System and methods for automated fish handling and monitoring

Abstract

An automated system for, and method of, harvesting and crowding fish in various environments, equipped with a combination of sensing elements to detect fish stress indicators and mechanical performance metrics is provided. An automated system for fish harvesting and crowding, equipped with AI-powered method that processes data from sensors, cameras, and other technologies to provide reactive and predictive control of the crowding process in various environments is also provided.

Description

SYSTEM AND METHODS FOR AUTOMATED FISH HANDLING AND MONITORINGCROSS-REFERENCE TO RELATED APPLICATION(S)5

[0001] This application claims the benefit of United States Provisional Patent Application No. 63 / 742,616, filed on January 7, 2025, the entire contents of which are incorporated herein by reference.FIELD

[0002] The present disclosure generally relates to the field of aquaculture, and in particular, to systems and methods for automated fish handling and monitoring.BACKGROUND

[0003] During harvesting and crowding processes, fish biomass may be caught in a large 15 aquiline net pen and slowly crowded into smaller spaces, wells, pipelines or other infrastructure, in order to move the fish to another location, treat the fish (e.g., delousing) and / or harvest the fish. During this process, the fish may get damaged.SUMMARY

[0004] In at least one broad aspect, there is provided a computer-implemented method for monitoring and handling fish in an aquaculture treatment system, the method comprising: (a) receiving at least one sensor data stream from at least one sensor assembly deployed in the treatment system, wherein the at least one sensor data stream comprises one or more' of measured environmental data, measured physical data, and image data; and (b) 25 applying a trained machine learning model to the at least one sensor data stream to generate one or more outputs comprising: (i) an identified stress condition associated with an increased risk to fish welfare in the treatment system; (ii) at least one contributing factor toWSLEGAL\097387\00003\42560596v2the condition; and (iii) an operational adjustment for a controllable device of the treatment system to mitigate the at least one contributing factor

[0005] In another broad aspect, there is provided a system for monitoring and handling fish in an aquaculture treatment system, the system comprising: (a) one or more sensor 5 assemblies deployed in the aquaculture treatment system, each sensor assembly configured to generate sensor data streams comprising one or more of measured environmental data, measured physical data, and image data; (b) at least one processor; and (c) a memory storing computer-executable instructions, which when executed by the at least one processor, configure the at least one processor to implement the method comprising: receiving the sensor data streams; and applying a trained machine learning model to the sensor data streams to generate one or more outputs comprising: (i) an identified stress condition associated with an increased risk to fish welfare in the treatment system; (ii) at least one contributing factor to the condition; and (iii) an operational adjustment to a controllable device of the treatment system to address the at least one contributing factor.15

[0006] In accordance with an embodiment, there is provided an automated system for harvesting and crowding fish in various environments, equipped with a combination of sensing elements to detect fish stress indicators and mechanical performance metrics.

[0007] In accordance with another embodiment, there is provided a method for automating the process of fish harvesting and crowding, utilizing a broad array of sensory data to optimize both fish welfare and mechanical efficiency across different harvesting environments.

[0008] In accordance with another embodiment, there is provided an automated system for fish harvesting and crowding, equipped with Al-powered method that processes data from sensors, cameras, and other technologies to provide reactive and predictive control of 25 the crowding process in various environments.

[0009] In accordance with another embodiment, there is provided a method for automating fish harvesting and crowding using artificial intelligence to analyze data from multiple sources, enabling predictive and reactive adjustments for optimal fish welfare and system efficiency.WSLEGAL\097387\00003\42560596v2

[0010] In some embodiments, enhanced analytics and statistical analysis capabilities are developed to monitor fish crowd behavior, compare treatment and transfer data, and automate the real-time detection of problem areas. One focus is on optimizing the crowding and de-lousing processes using Al algorithms for predictive modeling, improving 5 equipment performance analysis, and advancing a web dashboard for better usability and data interpretation.

