Difference between revisions of "Projects:2019s2-20001 Using Machine Learning to Determine Deposit Height and Defects for Wire + Arc Additive Manufacture (3D printing)"

Revision as of 20:26, 7 October 2019

Abstract

Additive manufacturing is an emerging technology that is complementary to subtracting manufacturing. AM is able to create complex model at the cost of production time. An example of AM is Wire + Arc Additive Manufacture. After slicing the manufacturing process of one model to multiple layers, WAAM uses electric arc as a source of heat to feed metal wire and create each layer. However, a slight height inaccuracy in this process will result in an expensive faulty, as error can be multiplied across many layers. Thus, in-process height measurement and faulty detection are critical to the production process. In this thesis, we will look into a machine learning approach to develop an on-the-line algorithm that automatically detects faulty processes and make height measurement by analysing electrical signal (current and voltage) of a Gas Metal Arc Welding machine. This project will introduce the approaches to tackle the same problem by other researchers, describe the design of our experiment, development of our machine learning model and discuss the final result.

Keywords: Additive manufacturing, Gas metal arc welding, Metal inert gas, Machine learning, Fault detection, Signal analysis

Introduction

3D printing is an emerging technology that has the potential to significantly reduce material usage through the production of near net-shape parts. Many of the systems for 3D printing are based on lasers and powders; however the deposition rate with such systems is very low making the production of large-scale parts difficult. AML Technologies specialises in the use of Wire + Arc Additive Manufacture (WAAM) where deposition is based on arc welding processes and the deposition rates are an order of magnitude greater. When building 3D printed parts, even a relatively small layer height error of only 0.1 mm can produce large build height errors when multiplied across the many layers of a typical build. This can make path planning difficult, so in-process layer height measurement is an essential building block of any production 3D printing system. A variety of techniques can be used for monitoring the layer height including laser scanners, and arc monitoring. It is the latter technique that will be explored in this project due to its robustness, and low cost of implementation – it only requires the measurement of arc current and voltage. Furthermore, it can potentially be used to detect defects by identifying waveform irregularities.

Project team

Project students

Anh Tran
Nhat Nguyen

Supervisors

Dr. Brian Ng
Dr. Paul Colegrove (AML3D, sponsored company)

Objectives

The objective of this project is to increase the efficiency of the manufacture process at AML3D. In order to do so, the team will investigate into the possibility of automating and optimising the quality control processes. The two quality control processes that are currently being implemented at AML3D are measuring layer height using laser sensors, and human supervision for detecting defects. These processes add overhead into production time and usage of human resource, which is not desired. To achieve the goal, it is expected that machine learning methods will be used extensively to analyse the electrical signatures of the weld process.

Background

Wire Arc Additive Manufacturing

Wire and Arc Additive Manufacturing (WAAM) is a type of additive manufacturing that uses electric arc as the heat source and material wire to feed the manufacture process ^[1]. WAAM has been investigated since the 1990s ^[1], but only recently that it received more attention from the manufacture world.

Its significant comes from the ability to manufacture complex model with less time and less material. Figure 1a and 1b show real parts that was manufactured by AML3D. Such custom made parts might takes months before be ready to be shipped, but with WAAM, the production time can reduce down to weeks. Currently, the industries that benefit the most from WAAM are maritime and aerospace.

Similar to other additive manufacture methods, WAAM achieves such results by sliding the models into multiple layers, and then build the model layer by layer. The movement control is normally handled by a robotic arm (AML3D uses ABB's Arc Welder robot), and the welding path is generated by a Computer Aid Manufacture (CAM) software.

Figure 1a: A propeller

Figure 1b: Another part of the ship

Figure 2: ABB's IRB 1520ID Arc Welder

Currently at AML3D, the manufacture process for a part can take anywhere from days to weeks. However, the contact tip (where the welding gun deposits the wire) needs replacement every couple of hours. A worn contact tip such as figure 3 will lead to unstable weld and the this effect will propagate and accumulate through multiple layers. Such defects in the manufacture process will no doubt be financially taxing for AML3D. Apart from constant human supervision, another quality control process implemented at AML3D is height measurement of each layer, after the layer is built. This process is implemented to ensure the height of each layer is in the acceptable range (1-2mm). As mentioned before, these quality control processes are costing AML3D in term of time and finance. Our project will focus on replace them with a less time consuming, more automating process.

