Make your paper interactiveCreate

Text from the interactive explanation

Distributed Joint Source-Channel Coding-Based Adaptive Dynamic Network Coding

Abdulah Jeza Aljohani; Soon Xin Ng

Original paper

An interactive guide to Aljohani & Ng · IEEE Access · 2020

Two neighbouring sensors must radio their readings to a distant receiver through a sky that fades. Now and then a whole burst of data, a frame, arrives too weak to save. In the authors’ computer simulations, letting a relay send mixtures of the sensors’ frames, and only when the receiver asks, reached the same reliability for closely matching sensors with 19.5 dB less power per bit , about 89× less .

Turn the power down. Switch the relay off. Watch which frames survive.

1Sensors broadcast

2Destination’s flag

3Relay sends mixtures

4Destination solves

Strategy

Long-run estimate · 20,000 simulated rounds of this model

This run

A sensor’s frame (brighter = weaker fade)

Grey dust = lost in a fade

Striped = relay’s mixed frame

Gold ring = recovered by solving

Mint / amber ring = destination’s flag (“all received” / “send help”)

Ledger: solid = arrived, gold rim = rescued, dark = lost

Live scene: a simplified, illustrative frame-loss model built on the paper’s set-up, not the paper’s full simulator. The paper’s own results are in What the simulations found.

1 · Correlation as a clue

Two sensors that almost agree

Picture two thermometers a few metres apart. Their readings, and the streams of ones and zeros that encode them, mostly match. The paper models this simply: sensor 1’s bit equals sensor 2’s bit, except that each bit is flipped with a small probability pe. The correlation coefficient is ρ = 1 − 2·pe , so ρ = 0.8 means one bit in ten differs.

Why would the receiver care? Because once it knows one stream, the other carries very little news.

Two bit streams

How many bits do they really need to send?

In 1973 David Slepian and Jack Wolf proved the result this paper builds on: even if each sensor compresses its own stream without talking to the other, a receiver that decodes both together only needs

R1 ≥ H(b1|b2), R2 ≥ H(b2|b1), R1 + R2 ≥ H(b1, b2)

Here R1 and R2 are bits sent per source bit, and H is the information content. Drag the point (or type values) to choose how much each sensor sends.

shaded: decodable togetherR1 + R2 = H(b1, b2)A, B: corner points

Redrawn from the paper’s Eqs. (1)–(3) / Fig. 2.

Predict: with ρ = 0.8, can both sensors send at only 0.75 bits each ?

A clue in action

Real channels add noise. Below, sensor 1’s bits cross a noisy channel. One decoder guesses each bit on its own. The other also looks at sensor 2’s matching bit, as a hint about what sensor 1 probably sent, weighted by how often the two disagree.

Guessing alone vs. with a clue

Illustration of the idea (per-bit guess using the correlation set above); the paper’s joint decoder is an iterative turbo decoder.

Takeaway. Correlation lowers the power needed to get the data through: that is the ρ slider in the live network above. In the paper’s own simulations (Fig. 7 below), each step up in correlation shaves roughly a decibel off the power needed under fast fading, and with slow fading added, ρ = 0.8 needs about 3 dB less than ρ = 0.6 at a bit error rate of 10−4.

2 · Fast vs. slow fading

Two ways the sky can fade

Radio waves bounce, interfere and fade. The paper’s channel combines two kinds at once, h = hs·hf (Eq. 11):

Fast fading changes from symbol to symbol inside a frame. A channel code spreads each bit’s information over many symbols, so good symbols make up for bad ones.

Slow (quasi-static) fading stays constant over a whole frame. If the frame lands in a deep fade, every symbol is weak together, and no code inside that frame can save it.

Frames arriving through the fade

Each bar is one symbol’s usable information, log2(1 + SNR·|h|2). A frame decodes if its average (solid line) reaches 1 bit per symbol (dashed line, a rate-½ code with QPSK). A simplified, illustrative decoding rule.

Long-run frame loss vs. power (this model)

Predict first. Under slow fading, raising the power by 10 dB (ten times more power) cuts the losses by about…

Takeaway. A frame that sinks as a whole cannot be rescued from the inside. The paper reports that with a single antenna at each end, even 30 dB was not enough to reach a bit error rate of 10−3 (its Fig. 9). The rescue has to come from other frames that faded differently, and that is the relay’s job.

