PaperlayerView exact revision on arXiv

Source-linked AI summary

A Novel Experimental Platform for In-Vessel Multi-Chemical Molecular Communications

Nariman Farsad, David Pan, Andrea Goldsmith

arXiv:1704.04810v1cs.ETcs.AIcs.ITcs.LGmath.IT

TL;DR

The paper addresses detection in multi-chemical communication, where analytical channel models are difficult to obtain and nonlinearities and long memory complicate the problem. It presents an experimental platform and trained machine-learning detectors, reporting lower BER than slope detection and data rates about 10 times higher than previous platforms.

  • Problem

    Detection is challenging in multi-chemical channels because analytical models are difficult to obtain and chemical reactions can produce nonlinearities and long memory.

    Source: §I · p. 1, B. Symbol Detection Based on Rate of… · p. 3

  • Method

    The paper builds a multi-chemical experimental platform and trains SVM- and RNN-based detectors on known transmission sequences.

    Source: §I · p. 1

  • Results

    The RNN detector's BER is about 10 times lower than slope detection and about two times lower than the SVM detector, while data rates reach 4 bps, about 10 times higher than previous works,.

    Source: §I · p. 2

  • Takeaways & Limitations

    The platform can be expanded to represent cardiovascular systems and complex pipe networks, and the trained detectors are especially useful when an analytical channel model is unknown.

    Source: §I · pp. 1–2

Abstract

from arXiv · show

This work presents a new multi-chemical experimental platform for molecular communication where the transmitter can release different chemicals. This platform is designed to be inexpensive and accessible, and it can be expanded to simulate different environments including the cardiovascular system and complex network of pipes in industrial complexes and city infrastructures. To demonstrate the capabilities of the platform, we implement a time-slotted binary communication system where a bit-0 is represented by an acid pulse, a bit-1 by a base pulse, and information is carried via pH signals. The channel model for this system, which is nonlinear and has long memories, is unknown. Therefore, we devise novel detection algorithms that use techniques from machine learning and deep learning to train a maximum-likelihood detector. Using these algorithms the bit error rate improves by an order of magnitude relative to the approach used in previous works. Moreover, our system achieves a data rate that is an order of magnitude higher than any of the previous molecular communication platforms.

I. INTRODUCTION

The paper introduces an accessible platform for multi-chemical molecular communication and addresses detection when the channel is nonlinear, has long memory, and lacks an analytical model. Its trained detectors improve on slope detection, while the platform achieves data rates about 10 times higher than previous works.

I. INTRODUCTION

  • The platform uses peristaltic pumps to inject different chemicals into fluid flow and can be extended to model cardiovascular systems and complex pipe networks.It is designed to be inexpensive and replicable without wet labs, and can automate experiments with minimal supervision.Source: §I · p. 1
  • Because the multi-chemical channel has nonlinearities and long memory but no known analytical model, the paper asks how to detect transmitted symbols reliably.The rate-of-pH-change, or slope, detector is one candidate, but its optimality is unclear.Source: §I · p. 1
  • The paper trains SVM- and RNN-based detectors on known transmission sequences, with the RNN behaving similarly to a maximum-likelihood detector that learns detection and equalization.The motivation is to address the shortage of experimental data used to validate molecular-communication models.Source: §I · pp. 1–2
  • The RNN detector's BER is about 10 times lower than slope detection and about two times lower than the SVM detector; data rates reach 4 bps, about 10 times higher than previous works,.The detector comparisons are reported for short symbol intervals, where intersymbol interference is severe.Source: §I · p. 2

II. MULTI-CHEMICAL EXPERIMENTAL PLATFORM

The platform uses a pH sensor to detect hydrogen-ion signals from acid and base transmissions, while its transmitter can be extended to release additional chemicals.

