Content uploaded by Emmanuel Oyekanlu
Author content
All content in this area was uploaded by Emmanuel Oyekanlu on Mar 14, 2019
Content may be subject to copyright.
Towards Low-Cost, Real-Time, Distributed Signal and Data Processing for Artificial Intelligence
Applications at Edges of Large Industrial and Internet Networks
Emmanuel Oyekanlu
Electrical and Computer Engr. Department,
Drexel University, Philadelphia,
Pennsylvania, USA.
eao48@drexel.edu
Kevin Scoles
Electrical and Computer Engr. Department
Drexel University, Philadelphia
Pennsylvania, USA
Abstract—
Digital Signal Processors (DSP) are vital system
components in industrial Artificial Intelligence (AI) applications. In
this paper, FIR filters that could be used for industrial AI applications
are designed from the Spline Biorthogonal 1.5 (SB1.5) mother wavelet
using a real-time, low-cost, generic industrial IoT (IIoT) hardware: the
C28x DSP
and low-level, Embedded C, system software. Our
contribution in this paper is the first reported application of the C28x
for SB1.5 wavelet construction. The significance of this approach is to
be able to repurpose low-cost, readily available hardware for
distributed AI applications. Our approach is different from the state of
the art, in which specialized hardware are
always manufactured for
implementing AI applications at large network edges. Our approach
supports low-cost and fast single-stage FIR implementation suitable for
use in real-time, distributed AI application at network edges, since in
our case, successive recursion of FIR filters leading to a full
implementation of Pyramid Algorithm is not implemented. The
designed FIR filter is evaluated and found capable of both low-
pass
and
high pass filtering operations. Results of this paper indicate that the
C28x real-time DSP, which exists in many IoT devices, could have
improved scalability by being deployed for other important AI and IoT
network edge analytic applications, different from its present uses.
Keywords- embedded systems, real-time, AI, FIR filter, wavelet
I. INTRODUCTION
DSPs are used extensively in AI applications. Real-time
DSPs are useful in preprocessing data and signals, in filtering
applications, in signal denoising and in feature extraction
applications for AI systems [1], [2]. In the past few decades,
wavelet transform methods, fully implemented on DSPs have
been used as integral part of several AI preprocessing steps
such as in image compression, computer vision and
audiovisual applications especially in the cloud layer of
Internet of Things (IoT) systems. The Mallat’s Pyramid
algorithm, a fast implementation of the Discrete Wavelet
Transform which involves successive recursion of FIR filters
with increasing resolution is often used in most wavelet-based
AI applications [3]. An illustration of Pyramid algorithm for a
two-step recursion is shown in Fig. 1. However, most edge
computing IoT devices always have low computing po
wer
with small memory, and as a result, the amount of analytics
works they can accomplish is often minimal. Hence most AI
algorithms due to their needs for excessive computing power
are often deployed in IoT clouds [4]. To deploy more AI
applications to the edge and fog layers of IIoT systems, our
contribution in this paper is to facilitate distributed computing
approach by which FIR filters, useful for pre-processing
applications in AI can be designed at the
edge and fog layer of
IIoT systems using generic, widely available IoT DSPs.
Towards this end, similar to the approach used in [3], we have
used the widely available C28x DSP, manufactured by Texas
Instruments (TI), and deployed in estimated billions of edge
computing devices [5],
to construct the SB1.5 mother wavelet
using low-level Embedded C programming language.
Embedded C is suitable for computing applications for edge
devices that have low computing power [6]. An advanced
signal processing method involving construction of sine wave,
then signal clipping, sequencing and concatenation as shown
in Fig. 2 is used to construct the SB1.5 in TI’s real-time Code
Composer Studio (CCS) programming environment.
Different from the approach used in [3]
however, as shown in
Fig. 1, we have not implement the full Pyramid algorithm.
Block diagram showing our contribution in this paper is
shown in Fig. 3. After constructing the SB1.5 wavelet,
samples of the SB1.5 wavelet are then exported to Matlab, and
used to construct FIR filters, useful for AI preprocessing
applications. The main reason for selecting the SB1.5 wavelet
for our design in this paper is that, it is a biorthogonal and
symmetric wavelet, which gives it the desirable property of
being a linear phase wavelet [7]. Hence, FIR filter that results
from the SB1.5 wavelet will be linear phase. Phase linearity is
an excellent property for filters deployed in many applications
in AI including data compression and modulation, computer
vision, image processing and in AI robot vision and target
tracking applications [8]. Section II of the paper shows our
Clipped signal
Sequencer
=
Fig. 2 Schematic of the Spline Biorthogonal 1.5
wavelet design for the C28x DSP
C
1
(k)
C
0
(k)
d
0
(k)
Our FIR filter derivation in this paper
x(n)
h(n)
w(n)
h(n)
w(n)
Conventional full Mallat’s Pyramid Algor ithm (DWT) filter bank for J = 2 octaves
Fig. 1. Comparison of our SB1.5 wavelet-based FIR filter for AI
application discussed in this paper compared to full Pyramid algorithm
166
2018 IEEE First International Conference on Artificial Intelligence and Knowledge Engineering (AIKE)
978-1-5386-9555-5/18/$31.00 ©2018 IEEE
DOI 10.1109/AIKE.2018.00037
results and the design of FIR filter for AI applications. Section
III is conclusion of the paper.
