Upload
others
View
0
Download
0
Embed Size (px)
Citation preview
1
EC 6014 COGNITIVE RADIO
Unit IV
COGNITIVE RADIO ARCHITECTURE
Architecture is a comprehensive, consistent set of design rules by which a specified set of
components achieves a specified set of functions in products and services that evolve
through multiple design points over time. Introduction of fundamental design rules by
which software-defined radio (SDR), sensors, perception, and automated machine
learning (AML) may be integrated to create aware, adaptive, and cognitive radios
(AACRs) is discussed. These SDRs will have better quality of information (QoI) through
capabilities to observe (sense, perceive), orient, plan, decide, act, and learn (the so-called
OOPDAL loop) in radio frequency (RF) and in the user domains. By performing this
integration, transition from adaptive to a demonstrably cognitive radio (CR) is obtained.
There five complementary perspectives of CR architecture (CRA), namely CRA I, CRA
II, CRAIII, CRA IV and CRA V.
CRA I - Functions, Components, and Design Rules
CRA II - The Cognition Cycle
CRA III - The Inference Hierarchy
CRA IV- Architecture Maps
CRA V- Building the CRA on SDR Architectures
CRA I perspective defines six functional components, black boxes to which are first-level
decomposition of AACR functions and among which important interfaces are defined.
One of these boxes is SDR, a proper subset of AACR. One of these boxes performs
cognition via the Self which is a self-referential subsystem that strictly embodies finite
computing.
CRA II perspective examines the flow of inference through a cognition cycle that
arranges the core capabilities of ideal CR (iCR) in temporal sequence for a logical flow
and circadian rhythm for the CRA. The
CRA III perspective examines the related levels of abstraction for AACR to sense
elementary sensory stimuli and to perceive QoI relevant aspects of a scene consisting of
the user in an environment that includes RFsection.
The CRA IV perspective examines the mathematical structure of this architecture,
identifying mappings among topological spaces represented and manipulated to preserve
set-theoretic properties.
2
CRA V perspective reviews SDR architecture, sketching an evolutionary path from the
Software Communications Architecture/Software Radio Architecture (SCA/SRA) to the
CRA. The CRA is expressed in Radio eXtensible Markup Language (RXML).
4.1 CRA I: Functions, Components, and Design Rules
The functions of AACR exceed those of SDR. Reformulating the AACR with self as a
peer of its own user where it establishes the need for added functions by which the self
accurately perceives the local scene including the user and autonomously learns to tailor
the information services to the specific user in the current RF and physical scene.
AACR Functional Component Architecture
The SDR components and the related cognitive components of iCR appear in Figure 4.1.
The cognition components describe the SDR in Radio eXtensible Markup Language
(RXML).RXML so that the self can know that it is a radio and that its goal is to achieve
high QoI tailored to its own users. RXML intelligence includes a priori radio background
and user stereotypes as well as knowledge of RF and space–time scenes perceived and
experienced. This includes both structured reasoning with iCR peers and cognitive
wireless networks (CWNs), and ad hoc reasoning with users, all the while learning from
experience.
Fig.4.1 CRA Augments with SDR
3
SDR Components
SDRs include a hardware platform with RF access and computational resources, plus at
least one software-defined personality. The SDR Forum has defined its Software
Communication Architecture (SCA) and the Object Management Group (OMG) has
defined its Software Radio Architecture(SRA). These are similar fine-grained
architecture constructs enabling reduced-cost wireless connectivity with next-generation
plug-and-play. These SDR architectures are defined in Unified Modeling Language
(UML) object models, Common Object Request Broker Architecture (CORBA),
Interface Design Language (IDL), and extensible Markup Language (XML) descriptions
of the UML models. The SDR Forum and OMG standards describe the technical details
of SDR both for radio engineering and for an initial level of wireless air interface
(waveform) plug-and-play. The SCA/SRA was sketched in 1996 at the first US
Department of Defense (DoD) inspired modular multifunctional information transfer
system (MMITS) Forum, was developed by the DoD in the 1990s and the architecture is
now in use by the US military. This architecture emphasizes plug-and-play wireless
personalities on computationally capable mobile nodes where network connectivity is
often intermittent at best.