II. MULTI-CHEMICAL EXPERIMENTAL PLATFORM

  • The same general platform can use other chemicals and sensors, including glucose or proteins and sensors that detect them.The described implementation assumes hydrogen ions as information-carrying particles and uses a pH sensor.Source: §II · p. 2

A. The Transmitter

  • The transmitter uses peristaltic pumps to move liquids, with aquarium dosing pumps available for as little as $15 each.A custom MOSFET switch board and Arduino Uno control the pumps.Source: A. The Transmitter · p. 2
  • Adding pumps and chemical containers extends the transmitter to release more chemicals than the acid-and-base configuration shown.Source: A. The Transmitter · p. 2

B. The Propagation Channel

The propagation channel carries injected chemicals through flowing water in a small silicon tube, and the receiver measures pH with a probe and Arduino-based digitization.

B. The Propagation Channel

  • The channel uses 2.5 mm silicon tubing, filtered water flowing at about 1.3 mL/s, and can substitute another liquid such as blood.The water flow is generated by a continuously running peristaltic pump, and measured water pH varied from 7 to 8.5 across different months.Source: B. The Propagation Channel · pp. 2–3
  • Acids and bases travel with the water to the pH-probe receiver, and the system can be modified as a closed loop where these chemicals react and cancel out.The paper notes that the average concentration, such as pH, would remain constant in such a loop.Source: B. The Propagation Channel · p. 3

C. The Receiver

  • The receiver digitizes pH-probe readings through a pH meter and Arduino Uno, while the probe's voltage output is linear with pH under normal conditions.Acidic and basic solutions produce positive and negative voltages, respectively.Source: C. The Receiver · p. 3

III. IMPLEMENTING A COMMUNICATION SYSTEM ON THE PLATFORM

The communication implementation is divided into modulation and protocol design, followed by receiver detection. Fig. 2 depicts the received pH signal for a specified bit sequence.

III. IMPLEMENTING A COMMUNICATION SYSTEM ON THE PLATFORM

  • The implementation has two parts: modulation and communication protocols, then detection at the receiver.Source: §III · p. 3
  • Fig. 2 presents the received pH signal for the bit sequence 110011010001001.Source: Figure 2 · p. 3

A. Modulation and Protocol

The platform uses time-slotted binary signaling, with acid and base pulses encoding the two bit values. Each message begins with a synchronization pulse, and a termination sequence can mark its end.

A. Modulation and Protocol

  • Bit-0 uses a 30 ms acid injection and bit-1 a 30 ms base injection at the start of each symbol interval.The silence after injection sets the symbol interval.Source: A. Modulation and Protocol · p. 3
  • In binary time-slotted chemical communication, acid or base is injected for 30 ms at each symbol’s start, then silence; pauses of 220, 304, 350, and 470 ms represent 4, 3, 2.6, and 2 bps, respectively.Source: A. Modulation and Protocol · p. 3
  • A 100 ms initial acid pulse followed by 900 ms of silence synchronizes the receiver; a special termination sequence can signal the end of transmission.Source: A. Modulation and Protocol · p. 3

B. Symbol Detection Based on Rate of Change

Analytical detection is difficult because multi-chemical channels can have nonlinearities, long memory, and pump- and sensor-induced noise. The rate-of-change approach uses binned pH readings, but its threshold performance becomes harder to separate as the data rate rises.

B. Symbol Detection Based on Rate of Change

  • A rate-of-change detector is motivated by the expected pH drop after acid and rise after base, but analytical models are difficult when reactions cause nonlinearities and long memory.Pump and sensor noise are also difficult to model.Source: B. Symbol Detection Based on Rate of… · pp. 3–4
  • The seventh pH rate of change is the best threshold-detection metric, while increasing the rate from 2 to 4 bps increases ISI and makes detection harder.The eye diagrams show rate-of-change values for many randomly transmitted sequences at 2 and 4 bps.Source: Figure 3 · p. 4, B. Symbol Detection Based on Rate of… · p. 4
  • The method divides each symbol interval into eight equal time bins, averages pH within each bin, and computes seven successive pH differences.Source: B. Symbol Detection Based on Rate of… · p. 4

C. Symbol Detection Using Machine Learning

When analytical channel models are unknown, the paper trains machine-learning detectors from known transmissions and received-signal features. It compares symbol-by-symbol detection with an approach that uses current and previous received features to account for ISI.

