CONTROL SYSTEM FOR LOW-COST MINIATURE ISOPYCNAL FLOATS (MINIONS)' TARGETING DISTRIBUTED OBSERVATIONS OF THE OCEANIC BIOLOGICAL CARBON PUMP

The flux and attenuation of Particulate Organic Carbon (POC) export through the mid-water column has been a subject of oceanographic research for decades. The methods to sample sinking particles have remained largely unchanged, and interpretation of the data is challenged by uncertainty and sparse measurements. This need prompted our efforts to develop a standalone sensor platform for subsurface timclapsc photography of accumulated particles as a proxy for carbon export, combined with simultaneous measurements of dissolved oxygen, temperature and pressure. The use of a flexible Linux single board computer with power cycling enabled the float to be; low power (for long term deployments), low cost (to be deployed cn-massc), easy to operate (to be deployed opportunistically by other scientists) and flexible for future expansion and improvements.


Introduction
Studying the transport of organic carbon from the ocean's surface to its depthsoften referred to as carbon export or the biological carbon pump (BCP) -is crucial to understanding the global carbon cycle and building better predictions of how this cycle will be modified under a changing climate. The primary mechanism of export is through sinking particles. The "efficiency" of the BCP is dictated by the rate that the carbon is rcrruncralized by marine bacteria in the mesopclagic and the sinking speed of the particles. The crucial measurement that integrates these two rates, is the flux of sinking organic carbon as a function of depth. Current in-situ direct observations of BCP fluxes arc largely limited to analyses of the contents of particle-intercepting traps (PITs) -a proven but costly method limited by spatial breadth and longevity of measurements.
This thesis describes the development and application of the control system for a MINION -a lvlINiaturc lsOpycNal float, an instrument designed to supplement and mitigate the limitations of PITs traps through broad distributions of measurements. The overarching goal for these platforms is to help answer a set of research questions identified as essential by a group of scientists and engineers at a 'Pump it up ' workshop held at WHOI in Sept 2017(doi: 10.1575/1912 [l]: What controls BCP export flux: physical mechanisms? biological/community structure? particle morphologies? chemical composition and density? 'What arc the particle sinking rates? What arc the associated elemental fluxes that arc important for geochemical mass balances? In order to address the questions defined during the 'Pump it up' workshop, the MINION floats were designed to be deployed in large numbers -seeding upward facing time lapse cameras and sensors throughout the upper mesopclagic ocean to obtain spatially distributed timclines of carbon export proxies and other biogcochcmical parameters. The MINIONs must be water-following (Lagrangian), allowing them to image sinking marine snow that settles on an upward-facing camera (Section 3.1) while minirruzing turbulence and sensor displacement relative to the water column fluctuation. They must also include basic sensors like temperature and pressure as well as an oxygen sensor (Section 3.3) that will enable calculation of mesopclagic respiration rates from the timc-rat=f-changc of the data. The system must be able to receive data from a ROAM acoustic tag to enable underwater gcolocalization, and must be able to be deployed for durations of weeks to months.
In order to achieve the broadest possible impact, the MINION control system was designed with these specific goals and rrussion constraints in mind: 1. Low cost for mass deployment.
2. Low power for long term (weeks to months) data collection.
3. Easy to use so that they can be deployed by other scientists opportunistically. This thesis first reviews the history of Lagrangian subsurface measurements with an emphasis on the RAFOS float -a direct precursor to this design, and the Carbon Flux Explorer that uses a similar imaging approach. Next, I provide an overview of the float, with a detailed description of the MINION control system (the major topic of this thesis). I then describe some of the systems that I developed as an undergraduate intern in Dr. Omand's lab. That directly informed and contributed to the MINION. Lastly, I describe some of the results from research cruises that allowed me to test and demonstrate various aspects of the overall design.  RAFOS floats incorporated a high accuracy clock allowed for long term location measurements while reducing accumulated error. The floats listen for transmissions at scheduled times and determine their location via on board processing. This location data was included in sensor data sent to shore while transmitting at the surface. This enabled float deployments in larger numbers, providing observations of mesoscalc variability, and illuminating the impact of eddies and jct meanders on vertical velocity (Bower and Rossby, 1989) [l].

