Content uploaded by J.J. Rodriguez-Andina
Author content
All content in this area was uploaded by J.J. Rodriguez-Andina on May 27, 2015
Content may be subject to copyright.
BELOW YOU MAY FIND THE ACCEPTED VERSION OF THE ARTICLE:
J.Rodríguez Araújo, J.J.Rodríguez Andina, J.Fariña, F.Vidal, J.L.Mato, M.Á.Montealegre,
“Industrial laser cladding systems: FPGA-based adaptive control”, IEEE Industrial
Electronics Magazine, vol. 6, no. 4, pp. 35-46, Dec. 2012.
DOI: 10.1109/MIE.2012.2221356
THE PUBLISHED VERSION IS AVAILABLE IN IEEE Xplore, AT:
http://ieeexplore.ieee.org/xpl/articleDetails.jsp?arnumber=6378531
INDUSTRIAL LASER CLADDING SYSTEMS͗&W'Ͳ^Wd/sKEdZK>
1. INTRODUCTION
Laser cladding is a deposition process, depicted in Fig. 1, which uses a laser beam to melt the
surface of a metallic substrate and a substance added in the form of powder flow, so that the two
materials are fused by metallurgical bonding. Usually, laser cladding consists of a processing head,
where the laser beam and powder flow are generated, which moves over the target part. The
melted powder particles and a thin layer of the substrate form a clad whose thickness and
penetration depend on the control parameters of the process.
The aim of this process is to modify the surface properties of a metallic material, thus improving
some of its characteristics, like hardness or corrosion resistance. It can also be used for
reconstructing or fixing surface defects in parts damaged because of fracture, wearing, or other
effects. This technique allows the bond strength between the part under treatment and the
material provided by the powder flow to be controlled, thus offering significant advantages with
regard to other alloying and hardening methods. Some sample applications are in aerospace
technology [1], repairing of turbine blades [2], or rapid part prototyping or repairing of fabrication
molds [3] (taking advantage of the possibility of stacking several layers of metallic powder).
In the technology center AIMEN there is a so-called Laser Application Center that provides facilities
and expertise for welding and cladding process development and validation, supporting more than
400 companies of the manufacturing industry, in areas such as metallurgy, automotive, civil
engineering, mechanics, or shipbuilding, among others. Currently available industrial laser
cladding systems suffer from several practical limitations, in particular regarding process
monitoring and control. The aim of the work reported in this paper is to optimize the industrial
system at the Laser Application Center, not only overcoming these limitations (that are analyzed in
next section) but also offering additional advantages in terms of flexibility to adapt to different
operating conditions as well as to different types of lasers. The key factors to achieve these goals
have been the use of FPGAs, for both monitoring and control purposes, and the development of
an adaptive controller, whose parameters are updated in real time according to a set of fuzzy
rules. The suitability of FPGAs for the implementation of instrumentation and control systems is
currently a well-established fact [4]-[8]. In this context, one of the most interesting features of
FPGAs is the possibility of combining embedded microprocessor cores with high-performance
specialized custom hardware peripherals.
This work has been carried out in several steps. First, process monitoring was enhanced by using a
vision system based on a CMOS camera and an FPGA-based image processing system [9], which
was later further improved in terms of its robustness to optical defects [10]. The FPGA system was
then extended to incorporate a digital PI controller [11]. Both the monitoring and control systems
presented in [9]-[11] outperformed preexisting ones used in industry, and overcame some of their
limitations, as discussed in sections 3 and 4. However, having a fixed control structure presents
some practical limitations. On one hand, its behavior can be significantly influenced by the
operating conditions, which vary depending on environmental conditions, as well as on the
materials to be treated and the geometry of the parts. On the other hand, adjusting control
parameters for each type of laser system used is a complex and critical task. These limitations can
be overcome by using an adaptive control system, as proposed here.
The remainder of the paper is structured as follows. The limitations of currently available
industrial laser cladding control systems are identified in Section 2. The previously developed
FPGA-based monitoring system is described in Section 3. Section 4 analyzes the FPGA-based
controller initially developed and its extension for adaptive behavior. Experimental results are
presented and discussed in Section 5. Finally, the current conclusions of the work, as well as future
improvements to the system, are summarized in Section 6.
2. LIMITATIONS OF CURRENT INDUSTRIAL LASER CLADDING SYSTEMS
Current industrial laser cladding systems are usually based on monitoring the melt pool
temperature [12] through the radiation measured by a pyrometer using Planck radiation law, and
controlling the process with a PC.
