A Procedure for Building Reduced Reliable Training Datasets from Real-World Data

Abstract

Dimensionality reduction and anomalous data detection are important tasks in machine learning and data mining applications. Many real-world datasets are affected by errors and variable redundancy and this fact can generate problems when the data are used to develop accurate models exploiting some training procedures for parameters tuning. In this paper an automatic procedure is proposed combining detection of unreliable data and reduction of dimensionality to be adopted before exploiting the data to develop a model for prediction purposes. The method has been tested on several datasets belonging to the UCI repository and industrial fields. The results of tests are showed and discussed in the paper. The proposed approach provide a good prediction accuracy providing a minimal but essential dataset.

A Deep Learning‐Based Approach to the Estimation of Jominy Profile of Medium‐Carbon Quench Hardenable Steels

“…This configuration of the predictor achieves an average MAE value over the profile on DS A of 0.96 HRC, which is 16% lower than the sequential model proposed in ref. [37], for which an average MAE value achieved is 1.10 HRC.…”

“…The achieved results are compared to those obtained by the sequential predictor described in ref. [37], by a NN of the MLP type with one hidden layer (indicated in the following as BaseMLP) including neurons with standard sigmoidal activation function, analogous to the one proposed in ref. [30], and by a fully connected DNN whose architecture has been optimized (in terms of number of layers and neurons in each layer) like DNN_Jom.…”

“…As a result, the proposed approach is encouraged, and part of the DNN_Jom that was previously trained on DS A was retrained according to the TL approach described in Section 2.5. Such approach was compared to the sequential model [ 37 ] and to a fully connected DNN trained exclusively on DS B . The DNN architecture was optimized in a similar manner to DS A , resulting in a smaller network consisting of two layers with ten neurons each, activated by the ReLU function.…”

“…Based on the variable section analysis performed in ref. [37], the contents of all the 15 chemical elements reported in Table 1 are fed as input of all the models.…”

“…To overcome such issue, a parametric method was proposed, [ 36 ] where the hardenability profile was approximated via a parametric analytical relationship linking hardness to distance from the quenched end, and wavelet NNs correlated the steel chemical composition to the function parameters. More recently, the same authors presented a hierarchical NN‐based Jominy profile predictor, [ 37,38 ] which showed remarkable features in terms of reliability, robustness, limited computational burden, and maintainability. Such model estimates the hardenability profile in a pointwise way through a “cascade” of simple NNs, each one outputting one hardness value of the Jominy curve and being fed with a few input variables, which accounts for both influence of the microalloying elements on the single profile points and correlation among the neighboring hardness values, as previously computed hardness values are used to calculate the following ones.…”