**1. Introduction**

Market requirements, along with the development of the industrial sector, have led to parts being demanded under restrictive specifications. As a result, surface finish has emerged as a key aspect of many products in which this feature is critical.

Single Point Incremental Forming (SPIF) is a rapid manufacturing technique based on the localized plastic deformation of sheet metal by a hemispherical tool [1]. For that, the sheet is fixed to a rigid frame and the deformation is produced, moving the tool according to an helicoidal path and establishing a pitch value to feed the operation. The method has generated considerable research interest because it overcomes some of the disadvantages of conventional forming. The SPIF process is characterized by its ability to economically develop complex geometries by means of reduced tooling costs if short series or prototypes are being produced, notably in biomedical and aeronautical applications [2–4]. The process has been applied to commercial copper, some aluminum alloys, and high formability mild steel. Some application with stainless steel AISI 304 has been also reported [5]. With regard to titanium alloys, research is mainly recent, and it supposes that their application has the most potential. However, titanium alloys, particularly Ti6Al4V, have a poor formability at room temperature which classifies them into the hard-to-form materials group. Thus, it is necessary to heat the sheet to carry out the SPIF process, also called heat assistant SPIF or HA-SPIF. Consequently, substantial efforts have been undertaken in recent years to develop heating methods. Götmann et al. [6] worked on the process of dynamic heating by laser, previously proposed by Duflou et al. [7]. Fan et al. [8] developed a heating procedure based

**Citation:** Bautista-Monsalve, F.; García-Sevilla, F.; Miguel, V.; Naranjo, J.; Manjabacas, M.C. A Novel Machine-Learning-Based Procedure to Determine the Surface Finish Quality of Titanium Alloy Parts Obtained by Heat Assisted Single Point Incremental Forming. *Metals* **2021**, *11*, 1287. https://doi.org/ 10.3390/met11081287

Academic Editor: Umberto Prisco

Received: 21 June 2021 Accepted: 13 August 2021 Published: 16 August 2021

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2021 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

on Joule heating by electric power from the tool to the sheet, which has also been adopted by other authors [9]. Ambrogio et al. [10] designed an external heating system using electric bands. Similar systems have been applied by Naranjo et al. [11] and Ortiz et al. [12]. There have been attempts to improve formability using the local heating generated between the tool friction on the sheet metal when it rotates at high speed [13]. This procedure is known as Stir-SPIF. Finally, some methods mixing friction and external sources of heat can be found in literature [14].

Despite the improvement reached by HA-SPIF methods, the obtained surface finishes are low quality. There have been some efforts to improve the surface finish after the SPIF process trying to reduce the friction tool sheet without lubricant [15] or considering the effect of lubricant on the process [16], but with a few exceptions, the research has been focused on only room temperature and not on titanium alloys.

Additionally, the influence of surface quality on the operability of the parts promotes the constant development of measurement methods. In recent decades, this sector has undergone successive updates in order to obtain a reliable, precise, and flexible technique to measure surface finish.

Contact profilometry is one of the most common methods in surface finish measuring despite having several drawbacks. Its use is often restricted to controlled environments because it is highly sensitive to vibrations and the stylus can cause flaws on the surface of the part. These limitations, together with its slowness, render it an unsuitable method for measuring online surface roughness.

Optical non-contact methods arose as a necessary alternative to overcome these drawbacks. Of these techniques, the procedure developed by Moreas et al. [17] should be mentioned. In this work, an optical microscope was built to capture images and topography was studied using the principle of triangulation. From another perspective, in the research conducted by Lehmann [18], interferometry techniques were employed in order to analyze the statistical properties of speckle patterns to obtain significant surface roughness parameters. However, non-destructive optical methods do not always achieve the level of precision provided by contact profilometry and samples must be adapted for each of them. Particularly, optical microscopes equipped with profilometer function could be appropriate with higher accuracy in the measurement of surface roughness, but the measuring field is too small and the sample must be split from the part. The dominant key issues for future research include systematic process planning, tool-path optimization, compensating unwanted deformations, advanced feature recognition techniques, forming reliable and robust communication interfaces, flexible hardware, adaptive control methods, the ability to form newer materials and a wider range of materials, complex shapes with ease and accuracy, and removing barriers on sizes of components.