AACR Node Functional Components
A simple CRA includes the functional components are shown in Figure 4.2. A functional
component is a black box to which functions have been allocated, but for which
implementation is not specified. Thus, while the applications component is likely to be
primarily software, the nature of those software components is yet to be determined. User
interface functions, however, may include optimized hardware (e.g., for computing video
flow vectors in real time to assist scene perception).
Fig.4.2 Minimal AACR Node Architecture
At the level of abstraction of this figure 4.2, the components are functional, not physical.
These functional components are as follows:
1. The user sensory perception (SP), which includes haptic, acoustic, and video
sensing and perception functions.
2. The local environment sensors (location, temperature, accelerometer,
compass, etc.).
4
3. The system applications (sys apps) media-independent services such as playing
a network game.
4. The SDR functions which include RF sensing and SDR applications.
5. The cognition functions (symbol grounding for system control, planning, and
learning).
6. The local effector functions (speech synthesis, text, graphics, and multimedia
displays).
These functional components are embodied on an iCR platform, a hardware realization of
the six functions. To support the capabilities, these components go beyond SDR in
critical ways. First, the user interface goes well beyond buttons and displays. The
traditional user interface has been partitioned into a substantial user sensory subsystem
and a set of local effectors. The user sensory interface includes buttons (the haptic
interface) and microphones (the audio interface) to include acoustic sensing that is
directional, capable of handling multiple speakers simultaneously, and able to include full
motion video with visual scene perception. In addition, the audio subsystem does not just
encode audio for (possible) transmission; it also parses and interprets the audio from
designated speakers, such as the user, for a high-performance spoken natural language
(NL) interface. Similarly, the text subsystem parses and interprets the language to track
the user’s information states, detecting plans and potential communications and
information needs unobtrusively as the user conducts normal activities. The local
effectors synthesize speech along with traditional text, graphics, and multimedia displays.
Sys apps are those information services that define value for the user.
Design Rules Include Functional Component Interfaces
The six functional components in Tables 4.1(a) and 4.1(b) imply associated functional
interfaces. In architecture, design rules may include a list of the quantities and types of
components as well as the interfaces among those components. This section addresses the
interfaces among the functional components.
The AACR N-squared diagram of Table 4.1(a) characterizes AACR interfaces. These
constitute an initial set of Aware Adaptive Cognitive Radio (AACR) Application
Programming Interface (API)s. In some ways, these APIs augment the established SDR
APIs. This is entirely new and much needed in order for basic AACRs to accommodate
even the basic ideas of the Defense Advanced Research Projects Agency (DARPA)
NeXt-Generation (XG) radio communications program.
5
6
7
8
9
4. 2 Cognition Cycle
Cognitive Radio Architecture (CRA) comprises a set of design rules by which the
cognitive level of information services may be achieved by a specified set of
components in a way that supports the cost-effective evolution of increasingly
capable implementations over time. The cognition subsystem of the architecture
includes an inference hierarchy and the temporal organization and flow of
inferences and control states—the cognition cycle.
Fig. 4.3 a) Simplified Cognitive Cycle
A cognition cycle by which a cognitive radio may interact with the environment is
illustrated in Figure 4.3a. Stimuli enter the cognitive radio as interrupts, dispatched to the
cognition cycle for a response. Such a cognitive radio continually observes the
environment, orients itself, creates plans, decides, and then acts. In addition, machine
learning is structured into these phases. Since the assimilation of knowledge by machine
learning can be computationally intensive, cognitive radio has sleep and prayer epochs
that support machine learning. A sleep epoch is a relatively long period of time (e.g.
minutes to hours) during which the radio will not be in use, but has sufficient electrical
power for processing. During the sleep epoch, the radio can run machine learning
algorithms without detracting from its ability to support its user’s needs. Learning
opportunities not resolved in the sleep epoch can be brought to the attention of the user,
the host network, or a designer during a prayer epoch.
During the wake epoch, the receipt of a new stimulus on any of its sensors initiates a new
primary cognition cycle. The cognitive radio observes its environment by parsing
incoming information streams. These can include the monitoring of radio broadcasts, e.g.
