During this webinar sponsored by Aurora Scientific, Dr. Burgués will discuss how he and his team are leveraging various signal processing and machine learning techniques in order to decode the fine-scale structure of turbulent chemical plumes using low-cost chemical sensors. Specifically, he will discuss three signal processing methods they developed to improve MOX sensor dynamics, and share the experimental setups they used to test their theories. Finally, he will share data from recent experiments and elaborate on the conclusions of their studies and how robotic plume tracking technology might apply to industrial and air quality monitoring, research and more.
For more information, please visit: https://insidescientific.com/webinar/decoding-turbulent-chemical-plumes-improved-signal-processing-machine-learning-aurora-scientific
Decoding Turbulent Chemical Plumes With Improved Signal Processing and Machine Learning
1. Decoding Turbulent Chemical Plumes
with Improved Signal Processing and
Machine Learning
Javier Burgués, MSc, PhD
Post Doctoral Researcher
Institute for BioEngineering
of Catalonia
Chris Rand
Senior Product Consultant
Aurora Scientific
2. Decoding Turbulent Chemical Plumes
with Improved Signal Processing and
Machine Learning
Join Dr. Javier Burgués as he demonstrates
how he has built faster and more reliable MOX
sensors for tracking turbulent chemical plumes.
3. To access webinar content,
Q&A reports, FAQ documents,
and information on lab
workshops,
subscribe to our mail list
4. InsideScientific is an online educational
environment designed for life science
researchers. Our goal is to aid in the
sharing and distribution of scientific
information regarding innovative
technologies, protocols, research tools
and laboratory services
5. Javier Burgués, MSc, PhD
Decoding Turbulent Chemical Plumes
with Improved Signal Processing and
Machine Learning
Post Doctoral Researcher
Institute for BioEngineering of Catalonia
Copyright 2020 J. Burgués and InsideScientific. All Rights Reserved.
7. Olfactory robots
Applications
• Source localization (SL)
• Concentration mapping (CM)
• Emission quantification (EQ)
• Robots equipped with one or more chemical
sensors.
• First prototypes in the early 1990s using
terrestrial robots.
• In recent years, nano-drones with miniature
sensors have become more popular.
Scenarios
• Industrial emission monitoring
• Air quality monitoring in cities
• Atmospheric research, volcanology
• Search and rescue
• …J. Burgués et al., Sensors, 2019, 19(3), 478
8. Structure of a turbulent plume
Celani et al. , Physical Review X, 2014
• Chaotic collection of gas patches
• Smooth gradients only observable after
long-term averaging (~10 min)
Spatial structure of plume
Dekker et al., Jnl Exp Bio., 2011
1 s
Temporal structure of plume
• Fast fluctuations of the instantaneous
concentration (up to 10 KHz)
• Intensity and duration of peaks are indicative of
the source-receptor (SR) distance.
Sensors with high
temporal resolution
Key requirement:
9. 5 s
Metal oxide (MOX) gas sensors
• Olfactory robots typically mount metal oxide
(MOX) sensors due to small form factor, simplicity,
cost and high sensitivity.
• MOX sensors are based on physicochemical
reactions happening on an active sensing layer.
• These reactions are slow.
• MOX sensors can be modelled as leaky integrators
(1st order low-pass filter) of the input stimuli.
330 Hz
0.1 Hz
y 𝑡 = 𝛼 ∙ 𝑦 𝑡 − 1 + 1 − 𝛼 ∙ 𝑥(𝑡)
MOX𝑥(𝑡) 𝑦(𝑡)
Related to time constant of the sensor
10. Improving the sensor dynamics by signal processing
• Key idea: If the transference function ℎ(𝑡) of the sensor is known, it could be inverted to
reconstruct the (fast) input stimuli based on the (slow) sensor output.
D. Martinez et al., Sensors, 2019, 19, 4029
𝐻(𝑓) 𝐻(𝑓)−1
ℎ(𝑡) ℎ(𝑡)−1
• The challenge is to determine the impulse response, ℎ(𝑡)
Very difficult to
produce in the lab
11. Goals of this work
1. Develop signal processing methods to improve the MOX sensor dynamics.
a. Low-pass differentiator (LPD) filter
b. Linear deconvolution model
c. Blind deconvolution
2. Extract features from the fast response that are informative of the source-
receptor (SR) distance in a turbulent plume.
3. Evaluate the performance of such features in real data acquired with mobile
robots and fixed sensor networks.