Artificial intelligence has become very popular in this field because machine learning techniques are capable of developing predictive models that provide better results than analytical ones. Machine learning algorithms have been shown to help reduce the number of experiments required to establish a correlation between surface roughness and manufacturing parameters with great effectiveness.

Thus, machine learning has been used to analyze the influence on surface finish of the parameters involved in the manufacturing process. The aim is to optimize the process by maximizing the number of suitable parts. In this regard, Abd et al. [19] developed a gradient boosting regression tree (GBRT) algorithm to establish a correlation between the operating parameters and surface quality of aluminum/stainless steel (Al/SUS) bimetal sheets obtained by SPIF. This research used artificial intelligence to expand on bimetal sheets the results obtained by Echrif and Hrairi [20] in single-layer sheets of aluminum alloy AA105-0.

The use of machine learning algorithms for measuring surface quality allows the limitations of traditional methods to be overcome. Models built with this methodology have proven to be accurate and able to adapt production speeds. In this context, the work by Mulay et al. [21] focuses on a feedforward backpropagation network that was implemented to predict the average surface roughness and maximum forming wall-angle of parts of aluminum alloy AA5052-H32 obtained by SPIF. This experiment showed that artificial neural networks are a promising tool with great economic benefits, greater predictability, and low simulation time.

The work developed by Abu-Mahfouz et al. [22] considered a classifier based on a support vector machine to obtain surface roughness from vibration signals collected during the milling of aluminum parts. In addition, this work analyzed the effect of different types of kernels and sets of features on the performance of the model. Their proposed classification model was trained with 32 different combinations of the milling process conditions and managed to predict the surface roughness with an accuracy of 81.25%.

Furthermore, the research by Koblar et al. [23] developed a computer vision method based on machine learning for surface roughness measurement. Decision trees were used to solve classification and regression problems on a database formed by 300 commutator images taken in 8-bit grayscale. This experiment offers new possibilities for evaluating surface quality and develops a suitable method for applying online roughness measurement without contact. However, the nature of the work calls for the use of a broad set of samples with respect to inputs based on manufacturing parameters.

The research conducted by Kurra et al. [24] developed predictive models for the average and the maximum surface roughness measurement (Ra and Rz) after the SPIF process based on different techniques: Artificial Neural Networks, Support Vector Regression, and Genetic Programming. In the training of these models, the effect of tool diameter, step depth, wall angle, feed rate, and lubricant type were taken into consideration as process variables. Each parameter was varied over three levels to study different combinations of manufacturing parameters and achieve a uniform distribution of roughness values. Roughness was thus analyzed from a technical viewpoint, so this method requires knowing in advance the conditions in which the parts were manufactured.

Other works, such as that by Lin et al. [25], have used deep neural networks, long short-term memory networks, and one-dimensional convolutional neural networks for the prediction of arithmetic mean roughness from vibration signals obtained in the milling process of S45C steel parts. Li et al. [26] compared the results from different classifiers to predict the surface roughness of polylactic acid intake flanges obtained by 3D printing. In both studies, Fast Fourier Transform (FFT) was used to extract the features' raw data and different techniques were evaluated to build an accurate predictive model.

The literature review showed that several studies have developed predictive models for surface quality measuring of work pieces produced by different processes. However, no research has focused on the application of machine learning techniques to predict the surface quality from images of pieces created by incremental forming, among other reasons, because of the large number of samples required for the model training and the amount of time needed to acquire the features. To fill this gap, the aim of this work was to develop a new methodology for the evaluation of surface quality using photographs of twenty-one parts formed by SPIF as inputs in machine learning classifiers. This procedure would allow the surface finish to be determined without the need to previously know the manufacturing conditions. In this way, surface quality could be studied in real time during the production process.

The remainder of the paper is organized as follows. Section 2 contains an introduction on machine learning and explains the experimental setup in detail. Section 3 describes the data preparation and the training of predictive models. Section 4 presents an evaluation of the predictive model performance and discusses the experimental results. Finally, Section 5 provides findings and future directions.
