimized and trained PSO-BPNN neural network model (the establishment flow chart is shown in the appendixFigure 5 ).

[0093] Step S51: Collect 200 groups of multi-dimensional parameter data from the production process of HDH-Ti raw materials. Clean the collected data. Divide the cleaned data into 160 groups of training sets, 20 groups of validation sets, and 20 groups of test sets. Use the training sets to train the PSO-BPNN neural network model, and use the test sets to evaluate the prediction performance of the trained PSO-BPNN neural network.

[0094] Step S52: Construct a PSO-BPNN network model. Use the training sets as the input parameters of the BP neural network model. Set the number of neurons in the hidden layer of the PSO-BPNN neural network model to 18. The hidden layer uses the tansig activation function, and the output layer uses the purelin activation function. Set the maximum number of training times of the model to 1000 times, the learning rate to 0.01, and the training target error to 0.000001.

[0095] Step S53: Initialize the particle swarm, assign random weights and thresholds to the neural network, including the initial position and velocity of the particle swarm, and calculate the fitness function of the particle. The fitness of each particle is determined by the error function of the neural network. The commonly used error function is the mean square error (MSE), that is: ………… Formula (VII); where n is the total number of samples, is the true value of the i-th experimental sample, is the predicted value of the i-th experimental sample.

[0096] Step S54: Take the fitness value of the best position of each particle in each iteration. If this fitness value is better than the previous iteration, then it is the individual best position of the particle; otherwise, the individual best position of the particle takes the position of the previous iteration; Take out the particle position with the best individual best position from all particles, which is the global best position.

[0097] Step S55: The particle swarm continuously iteratively optimizes the position and velocity. When the error accuracy reaches the requirement or the maximum number of iterations is reached, the optimal weight and threshold parameters of the BP neural network are obtained; Particle velocity update formula: ………… Formula (VIII); where, is the velocity of particle (i) at time (t), and this parameter can determine the direction and step size of the particle's next movement; w is the inertia weight, and this parameter simulates "inertia" in physics and controls the tendency of the particle to maintain its original velocity; c 1 and c 2is the acceleration factor, which adjusts the step sizes for the particle to learn from its individual historical best position and the global historical best position respectively; r 1 and r 2 are random numbers uniformly distributed within [0, 1], which have the physical meaning of introducing search randomness; is the historical best position of particle (i), and its physical meaning is the best solution found by the particle individual during the search process; is the global historical best position, and its physical meaning is the best solution found by the entire population so far; is the position of particle (i) at time (t), representing the coordinates of the current solution in the search space; is the velocity of particle (i) at time (t + 1).

[0098] Particle position update formula: ………… Formula (IX); where, is the position of particle (i) at time (t + 1).

[0099] Step S56, through the above PSO algorithm optimization process, select the optimal weight and threshold parameters, and use them as the weight matrices between the input layer and the hidden layer, and between the hidden layer and the output layer, as well as the bias threshold matrices in the hidden layer and the output layer. Finally, initialize the PSO - BPNN neural network model with the optimized weight and threshold parameters, and then input data for training.

[0100] Step S57, test the performance data of MIM titanium - based formed products; Step S58, to ensure the accuracy and reliability of the PSO - BPNN predicted data, calculate the mean absolute error and root mean square error between the predicted value of the MIM titanium - based formed product performance data and the true value . The calculation formulas are as follows: ………… Formula (XIV); ………… Formula (XV); In the above formulas, RMSE is the root mean square error, MAE is the mean absolute error, n is the total number of samples, is the true value of the i - th experimental sample, is the predicted value of the i - th experimental sample.

[0101] Finally, the present invention also provides a quality and performance prediction system for hydrogenated dehydrogenated titanium-based raw materials for MIM, including a memory 41 and a processor 42. The memory includes a data input module 411, a normalization module 412, a raw material parameter library 413, and a performance database 414; the processor 42 includes an output module 421, a grading determination module 422, a prediction data module 423, and a simulation training optimization module 424. Attached Figure 4 is a block diagram of the quality and performance prediction system.