Temperature measurements present some practical problems, because they work with average
values of radiation, so it is necessary to adapt the pyrometer optics to the dimensions of the clad
area. Therefore, geometry variations of the target surface (such as holes, steps, or protrusions)
negatively affect the behavior of the control system and, in turn, the quality of the results
obtained. These problems can be overcome using systems that measure by optical means and
image processing techniques the geometry (e.g., height [13] or width [9], in two different
applications) of the melt pool created by the laser beam. Clad width measurements allow the
degree of dilution (the relation between the melted area of the substrate and the amount of
added material) to be estimated, as shown in Fig. 2, so this has been the approach followed in this
work. Dilution is a fundamental parameter to determine the quality of the process, because the
composition and properties of the clad are affected by the degree of dilution achieved. However,
dilution cannot be directly measured in real-time processing [14]. Hybrid systems have also been
recently proposed, which combine temperature and optical measurements [15], but targeting a
different application domain (direct deposition).
One of the main problems of laser cladding is the relatively poor repetitiveness of the results
obtained, i.e., variations in the quality achieved in different runs under the same process
parameters and operating conditions. Small variations in the powder injection rate, the laser
speed, or the shielding gas, heating of the part as the process is carried out, or interference of
powder particles with the laser beam, may cause noticeable geometry changes. Therefore, a very
fast and accurate control is required in order for the final results to comply with the strong quality
requirements of the target applications.
Given the fast variations in the shape of the melt pool, the data acquisition rate is a critical
parameter of the control system. When cameras are used for measuring the size of the melt pool,
frame rate must be maximized, and image processing time minimized. On the other hand,
increasing the resolution leads to increasing accuracy of each sampled image, at the expense of
frame rate and processing time. Therefore, a tradeoff must be achieved between these two
factors. In current industrial systems the control algorithm is implemented in a PC. Because of its
slow response, only low-resolution images can be used, resulting in low accuracy of the
measurements, which negatively impact the performance of the control system. On the other
hand, the use of microcontrollers is discarded because they would not provide good enough
performance (frame rate would be limited) at the required resolution.
3. FPGA-BASED MONITORING SYSTEM
The block diagram of the laser cladding system installed in the Laser Application Center is shown in
Fig. 3. The system consists of three main blocks: laser beam generation, powder injection, and
robot. The properties of the laser beam have to be chosen according to the quality requirements
of the cladding process. Therefore, several laser systems are available, which can be
interchangeably mounted on a 6-axis IRB6600 robot system from ABB. Two of these systems have
been used for the development of this work: a 3.3kW Laserline LDL 160-3300 direct diode laser
and a 4.4kW Rofin Nd:YAG laser. The delivered beams are focused on a spot on the surface of the
target part. The powder injector generates a constant flow of material that concentrates (using a
coaxial nozzle COAX8 from Fraunhofer IWS) on the target substrate area together with the laser
beam focus (see Fig. 1). The 6-axis robot is used to produce the relative movement between the
substrate and the laser beam (at 8 mm/s in all experiments reported in this paper), in order for the
target shape of the part to be treated. The material in the substrate and in the powder was 1.2344
(H13) tool steel in all cases.
The optical system (Fig. 4) is placed in a coaxial arrangement with the laser beam. It consists of a
CMOS camera with 8-bit (256 grayscale levels) CameraLink interface and dynamic range up to
120dB, and auxiliary optical components: mirrors, a lateral augmentation telescopic system, and
two filters, one to protect the camera from the laser beam and another to eliminate all
components outside the visible spectrum.
The FPGA system, whose block diagram is shown in Fig. 5, is in charge of image processing (as well
as of the control algorithm, described later, and the generation of the corresponding actuation
signals). It is based on a MercuryCode board from EBV, including an EP3C40F484C7N Cyclone III
FPGA from Altera, a 64-Mbit EPCS4 serial Flash, 16MBytes of SDRAM and 1MByte of SRAM. This
board is connected to the CameraLink interface and to a TFT display (to visualize the captured
images for debugging purposes). The well-known capabilities of FPGAs for concurrent operation of
the different functional blocks implemented, as well as the use of the specialized embedded
resources and optimized library blocks available (soft processors, arithmetic blocks, embedded
memory) resulted in a system that, as demonstrated in subsequent sections, not only outperforms
currently available industrial ones, but also has additional advantages.
A Nios-II soft processor core is used to coordinate all tasks performed inside the FPGA, using
Avalon Memory Mapped and Avalon Streaming (for image data transfer) Interfaces. This implies
that the acquisition, processing, and actuating systems have been designed (describing them in
Verilog, and taking advantage of specialized FPGA resources, for instance extensively using
embedded arithmetic blocks) as peripherals of this main controller, and are accessible through its
standard buses.
The image processing peripheral is a custom-designed block that implements image processing
(binarization and erosion) and extracts from the image the information regarding melt pool width.