12. Method 1: Low pass differentiation (LPD)
𝑓
H−1
(𝑓)
𝑓𝑐
𝑈(𝑓)
derivative
Low-pass derivative (LPD)
noise
𝒇 𝒄 is the key
parameter to
be estimated
• Assuming that h(t) is an integrator of the concentration,
a differentiator filter will invert that operation.
MOX
𝑥(𝑡) 𝑦(𝑡) 𝑥(𝑡)
𝑑
𝑑𝑡
• However, differentiation is a risky signal processing operation that amplifies high-frequency noise.
• Differentiation must be restricted to the signal bandwidth Low-pass differentiator (LPD) filter
• How to estimate the signal bandwidth? Fast photo-ionization detector (PID)
13. Method 2: Empirical modelling of h(t)
• Generating a delta in concentration is very difficult to achieve.
• However, we can easily generate a step function.
• The derivative of a step function is a delta function.
• Therefore, the impulse response can be obtained as the derivative of the step response.
Stimulate the sensor with a
step-like function
Model the response and
differentiate it to obtain ℎ(𝑡)
𝑑
𝑑𝑡
14. Method 3: Blind deconvolution
D. Martinez et al., Sensors, 2019, 19, 4029
• The previous methods require the PID signal (ground truth) to optimize the model parameters.
• However, in a real application the PID signal may not available.
• Key idea: If two sensors are excited by the same input 𝑥(𝑡), the deconvolved
outputs 𝑥1(𝑡) and 𝑥2(𝑡) should converge to the same result.
MOX1
𝑥(𝑡)
𝑦1(𝑡)
𝑦2(𝑡)MOX2
+
−
𝑥1(𝑡)
𝑥2(𝑡)
𝑒(𝑡)
MOX1
-1
MOX2
-
1
• Minimization of 𝑒(𝑡) should provide the best signal reconstruction.
16. 2.1. Dataset #1 – Open-environment
• Chemical source is a beaker with liquid analyte
(ethanol or acetone)
• Pressurized air stream creates a turbulent plume.
• Sensing platform integrates 4 MOX sensors and a
fast photoionization detector (PID) with 330 Hz
bandwidth (for ground truth).
• 9 receptor locations (distance range : 15 – 135cm)
• 5 minutes of measurement at each receptor.
J. Burgués et al., ISOEN, 2019, doi:
10.1109/ISOEN.2019.8823158
17. • 5 receptor locations (SR distance: 25 – 150 cm)
• 3 wind speeds (10, 21 and 30 cm/s)
• 10 chemical substances (Acetone, Ethanol, etc.)
• 4 minutes of measurement at each receptor.
• 20 trials per distance/wind/gas combination.
• No ground truth available.
A.Vergara et al., Sens. Act. B:Chem, 2013, 185, p. 462-477
2.2. Dataset #2 –WindTunnel
18. 2.3. Dataset #3 – Small Office
Room
• 25m2 office room
• 27 MOX sensors deployed in a
3x3x3 grid
• SR distance range: 0.25 – 5.5 m
• 10 experiments of 90 min
duration each.
• Variations in source position,
release rate and wind.
J. Burgués et al., Sens. Act. B:Chem, 2020,
304, p.127309.
19. 2.4. Dataset #4 – Large Robotics Laboratory
• Robotics laboratory (160 m2)
• Nano-drone equipped with 2 MOX
sensors.
• Pre-defined exploration strategy.
• 1 chemical substance (Ethanol)
• 3 experiments of ~3 min duration
• Variations in source position,
release rate and wind.
J. Burgués et al., Sensors, 2019, 19(3), 478, doi.org/10.3390/s19030478
22. 3.1.1. Optimization of the low-pass differentiator (LPD) filter
Filtered MOX
99% confidence
99% confidence
Raw (15 cm)
Raw (135 cm)
Filtered (15 cm)
Filtered (135 cm)
7.5 s
0.4 s
20-fold
improvement in
response time
Dataset #1 (open-environment)
Optimize 𝒇 𝒄 for maximum cross-covariance between the
filtered signal and the ground truth (PID):
𝐶𝑜𝑣 𝑥𝑦 𝑚 = 𝐸 𝒙[𝑛] − 𝜇 𝒙 𝒚[𝑛 + 𝑚] − 𝜇 𝒚
𝑇
Test multiple LPD filters with different values of 𝒇 𝒄
J. Burgués & S. Marco, IEEE Access, 2019
J. Burgués et al., ISOEN, 2019
23. 3.1.2. Empirical modelling of h(t)
D. Martinez et al., Sensors, 2019, 19, 4029
Blind deconvolution using two MOX sensors
Supervised deconvolution
using a PID as ground truth
ℎ 𝑡 = 1 − 𝑒−
𝑡
𝜏
Step response
Time constant
Sensor stimulation
25. 3.2.1 Transient feature extraction for SR distance prediction
Filtered MOX
Bouts
• The “bouts” are the rising edges of the LPD-filtered signal
with amplitude higher than the noise threshold (𝒃 𝒕𝒉𝒓).