List of
The basic principle of operation for the MINION is similar to its deeper, longer endurance, longer range, predecessor, the RAFOS float (Figure 1, Rossby ct al., 1986) [2]. Both RAFOS and MINIONs arc enclosed in glass tube housings, they measure temperature and pressure, and use embedded hydrophones to records sound signals from at least two acoustic beacons. Their subsurface position is re-constructed based on triangulation of the distances between the source and receiver estimated from arrival times (sources and floats have synchronized clocks) and the sound speed. Both styles arc passively ballasted ( without active buoyancy control) and designed to drift at a prescribed ambient density layer for a prcdcternuned duration. At the end of their mjssion, a weight is dropped and the floats rise to the surface for data transmission. A key distinction between these two systems arises from their different data objectives. lvlINIONs will carry BCP-specific sensors, oxygen and a marine snow flux camera, and will target the upper water column (between the mixed layer and about 300 m depth) for durations of days to months.
In contrast, RAFOS floats arc designed for physical measurements, and deep deployments of up to two years. Other significant differences include physical size and the cost to produce each. lvlINIONs range from 16 to 22" in length, small in comparison to the typical RAFOS float at 5 to 6 feet long depending on payload.
RAFOS floats typically cost S6K whereas the lvlINIONs arc assembled at a cost of roughly S2K each including labor. In a study published in 2016, both Lagrangian CFE vehicles with OSR cameras and moored OSR platforms were deployed to compare if particle size and distribution varied between sampling methods. Comparison of the resulting image data revealed that the Lagrangian CFE captured particles from 5mm to greater than a centimeter while the moored OSR images contained "only fragments of these aggregates and few particles larger than 2 nun" (Section 3.3 of [3]). This finding was attributed to turbulence generated from currents as weak as 2 cm/s at the particle baffle, that can still be much faster than the typical carbon export sinking rate of .1 cm/sand slower. These results demonstrate the advantage of Lagrangian imaging observations by mitigating surrounding fluid disturbance.
Previous studies have clearly defined the need for high volume, spatially dispersed, long term vertical timclapse imagery of carbon flux. MINIONs combine Lagrangian sampling methods, informed by the CFE, with RAFOS positioning capabilities to add a new layer of context to observations of the oceanic biological carbon pump.

Methodology
The design of the MINION control system began with selection of a low cost, high resolution camera for marine snow flux imaging. The Raspberry Pi Zero W with the 8 MP camera module offers the same functionality as a full size RPi while consuming a tenth of the power due to its much less powerful processor. Since the Pi was capable of communicating with all other required sensors, it was made the primary sampling device. The Zero vV variant also features on board vViFi allowing wireless access for configuring and data recovery. This is critical to avoid opening the housing after the MINION is ballasted and vacuum scaled. A thick, cast acrylic sheet was selected for the single end-cap, because it is sufficiently optically clear and also easy to machine ( Figure 4). The thickness (1.5") was found to be the optimal distance between the focal plane and flux camera, providing a ~10 cm 2 imaged area. Thus, the camera and internals can be mounted directly onto this inner face. Fully ballasted, the MINION weighs 2 kg. Slocum gliders by Teledyne 'Webb Research) is used to pull a vacuum inside the float before deployment. The vacuum reduces the humjdity and prevents condensation on the lens, and also holds the endcap tightly in place. An alumjnum pasi:rthrough (Blue Robotics), is used to construct a burn-wire ( Figure 5). An Inconel loop anode is epoxied into the pass-through opening. The anodizing is sanding off of the outer surface, and a ground wire is fixed on the inside, press fit between an aluminum tube and the inner surface of the shaft of the pasi:rthrough.
These expendable burn wire units can then be readily replaced by unthreading from the endcap. At the end of the mjssion, a ballast weight is released, the orientation of the centers of mass and buoyancy reverse, and the Minion flips over (Figure 6, E).
At the surface, the u ppcr few inches of the domed part of the housing with a strobe and embedded Iridium and GPS antenna emerges and begins transmjtting for recovery and/or data transmission(Figurc 6, F).