In the initial FPGA implementation [9], the image is first captured and stored in memory, and later
binarized and eroded. The binarization of the image is performed by defining a gray level
threshold with which pixel values are compared. Further processing can be accurately carried out
using the resulting black and white image (which significantly simplifies processing) because the
high temperature of the melt pool produces a high-intensity image (Fig. 6). In addition, exposure
time is adjusted to cause image saturation (white pixels) in the area of the melt pool. Erosion is a
process that allows incorrectly saturated pixels (usually produced by noise) to be removed from
the image. The image is scanned with a 3x3 pixel matrix. If all pixels in the matrix are saturated,
the central pixel is kept saturated; otherwise, it is set to black (9 1-bit logic products are required
per pixel). An example of erosion is shown in Fig. 7. Finally melt pool width is easily determined by
reading the rows of the eroded image and computing the difference between the first and the last
rows for which saturated pixels exist (counting and comparison operations are used with this
purpose).
In spite of its simplicity, this solution allows a much better performance than that of the existing
PC-based solutions to be achieved [9]. However, it still has a practical drawback associated to
possible optical defects. Because of damages to the nozzles or variations in the laser beam power
(the main control parameter, as discussed later), non-fused powder can remain suspended
between the optical system and the cladding area, causing (part of) this area to be hidden for the
optical system. For instance, when stacking several layers of metallic powder to reconstruct a part,
surface temperature increases in successive iterations. Laser power is reduced accordingly, then
increasing the probability of non-negligible amounts of non-fused powder to appear. In addition,
incandescent powder projections may be incorrectly identified to be part of the melt pool.
Moreover, because of the harsh operating environment associated to laser cladding, it is likely
that the optical system is damaged at some point of its life cycle. Obviously, detecting the
occurrence of such problems and mitigating their effects is of paramount importance because of
both functional and economic reasons.
The aforementioned defects can be identified in Fig. 6, which shows
that binarized images may not
be perfect. In this case, there are on one hand black lines associated to damaged pixels, which are
not a problem in this particular case, but may be if the defects affect a larger area of the sensor.
On the other hand, there are white pixels outside the actual melt pool area, which cause the
original FPGA image processing system to provide a wrong measurement (rectangle). These white
pixels can be due to incandescent suspended powder (this case) or to wrong white areas
associated to damaged pixels. Because of the erosion processing step, only white areas greater
than 3x3 pixels will be present in the final image. Unfortunately, such areas actually happen in
practice. Wrong black areas (shadows) can also appear due to non-fused suspended powder. This
is most likely to occur at low and moderate power, as can be seen in the top and middle images in
Fig. 6.
As it can also be seen in Fig. 6, the shape of the melt pool can be approximated to an ellipse.
Taking advantage of this fact, the image processing peripheral was modified [10] to obtain melt
pool width (the minor axis of the ellipse) from the computation of moment functions [16]. This is
an extension of the well-known concept used in physics to determine the distribution of mass in
an object. It can be extrapolated to image processing, using the value of pixels in grayscale or
binarized images instead of mass. The computation of the moments of an image allows it to be
approximated to a geometric shape mathematically equivalent in terms of moments, for which
parameters such as center, orientation or size can be determined. For this, two types of
computations are required. Integer multiplications and additions, to compute the so-called
general moments, are carried out by a new dedicated hardware peripheral, using embedded
hardware arithmetic blocks to optimize performance. Central moments and all other values
required to determine melt pool width (and actually also length) are computed with a driver
implemented in the Nios-II processor, because the floating-point operations required cannot be so
efficiently implemented using FPGA’s hardware resources. This highlights the fact that decisions
had to be made in the design process about functionalities to be implemented in hardware or in
software. Computation of melt pool width and length by the driver is performed after that of
general moments (these two steps cannot be overlapped in time). The number of operations (and
their complexity for a software implementation) is very low [10], so they only add a little
processing time which does not significantly impact performance, even for the larger image
resolutions used.
Once the design was to be modified, it was decided to also optimize other parts of the system to
reduce resource usage and decrease latency. This has been accomplished by implementing
streaming versions of the binarization and erosion algorithms, so performance was further
improved. In streaming processing pixels are binarized as they are received from the camera, and
sent to erosion through an intermediate FIFO buffer (defined in Verilog to optimally perform
erosion operations in a single clock cycle), from where they are taken as soon as they can be
eroded. On one hand, this reduces the latency of the system, since the three main steps of image
processing are overlapped in time, which is a unique advantage of FPGA implementation (in
computer / microcontroller / Graphic Processing Unit-based image processing, the whole image is
first stored in memory and later processed). On the other hand, there is a significant saving in the
amount of memory required, since now there is no need for the image to be stored in FPGA
memory (since it is processed “on the fly”). In addition, logic resource usage is also reduced,
because most resources related to memory access management are no longer needed. As
demonstrated in [10], this solution is not only robust to defects affecting the optical system, but
also exhibits a reduced latency that results in a better dynamic response.