10
the weather channel, stock ticker tapes, etc. Any RF-LAN or other short-range wireless
broadcasts that provide environment awareness information are also parsed. In the
observation phase, it also reads location, temperature, and light level sensors, etc. to infer
the user's communications context. The cognitive radio orients itself by determining the
priority associated with the stimuli. A power failure might directly invoke an act
(―Immediate‖ path in the figure). A nonrecoverable loss of signal on a network might
invoke reallocation of resources, e.g. from parsing input to searching for alternative RF
channels. This is accomplished via the path labeled ―Urgent‖ in the figure. However, an
incoming network message would normally be dealt with by generating a plan (Normal
path). Planning includes plan generation. As formal models of causality are embedded
into planning tools, this phase should also include reasoning about causality. The Decide
phase selects among the candidate plans. The radio might have the choice to alert the user
to an incoming message (e.g. behaving like a pager) or to defer the interruption until later
(e.g. behaving like a secretary who is screening calls during an important meeting).
―Acting‖ initiates the selected processes using effector modules.
Learning is a function of observations and decisions. For example, prior and current
internal states may be compared with expectations to learn about the effectiveness of a
communications mode. The cognition cycle implies a large scope of hard research
problems for cognitive radio. Parsing incoming messages requires natural language text
processing. Scanning the user’s voice channels for content that further defines the
communications context requires speech processing. Planning technology offers a wide
range of alternatives in temporal calculus, constraint based scheduling, task planning ,
causality modeling, and the like. Resource allocation includes algebraic methods for
wait-free scheduling protocols, Open Distributed Processing (ODP), and Parallel Virtual
Machines (PVM).
Observe (Sense and Perceive)
The iCR senses and perceives the environment (via ―observation phase‖ code) by
accepting multiple stimuli in many dimensions simultaneously and by binding these
stimuli all together or more typically in subsets to prior experience so that it can
subsequently detect time-sensitive stimuli and ultimately generate plans for action. Thus,
iCR continuously aggregates experience and compares prior aggregates to the current
situation. A CR may aggregate experience by remembering everything.
Orient
The orient phase determines the significance of an observation by binding the observation
to a previously known set of stimuli of a scene. The orient phase contains the internal
data structures that constitute the equivalent of the short-term memory (STM) that people
use to engage in a dialog without necessarily remembering everything with the same
degree of long-term memory (LTM). Typically people need repetition to retain
information over the long term. The natural environment supplies the information
11
redundancy needed to instigate transfer from STM to LTM. In the CRA, the transfer from
STM to LTM is mediated by the sleep cycle in which the contents of STM since the last
sleep cycle are analyzed both internally and with respect to existing LTM.
Stimulus Recognition
Stimulus recognition occurs when there is an exact match between a current stimulus and
a prior experience. The CR I prototype is continually recognizing exact matches and
recording the number of exact matches that occurred along with the time measured in the
number of cognition cycles between the last exact match. By default, the response to a
given stimulus is to merely repeat that stimulus to the next layer up the inference
hierarchy for aggregation of the raw stimuli. But if the system has been trained to respond
to a location, a word, an RF condition, a signal on the power bus, or some other
parameter, it may either react immediately or plan a task in reaction to the detected
stimulus. If that reaction were in error, then it may be trained to ignore the stimulus,
given the larger context, which consists of all the stimuli and relevant internal states,
including time.
Binding
Binding occurs when there is a nearly exact match between a current stimulus and a prior
experience and very general criteria for applying the prior experience to the current
situation are met. One such criterion is the number of unmatched features of the current
scene. If only one feature is unmatched and the scene occurs at a high level such as the
phrase or dialog level of the inference hierarchy, then binding is the first step in
generating a plan for behaving in the given state similar to the last occurrence of the
stimuli.
Plan
Most stimuli are dealt with deliberatively rather than reactively. An incoming network
message would normally be dealt with by generating a plan (in the plan phase, the normal
path). Such planning includes plan generation. In research quality or industrial-strength
CRs, formal models of causality must be embedded into planning tools. The plan phase
should also include reasoning about time.
