MACRo 2015- 5th International Conference on Recent Achievements in

Mechatronics, Automation, Computer Science and Robotics

Hardware Implementation of a Neuro-Fuzzy Controller

Using High Level Synthesis Tool
Tibor TMAS1, Sndor Tihamr BRASSAI2
1 Department of Electrical Engineering, Faculty of Technical and Human Sciences,
Sapientia University, Tg. Mure, e-mail: tamas_tibor2000@yahoo.com
2 Department of Electrical Engineering, Faculty of Technical and Human Sciences,

Sapientia University, Tg. Mure, e-mail: tiha@ms.sapientia.ro

Manuscript received January 25, 2015, revised February 9, 2015.

Abstract: The purpose of this work is to present the design flow and the
implementation of a neuro-fuzzy controller Intellectual Property (IP) core, using High
Level Synthesis (HLS) tool. The realized IP core is designed for FPGA based embedded
system architectures. The implemented control algorithm is a Sugeno model based
Adaptive Neuro-Fuzzy Inference System (ANFIS). The optimization possibilities using
the HLS tool and the designing of the interfaces for the IP core are presented.

Keywords: ANFIS, FPGA HLS, Fuzzy, Adaptive, neuro-fuzzy

1. Introduction
For the control engineering, the neuro-fuzzy system has many favorable
features. First of all it combines both the neural network structures and the
fuzzy inference systems advantages. The neural network structure gives the
adaptability of the system, and the fuzzy logic part makes it possible to use a
linguistic rule base to describe the ideal behavior of the controlled system.
Beside the advantages of this kind of systems, there are some difficulties with
the real time implementation, because the algorithm requires complex
operations under time constrains. The sequential implementations in most of the
cases cannot work in real time, but a parallelized structure of the operations
could be a solution. There are many related works to the implementations of
neuro-fuzzy models, using different techniques [1], [2], [3]. The best platform
to realize embedded systems with parallel architectures is the Field
Programmable Gate Array (FPGA). With the flexibility of these devices it is
possible to create an efficient hardware implemented neuro-fuzzy controller.
The FPGA devices can be programmed using Hardware Description Languages
(HDL), this is an efficient way to create embedded hardware, but also requires

183 10.1515/macro-2015-0018

Brought to you by | provisional account

Unauthenticated
Download Date | 10/27/16 10:34 AM
184 T. Tamas, S. T. Brassai

deep electrical understanding of the circuit. Nowadays, we have the possibility

to create FPGA based hardware by describing the behavior of the circuit using
high level programming languages, like C/C++/SystemC. This type of codes
can be synthetized to HDL with HLS tools, like Xilinx Vivado HLS [4]. This
type of designing hardware architectures has many advantages, like shorter
design time, easier debugging solutions and the detailed knowledge of the
Register Transfer Level (RTL) circuit is not necessary. Until few years ago, the
HLS tools hadnt been advanced enough to use them to create efficient
hardware architectures, but with the recent improvements and with the new high
capacity FPGA boards this could be a reasonable option [5], [12], [13].

2. Sugeno model based ANFIS structure

The Sugeno fuzzy model has many similarities with the Mamdani fuzzy
model, the main difference between them is the defuzzification part. While the
Mamdani model uses an output space covered with membership functions, the
Sugeno model uses a linear function for each rule. From these active function
values a weighted average is calculated. Because the Mamdani defuzzification
requires the calculation of the area and the center of gravity of the active output
membership functions it demands much more resources then the Sugeno
defuzzification. The general structure of the Sugeno model based ANFIS can be
seen on Figure 1. We can see that the neural network has two inputs and one
output, and also has five layers. The first layer is actually the fuzzification layer
(1), (2), the nodes of this layer are the input membership functions. Each
membership function node is connected to only one input, in this case five
functions for each input. The input value activates maximum two membership
functions, for each input.
1, = (), = 1. .5 (1)
1, = (), = 6. .10 (2)
On the second layer the active membership functions from the two inputs are
combined, and a product is calculated from each pair of the membership values
(3).
2, = = () () (3)
The third layer should be a normalization for the calculated values in the
previous layer, but for resource saving reasons this operation is done just after
the summation of the active output values. In this case the third layer is part of
the defuzzification, these nodes contain the functions corresponding to the
active rules (4).

Brought to you by | provisional account

Unauthenticated
Download Date | 10/27/16 10:34 AM
 Hardware Impl. of a Neuro-Fuzzy Controller Using High Level Synthesis Tool 185

O3,k = wk *fk = wk *(pk *X + qk *Y + rk ) (4)

The fourth and the fifth layers are a summation of the active rule function
values and the normalization of the weights (5), (6).
O4 = k wk *fk (5)
O4
O5 = k wk
(6)

Figure 1: Sugeno model based ANFIS structure.