3 · Network coding

Mix frames so that any few will do

The protocol has two phases. In the broadcast phase , each sensor sends its information frame (IF) to both the relay and the destination. In the cooperative phase , the relay sends parity frames (PFs): mixtures of the frames it decoded, each multiplied by its own coefficient (arithmetic in a small finite field, GF(q)). Every frame suffers its own independent slow fade, so the destination can treat lost frames as unknowns and solve for them from whatever arrived.

The four-tile puzzle · one relay with two antennas, Eq. (14)

Tap a tile to sink it in a fade; tap again to bring it back.

Challenge: sink exactly two tiles so the destination can’t recover both frames. Pairs tried: 0 of 6

Eight frames, any four will do · two rounds, Eq. (55)

The paper’s larger code (its Fig. 12) runs two rounds: each sensor sends twice, giving four information frames, and the relay’s two antennas send four parity frames. In the results it is also used with two relays. The coefficients below are the paper’s, over GF(8). Sink as many tiles as you like.

Any 4 of these 8 frames recover everything. This is the “maximum distance separable” property of the Reed–Solomon code the paper uses (its transfer matrix was generated with the SAGE software).

Watch the destination solve · the paper’s worked example, Eqs. (35)–(43)

Eight frames, as in the larger code above. Only one information frame arrived directly, plus two parity frames. In one of those mixtures the first two frames’ coefficients are 0, as happens when the relay failed to decode them.

As in the paper’s example, ordinary arithmetic is used for readability; real network coding does the same algebra in the finite field GF(8).

Why mixing beats shouting: diversity

The paper measures robustness by the diversity order D: roughly, how many independent fades must go wrong at the same time before data is lost. With D = 1, losses fall ten-fold for every extra 10 dB, as in the slow-fading curve above. With higher D, losses fall about 10D-fold for the same extra power. The bounds are (K·L) + n2 ≤ D ≤ (K·L)·n2 + 1 (Eq. 56), with K relays of L antennas each.

Loss now requires several independent fades to coincide, so losses fall steeply as power rises.

4 · Adaptive feedback

Only ask for help when it’s needed

Parity frames cost airtime. If the destination already decoded every information frame in the broadcast phase, the relay’s mixtures are wasted. In the adaptive scheme ( ADNC ), the destination sends a one-bit feedback flag after the broadcast phase: “all received, stay quiet” or “send help”.

With two sensors and one two-antenna relay, relaying every round halves each sensor’s throughput to 0.5 bits per symbol (Eq. 50). Relay-on-request delivers 1 bit per symbol in rounds the flag says “all received”, and 0.5 otherwise (Eqs. 52–54).

The same airtime, two strategies

Throughput per sensor vs. power (this model)

Our simplified model climbs at somewhat higher power than the paper’s two-antenna system, but the shape is the same: as the sky clears, the relay is asked less and less.

Why asking pays twice

Fewer parity frames mean more useful data per unit of airtime, less decoding work at the destination and less delay.

They also save energy. Because power is only spent on frames that matter, the adaptive scheme needs about 1.5 dB (one relay, Fig. 13) to 2 dB (two relays, Fig. 14) less power per information bit than relaying every time, for the same reliability.

5 · Evidence

What the simulations found

The authors simulated the full system on a computer: their turbo-coded joint decoder, 120,000 symbols per frame, 100,000 simulated frames, with fast and slow Rayleigh fading. They asked how much transmit power per information bit each scheme needs to lose only one frame in a thousand (FER = 10−3), compared with the same code sent without network coding. Pick a set-up:

Antennas

Relays & code

Strategy

Correlation

All the gains · dB saved at FER 10−3 vs. no network coding (Tables 8 & 9)

Select any cell to show it above. Every set-up gains well over 14 dB, at every correlation level. Higher correlation lowers the power needed both with and without relays, so the gap between them stays similar.

6 · Return

Back to the whisper

1 Correlation is a clue. Two similar sensors carry less news than two separate ones, and a joint decoder turns that similarity into power savings.

2 Mixing across frames beats whole-frame fades. A relay’s mixtures give the destination equations to solve, so a lost frame can be recovered from frames that faded differently.

3 Ask only when needed. A one-bit flag keeps the relay quiet when nothing is missing, saving airtime, decoding work and energy.

Now turn the power down towards a whisper and watch the relay refill most of the gaps a direct-only link would leave.

What it brings closer

In the authors’ framing: correlated sensors can spend far less transmission energy while staying reliable over hostile fading channels. Because the relays mix frames with non-binary coefficients, the design can scale beyond two users (earlier schemes that combined frames with a simple bit-wise XOR could not, the paper notes). Adaptive feedback keeps the extra traffic, delay and decoding complexity low when the sky is clear.