[0102] The data input module 411 is an input unit for the particle size D90, average roundness, angle of repose, Carr index, oxygen content, Ti 4+ content ratio, and powder critical loading amount of HDH-Ti raw materials, their mixtures with gas atomized titanium powder of the same composition in any ratio, and subsequent new samples, as well as an input unit for the performance data of MIM titanium-based formed products.

[0103] The normalization module 412 is a normalization processing unit for all parameters of HDH-Ti raw materials.

[0104] The raw material parameter library 413 is used to store the parameter data of HDH-Ti raw materials, their mixtures with gas atomized titanium powder of the same composition in any ratio, and the parameter data of subsequent new samples.

[0105] The performance database 414 is used to store the performance data of MIM formed products.

[0106] The grading determination module 422 is an attached Figure 3 random forest algorithm operation module shown in the figure, which judges the quality grading of samples according to the specified program.

[0107] The prediction data module 423 predicts the performance index grade of raw material samples based on a data prediction model.

[0108] The simulation training optimization module 424 optimizes the model based on machine learning training to control the output accuracy.

[0109] When the processor 42 executes the program, the following steps are implemented: (1), Input the raw material parameters through the data input module 411 and store them in the raw material parameter library 413; (2), Normalize the raw material parameters through the normalization module 412 to obtain a raw material parameter data set; (3), Input the performance data into the performance database 414 to obtain an initial performance data set; (4), Based on the above data sets, establish a prediction data module 423 loaded with a grading determination module 422 through machine learning random forest algorithm analysis; (5) Optimize the neural network training through the simulation training optimization module 424 to optimize the basic prediction data model in the prediction data module 423. (6) Input new sample parameters, use the basic prediction data model to predict the performance index level of the raw material, and output and determine whether it meets the requirements for MIM molding.

[0110] In S4, no less than 100 groups of HDH titanium-based raw materials are used to obtain their parameters as independent variables, and no less than 100 groups of performance data of MIM titanium-based molded products corresponding to them are input as dependent variables. The system uses all the raw material parameters and the marked performance data of their MIM titanium-based molded products to obtain the basic prediction data model through model training.

[0111] Further, the basic prediction data model evaluates the raw materials as first-class, second-class, third-class, and unqualified products. Among them, the first-class raw materials can be directly put into MIM molding for use, the second-class raw materials need to be tested again after MIM molding, the third-class raw materials need to adjust the MIM process to obtain MIM molded products for testing, and the unqualified products cannot be used for MIM molding.

[0112] After inputting the new sample parameter values, the sample rating of the HDH-Ti raw material and the predicted tensile strength value of its corresponding MIM molded product can be obtained through the basic prediction data model.

[0113] In this solution, the quality and performance prediction and evaluation system obtained after inputting and training more than 100 groups of data can better predict whether the HDH-Ti raw material meets the MIM molding requirements, and more accurately predict its tensile strength value, which can be used as a reference for production practice and greatly reduces the production verification cost.

[0114] It should be noted that the term "including" and any of its variations in the description and claims of the present invention are intended to cover non-exclusive inclusion. For example, a series of components included do not necessarily have to be limited to those clearly listed components, but may include components not clearly listed or other components inherent to the components.

[0115] The above is only the preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention. For example, those skilled in the art can operate by referring to other indicators of MIM molded products (including but not limited to tensile strength, reduction of area, yield strength, etc.) according to this method. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A method for predicting the quality and performance of a hydrogenated dehydrogenated titanium-based raw material for MIM, characterized in that, Including the following steps: S1. Detect the parameters of HDH-Ti raw materials to obtain a raw material parameter dataset; S2. Detect the performance data of the MIM titanium-based formed products of the HDH-Ti raw materials to obtain an initial performance dataset; S3. Use the random forest machine learning algorithm and the particle swarm optimization-based backpropagation neural network algorithm to process the datasets of S1 and S2, construct a model and conduct training and testing; S4. Input the parameters of the raw material to be tested, and use the random forest prediction data model to obtain the predicted quality grade of the raw material to be tested; S5. Input the parameters of the raw material to be tested, use the trained particle swarm optimization neural network model to predict the performance evaluation data of the MIM titanium-based formed products, and judge whether it meets the requirements for MIM forming; The HDH-Ti raw material parameters include particle size D90, average roundness, angle of repose, Carr index, oxygen content, Ti 4+ content ratio, and powder critical loading; the performance data of the MIM titanium-based formed product is tensile strength.

