Team:UiOslo Norway/Integrated Human Practices

Integrated Human Practices

We have gathered information from experts in their respective fields, which has provided new perspectives and valuable insight to our project.

At the start of our project, we intended to build a stochastic model with enough tunable parameters to represent individual fish. We then wanted to use monitoring data from aquaculture ponds, for example from cameras, to tune the model by adjusting these parameters. We would then be able to use this model to aid in the design of a detection system.

Here is an example of what we refer to as a detection system:

The system consists of a set of sensors gathering data, followed by feature extraction, finally we would feed the features to a classifier that assigns our data to a predefined category. This can of course be extended, by for example assigning a probability to a classification. An example of this system could be cameras filming a fish school, then estimating the total velocity giving a scalar for some given resolution, then feeding this time series into a neural net.

In a meeting with Hans Jonas Fossum Moen, Associate Professor, section for Autonomous Systems and Sensor Technologies at the University of Oslo, we were advised that adjusting a model by use of tracking individual fish (e.g. from camera data) would be very difficult. This would require a whole set of experiments and modeling on its own, since the fishes position needs to be estimated due to the 3 dimensions of freedom. Setting up this system for estimating positions, and running experiments to validate it, would take more time than we have and would not be interesting to our end goal. Instead he advised that we just focus on modeling fish, which we can base on previous research, and where lots of information is already easily accessible. It is also probably easier to validate the system we arrive at with this approach, since we can test it directly on or with data from a fish farm. So, we changed the project to simply focus on modeling of fish behavior and their environment based on previous publications.

A problem we had realized at this point was that fish have too many and complicated behaviors for us to account for them all. To help alleviate this problem we had the idea to simply represent behaviors deemed unimportant or unfeasible as noise. The noise here is then just random changes to an individual fish’s position.

In email exchanges with Kai Olav Ellefsen, Associate Professor, Research Group for Robotics and Intelligent Systems, UiO. we laid out the idea given above, asking for feedback on how we could make our model as realistic as possible. Our system needs to operate with a lot of noise. The idea was to make the model more noisy than in the real world, and then make sure our detection system works with different noise levels. This might be a good indication that this detection system would be the best in the real world as well, since the real world would contain many complicated behaviors that might seem random and act as noise to our system. Prof. Ellefsen proposed that we might be able to prove this mathematically or show it by successively noisy simulations. He thought we would be able to say a lot about the detection system from a model if it exhibits similar characteristics to those present in real fish.

Lastly, we talked to Simen Halaand from Fredrikstad Seafoods. We sent him emails asking about the sensors, procedures and technologies they use at their fish farming facilities to monitor fish behavior. Questions and responses can be seen below (translated to English from Norwegian).

Question	Response
List of sensors (cameras and all measurement devices)	Mostly water quality parameters (O2, CO2, Temperature, Salinity, pH, redox, etc.)
Sensor (cameras and all measurement devices) layout	Measured in the fish tank and Recirculating aquaculture system(RAS)
Tank design, layout, the best would be a cad model or picture of a model.	RAS2020 Kruger design.
Disease recognition procedure (Gill score etc.)	We use the handbook from Nofima/Vet.inst/CtrlAqua to score well being, including gill health
Any information about an automatic system that detects disease. Do they have any automatic recognition of altered swimming patterns?	Nothing like that in our facility.

They unfortunately listed no sensor that we can directly use our approach. But it is very interesting that they use no automatic sensing techniques, leaving us with no competition in our project, and showing that we might generate a product with great value to the industry.

Conclusion

The interactions with Prof. Fossum and Prof. Ellefsen defined the path for both our model and for the final detection system. These interactions led us to base our behavioral modeling on published data and simplify our approach, making the project feasible. After having created an appropriate model we can then test our detection system’s robustness by increasing the noise. This increases the likelihood that it will work in the real world, with all the complex behaviors that fish exhibit. The interaction with these two experts was thus fundamental to our project.

The interaction with Halaand further validated the value of our project and led us to be more considerate with the proposed implementation.