3. Implementation of the algorithm
The target device for this architecture is a Zynq System on Chip type
development board, with a dual-core ARM Cortex A9 hard-processor and a
programmable logic part. The ANFIS IP core is implemented into the
programmable logic, and the learning algorithm will run on the processor. The
ANFIS algorithm was implemented in C language, using the Xilinx Vivado
High Level Synthesis tool. The HLS tool, during the synthesis generates an
FSM based hardware architecture from the C code. During the synthesis the
technical constraints of the target device are taken in account, the timing
constraints, and the directives assigned to some variables or program structures.
Only integer operations are used, because the floating point operations require
much more resources. The program has three main parts, the ANFIS function,
the Fuzzification function, and the Bell_Curve function. The ANFIS function is

Brought to you by | provisional account

Unauthenticated
Download Date | 10/27/16 10:34 AM
186 T. Tamas, S. T. Brassai

the main function, the inputs are the two inputs of the controller (X, Y) and the
third layers parameters (L3_par), and the outputs are the calculated control
signal (U), and the Ret parameter structure which contains data for the learning
algorithm. The Fuzzification function (L1) calculates the membership values for
each input value and returns to the main function the indices of the active
membership functions and the corresponding membership values.

Figure 2: ANFIS IP core block diagram.

As membership functions, we use bell curves which have the formula (7),
but because we can use only integer operations, we rescaled the [0, 1] interval
to [0, 255] with the formula (8).
1
(x) = 2b (7)
x-c
1+| |
a

255*|a|2b
(x) = (8)
|a|2b +|x-c|2b

Figure 3: Membership functions of one input space

Brought to you by | provisional account

Unauthenticated
Download Date | 10/27/16 10:34 AM
 Hardware Impl. of a Neuro-Fuzzy Controller Using High Level Synthesis Tool 187

The Bell_Curve function is used by the Fuzzification function, and calculates

the membership value from the input value and the parameters of the respective
membership function. The membership functions of one input space could be
seen on Figure 3. The [0, 1024] intervals of the two input spaces are covered by
five-five membership functions.

To optimize the code, optimization directives are applied to some elements of

the program. Pipeline and unroll directives for some loops or DATA_PACK
and ARRAY_MAP directives are used for resource saving reasons. With this
optimization, the synthesis resulted in a latency of 68 clock cycles at 100 MHz
which means 680 ns for the calculation of an output. In the 1. Table the
quantities of the utilized resources can be seen.
1. Table Resource utilization
Name BRAM_18K DSP48E FF LUT
Expression - 12 0 869
FIFO - - - -
Instance - 36 4824 4792
Memory 0 - 64 8
Multiplexer - - - 43
Register - - 1808 -
Total 0 48 6696 5712
Available 280 220 106400 53200
Utilization (%) 0 21 6 10

We can see, that only 21% of the available DSP-s, 6% of the available flip-flops
and 10% of the available LUTs were utilized to generate the ANFIS IP core.
There are many training algorithms for this kind of systems [6], [7], [10]. The
learning algorithm implemented also in C language is running on the processor.
The Ret structure contains the indices of the active rules and the weights of the
second layer (w) for each active rule. The processors software refreshes the
parameters with the following formulas:
pindexX[i],indexY[i] = pindexX[i],indexY[i] + *error*X*w
[i] (9)
qindexX[i],indexY[i] = qindexX[i],indexY[i] + *error*Y*w
[i] (10)
rindexX[i],indexY[i] = rindexX[i],indexY[i] + *error*w
[i] (11)

Brought to you by | provisional account

Unauthenticated
Download Date | 10/27/16 10:34 AM
188 T. Tamas, S. T. Brassai

Where indexX and indexY contains the indices of the active membership
functions on the two inputs, is the learning coefficient,
is the normalized
weight of the second layer.

4. Interfaces
The ANFIS IP core has two types of interfaces, AXI4Lite slave [8] interface
and BRAM interface. The whole architecture, containing the ANFIS IP core,
and the processor can be seen on the Figure 4. The X and Y inputs and the U
control signal, and the Rett data structure outputs of the ANFIS IP core are
using the AXI4Lite interface.

Figure 4: Interface architecture of the processor base embedded system

Each of these four inputs/outputs has a handshake type protocol assigned,
and each of them can be accessed through a register. The third layers parameters
are stored into a dual-port BRAM memory. The Dual-port BRAM has the
advantage that the contents can be read through the port A and can be written
through the port B at the same time. So the ANFIS reads the parameters for the
current calculation, and at the same time, the processor calculates the new
adapted parameters for the previous inputs and writes them into the BRAM. The
BRAM is connected to the processor via an AXI BRAM controller. The AXI
channel from the ANFIS IP core, and the AXI BRAM controller are connected
to an AXI Interconnect module and this module makes the connection to the
processor.
5. Simulation results
For the training simulation we choose a reference surface (Figure 5), which
was the result of a simulation with a fuzzy controller. The surface is generated

Brought to you by | provisional account

Unauthenticated
Download Date | 10/27/16 10:34 AM
 Hardware Impl. of a Neuro-Fuzzy Controller Using High Level Synthesis Tool 189

from the control signal values calculated according to the input values from the
X and Y axis.