4. FPGA-BASED CONTROL SYSTEM
Although laser cladding is sometimes applied as an open-loop technique, closed-loop control is
mandatory in applications where homogeneous results are required, because the type, size, and
geometry of the part being treated influence the process. The inherent non-repetitiveness of the
technique makes it necessary a control of process parameters fast and accurate enough according
to the features of the melt pool.
The control system has to fulfill three main conditions in these
cases. Settling time must be short enough as to allow small parts to be adequately treated. In
addition, it has to be able to adequately respond to heating effects caused by the power applied to
the target metallic part. Finally, it has to be fast enough to track complex geometries.
There are three main process parameters that can be used to control a laser cladding system: laser
beam power, powder flow rate and relative speed between laser beam and substrate. Laser power
is usually the variable being controlled, because it is the only one which allows real-time control to
be implemented using this kind of industrial equipment, so this has also been our choice. On one
hand, accurate powder flow rate control is very difficult to achieve, because powder flow cannot
be measured in the nozzle but in the powder feeder upstream. On the other hand, the accuracy
and the speed change rate required if speed was the main control variable would also result in
very complex solutions, which would imply the need for significant changes in the plant (e.g., by
implementing a visual servoing system to control the movement of the robot). It is worth noting
that the response of commercial solutions for real-time control of robots is much slower than that
of the laser system.
For the development of the control algorithm, the target plant has to be first modeled. In this
sense, laser cladding is a complex process, because of its inherent uncertainties and also to the
fact that the geometry of the melt pool is affected by a large amount of external factors, some of
which (like powder flow) are very difficult to model. Due to this, it was decided to initially use a PI
controller, whose parameters were empirically obtained using the Ziegler-Nichols method, for one
of the laser systems available at the Laser Application Center (the diode one) under usual
operating conditions. Because of the “noisy” nature of the process (discussed below and
highlighted by the experimental results presented in next section) the use of a derivative term is
not a good option, so a PID controller is not suitable for this kind of applications.
In the FPGA implementation, the Nios-II reads the data regarding the geometry of the melt pool
provided by the image processing peripheral and, after applying the digital PI control algorithm,
generates the corresponding duty cycle that configures the behavior of the PWM generator (see
Fig. 5). Laser power can be controlled with a dedicated analog input to the laser systems, in the
range 0-10V. As the MercuryCode board used does not include Digital-to-Analog converters, the
control signal is obtained using a 15kHz digital PWM generator with 10-bit resolution, whose
output is applied to an external Sallen-Key low-pass filter (whose cutoff frequency is 1kHz),
specifically designed for this application.
As experimentally demonstrated in [11], this approach performs significantly better than existing
PC-based solutions, particularly when targeting short or narrow areas or complex geometries
(found in many practical applications), for which PC-based solutions are not suitable because of
their low speed and large settling time. However, some practical problems and issues still affect
this “static” solution. Heating of the parts as they are being treated changes in real time the plant
to be controlled with regard to the one the controller has been empirically adjusted for. This may
cause loss of homogeneity in the treatment along the target part. This is a very significant effect in
some important cases, e.g. when producing overlapping clad tracks (a usual practical solution to
treat large areas), which changes not only substrate temperature because of heating, but also
thermal conductivity, because successive tracks are applied in part to areas already covered by
previous ones. Also, since different types of lasers have to be used for different applications,
different adjustments are needed for each one of them. Moreover, “noisy” measurements are
inherently associated to the laser cladding process, because of the unavoidable uncertainties due,
among other factors, to the use of powder or to variations in the surface under treatment.
Reducing the control action when the system is closer to its set point helps mitigating problems
associated to this process “noise”.
Because of these facts, an adaptive version of the controller has been developed. Obtaining a
model of the plant valid for all possible configurations and operating conditions is not feasible.
Therefore, an adaptation of the control parameters is performed in real time from a set of fuzzy
rules obtained from the empirical knowledge of the process, according to the block diagram of Fig.
8, using Mamdani’s method [17]. The discrete incremental formulation of the PI controller is:
߂ݑ
ூ
[݇] = ܭ
⋅ ߂݁[݇] + ܭ
⋅ ܶ
௦
⋅ ݁[݇]
ݑ
ூ
[݇] = ݑ
ூ
[݇ − 1] + ߂ݑ
ூ
[݇]
where e is the error (in this case, the difference between the measured melt pool width and the
set point, SP), ߂݁ is the difference between the error in the current and the previous sample, K
p
and K
i
are the proportional and integral gains of the controller, respectively, T
S
is the sampling
period, and u is the control action that defines the laser power command (p).
