Introduction
In this abstract, methods and technical solutions for real-time simulation and monitoring of farmed salmon (Salmo salar ) in flexible net cage are introduced. Machine learning -based estimation methods are implemented to combine an individual-based fish model with field measurement data and quantify behavioural changes of the fish when exposed to dynamically changing environments. Fish density distributions in the cage are determined by the simulation results and assimilated echo- sounder data. Fish distributions around a submerged structure such as an underwater vehicle can also be simulated according to the 360-degree sonar data ( BlueRobotics - Ping360 Scanning Image Sonar). Field trials and demonstrations have been conducted in an industrial-scale fish farm for instrument testing and verifications of the integrated simulation and monitoring system .
This work was financed by the Research Council of Norway through the project: CHANGE [1] .
Materials and methods
A flexible net cage and fish were modelled in the simulation framework, FhSim, which allows a high degree of flexibility to combine different mathematical models, numerical solvers, sensors/observers and relevant estimation techniques for time-domain representation of a complex system [2]. The fish model in FhSim is individual based [3] , able to simulate full-scale fish populations (e.g., 200,000 individuals in one cage) in real time. The spatial and temporal fish behavioural responses towards the cage, feed, temperature, light, water currents, waves and other individuals are considered. However, there are several influence factors (e.g., oxygen level and consumption) that are not yet included in the model, and it is notoriously difficult to measure all the environmental conditions and complex interactions in the field.
Therefore, several simulations have been conducted with randomised environmental and behavioural parameters (more than 10,000 samples). These simulation data were used to train machine learning -based surrogate models [4 ] which were then able to find the most possible behavioural parameters for the resulting fish distributions. As shown in Figure 1, the machine learning -based estimation methods are incorporated into the simulation model to continuously update the behavioural parameters according to measured fish distribution data. The integrated simulation and monitoring system will not rely on accurate environmental inputs that might be difficult to obtain, but can still use the environmental and structural monitoring data when available.
Results
A real-time data collection and communication system has been tested in an industrial-scale fish farm (as shown in Figure 2) . The measured fish densities from a multi-beam echo sounder were assimilated with the simulation model for real-time monitoring of vertical fish distributions. The simulation results were shown to coincide with the echo-sounder data and provide more information about the fish, such as the influence of the current and cage deformation on fish distributions and changes in the fish swimming speeds. An extended setup of several echo sounders (both single-beam and multi-beam) in the cage will be tested, as well as the integrated simulation and monitoring system .
Machine learning -based estimation methods have also been implemented to identify and quantify fish distributions around an underwater vehicle or a generalised object, where the 360-degree sonar data were used to parameterize the behavioural changes. These could be extended for real-tim e fish detection and characterization of the possible interferences to the fish. In combination with the estimated fish distributions in the cage, a bio-interactive control routine can be implemented to ensure fish welfare and increased efficiency in relevant underwater operations ( e.g., autonomous underwater data acquisition).
Conclusions and future work
An integrated simulation and monitoring system is being developed and shown to be suitable for real-time applications in an industrial-scale fish farm. The tested instrument setup is preliminary and will be extended for actual use. The machine learning -based estimation methods incorporated into the simulation framework form a basis for hybrid analysis and modelling [5] and relevant digital twin implementations. A holistic digital twin solution for aquaculture structures and farmed fish will improve the ability to monitor, control and document aquaculture productions and facilitate knowledge-based decision making, thereby contribute to the realisation of precision fish farming.
References
[1] CHANGE . https://www.sintef.no/en/projects/2021/change-an-underwater-robotics-concept-for-dynamically-changing-environments/.
[2] Su et al., 2019. A multipurpose framework for modelling and simulation of marine aquaculture systems. In Proceedings of the 38th International Conference on Offshore Mechanics and Arctic Engineering (OMAE 2019).
[3] Føre et al., 2009. Modelling of Atlantic salmon (Salmo salar L.) behaviour in sea-cages: A Lagrangian approach. Aquaculture 288, 196-204.
[4] Bouhlel et al., 2019. A Python surrogate modeling framework with derivatives . Advances in Engineering Software 135, 102662.
[5] Pawar et al., 2021. Hybrid analysis and modeling for next generation of digital twins . Journal of Physics: Conference Series 2018, 012031.