Decide
The decide phase selects among the candidate plans. The radio might have the choice to
alert the user to an incoming message (e.g., behave like a pager) or to defer the
interruption until later (e.g., behave like a secretary who is screening calls during an
important meeting).
Act
Acting initiates the selected processes using effector modules. Effectors may access the
external world or the CR’s internal states.
12
Externally Oriented Actions
Access to the external world consists primarily of composing messages to be spoken into
the local environment or expressed in text form locally or to another CR or CN using the
Knowledge Query and Manipulation Language (KQML), Radio Knowledge
Representation Language (RKRL), Web Ontology Language (OWL), Radio eXtensible
Markup Language (RXML), or some other appropriate knowledge interchange standard.
Internally Oriented Actions
Actions on internal states include controlling machine-controllable resources such as
radio channels. The CR can also affect the contents of existing internal models, such as
adding a model of stimulus experience response to an existing internal model structure.
The new concept itself may assert-related concepts into the scene. Multiple independent
sources of the same concept in a scene reinforce that concept for that scene. These
models may be asserted by the self to encapsulate experience. The experience may be
reactively integrated into RXML knowledge structures as well, provided the reactive
response encodes them properly.
Learning
Learning is a function of perception, observations, decisions, and actions. Initial learning
is mediated by the observe phase perception hierarchy in which all SP are continuously
matched against all prior stimuli to continually count occurrences and to remember time
since the last occurrence of the stimuli from primitives to aggregates. Learning also
occurs through the introduction of new internal models in response to existing models
and case-based reasoning (CBR) bindings. In general, there are many opportunities to
integrate ML into AACR.
Self-monitoring
Each of the prior phases must consist of computational structures for which the execution
time may be computed in advance. In addition, each phase must restrict its computations
to not consume more resources than the precomputed upper bound. Therefore, the
architecture has some prohibitions and some data set requirements needed to obtain an
acceptable degree of stability of behavior for CRs as self-referential self-modifying
systems.
13
Fig. 4.3 b) Simplified Cognitive Cycle
4. 3 CRA III Inference Hierarchy
The phases of inference from observation to action show the flow of inference, a top-
down view of how cognition is implemented algorithmically. The inference hierarchy is
the part of the algorithm architecture that organizes the data structures. Inference
hierarchies have been in use since Hearsay II in the 1970s proposed it, but the CR
hierarchy is unique in its method of integrating markup language (ML) with real-time
performance during the wake epochs. An illustrative inference hierarchy includes layers
from atomic stimuli at the bottom to information clusters that define action
contexts, as shown in Figure 4.4.
Fig. 4.4 Standard Inference hierarchy
The pattern of accumulating elements into sequences begins at the bottom of the
hierarchy. Atomic stimuli originate in the external environment including RF, acoustic,
14
image, and location domains, among others. The atomic symbols extracted from them are
the most primitive symbolic units in the domain. In speech, the most primitive elements
are the phonemes. In the exchange of textual data (e.g., in e-mail), the symbols are the
typed characters. In images, the atomic symbols may be the individual picture elements
(pixels) or they may be small groups of pixels with similar hue, intensity, texture, and so
forth.
A related set of atomic symbols forms a primitive sequence. Words in text, tokens from a
speech ―tokenizer,‖ and objects in images (or individual image regions in a video flow)
are primitive sequences. Primitive sequences have spatial and/or temporal coincidence,
standing out against the background (or noise), but there may be no particular meaning in
that pattern of coincidence. Basic sequences, in contrast, are space–time–spectrum
sequences that entail the communication of discrete messages.
These discrete messages (e.g., phrases) are typically defined with respect to an ontology
of the primitive sequences (e.g., definitions of words). Sequences cluster together
because of shared properties. For example, phrases that include words such as ―hit,‖
―pitch,‖ ―ball,‖ and ―out‖ may be associated with a discussion of a baseball game.
Knowledge Discovery in Databases (KDD) and the Semantic Web offer approaches for
defining, or inferring, the presence of such clusters from primitive and basic sequences.
A scene is a context cluster, a multidimensional space–time–frequency association, such
as a discussion of a baseball game in the living room on a Sunday afternoon. Such
clusters may be inferred from unsupervised ML (e.g., using statistical methods or
nonlinear approaches such as Support Vector Machines (SVMs).