The fuzzy adaptive block in Fig. 8 implements the fuzzification, inference, and defuzzification
steps. The same type of membership functions are used for input and output variables. As shown
in Fig. 9, they are triangular functions, symmetrical with regard to the central value of the fuzzy set
and with variables uniformly distributed with 50% overlapping. Three fuzzy sets (‘Negative’, ‘Zero’,
and ‘Positive’ for inputs; ‘Small’, ‘Medium’, and ‘Large’ for outputs) have been defined and
normalized in the [-1.0, 1.0] range. The rule base used for the inference step is listed in Table I. The
resulting fuzzy control surface is shown in Fig. 10.
Same as the initial implementation, the adaptive PI controller is executed in the Nios-II processor.
Its relative simplicity causes the impact on performance of this design decision not to be
significant, although a hardware implementation is considered an interesting future work, as
mentioned in section 6. The improvements it provides are demonstrated by the experimental
results presented in next section.
5. EXPERIMENTAL RESULTS
As stated above, the FPGA-based monitoring system had already been experimentally proven to
have a very good performance in an actual industrial environment. The same applies to the
“static” PI controller in some operating conditions. Therefore, this section concentrates on the
situations where improvements were required, to demonstrate that such improvements can be
achieved with the proposed adaptive controller. Previous results were obtained with the diode
laser available at the Laser Application Center for separate tracks. The experiments reported here
have been obtained with the Nd:YAG laser for overlapping tracks. In each experiment, 13 linear
clad tracks were created on the substrate (a rectangular plate with 100 mm length, 20 mm width,
and 8 mm thickness), with an overlapping of 45% between consecutive tracks. At the end of each
track, the laser beam is turned off. The robot then moves to the starting point of the next track,
and once there the laser beam is turned on again. Because of the geometric characteristics of the
beam generated by the Nd:YAG laser, in this case length (the major axis of the ellipse) is the target
dimension, instead of width. This is not a problem since, as stated in Section 3, the image
processing peripheral actually computes both dimensions.
Experiments have been conducted with both the “static” PI controller and the adaptive one for
two different set points. It is interesting to note that the adaptive controller starts operation with
its last settings (i.e., in the first experiment, the initial settings were the final ones of an
experiment conducted with the diode laser).
The camera and the image processing peripheral have been configured at 800x600 pixel
resolution, resulting in a sampling period of 10 ms (i.e., the control system works at 100 Hz). It is
important to note that the limiting factor is the exposure time of the camera. By improving the
gain of the optical system, the performance of the control one can be improved. This is a subject
of current work.
Figure 11 shows the instantaneous laser power and length measurements in experiments where
the target length was 2.35mm. In this and subsequent figures, blue plots correspond to the
“static” controller and green plots to the adaptive one. It can be seen that both controllers provide
similar results, although the average power is slightly higher (≈1%) for the adaptive controller
(810W) than for the “static” controller (802.1W). In spite of the “noisy” (as expected)
measurements, the average length in both cases (2.34mm) is very close to the set point. It can be
noticed that both controllers react correctly in this case to the heating of the parts as the
experiment goes on, by gradually reducing the power command. It can be also noticed that the
control system keeps providing a power command during the intervals between tracks (where the
laser system is actually off). This is due to the fact that the laser systems require an initial power
command when turned on.
Figure 12 shows pictures of the actual parts after treatment, where the 13 tracks can be identified.
Below them there is the area affected by the treatment (the important one), whose lower limit is
identified by a change in color, which separates it from a second area subject to thermal
affectation. The corresponding dilution measurements are shown in Fig. 13, where target ranges
(according to the expected result of the treatment) are defined by horizontal dashed lines. An
excessive dilution is obtained in all cases in the first track. This is due to the fact that, at the
(relatively low) power range corresponding to the target length, heating of the part is (relatively)
slow, which means that initial length measurements do not correspond as accurately as expected
to resulting dilution (i.e., illumination is initially weaker than expected given the power density
applied to the part). Results have been obtained this way to highlight the problem, but it is
important to note that it can be solved by having a different (shorter) length set point (or a
different threshold for binarization) at the initial stages of low-power runs.
In the case of the adaptive controller, all other results are within limits, except one slightly over
the upper limit for the start point of track #12. In the case of the “static” controller, most of the
end points have a dilution below the lower limit. These deviations being very small, results in both
cases are in principle good enough from a metallographic point of view. Anyway, a slightly better
performance can be claimed for the adaptive controller.