• The bout frequency (BF) encodes the SR distance:
• The higher is the BF, the closer is the source.
• Figure of merit: Root mean squared error in prediction:
J. Burgués & S. Marco, Sens. Act. B:Chem, 2020
26. 3.2.2. SR distance prediction results
Effect of window size Prediction error under different wind speeds
J. Burgués & S. Marco, IEEE Access, 2019
J. Burgués & S. Marco, Sens. Act. B:Chem, 2020
Effect of 𝒇 𝒄 and 𝒃 𝒕𝒉𝒓
• Joint optimization of 𝑓𝑐 and 𝑏𝑡ℎ𝑟 leads to the best RMSEP.
• Predictive models based on BF require longer
measurement windows than those based on statistical
descriptors of the signals.
• Optimized BF models can be relatively
insensitive to variations in wind speed.
Algorithm is publicly available
28. 3.3.1. SR distance prediction in dataset #3 (small office room)
J. Burgués & S. Marco, Sens. Act. B:Chem, 2020, 304, p.127309. (Dataset & code publicly available)
• Real-time signals of 27 MOX sensors during 90 minutes of gas release.
• Gas distribution maps (mean, variance, bout frequency) computed every 5 minutes.
• Gas source localization using the cell with maximum value in each of the maps.
29. 3.3.2. SR distance prediction in dataset #3 (small office room)
• Maximum
• Variance
• Bout Frequency
Localization error in Experiment 2
• BF provides lowest error considering all
experiments.
• BF reduces overall error by 25% w.r.t.
variance estimator.
Overall results
Max
Var
BF
• Statistical descriptors of the signal (mean,
variance, etc.) are sensitive to timestamp.
• Bout frequency (BF) continuously provides low
error estimates (1.3 m)
31. 3.4.1. SR distance prediction in dataset #4 (large indoor lab)
J. Burgués et al., Sensors, 2019, 19(3), 478, doi.org/10.3390/s19030478
• Scenario: 160 m2 indoor laboratory.
• Predefined navigation path (3D).
• Average flight speed: 1 m/s
• External localization system: RF beacons.
• 3D map of instantaneous concentration.
• 3D map of bouts (blue circles)
32. 3.4.2. SR distance prediction in dataset #4 (large indoor lab)
• Localization errors using optimized BF: 0.7 - 2.2 m
• Measuring near the source is key.
• Proper selection of 𝑏𝑡ℎ𝑟 is required for maximum
sensitivity and noise rejection.
J. Burgués et al., Sensors, 2019, 19(3), 478, doi.org/10.3390/s19030478
Impact of 𝒃 𝒕𝒉𝒓 on the bout map (Blue: low 𝒃 𝒕𝒉𝒓 ; green: high 𝒃 𝒕𝒉𝒓)
33. • Improving the bandwidth of low-cost chemical sensors is necessary to successfully apply them in
mobile robots for gas sensing tasks.
• In this work, we developed signal processing techniques to boost the bandwidth of MOX sensors
by a factor 20 and machine learning algorithms to extract dynamic features which are indicative of
the source-receptor distance.
• The results have been validated in four different scenarios, including wind tunnels and 3D indoor
spaces, using fixed and mobile sensing platforms (nano-drones)
• The proposed algorithms enable the use of low-cost gas sensors for source localization,
concentration mapping and emission quantification in environmental and safety applications.
Summary and conclusions
34. Javier Burgués, MSc, PhD
Decoding Turbulent Chemical Plumes
with Improved Signal Processing and
Machine Learning
Post Doctoral Researcher
Institute for BioEngineering of Catalonia
Copyright 2020 J. Burgués and InsideScientific. All Rights Reserved.
jburgues@ibecbarcelona.eu
/Javier_Burgues
/jburgues8
35. Javier Burgués, MSc, PhD
Post Doctoral Researcher
Institute for BioEngineering
of Catalonia
Chris Rand
Senior Product Consultant
Aurora Scientific
Thank You! CLICK HERE to learn
more and watch the
webinarTo learn more about olfaction and plume tracking
research solutions from Aurora Scientific, please visit:
https://aurorascientific.com/products/neuroscience/