Minion Ballasting
In order to properly ballast each MINION, a custom tank (Figure 7) was constructed in the Watkins Lab by Professor Omand and Ben Hodges (WHOI). At 4 feet deep, 7 inch inner diameter and inch thick walls, the chamber can be pressurized up to 500 psi with the tank filled with fresh or saltwater from a holding reservoir (allowing the water to reach a stable room temperature and to r<7USC saltwater).
A 'chain method' is employed to accurately determine the amount of added weight needed for each Minion to hit their desired isopycnal target. A metal chain of known, uniform density is hung from the float ( additional ballast is needed) such that the vehicle is suspended in the water column with part of the chain resting on the bottom of the tank. Before pressurizing, the links suspended in water arc counted. The added ballast should be sufficient to allow for an initial descent (as trapped air and other compressible parts arc squcczcd, typically over the first 50 psi). After this small descent, the Minion become less compressible than water and it will rise gradually as the pressure is increased. By counting the number of linkages in the d1ain at various pressure levels (ideally encompassing the target pressure) the density and compressibility of the float can be accurately dctermjned to within about 0.05 g/rnl. During sleep, the MINION consumes ~.07 Watts and roughly 2 Watts for 80 seconds during sampling. If set on an hourly sleep cycle, the batteries would be fully depicted after about 30 days, however, we reserve about 50% of this power for recovery (burn-wire, strobe and Iridium communications).

Description of the MINION Control System
The following sections provides a detailed description of the Minion platform: electronics, software and operation. and is easily enabled as a device tree overlay on the RPi. 'Time is kept with a 3V CR1220 coin cell battery backup on the HA'T.

Accelerometer
Vehicle orientation used to analyze timclapsc results is monitored by an on-HA'T Analog Devices ADXL345 i2c accelerometer. 'The 10bit 3 axis device samples at 100 flz to add another layer of conte:,.-t to sensor readings. In previous experiments this data was used to determine vehicle orientation while sampling.

Analog to Digital Converter
For simple analog to digital conversions on-board, a 'Texas Instruments ADSl 115 i2c ADC is board mounted for creative us~ascs. Equipped with 4 single ended or 2 differential inputs this ADC only consumes 150 flA in continuous-conversion mode.

RPi Camera
The Raspberry Pi Camera Module V2 was demonstrated to be a good tool for marine snow imaging during the 2018 EXPORTS cruise (Section 4.3), and was again chosen for the flux camera. The Sony IMX219 sensor images just above standard 2K resolution (2592 X 1944) using the V4L2 open source driver. The RPi camera is adapted to shoot macro photography by manually unscrewing the included lens from the sensor, about 1.5 rotations, until desired depth of focus is achieved, resulting in high quality macro images for a $35 camera.
A 12V automotive headlight halo LED ring was determjned to provide even and diffused illumjnation to the imaging surface. While the halo is on, the MINION takes advantage of the RPi's built in white balance and automatic exposure settings to improve imaging performance in the changing lighting conditions of the twilight zone. This comes at the cost of consistent imaging conditions for automated image post-processing, though at this stage imagery is used for qualitative purposes.

MS5837-30BA Pressure Sensor
This pressure sensor is sourced from Blue Robotics. Based on the MS5837-30BA sensor by TE Connectivity, the Blue Robotics package comes pr<7potted inside a standard Blue Robotics pass-through, thus saving time during assembly. The MS5837-30BA sensor consumes 1.25 mA of current while communicating data via i2c, with a maximum depth reading of300 meters, reported+/-200 mbar accuracy and a temperature range of -20 to +85.C the pr<7potted device fits the MINION's needs for a low price of $72. Recently we became aware of a potential drift problem in this sensor during multi-day deployments which will be tested in our ballast tank before we can determjne if a replacement sensor is necessary.

TSYS0l Temperature Sensor
This $60 temperature sensor and housing is also sourced from Blue Robotics. Based on the TSYS0l from TE Connectivity it comes in a similar form factor to the Blue Robotics pressure sensor above with the added feature of suspending the potted chip inside an exposed metal cage for increased water flow over the sensor. The sensor consumes 1.4 mA while in use with a temperature range of -40 to +125•c and accuracy of+/-0.5•C. Though the MS5837 docs include a temperature reading to calibrate pressure, this sensor provides higher accuracy for minimal additional investment.