C. Symbol Detection Using Machine Learning

1) Support Vector Machine Detector:

The SVM detector extends linear regression to nonlinear multi-chemical signals by mapping features into a larger space and using a kernel. It uses training samples in its decision process, with the Gaussian kernel selected for the model.

1) Support Vector Machine Detector:

  • SVM maps received features into a typically larger space to address the nonlinear relationship between received and transmitted signals.Source: 1) Support Vector Machine Detector · p. 4
  • The detector uses the training dataset directly in its decision process, with sample-associated weights and a bias found through convex optimization.Most sample weights are zero, so only a small fraction of training samples are used during detection.Source: 1) Support Vector Machine Detector · pp. 4–5
  • The Gaussian, or RBF, kernel uses Euclidean feature distance, maps to an infinite-dimensional space, and performs linear regression there.Its σ^2 parameter can be optimized numerically using training data.Source: 1) Support Vector Machine Detector · p. 5

2) Recurrent Neural Network Detector:

The RNN detector uses sequential received features to capture intersymbol interference and system memory that the SVM detector misses. It is trained on known transmission sequences to estimate symbol probabilities and maximize log-likelihoods.

2) Recurrent Neural Network Detector:

  • The RNN captures intersymbol interference and system memory by training on sequential data, unlike the SVM detector, which does not capture these effects.The hidden state carries information about received features from previous symbol intervals.Source: 2) Recurrent Neural Network Detector · p. 5
  • Training uses known symbol sequences and received feature vectors, with one-hot vectors representing the transmitted symbols.The training dataset consists of n samples of length-K sequences.Source: 2) Recurrent Neural Network Detector · p. 5
  • The RNN estimates symbol probabilities from current and previous received features, and its training objective maximizes log-likelihoods, making it similar to a maximum-likelihood detector.The network structure determines the complexity of the model relating transmitted and received signals.Source: 2) Recurrent Neural Network Detector · p. 5
  • The platform uses long short-term memory networks, which have more nodes than the generic RNN described and better training performance.Source: 2) Recurrent Neural Network Detector · p. 5

IV. RESULTS

Both machine-learning detectors outperform the slope baseline, with the RNN performing especially well at short symbol intervals where intersymbol interference is severe. The evaluation uses experimental transmissions across four symbol intervals and a held-out test set.

IV. RESULTS

  • The evaluation uses 194 experiments with 120 random bits each, totaling more than 23,000 transmitted bits across intervals of 250, 334, 380, and 500 ms.Twenty percent of experiments at 250, 334, and 380 ms were reserved for testing; the remainder supported training and parameter tuning.Source: §IV · p. 6
  • The SVM uses pH-bin values and summary statistics alongside rates of change, while the RNN uses a single-layer LSTM with 16 internal-state nodes and average pH and rate-of-change features.The selected features and hyperparameters were those reported to give the best detection performance.Source: §IV · p. 6
  • Across symbol intervals, both machine-learning detectors significantly outperform the slope baseline, and the RNN performs better than the SVM especially at shorter intervals with severe intersymbol interference.All detectors improve as symbol intervals increase and intersymbol interference is reduced.Source: §IV · p. 6
  • The RNN detector’s BER is almost half that of the SVM detector and about 10 times better than the baseline detector because it captures channel ISI and “equalizes” its effects.Source: §IV · p. 6

V. CONCLUSIONS

The paper presents a multi-chemical communication platform using acid and base signals carried by pH, and reports improved detection performance and data rates relative to prior platforms.

V. CONCLUSIONS

  • The platform supports multi-chemical communication in small tubes and can be expanded to emulate cardiovascular and complex pipe-network environments.Source: §V · p. 6
  • The implemented system transmits acid and base signals and carries information through pH measurements.Source: §V · p. 6
  • The proposed machine-learning and deep-learning detectors improve performance tenfold over slope-based detection used in previous works, while the platform achieves data rates ten times higher than previous platforms.Source: §V · p. 6

Related papers

Loading 1704.04810v1…