2. The quality and performance prediction method according to claim 1, characterized in that Calculating the Ti content ratio of the HDH-Ti raw material according to the XPS test spectrum 4+ The specific steps are as follows: Using the monochromatic Al Ka spectral line of the x-ray source, ion etching the raw material powder at a voltage of 15 kV and a current of 20 mA, and measuring the photoelectron spectrum of titanium under 5 nm on the powder surface; calibrating the spectrum according to the binding energy of C 1s; finally, performing peak fitting on the XPS and calculating the ratio of tetravalent titanium according to the integral area.

3. The quality and performance prediction method according to claim 1, characterized in that, Adopt the method of combining scanning electron microscope photos with image analysis software to test the average roundness of HDH-Ti raw materials, and calculate according to the following formula (I): …………Formula (I); Where: R - roundness; A - cross-sectional area of the particles on the observation surface, mm 2 ; d max - longest diameter of the particles on the observation surface, mm; the average roundness should be at least the average value of the roundness of 200 powder particles.

4. The quality and performance prediction method according to claim 1, wherein The calculation of the Carr index is carried out according to the following formula (II): …………Formula (2); where ρ t is the tapped density of the raw material powder, and ρ a is the bulk density of the raw material powder.

5. The quality and performance prediction method according to claim 1, characterized in that The calculation formula of the powder critical loading amount is formula (III): ………… Formula (III); Where ρ b and ρ p are the density of the binder and the theoretical density of the raw material powder, and w b and w p are the masses of the binder and the raw material powder; Among them, the calculation formula of the binder density is formula (IV): …………Formula (IV); where w i and ρ i are the mass fraction and density of the i-th component in the binder, respectively.

6. The quality and performance prediction method according to claim 1, wherein After obtaining the raw material parameters in S1, it further includes the step of normalizing the raw material parameters; the raw material parameters in S1 include positive indicators and negative indicators, the positive indicators include average roundness and powder critical loading amount, and the negative indicators include particle size D90, angle of repose, Carr index, oxygen content and Ti 4+ content ratio; The normalization method for positive indicators is carried out according to formula (V): …………Formula (V); The normalization method for negative indicators is carried out according to formula (VI): …………Formula (VI); In the normalization methods of the above positive and negative indicators, max a refers to the maximum threshold of any parameter a, and min a refers to the minimum threshold of any parameter a, and m n refers to the actual value of any parameter a, and x n refers to the value of the actual value of any parameter a after normalization.

7. The quality and performance prediction method according to claim 1, characterized in that The tensile strength of the MIM titanium-based formed product is obtained by testing in accordance with "GB / T 228.1-2021 Metallic materials - Tensile testing - Part 1: Method of test at room temperature"; the preparation method of the MIM titanium-based formed product is as follows: The HDH-Ti raw material and the binder are mixed and loaded in accordance with the proportion of 97% of the powder critical loading amount, and after mixing on an argon-protected mixer, the temperature is raised to 180 °C for mixing. After granulation, a green body is injection-molded on an injection molding machine, the injection temperature is 180 °C, and the injection pressure is 80 Mpa; after completion, debinding is carried out, the debinding temperature is 120 °C, and after debinding, it is sintered in a high-vacuum sintering furnace at 10 -3 Pa or above, sintered at 600 °C for 2 h to remove the residual polymer, and then sintered at 1500 °C for 2 h to obtain the finished product.

