數據挖掘
In powder metallurgy high velocity compaction is an effective technology with the advantage of high and uniform density [1-3] with low spring back [4-6]. It is widely used in aerospace such as filters and some military components. The green density of HVC products is critical which significantly affects the mechanical and functional properties of the subsequent sintering phase. The processing parameters including the compacting speed and frequency, coupled with the chemical composition and properties of raw powders have a profound impact on the green density. There exists some compaction theory formulas for green density prediction, such as the famous Balshin’s theory of compaction and Huang Peiyun Double Logarithmic Equation[7]. However, they all treated the powder as an elastomer and neglected the influence of pressing time, which had a great effect on the precision and applicability. Therefore, the existing compaction theory formulas are not suitable for green density prediction of HVC. Most of the green density predictions rely on experts, prior knowledge, trial and error during experiments, which is costly and time-consuming with few accurate judge rules during alloy design.
An input training dataset is needed in order to build a quantitative statistical learning model. Each material we are interested must be characterized via a representation in terms of one or more material descriptors. The dataset used in this work consists of chemical composition, raw materials property, processing parameters and green density information from several metal powders used for HVC, such as Fe, Cu, iron-based alloy, aluminum alloy, titanium alloy and composite. There are totally 223 entries, as shown in Table1, and all the entries are collected from the previous experiments accumulation in our laboratory on an HYP 35-2 HVC machine whose maximum capacity is 2 kJ per stroke, the highest velocity is 10m/s and the mass of the hammer is 42 kg.
The green density is the target which we aim to construct predictive models consisting of the details of chemical composition, raw material properties and the HVC processing parameters. We construct a database to curate this dataset, which is free-access by http://www.fyniu.com/search2.php?tty=901. This dataset is also subjected to National Materials Scientific Data Sharing Network (http://www.fyniu.com). All the data in the database is verified by the peer review and contains literature derivation information to assure the quality and reliability. In this work, the features in the dataset is categorized into the following groups:
l Chemical composition: %C, %Si, %Mn, %Ni, %Cr, %Cu, %Mo (in wt. %)
l Raw material property: mean particle size, apparent density, Young’s modulus of the powder mixture, Vickers hardness
l HVC processing details: Compaction velocity/energy, compaction times.
2. Machine learning method
Machine learning is an inevitable result of the development of artificial intelligence research to a certain stage which mainly uses induction, synthesis rather than deduction[22]. Machine learning is a field that deals with the design, development and implementation of techniques that permit computers to learn based on data. And machine learning is an algorithm that automatically analyzes and obtains rules from data and uses rules to predict unknown data[23,24]. Of course, with the rapid increase in the ability of people to collecting and processing data, machine learning has received extensive attention from different disciplines. Machine learning provides many methods and models such as Linear regression, Decision tree, Bayesian decision and so on aiming to different kind of datasets[25]. Scholars in different fields can pick the best method that fits well with themselves dataset and has a good prediction in research.
3. Feature Selection
Feature selection is critical for machine learning model accuracy. If some of the features used in statistical learning models are not directly associated with the targeted property, the dimensionality of the feature space will be too high and may reduce the training accuracy.
Powders are densified by the application of pressure, initially by the particles sliding past one another and then by particle deformation at higher pressures [26,27]. There are many factors that affect the process of the compaction, powder’s properties such as the hardness of the powder, friction performance, the particle and the shape of the powder. Also the Young’s modulus and the compressibility are critical to compaction. Moreover the Lubricants and the way of compacting have a certain degree of influence [28,29].At this stage, it is difficult for us to quantify all the influencing factors of the process of the compaction, so we are trying to sort out more relative parameters and combine with the machine learning to explore a novel mothed for powder metallurgy.
There is no specific physical meaning for chemical composition of powder materials, so we map the chemical composition information to Young’s modulus of powders mixture. As we know, during the compacting procession, the powders will lead to plastic deformation and elastic deformation. When at the end of the compacting procession, the green compact will swell to some extend[28]. And the Young’s modulus has great influence on the procession. In homogeneous materials, stress and strain can be defined at any point within the object. However, in the composite material, it can’t determine the stress and strain of a point because it is not sure which material of the point. Therefore, the volume average strain can be used to define the strain in the composite. According the method of the idea. In this paper, we utilize the elastic modulus of each element to express the elastic modulus of the material [7].So we take all the elements into consideration and utilize the Young’s modulus to substitute chemical composition. The Young’s modulus of powders mixture is calculated according to the linear addition method.
4. Model Training
Due to the small number of dataset in the research and prevent over-fitting and under-fitting of the models, so cross-validation is used in the paper. In order to build prediction model, we will try to find out a best model aimed to our target through Weka3.8 and Python including Support Vector Regression with poly kernel (SVR /poly), Support Vector Regression with RBF kernel(SVR/rbf), Multilayer Perceptron (MLP) and Kernel Ridge Regression. Support vector Regressors (SVRs) are kernel-based statistical learning approaches that are widely used in pattern recognition and imaging [33-35]. Multilayer Perceptron is one type of neural network to model highly non-linear functions, without making prior assumptions concerning the data distribution, and it can be accurately generalized when dealing with new, unseen data. This advantage makes MLP widely used in forecasting complicated nonlinear problems [36-38]. Kernel ridge regression (KRR) combines Ridge Regression with the kernel trick. It learns a linear function in the space induced by the respective kernel and the data. Kernel ridge regression is more flexible because it uses kernel techniques and is suitable for complex fitting. And for reducing the error of models and taking full advantage of all data we collected. Cross validation is used in the research and figure out the average score of 10 random training and test data where test data should be 20% and training data should be 80%.
5. Model validation
To validate the accuracy of the green density prediction models, we compare the predicted green density with the experimental green density. The validation dataset is obtained by experiment, which is not contained in our training set of 223 instances. The validation dataset consists of 14 instances that not included in the training set of 223 instances listed in Table 2. We choose such samples for their large scale variability of the chemical composition and different compacting energy for making sure the accuracy of prediction. Under comparison the predicted values of the four models, the MLP model out-performs all the other models. There are larger error due to overfitting occurring in the other three models. So MLP is selected for further model validation.
相關工作發表在：
Kai-qi Zhang, Hai-qing Yin*，etc.A novel approach to predict green density by HVC based on Materials informatics method. International Journal of Minerals, Metallurgy and Materials. 2018. [Accepted]