NL in CRA Inference Hierarchy
The issues required to integrate existing Natural Language Processing (NLP) tools, the
discussion does not pretend to present a complete solution to this problem of Inference
hierarchy.
15
Fig. 4.5 NL in Inference Hierarchy
NLP systems work well on well-structured speech and text, such as the prepared text of a
news anchor. But they do not yet work well on noisy, nongrammatical data structures
encountered, for example, when a user is trying to order a cab in a crowded bar. Thus,
less-linguistic or meta-linguistic data structures may be needed to integrate core CR
reasoning with speech and/or text-processing frontends. The CRA has the flexibility
illustrated in Figure 4.5 for the subsequent integration of evolved NLP tools. The
emphasis of this version of the CRA is a structure of sets and maps required to create a
viable CRA. Although introducing the issues required to integrate existing NLP tools, the
discussion does not pretend to present a complete solution to this problem.
4.4 CRA IV: Architecture Maps
Cognition functions are implemented via cognition elements consisting of data structures,
processes, and flows, which may be modeled as topological maps over the abstract
domains identified in Figure 4.6.
16
Fig. 4.6 Architecture based on Cognition Cycle
The Self is an entity in the world, whereas the internal organization of the Self is an
abstraction that models the Self itself.
The hierarchy of words, phrases, and dialogs from sensory data to scenes is not
inconsistent with visual perception. Words correspond to visual entities; phrases to
detectable movement and juxtaposition of entities in a scene. Dialogs correspond to a
coherent sequence of movement within the scope of a scene, such as walking across the
room. Occlusion may be thought of as a dialog in which the room asserts itself in part of
the scene while observable walking corresponds to assertion of the object. The model
data structures may be read as generalized words, phrases, dialogs, and scenes that may
be acoustic, visual, or perceived in other sensory domains (e.g., infrared). These
structures refer to set-theoretic spaces consisting of a set X and a family of subsets Ox
that contain {X} and { }, the null set, and that are closed under union and countable
intersection. In other words, each is a topological space induced over the domain.
Proceeding up the hierarchy, the scope of the space (X, Ox) increases. A Scene is a
subset of space–time that is circumscribed by the entity by sensory limits. The cognition
functions modeled in these spaces are topology-preserving maps as given in Figure 4.6.
Data- and knowledge-storage spaces are shown as rectangles (e.g., dialog states, plans),
whereas processing elements that transform sets are modeled as homeomorphisms, or
topology-preserving maps, shown as directed graphs (e.g., π) in this figure.
CRA Topological Maps
The processing elements of the architecture are modeled topological maps, as shown in
Figure 4.6:
17
Behaviors in the CRA
CRA entails three modes of behavior: waking, sleeping, and praying. Behavior that lasts
for a specific time interval is called a behavioral epoch. The axiomatic relationships
among these behaviors are expressed in the topological maps of Figure 4.7
18
Fig. 4.7 Cognitive Behavior Model consists of Domains and Topological maps
Waking Behavior
Waking behavior is optimized for real-time interaction with the user, isochronous control
of Software Radio ( SWR) assets, and real-time sensing of the environment. The conduct
of the waking behavior is informally referred to as the awake-state, although it is not a
specific system state, but a set of behaviors. Thus, referring to Figure 4.7, the awake-state
cognition-actions α map the environment interactions to the current stimulus–response
cases. These cases are the dynamic subset of the embedded Stimulus Experience
Response Model (serModel). Incremental ML δ maps these interactions to integrated
knowledge, the persistent subset of the serModels.
Sleeping and Dreaming Behaviors
Cognitive PDAs (CPDAs) detect conditions that permit or require sleep and dreaming.
For example, if the PDA predicts or becomes aware of a long epoch of low utilization
(such as overnight hours), then the CPDA may autonomously initiate sleeping behavior.