Figure 3a: A worn contact tip
Figure 3b: Close up photo of layers

Relating Contact Tip to Work-piece Distance to layer height

At AML3D, the sensor captures the Contact Tip to Work-piece Distance (CTWD), and from there, we can deduce the layer height. A more detail explanation is of follow:

Figure 4: From CTWD to layer height

The CTWD can be set at the beginning of the weld, however, as the welding process running, the actual CTWD fluctuate around the set CTWD. For example, after finish constructing one layer, the laser sensor capture the model height at 3 different position: A, B and C. Let the 3 model height be m1, m2 and m3. Calculate n_mean and for argument sake, if the set CTWD is 15 mm, the actual CTWD at A is . The height of one layer can be calculated as . A worked example can be found at table 1.

Table 1: Worked example
CTWD	m1	m2	m3	m_mean	c1	c2	c3
15	20	22	25	22.3	17.3	15.3	17.7

Gas metal arc welding

Figure 5: GMAW welding process ^[2]

Gas metal arc welding (GMAW), or also known as metal inert gas (MIG), is the welding process used at AML3D for WAAM. Due to its high deposition rate and economic benefits, GMAW has became more popular. We will explore two variants of GMAW, which are Pulse Multi Control and Cold Metal Transfer. Both processes are developed by Fronius, an Austrian welding company. Note that it is not required to understand the welding physics to follow this wiki. The next two following sections' purpose is to highlight the high variability, high dynamic nature in welding.

Figure 6: Fronius logo

Pulse Multi Control

Pulse Multi Control (PMC) welding is Fronius' modified Pulsed GMAW. The advantage of Pulsed GMAW is the ability to control the metal droplet transfer in welding. Note that from figure 7, the metal droplet is characterised by a downward slope of the current pulse. Figure 8 shows the current signal captured from experiments conducted at AML3D (PMC is used).

Figure 7: Relationship between pulse current and droplet transfer^[2]
Figure 8: Current signal captured from experiments conducted at AML3D (PMC process)

Cold Metal Transfer

Cold Metal Transfer (CMT) is a complex welding process, also developed by Fronius. By detecting a short circuit (mode C and D in figure 9), CMT can make adjustment such as retracting the welding material to cool down, and therefore create a smoother, more stable weld. The complexity of CMT can be seen in the current signal of the process (figure 10). Compare to PMC, CMT's signal has more variation during one signal period.

Figure 9: Mechanism of metal droplet transfer mode
Figure 10: Current signal captured from experiments conducted at AML3D (CMT process)

Machine Learning

Machine learning (or statistical learning) refer to a set of tools to give data scientists an insight into the data and make better decisions based on those information ^[3]. In this project, we will explore the capabilities of classical machine learning methods (i.e anything but Deep Learning) due to the limitation of our available data. By the time of writing this wiki, we have only explored Support Vector Machine.

Support Vector Machine

The main idea of Support Vector Machine (SVM) is to draw hyperplanes that best separate multiple classes of data. The original SVM judges best decision hyperplane based on the separation margin between data classes (figure 11). However, there are multiple variants of SVM, such as hard/soft margin SVM, Nu-SVM, One class SVM (to solve anomaly detection problem). In our project, we choose to use a soft margin implementation of SVM. The benefit of soft margin SVM (as oppose to hard margin SVM) is that by allowing some misclassifications, we get a wider margin and therefore the solution becomes more "generalise" (i.e work better with a new set of dataset that is independent from the training dataset).

Figure 11: Separation margin ^[4]

SVM algorithm makes use of Lagrange multiplier. Let's the linear model of the dataset be:

where denotes the feature space transformation, w denotes the weights and b denotes bias constant.

Let be the input vector (where each index in the vector denotes a feature), and be the target (the class of input vector). It is showed that in order to maximise the (soft) margin, we will need to minimise the following:

with subject to:

where N is the total number of training data, C is the regularisation parameter to control the effect of slack variable . A more visual presentation of slack variables can be found in figure 12.

Figure 12: Slack variable ^[4]

If we were to apply the Lagrangian method to solve the above minimisation problem, we can find the decision hyperplane for the SVM algorithm. However, we shall not discuss the solution here as it is not the main focus of the project.

Kernel tricks

Both figure 11 and 12 are of linearly separable cases for SVM. For dataset that is not easy to separate with a hyperplane in their original features space (such as figure 13), the data often get projected into a higher dimension space where it is easier to separate (figure 14). Such technique is called the Kernel Tricks. It is important to note that Kernel Tricks does not actually "map" the data from X dimension to Z dimension, but rather pairwise compare the similarity of the input space in Z dimension. However, we shall not dive too deep into this as it is not the main focus of the project.

Figure 13: Nonseparable dataset ^[4]
Figure 14: Kernel tricks

Previous studies

There are many researches into defect detection in GMAW process. The three most dominant analysis methods are spectroscopic analysis, acoustic analysis, and electrical signal analysis.