Y
Figure 5: Reference surface
On the Figure 6 the surface generated from the simulations made with the
ANFIS IP core can be seen. We can see the evolution of the surface during the
training cycles. These plots are made after 5, 50, 75 and 100 training cycles.
There are used five membership functions for each input space with the
parameters shown in 2. Table. The reference surface contains 60 x 60 samples.
2. Table: Parameters of the membership functions
a - dispersion b - steepness c - center
MF1 128 1.5 0
MF2 128 1.5 256
MF3 128 1.5 512
MF4 128 1.5 768
MF5 128 1.5 1024

Brought to you by | provisional account

Unauthenticated
Download Date | 10/27/16 10:34 AM
190 T. Tamas, S. T. Brassai

Figure 6: Output surface at different number of training cycles

6. Conclusion
We have presented the hardware implementation of a neuro-fuzzy system
with high level synthesis tool. We proved the usability of the high level
synthesis tools in creation of efficient hardware architectures. The hardware
implemented with HLS requires less design time then other techniques, there
are many optimizing possibilities and a simple method to create interfaces.
Reasoning from the low latency, from the acceptable resource utilization and
the measurement results we can conclude that the ANFIS IP core can be used in
real-time control applications.

Acknowledgements

This publication/research has been supported by the European Union and

Hungary and co-financed by the European Social Fund through the project
TMOP-4.2.2.C-11/1/KONV-2012-0004 - National Research Center for
Development and Market Introduction of Advanced Information and
Communication Technologies.

Brought to you by | provisional account

Unauthenticated
Download Date | 10/27/16 10:34 AM
 Hardware Impl. of a Neuro-Fuzzy Controller Using High Level Synthesis Tool 191

References
[1] Echanobe J., del Campo I., Bosque G., "An adaptive neuro-fuzzy system for efficient
implementations," Information Sciences, vol. 178, pp. 2150-2162, 208.
[2] Nedjah N., da Silva R., de Macedo Mourelle L., "Compact yet efficient hardware
implementation of artificial neural networks with customized topology," Expert Systems with
Applications, vol. 39, no. 10, pp. 9191-9206, 2012.
[3] del Campo I., Echanobe J., Bosque G., Tarela J. M., "Efficient Hardware/Software
Implementation of an Adaptive Neuro-Fuzzy System," IEEE TRANSACTIONS ON FUZZY
SYSTEMS, vol. 16, no. 3, pp. 761-779, 2008.
[4] Cong, J., Bin Liu, Neuendorffer, S., Noguera, J., Vissers, K, Zhiru Zhang, "High-Level
Synthesis for FPGAs: From Prototyping to Deployment," Computer-Aided Design of Integrated
Circuits and Systems,IEEE Transactions on, vol. 30, no. 4, pp. 473-491, 2011.
[5] Skalicky S., Wood C., ukowiak M., Ryan M., "High Level Synthesis: Where Are We? A
Case Study on Matrix Multiplication," in Reconfigurable Computing and FPGAs (ReConFig),
2013.
[6] Oentaryo R. J., Joo M., Linn S., Li X., "Online probabilistic learning for fuzzy inference
system," Expert Systems with Applications, vol. 41, p. 50825096, 2014.
[7] Teshnehlab M., Shoorehdeli M.A. Sedigh A.K, "Novel Hybrid Learning Algorithms for
Tuning ANFIS Parameters as an Identifier Using Fuzzy PSO," in Networking, Sensing and
Control, IEEE International Conference, 2008.
[8] Xilinx, Vivado Design Suite - Vivado AXI Reference Guide, November 20, 2014.
[9] Xilinx, Vivado Design Suite User Guide, May 30, 2014.
[10] Silva M., Caminhas W., Lemos A., Gomide F, "A fast learning algorithm for evolving neo-
fuzzy neuron," Applied Soft Computing, vol. 14, pp. 194-209, 2014.
[11] Echanobe J., del Campo I., Basterretxea K., Martinez M.V., Doctor F., "An FPGA-based
multiprocessor-architecture for intelligent environments," Microprocessors and Microsystems,
vol. 38, pp. 730-740, 2014.
[12] Casseau E., Le Gal B., "Design of multi-mode application-specific cores based on high-
level synthesis," INTEGRATION, the VLSI journal, vol. 45, pp. 9-21, 2012.
[13] Prost-Boucle A., Muller O., Rousseau F., "Fast and standalone Design Space Exploration
for High-Level Synthesis under resource constraints," Journal of Systems Architecture, vol. 60,
pp. 79-93, 2014.

Brought to you by | provisional account

Unauthenticated
Download Date | 10/27/16 10:34 AM