Sleep occurs during planned inactivity, for example, to recharge batteries. Dreaming
behavior employs energy to retrospectively examine experience since the last period of
sleep. In the CRA, all sleep includes dreaming. In some situations, the CPDA may
request permission to enter sleeping/dreaming behavior from the user (e.g., if predefined
limits of aggregate experience are reached). Regular sleeping/dreaming limits the
combinatorial explosion of the process of assimilating aggregated experience into the
serModels needed for realtime behavior during the waking behaviors. During the
dreaming epochs, the CPDA processes experiences from the waking behavior using
19
nonincremental ML algorithms. These algorithms map current cases and new knowledge
into integrated knowledge β.
A conflict is a context in which the user overrode a CPDA decision about which the PDA
had little or no uncertainty. Map β may resolve the conflict. If not, it will place the
conflict on a list of unresolved conflicts (map γ).
Prayer Behavior
Attempts to resolve unresolved conflicts via the mediation of the PDA’s home network
may be called prayer behavior, referring the issue to a completely trusted source with
substantially superior capabilities. The unresolved-conflicts list γ is mapped ( λ) to
RXML queries to the PDA’s home CN expressed in XML, OWL, KQML, RKRL,
RXML, or a mix of declared knowledge types. Successful resolution maps network
responses to integrated knowledge ( µ). Many research issues surround the successful
download of such knowledge, including the set of support for referents in the unresolved-
conflicts lists and the updating of knowledge in the CPDA needed for full assimilation of
the new knowledge or procedural fix to the unresolved conflict. The prayer behavior may
not be reducible to finite-resource introspection, and thus may be susceptible to the
partialness of Turing Capable (TC), even though the CPDA and CWN enforce watchdog
timers
4.5 CRA V: Building the CRA on SDR Architectures
A CR is an SWR or SDR with flexible formal semantics-based entity-to-entity messaging
via RXML and integrated ML of the self, the user, the RF environment, and the situation.
This presents how SWR, SDR, and the SCA, or SRA, as they relate to the SRA.
Although it is not necessary for an Aware Adaptive Cognitive Radio (AACR) to use the
SCA/SRA as its internal model of itself, it certainly must have some model, or it will be
incapable of reasoning about its own internal structure and adapting or modifying its
radio functionality autonomously.
Review of SWR and SDR Principles
Hardware-defined radios such as the typical amplitude/frequency modulation (AM/FM)
broadcast receiver convert radio to audio using such radio hardware as antennas, filters,
analog demodulators, and the like. SWR is the ideal digital radio in which the analog-to-
digital converter (ADC) and digital-to-analog converter (DAC) convert digital signals to
and from RF directly, and all RF channel modulation, demodulation, frequency
translation, and filtering are accomplished digitally. For example, modulation may be
accomplished digitally by multiplying sine and cosine components of a digitally sampled
audio signal (called the ―baseband‖ signal, to be transmitted) by the sampled digital
values of a higher-frequency sine wave to upconvert it, ultimately to the RF spectrum.
20
Fig.4.8 SDR Principle applied to Cellular Base Station
Figure 4.8 shows how SDR principles apply to a cellular radio-base station. The ideal
Software Radio (SWR) would have essentially no RF conversion, just ADC/DAC blocks
accessing the full RF spectrum available to the (wideband) antenna elements. Today’s
SDR-base stations approach this ideal by digital access (DAC and ADC) to a band of
spectrum allocations, such as 75 MHz allocated to uplink and downlink frequencies for
3G services. In this architecture, RF conversion can be a substantial system component,
sometimes 60 percent of the cost of the hardware, and not amenable to cost
improvements through Moore’s law. The ideal SDR would access more like 2.5 GHz
from, say 30 MHz to around 2.5 GHz, supporting all kinds of services in TV bands,
police bands, air traffic control bands, and other bands. Although this concept was
considered radical when introduced in 1991 and popularized in 1995.
Fig.4.9 SWR Principle ADC and DAC at the Antenna
This ideal SWR may not be practical or affordable, so it is important for the radio
engineer to understand the trade-offs. In particular, the physics of RF devices (e.g.,
antennas, inductors, filters) makes it easier to synthesize narrowband RF and intervening
analog RF conversion and intermediate frequency (IF) conversion. Given narrowband
RF, the hardware-defined radio might employ baseband (e.g., voice frequency) ADC,
DAC, and digital signal processing. The programmable digital radios (PDRs) of the
21
1980s and 1990s used this approach. Historically, this approach has not been as
expensive as wideband RF (i.e., the cost of antennas, conversion), ADCs, and DACs.