Dissolved 02 Sensor
The incorporation of an optical PrcSENS OxyBasc WR-RS232 dissolved oxygen sensor further expanded the MINION payload. The OXYbasc was chosen for its accuracy and 12mm diameter form factor. The vVR-RS232 has a dissolved oxygen detection range between 0 -22.5 mg/L with accuracy down to 0.01 mg/L. A converter board was made to interface between RPi TTL serial bus and OXYbasc RS232 via MAX3232RL. The cost of the sensor at $1200 significantly adds to the overall cost of the MINION and as such we anticipate not all MINIONs will be equipped with this sensor. As the device name suggests, communication is handled via RS232 through custom TTL to RS232 converter (Figure 9).

Iridium and GPS
In previous iterations of the MINION control system, in order to save money and development cost, a Spot Trace standalone GPS tracker was responsible for relaying the MINIO N's position for recovery at sea. This off the shelf option left a lot to be desired, between the proprietary device firmware and less than effective internal antenna, resulting in the exploration of a custom solution ( Figure 10).  The magnetic recd switch located on the MINION endcap controls the enable pin of the system's primary voltage regulator. 'When the magnet, embedded in the MINION lens cover, is applied the system consumes no power for transit.
Removing the magnetic endcap delivers power to the micr~ontroller and begins the programmed MINION sample cycle. Post recovery, the recd switch is employed to 'wake up' the MINION during sleep cycles by restarting the micro-controller via power resetting and connecting to the MINION Hub WiFi network.

Batteries
As with all underwater vehicles, the batteries make up a portion of the total payload weight consideration. 20 LiMnO2 CR123A 3V batteries by 'Titanium Innovations (Figure 11) arc arranged into a 12V 7 Ah battery pack . 'The pack is shaped in an open ring and located against the cndcap so the traveling weight slide into the center hole keeping the weight as low as possible to maximize float stability during recovery and data transmjssion.

Software
All software is made publicly available at: https://github.com/jacksonsugar/MINION--2.9.git 'The repository contains scripts for both the A'TMEGA328P and Raspberry Pi as well as installation instructions for both devices.

ATMEGA328P
Shared by both the Arduino development boards from previous projects (Section 3.7), the A'TMEGA328P satisfies the MINION's needs for a variety of IO as well as low power sleep capabilities. 'Though the microcontrollcr is now embedded in a custom PCB, programming via Arduino IDE is still available for novice programmers. Rather than USB, the microcontroller is programmed via a standalone usbtiny SPI AVR progranuner. From the factory, the 328P needs fuses set in order to enable the internal crystal oscillator before the end user uploads the MINION C++ script.

Deployment
To deploy, remove magnetic lens cap and wait for the red LED on the HA'T to trigger to ensure the !\,[INION correctly began initial sampling. Once sampling, lower device into ocean and release. \,Ve found that a simple slip line with release pin (Figure 15) worked well while operating off a larger vessel. At the beginning of the MINION Life Cycle (Fig. 6, A), the Pi checks the data output directory for samples, if none arc present it will sample all sensors for the configured initial sample time to track the MINION's descent to its target isopycnal. Once complete, the !\,[INION begins the sleep cycle.
The MINION will continue to cycle power and sample for the configured mission duration (Fig. 6, C & D).
During the mjssion, while the RPi is active, the pressure sensor is sampled continuously in case the MINION surpasses its configured maximum operating depth.
In such case, the MINION will abort mission and begin recovery procedures early.

Recovery
After the last time lapse photo is taken, the Pi communicates to the 328P to begin recovery. The RPi releases the ballast and the MINION samples all sensors continuously while ascending to the surface. The RPi, informed by the pressure sensor, ensures the MINION reaches the surface before attempting to send location and sensor data via Iridium (Fig. 6, F). Due to the small size of the MINIONs, fish net style recovery from the side of larger research vessels can be performed safely ( Figure 16).

7 MINION Precursors
Before completion of the first Minion prototype, various pre-cursors enabled me to demonstrate and refine discrete aspects of the design. The following sections describe some of these activities I completed while I was an undergraduate working in Dr. Omand's lab.