It is worth noting the destructive nature of the tests that allow actual results to be accurately
assessed, since parts need to be cut at different places to measure dilution. This adds to the need
to stop processing of customer parts. It is therefore very important from the industrial point of
view to reduce the number of tests (and the related uncertainties) to the minimum. This is the
reason why the system has been evolved in step-by-step limited improvements, to overcome
problems as they are identified. This also justifies the simple fuzzy structure used in principle,
which became the final one given the good results it provides, as demonstrated below. The
flexibility provided by the reconfigurable nature of FPGAs has also been instrumental in the
evolution of the system through successive refinements.
Experiments were repeated for a target length of 3.055mm. The corresponding laser power and
length measurements are shown in Fig. 14. It can be clearly seen how the change in operating
conditions affects the “static” controller, whereas the adaptive controller performs significantly
better. For the last three tracks, changes in the plant, caused by increased heating, make the
“static” controller to have a wrong behavior. At first sight, it may seem that the power reduction
when applying those tracks is an advantage but, as discussed below, this is denied by length and,
more importantly, dilution measurements. Even with this reduction at the end of the experiment,
in this case the average power with the adaptive controller (1229 W) is lower than with the
“static” one (1238.2 W). If the last three tracks are not considered (so comparison is made when
both systems are behaving correctly), the difference is more significant (1237.8 W vs. 1313.7 W).
Absolute length measurements are shown in Fig. 14 together with 10-sample average plots
superimposed, which allow the behavior of the otherwise noisy absolute data to be better
analyzed. It can be clearly seen that, although the average length value for the whole experiment
is closer to the set point for the static controller (3.05 mm and 3.04 mm, both are actually very
close to the set point), length measurements are less variable for the adaptive controller.
Actual parts from this second set of experiments are shown (only middle points, for the sake of
conciseness) in Fig. 15, and dilution plots in Fig. 16. In this case it can be clearly seen that the
adaptive controller performs much better than the “static” one, where in addition the power
reduction at the end of the experiment results in an unacceptable decrease of dilution. Anyway, a
trend for dilution to increase above the upper limit can be identified in the results for the adaptive
controller. By analyzing width results (since, as stated above, both dimensions are computed by
the image processing block), it can be concluded that this trend corresponds to an increase in
width even if length is kept constant. Therefore, if more accurate results are required in high-
power applications, both dimensions have to be considered for controlling the process.
6. CONCLUSION
This paper demonstrates the suitability of FPGAs for optimizing the monitoring and control of a
complex real industrial process, namely laser cladding, at the Laser Application Center of the
technology center AIMEN. With this purpose, a Field Programmable System-on-Chip has been
developed, which combines an embedded microprocessor core (Nios-II) with custom high-
performance hardware peripherals.
The resulting system is capable of working at a control sampling rate of 100 Hz with an image
resolution of 800x600 pixels, and clearly outperforms existing alternatives, both from the
instrumentation and the control points of view. Given the inherent non-repetitiveness of the laser
cladding process, and the strong requirements in terms of homogeneity of the treatment along
the target parts, it is not only necessary to have a fast control system with enough resolution in
the optical part, but also it has to provide a good response from the “noisy” measurements
obtained in practice. This goal has been achieved by implementing an adaptive PI controller,
whose gains are continuously updated in real time according to a set of fuzzy rules. An additional
advantage of this approach is that the system can be used with different types of lasers, and under
significantly different operating conditions, without the need for any modification.
The experimental results presented demonstrate the claimed capabilities of the proposed system.
In addition, its analysis has allowed areas where further improvements can be obtained to be
identified. When heating effects go above a certain threshold, it has been detected that
considering only the effect of melt pool length (or width) on dilution results in overtreatment of
the parts. Under these conditions, the dependence on both dimensions (or on the area) can be
considered to obtain more accurate results. A complementary approach under study is the use of
grayscale images (instead of binarized ones) from which indirect information about temperature
can also be extracted (without the need for a temperature sensor, i.e., a pyrometer). This will only
imply a minor change to adapt moment functions in the image processing peripheral, but will not
result in any significant loss of processing speed.
The limiting factor to increase processing speed is the exposure time of the camera, which will be
reduced by increasing the gain of the optical system. Once this is accomplished, another possibility
for improvement will be the implementation of the adaptive controller in hardware.
ACKNOWLEDGMENTS
This research was partially supported under Xunta de Galicia 09DPI029CT project, and Spanish
Ministerio de Ciencia e Innovación TEC2010-21429-C02-01 project, including European FEDER
funds.