Spectroscopic analysis

In ^[5], the authors captured the intensity of radiation emission of electric arc. It is obvious from figure 15 that good weld's intensity stays stable while defect one's is characterised by a sudden peak. Therefore, a threshold is set to detect the defect welds.

Figure 15: Spectroscopic analysis ^[5]

Other research into using spectroscopic data to monitor that can be of interest is ^[6].

Acoustic analysis

With ^[7], the author discovered that arc ignition is characterised by a distinct sound and therefore it is possible to capture the frequency of ignition through acoustic sensors. A tolerance band was set and both ignition frequency and sound pressure level were used to monitor the quality of the weld.

Figure 16: Acoustic analysis ^[7]

Other research into using acoustic signal to monitor that can be of interest are ^[8], ^[9].

Electrical signal analysis

With ^[10], the author plot the probability density distribution of voltage signal and discovered that the stable weld's signal concentrate in one region while unstable signal spread out. While in the paper, the authors did not developed a monitoring method using their findings, it shows that there is a significant differences between stable and unstable weld's electrical signature.

Figure 17: Voltage PDD ^[10]

In ^[11], the authors shows that by plotting the signature image of voltage and current scatter plot, both arcing and short-circuiting region's signature image can notify a defects welds. A more visual representation of short circuiting and arcing region can be found in Gas metal arc welding section.

Figure 18a: Electrical voltage signature of stable weld ^[11]
Figure 18b: Electrical voltage signature of unstable weld ^[11]

Other research into using electrical signal to monitor that can be of interest are ^[12], ^[13],^[14],^[15], and ^[16].

Experiment setup

The welding system uses TPS 320i C Pulse (figure 15a), produced by Fronius, to supply power, feed wire, and system cooling. For movement control of the welding torch, the system use ABB’s IRB 1520ID (figure 15b), which is a dedicated arc welding robot. The Data Acquisition device is NI's DAQ. Each channel (voltage and current) is sampled at 25kHz while the dominant frequency for voltage and current signal are around 130Hz (for CMT process) and 190Hz (for PMC process).

Figure 19a: Fronius' Power Supply
Figure 19b: ABB's welding robot

Figure 16: NI's DAQ device

Statistical Analysis

Preprocessing

As the captured signal coming from Fronius' power supply, the signals also pick up switching power noise. The noise will interfere with the accuracy of the analysis if not filtered. In this project, we applied a 1000Hz digital Low Pass Butterworth filter. The effect of the filter can be seen in figure 20, while the comparation of using different filters can be seen in the gallery.

Figure 20: Effect of 1000Hz low pass filter
Figure 21: Frequency response of the in-use filter

Gallery: Comparison of different filtering method
In use filter vs 300Hz low pass filter
In use filter vs 500Hz low pass filter
In use filter vs 750Hz low pass filter
In use filter vs 2000Hz low pass filter

Power Spectral Density

Prerequisite knowledge

Looking at the spectrum is another way to analyse the fundamental information of the signal, in frequency domain ^[17]. Note that the term "Power" in Power Spectral Density does not refer to the convention power definition. In signal processing, energy of a discrete time signal is defined as follow:

The way to interpret this definition is if we consider the signal to be of "Voltage" unit, E is the total energy (in joules) dissipated over a 1 Ohm resistor ^[17]. The power spectral density of a signal over time interval [-M,M] is defined as follow:

where is the truncated Fourier transform:

Analysis

Our goal for this project is to detect the defect welds and predict the layer heights (the contact tip to work piece distance, CTWD, to be more accurate) in real time (refer to this section to understand the relationship between Contact Tip to Work-piece Distance and Layer height). An analysis of the current power spectral density (PSD) shows that there is an inverse proportional relationship between the CTWD of a weld and the frequency where its current signal PSD reach the peak, and also the magnitude of the peak. Since the peak of PSD can be fairly sensitive, we only consider the frequency of the peak PSD.

Figure 22 shows the PSD plot of voltage, current, and power signal of the weld using PMC process. In this experiment, we created multiple one-line, 8 seconds welds. These eight-second welds then get segmented into smaller two-second chunk to increase the number of data and decrease the processing time.

Figure 22: Power spectral density of PMC process
Figure 23: Simple one-line weld experiment
Figure 24: Simple one-line weld experiment

It seems like the inverse proportional relationship holds true for most cases. Other experiments showed that for a short period at the beginning and ending of the weld, the signal can behave differently relative to the rest of the weld. Currently, to migrate with this problem, data will not be collected and analysed (both offline and online) at these periods. However, the above results seem to only hold true for PMC processes. With CMT process, the current PSD plot indicates no relationship between the CTWD and the frequency at peak (figure 24).

For the next stage of the project, our plan is to develop a linear regression model to predict the CTWD for PMC process. With CMT process, we will have to look into different analysis techniques.