II. DESIGN OF FIR FILTERS BASED ON THE SPLINE
BIORTHOGONAL 1.5 WAVELET AND C28XIOT DSP FOR
ARTIFIC IAL INTELLIGENCE A
PPLICATIONS
A generic IIoT edge device, the TMS320C2000 C28x
powerline communication (PLC) modem is used to show our
distributed computing approach used in this paper. The C28x
PLC modem is attached to a PC running CCS. The advanced
signal processing method described in Fig. 2 is used to
construct the SB1.5 wavelet using Embedded C and the result
is shown in Fig. 4. This is used to signify the distributed edge
computing for AI application at the edge of the IIoT network.
To show distributed edge computing at the IIoT fog layer,
125
samples of the SB1.5 wavelet is exported to Matlab using
Excel.csv file. The exported wavelet samples are then used to
construct low-pass (LP) and high pass (HP) filters of order 20
suitable for use in AI preprocessing applications. Matlab ‘fir1’
(FIR filter) toolbox is used at this layer since it is assumed that
fog layer devices can run more compute intensive applications
such as Matlab. Kaiser widow is used. A Kaiser window
allows better control over passband and stopband ripples
when compared to other window types. The designed LP and
HP FIR filters using SB1.5 wavelet samples are obtained and
plotted in Fig. 5. Results shows that the filters have the desired
linear phases with LP and HP characteristics. To evaluate the
performances of the designed FIR filters, sine waves of 150
Hz, 600 Hz, 2.8 kHz and 3.4 kHz respectively are generated
and passed through the filters using their fast Fourier
transforms (FFTs). Results for LP and HP FIR filters are
shown in Fig. 6. The FIR LP performs by filtering out the
higher frequency 2.8 kHz and 3.4 kHz signals and allows
lower frequency signals. Similarly, the HP FIR filter performs
by filtering out the lower frequency 150 Hz and 600 Hz
signals and allows the high frequency signals.
III. CONCLUSION
In this paper, a distributed computing approach that will
enable generic, low-cost, low-memory, DSPs to participate in
distributed computing for AI applications at edges of large
scale IoT networks is discussed. Real-time C28x DSP is used
with Embedded C to design the SB1.5 wavelet at the edge
layer and Matlab is used at the fog layer of IIoT system to
construct FIR filter suitable for AI applications out of the
constructed SB1.5 wavelet. Future work will include complete
design of the both wavelet and filter at the edge layer for AI
applications.
REFERENCES
[1] J. Szalai, F. Mozes, “Intelligent Digital Signal and Feature Extraction
Methods”, Springer Int’l Publishing, 2016
[2] R. Oshana, “DSP Software Development Techniques for Embedded
and Real-Time Systems”, Elsevier, 2006
[3]
Y. Guo et’al, “VLSI Implementation of Fast Discrete Wavelet
Transform Algorith with Reduced Complexity”, IEEE, 2001.
[4] B. Dickson, “How Do You Bring Artificial Intelligence form the Cloud
to the Edge?”, online, available,
www.thenextweb.com/contributors/2017/08/21
[5] Texas Instruments, “Digital Control Developer Forum – Digital
Control Systems Group”, Texas Instruments, 2000
[6] M. J. Pont, “Embedded C”, Addison Wesley, 2002.
[7]
Wavelet Browser, “Wavelet Biorthogonal 1.5”, online, available,
http://wavelets.pybytes.com/wavelet/bior1.5/ March 2018
[8] H. Kennedy, “Multidimensional Digital Smoothing Filters for Target
Detection”, Journal of Sig. Proc., vol. 114, Sept. 2015
Edge Layer Computing for Artificial
Intelligence in large networks:
Distributed Task at Edge Layer:SB1.5
Mother Wavelet construction using
low-level machine language due to
computing resource constraint at
network edges
Fog Layer Computing for Artificial
Intelligence in large networks:
Distributed Task at Fog Layer:Adaptive
Pyramid Algorithm by generating only
FIR filter from constructed wavelet.
Filtering for AI applications. More
computing resource exist at this layer
Cloud Layer Computing for Artificial
Intelligence in large networks:
Distributed Task at Cloud Layer:
Complete Pyramid Algorithm
implementation: Full multi-resolution
analysis (MRA), not implemented in
this paper.
Fig. 3 Block diagram showing the design of the AI preprocessing FIR filter at the edge and fog layer of IIoT system discussed in this paper
Fig. 5. High Pass (left) and Low Pass (right) C28x DSP Spline
Biorthogonal 1.5 wavelet-based FIR filter showing linear phase
Fig. 6. Designed FIR filters work by allowing only high (HPF) and
low (LPF) frequency signals
DAC Voltage Amplitude
Fig. 4. Spline Biorthogonal wavelet constructed for AI applications using
low-level Embedded C for C28x real-time DSP at the edge of large scale
(SG) network
167