Handsets are less amenable to SWR principles than the base station as in Figure 4.9. Base
stations access the power grid. Thus, the fact that wideband ADCs, DACs, and DSP
(digital signal processor) consume many watts of power is not a major design driver.
Conservation of battery life, however, is a major design driver in the handset.
Thus, insertion of SWR technology into handsets has been relatively slow. Instead, the
major handset manufacturers include multiple single-band RF chip sets into a given
handset. This has been called the Velcro radio or slice radio. The ideal SWR is not
readily approached in many cases, so the SDR has comprised a sequence of practical
steps from the baseband DSP of the 1990s toward the ideal SWR. As the economics of
Moore’s law and of increasingly wideband RF and IF devices allow, implementations
move upward and to the right in the SDR design space as in Figure 4.10.
Fig.4.10 SDR Design Space
This space consists of the combination of digital access bandwidth and programmability.
Access bandwidth consists of ADC/DAC sampling rates converted by the Nyquist
criterion13 and practice into effective bandwidth. Programmability of the digital
subsystems is defined by the ease with which logic and interconnect may be changed
after deployment. Application-specific integrated circuits (ASICs) cannot be changed at
all, so the functions are ―dedicated‖ in silicon. Field-programmable gate arrays (FPGAs)
can be changed in the field, but if the new function exceeds some performance parameter
22
of the chip, which is not uncommon, then one must upgrade the hardware to change the
function, just like with ASICs.
The SDR Forum defined a very simple, helpful model of radio in 1997, which is shown
in Figure 4.11. This model highlights the relationships among radio functions at a tutorial
level. The CR has to ―know‖ about these functions, so this model is a good start because
it shows both the relationships among the functions and the typical flow of signal
transformations from analog RF to analog or (with SDR) digital modems, and on to other
digital processing, including system control of which the user interface is a part.
Fig. 4.11 SDR Forum
Software Communications Architecture (SCA)
The US DoD developed the SCA for its Joint Tactical Radio System (JTRS) family of
radios. The SCA identifies the components and interfaces shown in Figure 4.12 The
APIs define access to the PHY layer, to the MAC layer, to the logical link control (LLC)
layer, to security features, and to the input/output of the physical radio device. The
physical components consist of antennas and RF conversion hardware that are mostly
analog and that typically lack the ability to declare or describe themselves to the system.
Most other SCA-compliant components are capable of describing themselves to the
system to enable and facilitate plug-andplay among hardware and software components.
In addition, the SCA embraces the portable operating system interface (POSIX) and
CORBA.
23
Fig. 4.12 JTRS SCA Version
The model evolved through several stages of work in the SDR Forum and OMG into a
UML-based object-oriented model of SDR (Figure 14.19). Waveforms are collections of
load modules that provide wireless services, so from a radio designer’s perspective, the
waveform is the key application in a radio. From a user’s perspective of a wireless PDA
(WPDA), the radio waveform is just a means to an end, and the user doesn’t want to
know or to have to care about waveforms. Today, the cellular service providers hide this
detail to some degree, but consumers sometimes know the difference between CDMA
and GSM, for example, because first generation CDMA works in the United States, but
not in Europe. With the deployment of the 3G of cellular technology, the amount of
technical jargon consumers will need to know is increasing. So the CRA insulates the
user from those details, unless the user really wants to know.
24
Fig. 4. 13 SDR Forum UML model of radio services
SDR Forum of UML Radio Services
In the UML model shown in Figure 4.13, Amp refers to amplification services, RF refers
to RF conversion, interference management refers to both avoiding interference and
filtering it out of one’s band of operation. In addition, the jargon for US military radios is
that the ―red‖ side contains the user’s secret information, but when it is encrypted it
becomes ―black,‖ or protected, so it can be transmitted. Black processing occurs between
the antenna and the decryption process. The Figure 4.13 has no user interface. The UML
model contains a sophisticated set of management facilities, to which the human–
machine interface (HMI) or user interface is closely related.