8. The quality and performance prediction method according to any one of claims 1-7, characterized in that, The raw material parameter dataset consists of no less than 100 groups of HDH-Ti raw material parameters; the HDH-Ti raw materials are selected from HDH-Ti raw materials of different manufacturers, different models, different batches, HDH-Ti raw materials mixed in any proportion, or mixed Ti raw materials composed of HDH-Ti raw materials and gas atomized spherical titanium powder with the same composition in any proportion.

9. The quality and performance prediction method according to claim 1, wherein S4 specifically includes: S41. Collect multi-dimensional parameter data from the production process of HDH-Ti raw materials, including oxygen content, Ti 4+ Content ratio, particle size D90, average roundness, angle of repose, Carr index, and critical powder loading; clean the collected data, remove outliers, and label the quality level of the target variable; divide the cleaned data into training set, validation set, and test set in a ratio of 8:1:1; S42. Initialize the data model and set hyperparameters; perform bootstrap sampling on the training set to generate multiple different subsets, and each subset is used to train a decision tree; moreover, during the process of constructing the decision tree, randomly select some features at each node, and perform splitting operations according to the criteria of maximizing the Gini coefficient and information gain until the stopping condition is met; S43. Use the training set to construct multiple decision trees to form a random forest classifier; then, perform classification prediction on the test set data, and finally use a voting mechanism to comprehensively make a decision on the classification results of multiple decision trees to determine the final quality grade; S44. Use a confusion matrix to evaluate the classification results, calculate the accuracy, precision, recall rate, and F1 score; optimize the hyperparameters of the random forest model according to the error analysis of the classification results; S45. Use the feature importance evaluation function of the random forest model to quantify the influence degree of each input parameter on the classification result; based on the feature importance analysis result, optimize the production process and adjust the parameters that have a greater impact on quality; S46. Deploy the optimized random forest model to an online quality inspection system, perform real-time classification on the newly input HDH-Ti raw material parameters, and output the evaluation result of the quality grade.

10. A quality and performance prediction system for hydrogenated and dehydrogenated titanium-based raw materials for MIM, characterized in that, It includes a memory (41) and a processor (42). The memory includes a data input module (411), a normalization module (412), a raw material parameter library (413), and a performance database (414); the processor (42) includes an output module (421), a grading determination module (422), a predicted data module (423), and a simulation training optimization module (424); The data input module (411) is an input unit for the particle size D90, average roundness, angle of repose, Carr index, oxygen content, Ti 4+ content ratio and powder critical loading amount of HDH-Ti raw materials, mixtures thereof mixed with gas atomized titanium powder of the same component in any ratio, and subsequent new samples, as well as an input unit for the performance data of MIM titanium-based formed products; The normalization module (412) is a normalization processing unit for all parameters of the HDH-Ti raw material; The raw material parameter library (413) is used to store the parameter data of the HDH-Ti raw material, the mixture of it and the gas atomized titanium powder of the same component in any ratio, and the parameter data of subsequent new samples; The performance database (414) is used to store the performance data of MIM formed products; The predicted data module (423) predicts the quality grade of the raw material sample and the performance evaluation data of its MIM titanium-based formed product based on the data prediction model; When the processor (42) executes the program, the following steps are implemented: (1), Input the HDH-Ti raw material parameters through the data input module (411) and store them in the raw material parameter library (413); (2), Perform normalization processing on the raw material parameters through the normalization module (412) to obtain a raw material parameter data set; (3), Input the performance data of the MIM titanium-based formed product into the performance database (414) to obtain an initial performance data set; (4), Perform machine learning and neural network training through the simulation training optimization module (424) to optimize the basic prediction data model in the predicted data module (423); (5), Input new sample parameters, use the output module (421) to predict the quality grade of the raw material and the performance evaluation data of its MIM titanium-based formed product, and judge and output whether the raw material meets the requirements for MIM forming.

Images