BIOGRAPHIES
Jorge Rodríguez-Araújo (jorge.rodriguez@aimen.es) received the M.Sc. degree in electrical
engineering from the University of Vigo, Spain, in 2010. He has been working several years on
research and development of FPGA solutions, especially for visual monitoring and real-time
control. Currently, he works in AIMEN Technology Centre in the Design, Simulation and
Automation area. He has authored several papers dealing with the design of FPGA systems applied
to laser processing.
Juan J. Rodríguez-Andina (jjrdguez@uvigo.es) received the M.Sc. degree in electrical engineering
from the Technical University of Madrid, Spain, in 1990 and the Ph.D. degree (with honors) in
electrical engineering from the University of Vigo, Spain, in 1996. He is currently an Associate
Professor with the Department of Electronic Technology, University of Vigo. From August 2010 to
June 2011 he was a Visiting Professor at the Electrical and Computer Engineering Department,
North Carolina State University, Raleigh, NC. He has authored over 130 journal and conference
papers and is the holder of several Spanish, European, and U.S. patents. His research interests
include implementation of complex processing and control algorithms in FPGAs and concurrent
testing of complex systems, with emphasis in industrial applications. He is a Senior Member of
IEEE.
José Fariña (jfarina@uvigo.es) received the M.Sc. degree in electrical engineering from the
University of Vigo, Spain, in 1984 and the Ph.D. degree (with honors) in electrical engineering from
the University of Vigo, Spain, in 1989. He is currently an Associate Professor with the Department
of Electronic Technology, University of Vigo. He has authored over 90 journal and conference
papers and is the holder of several Spanish, European, and U.S. patents. His research interests
include troubleshooting industrial control through the implementation of complex processing
algorithms in FPGAs. He is a Member of IEEE.
Félix Vidal (fvidal@aimen.es) received the M.Sc. degree in electrical engineering from the
University of Vigo, Spain, in 2003. He has managed several industrial collaborative projects and
participated in several Spanish and European projects dealing with real-time process control,
robotics, hardware design, and NDT development. Currently, he works in AIMEN Technology
Centre in Design, Simulation and Automation area. He has authored several papers dealing with
the implementation of control systems applied to laser processing. He is an International Welding
Engineer (IWE).
José Luis Mato (jmato@aimen.es) received the M.Sc. degree in electrical engineering from the
University of Vigo, Spain, in 2005, where he is working toward the Ph.D. degree in the Department
of Electronic Technology. His research interests include the implementation of complex processing
algorithms in FPGAs. Currently, he works in AIMEN Technology Centre, in Design, Simulation and
Automation area, on system monitoring and RT-process control for many applications in different
industrial sectors.
Mª Ángeles Montealegre (amontealegre@aimen.es) received the M.Sc degree in Metallurgical
Chemistry from the Complutense University of Madrid, Spain in 1998 and the Ph.D. degree (with
honors) at the Department of Material Sciences and Metallurgy Engineering of the Complutense
University of Madrid, in 2004. She is currently a senior researcher in the Laser application Centre
of AIMEN Technology Centre. She has authored a number of papers in technical and scientific
journals. Her research interests includes laser surface engineering: laser hardening and surface
modification and coating by means of laser technologies. She is an International Welding Engineer
(IWE).
REFERENCES
[1] L. Sexton, S. Lavin, G. Byrne, and A. Kennedy, “Laser cladding of aerospace materials”, Journal
of Materials Processing Technology, Vol. 122, Issue 1, pp. 63-68, March 2002.
[2] L. Shepeleva, B. Medres, W. D. Kaplan, M. Bamberger, and A. Weisheit, “Laser cladding of
turbine blades”, Surface and Coatings Technology, vol. 125, pp. 45-48, Mar. 2000.
[3] J. Y. Jenga and M. C. Lin, “Mold fabrication and modification using hybrid processes of
selective laser cladding and milling”, Journal of Materials Processing Technology, vol. 110, pp.
93-110, Mar. 2001.
[4] E. Monmasson, L. Idkhajine, and M.W. Naouar, “FPGA-based controllers”, IEEE Industrial
Electronics Magazine, vol. 5, no. 1, pp. 14-26, Spring 2011.
[5] E. Monmasson, L. Idkhajine, M. N. Cirstea, I. Bahri, A. Tisan and M. W. Naouar, “FPGAs in
Industrial Control Applications”, IEEE Trans. on Industrial Informatics, vol. 7, pp. 224-243, May
2011.
[6] J. L. Mato, M. Pereira, J. J. Rodríguez-Andina, J. Fariña, E. Soto, and R. Pérez, “Distortion
mitigation in RF power amplifiers through FPGA-based amplitude and phase predistortion”,
IEEE Trans. on Industrial Electronics, vol. 55, pp. 4085-4093, Nov 2008.