Surface Tethered Trap (STT) Timelapse Gel Cameras
The STT (Surface Tether Trap) gel cameras were developed as a means to evaluate the suitability of the Raspberry Pi Foundation's Camera Module 2.0 for imaging marine snow particles (Figure 17a) (a) ST1' gel cameras on the bench during the cruise (b) An example of an Image captured by ST'T gel camera. The photo illustrates the size range and quality of particle images that can be obtained with this system, generally encompassing mru·ine snow aggregates, fecal pellets, swimm ers, and smaller particles that rru1ge from 500 um to l cm. Results from Research Cruises The following sections briefly describes cruises that enabled demonstration of the MINION or MINION precursors.

EN601
In the summer of 2017 the R/V Endeavor set sail in the Gulf of Mexico with the intention of validating a small form factor hydrophone/recorder for tracking of small fish and lobsters [l]. These recorders, named Fish Tag-s (Section 5.1.1), were confirmed capable to suitably resolve sub-surface gcolocation greater than 60 Km from the sound sources with an error between 70 and 560 meters [2]. Figure 19 details the experimental data collected during testing. Employing this underwater tracking solution for MINIONs will help constrain the trajectory of the vehicles underwater post recovery. Integration of these tags is planned, but not as part of this thesis.  deeper -to about 140m, so we have some refinements to make to our ballasting method. Regardless, since both of them settled on an isopycnal, took photos and measurements and were recovered, these deployments were considered a success.
The pressure records provided valuable insights into the descent rate of the floats.
A closer look at Minion 102's approach reveals an initial descent at about 6 m/s ( Fig. 23 lower panel). After a couple of mjnutcs, the lifesaver dissolved and dropped the extra ballast. The Minion then rose and toward its neutral density.
\Ne can then sec it move up and down with internal waves of about 5 minute period.
A more detailed analysis of this behavior, and an analysis of vehicle properties like drag, is planned for a future paper.  Rossby and Daniel Moonan, known as the fish tag [1] was designed with mjd water drifters in mjnd. Potted within the hydrophone exterior, the self contained unit houses accurate timjng, pressure sensor, temperature sensor, signal processing and on board memory to support standalone acoustic tracking based on RAFOS triangulation.
Integration of these acoustic tag'S would allow for gcolocalization while the MIN-IONs arc underway (though not in real-time) ( Figure 18). As the MINIONs arc Lagrangian, this mjd-water location data will help place MINION measurements into a spatial context. Dr. Om and is working with URI professor Dr. Godi Fischer and Dr. Tom Rossby to supply these for the MINION. Unlike traditional batteries where voltage measurement is a good indication of discharge level, the r;rc battery maintains output voltage until almost 10% power remains. This is beneficial to ensure ample power delivery to Iridium hardware during data recovery at the end of the mjssion.
In order to monitor the new battery supply, an i2c Texas Instruments BQ35100 will be embedded on the battery pack for coulomb counting during deployment.
During operation, the RPi resets the counter after new batteries arc installed and can receive warning when the batteries approach total depiction.

Expendable MINION Features
As MINIONs arc deployed in larger quantities and for longer time spans it becomes less feasible to physically recover the vehicles for data analysis. Opposed to the current recovery protocol, transmitting location for retrieval, the expendable MINION will process both sensor and image data for Iridium transmjssion.
The Raspberry Pi on board computer is compatible with many source image quan-tification/classification software packages for pre-transmission image processing i.e.
OpcnCV, Tensor Flow or other Linux compatible software. Once data transmission is complete, the expendable !v[[NION's final mission is to sink to avoid becomjng floating debris. Methods to reliably sink the vehicle post data transmjssion arc still begin discussed.

Conclusion
The MINION platform demonstrates the viability of employing low cost hardware and power saving strategics to collect dispersed, long term data sets for measurements in the ocean twilight zone. The control system can be configured to sample for days to months at a time. Autonomous, with failure modes in place to recover itself in the event of a malfunction. The platforms arc much smaller and cheaper to produce than previous sampling methods allowing case of shipping and cn-massc deployments. Sampling and telemetry is handled by a flexible Linux single board computer for future sensor inclusion and upgrades, with programming made simple via vViFi configuration. The first large scale deployment (up to 20 lvlINIONs) was scheduled for May of 2020 during the 2020 NASA EXPORTS cruise (Canceled due to Covid-19). We look forward to evaluating the performance of the most recent revision of the MINION during the rescheduled 2021 cruise.