What the study covered

Two correlated sensors; one or two relays placed at the midpoint.

Rayleigh fast + slow fading; channel state known at the receivers; a feedback flag available.

Two antenna set-ups: two antennas at relays and destination (A) or one everywhere (B).

Computer simulations over 100,000 frames of 120,000 symbols each.

Deeper

During the broadcast phase (BP), each of the M sensors sends its information frames to both the relays and the destination in its own time slot. During the cooperative phase (CP), each of the K relays sends parity frames built from the information frames it decoded correctly. The whole transmission is described by a transfer matrix G = [I | P]: the identity part for the information frames and the parity part P for the mixtures (Eqs. 26–27). If a relay misses a frame, the corresponding coefficients in P are set to zero (Eq. 28) and the destination solves the resulting modified system G′ by Gaussian elimination.

In Scenario A , relays and destination have two antennas (2×2 links), and each relay sends its two parity frames from its two antennas. In Scenario B , every node has a single antenna, a cheaper design; each relay squeezes both parity frames into one combined 16QAM signal. Scenario A performs better, at the cost of more hardware.

Received power falls with distance as d−α. With the relay halfway and a path-loss exponent α = 2, each hop is half as long, so it enjoys 22 = 4 times (6 dB) more received power than the direct link (Eq. 7). The live network uses the same 6 dB bonus on sensor-to-relay and relay-to-destination links.

With n1 information frames per sensor, n2 parity frames per relay antenna, M sensors and K relays of L antennas: RDNC = n1·M / (n1·M + n2·K·L). For two sensors, one relay and two antennas this is 1/2; with two relays it is 1/3. With adaptive feedback, a round where the flag says “all received” has rate 1, otherwise the DNC rate. The effective throughput η multiplies these by the rate of the underlying code and modulation.

Comparing schemes that send different numbers of frames at the same power per symbol would be unfair. The paper therefore compares the transmit energy per information bit: [Eb/N0]t = SNRt − 10·log10(η) in dB. This is why skipping unneeded parity frames (raising η) shows up as a power saving for relay on request.

The destination first runs the joint (turbo) decoder on each received frame, then solves the network code. The authors deliberately skip iterating between the network decoder and the joint decoder: it would add a lot of complexity, and in block-fading channels the extra gain is limited.

(K·L) + n2 ≤ DDNC ≤ (K·L)·n2 + 1 (Eq. 56), where n2 = 1 for G2×4 and n2 = 2 for G4×8. Tables 4–7 of the paper list the resulting rates and diversity orders for each scenario.

Source & credits

Aljohani, A. J. & Ng, S. X. (2020). Distributed Joint Source-Channel Coding-Based Adaptive Dynamic Network Coding. IEEE Access, 8, 86715–86731. DOI: 10.1109/ACCESS.2020.2993056

Author-hosted PDF: southampton.ac.uk/~sxn/paper/access2020_jscc.pdf

Figures 7, 13, 14, 15(a) and 19 are reproduced from the paper under the Creative Commons Attribution 4.0 licence (CC BY 4.0), cropped with no other changes. Fig. 2 is redrawn. All results quoted are from the authors’ computer simulations. The interactive models on this page are original, simplified illustrations built from the paper’s equations and set-up.

Topics in this iPaper

Engineering
Both systems are proposed for supporting correlated source transmissions over hostile channels experiencing both small-scale and large-scale fading

Abstract, page 1 (p. 86715)

Computer science
Network Coding (NC) is an efficient data transport technique leveraging network efficiency

Abstract, page 1 (p. 86715)

network coding
The NC transmission protocol can be fully characterised using a transfer generator matrix G

Section IV-A Network Encoder, page 8 (p. 86722)

distributed source coding
Distributed Source Coding (DSC) refers to the problem of compressing several physically separated, but correlated sources

Section I Introduction, page 2 (p. 86716)

joint source channel coding
Distributed Joint Source-Channel coding (DJSC) is associated with the scenario, where the correlated source signals are transmitted through a noisy channel

Abstract, page 1 (p. 86715)

rayleigh fading
the slow Rayleigh fading channel does not exhibit this time diversity

Section III, page 6 (p. 86720)

cooperative communication
when designing any cooperative scheme

Section V-A Transmission Rate, page 10 (p. 86724)

trellis coded modulation
Distributed Joint Turbo Trellis-Coded Modulation (DJTTCM)

Abstract, page 1 (p. 86715)

Distributed Joint Source-Channel Coding-Based Adaptive Dynamic Network Coding — Interactive explanation | iPaper.ai