[7] E. Monmasson and M. N. Cirstea, “FPGA design methodology for industrial control systems. A
review”, IEEE Trans. on Industrial Electronics, vol. 54, pp. 1824-1842, Aug. 2007.
[8] J. J. Rodriguez-Andina, M. J. Moure, and M. D. Valdes, “Features, design tools, and application
domains of FPGAs”, IEEE Trans. on Industrial Electronics, vol. 54, pp. 1810-1823, Aug. 2007.
[9] P. Colodrón, J. Fariña, J.J. Rodríguez-Andina, F. Vidal, J.L. Mato, and M.A. Montealegre,
“FPGA-based measurement of melt pool size in laser cladding systems”, in Proc. 2011 IEEE
International Symposium on Industrial Electronics, ISIE’2011, pp. 1503-1508..
[10] J. Rodríguez-Araújo, J.J. Rodríguez-Andina, J. Fariña, F. Vidal, J.L. Mato, and M.A.
Montealegre, “FPGA-based laser cladding system with increased robustness to optical
defects”, in Proc. 38th Annual Conference of the IEEE Industrial Electronics Society, IECON
2012 (in press).
[11] P. Colodrón, J. Fariña, J.J. Rodríguez-Andina, F. Vidal, J.L. Mato, and M.A. Montealegre,
“Performance improvement of a laser cladding system through FPGA-based control”, in Proc.
37th Annual Conference of the IEEE Industrial Electronics Society, IECON 2011, pp. 2814-2819.
[12] L. Song and J. Mazumder, “Feedback control of melt pool temperature during laser cladding
process”, IEEE Trans. on Control System Technology, vol. 19, no. 6, pp. 1349-1356, Nov. 2011.
[13] M. Iravani-Tabrizipour and E. Toyserkani, “An image-based feature tracking algorithm for real-
time measurement of clad height”, Machine Vision and Applications, vol. 18, no. 6, pp. 343-
354, Dec. 2007.
[14] J. T. Hofman, D. F. De Lange, and J. M. Meijer, “Camera based feedback control of the laser
cladding process”, in Proc. 25th Int. Congress on Applications of Lasers & Electro-Optics, pp.
456-460, 2006.
[15] L. Song, V. Bagavath-Singh, B. Dutta, and J. Mazumder, “Control of melt pool temperature and
deposition height during direct metal deposition process”, The International Journal of
Advanced Manufacturing Technology, vol. 58, nos. 1-4, pp. 247-256, Jan. 2012.
[16] R. Mukundan and K. R. Ramakrishnan, Moment Functions in Image Analysis: Theory and
Applications, World Scientific, 1998.
[17] E.H. Mamdani and S. Assilian, “An experiment in linguistic synthesis with a fuzzy logic
controller”, International Journal of Man-Machine Studies, vol. 7, no. 1, pp. 1-13, 1975.
FIGURES
Fig. 1. Laser cladding process
Fig. 2. Sample empirical verification of the linear dependency of dilution on melt pool width
0,15
0,2
0,25
0,3
0,35
0,4
0,45
2,6 2,8 3 3,2 3,4 3,6 3,8
Width (m m )
Dilution (mm)
Fig. 3. Block diagram of the laser cladding system at the Laser Application Center
Fig. 4. Optical system at the Laser Application Center
Fig. 5. Block diagram of the proposed FPGA system
Fig. 6. Sample images captured (left) and binarized (right) during one of the experiments with the diode laser system
Fig. 7. Sample original (left) and eroded (right) binary images; the red dot represents the pixel size.
Fig. 8. Block diagram of the proposed system based on an adaptive digital PI controller
Fig. 9. Fuzzy membership functions
Table I: Rule Base
AND
Negative
Zero
Positive
Negative
Large
Medium
Small
Zero
Medium
Small
Medium
Positive
Small
Medium
Large
Fig. 10. Fuzzy control surface
Fig. 11. Instantaneous laser power (top) and melt pool length (bottom) in experiments with the Nd:YAG laser system (length set point 2.35mm)
Fig. 12. Actual parts after treatment in the experiments of Fig. 11. Results for the “static” (left) and adaptive (right) controllers in three different
positions: start (A), middle (B), and end (C) points of the tracks
Track
Dilution
Area
Thermal
Affectation Area
Fig. 13. Dilution in the parts of Fig. 12.
Fig. 14. Instantaneous laser power (top) and melt pool length (middle and bottom) in experiments with the Nd:YAG laser system (length set point
3.055mm)
Fig. 15. Actual parts after treatment in the experiments of Fig. 14. Results for the “static” (top) and adaptive (bottom) controllers in the middle
points of the tracks
Fig. 16. Dilution in the parts of Fig. 15.