[0011] In some embodiments, data is collected through multiple sensor units during fish transfer and de-lousing operations. These units monitor environmental conditions such as dissolved oxygen and temperature, as well as vacuum pressure and acceleration data to ensure optimal equipment settings as the units travel with the animals through these stressful environments (e.g., pipes, pumps, and mechanical treatment infrastructure).

[0012] In some embodiments, validation of the data involves deploying sensor and other technology in everyday salmon aquaculture production activities. The data will be trained and validated through historical analysis of equipment performance, fish mortality, fish 15 quality all through standardized systems. In addition, access to internal systems which track quality loss, mortality, and fish welfare indicators to correlate data sets to training and testing our Al models ensures the robustness and accuracy of predictive models. Additionally, real-time data may be continuously monitored and compared against historical data sets to validate the effectiveness of proposed methods and insights.20

[0013] In some embodiments, systems and methods are provided that enhance fish welfare and operational efficiency in aquaculture operations.

[0014] In some embodiments, units engineered to optimize fish handling processes feature advanced sensors and intelligent monitoring capabilities, providing real-time data on crucial water quality and physical parameters. This ensures that a fish biomass is handled safely and efficiently, empowering aquaculture professionals to make informed decisions about equipment and overall performance, minimizing stress and promoting fish health.

[0015] In some embodiments, the units comprise customizable sensor designs and a spherical shape for transfer through pipes, pumps, and equipment. These units integrate into various aquaculture environments, offering flexibility and reliability. The units assist with 30 improving fish welfare to streamlining operational efficiency.WSLEGAL\097387\00003\42560596v2

[0016] In various further aspects, the disclosure provides corresponding systems and devices, and logic structures such as machine-executable coded instruction sets for implementing such systems, devices, and methods.

[0017] In this respect, before explaining at least one embodiment in detail, it is to be 5 understood that the embodiments are not limited in application to the details of construction and to the arrangements of the components set forth in the following description or illustrated in the drawings. Also, it is to be understood that the phraseology and terminology employed herein are for the purpose of description and should not be regarded as limiting.

[0018] Many further features and combinations thereof concerning embodiments described herein will appear to those skilled in the art following a reading of the instant disclosure.DESCRIPTION OF THE FIGURES

[0019] Embodiments will be described, by way of example only, with reference to the 15 attached figures, wherein in the figures:

[0020] FIG. 1A illustrates an example of a sea-based treatment system;

[0021] FIG. IB illustrates an example of a land-based treatment system;

[0022] FIG. 2 exemplifies a system for automated fish handling and monitoring, according to at least one example;

[0023] FIG. 3A illustrates a partially exploded view of a sensor assembly, in accordance with some embodiments;

[0024] FIG. 3B illustrates an example of sensor unit, in accordance with some embodiments;

[0025] FIG. 3C illustrates an example of a sensor array;25

[0026] FIG. 4A shows an example method for automated fish handling and monitoring;

[0027] FIG. 4B shows a further example method for operating a control subsystem of a treatment system;WSLEGAL\097387\00003\42560596v2

[0028] FIG. 4C shows an example method for training a machine learning model;

[0029] FIGs. 5 A - 5D illustrate, in screenshots, examples of a data dashboard;

[0030] FIG. 6 illustrates an example of a graph monitoring fish welfare throughout the treatment operation;5

[0031] FIGs. 7 A and 7B illustrate results for various conducted tests;

[0032] FIG. 8 illustrates, in a screenshot, an example of G-force readings through a smolt transfer;

[0033] FIG. 9 illustrates, in a graph, an example of results for sensor unit data through a smolt transfer;

[0034] FIG. 10A illustrates an example of results for a worst-case scenario for G-force readings through a delousing system;

[0035] FIG. 10B illustrates an example of results for a calibrated G-force readings through the delousing system;

[0036] FIG. 11 is a hardware block diagram of a computing device such as a server; and 15

[0037] FIG. 12 is a hardware block diagram of an example sensor unit.

[0038] It is understood that throughout the description and figures, like features are identified by like reference numerals.DETAILED DESCRIPTION

[0039] Disclosed examples generally relate to systems and methods for automated fish handling and monitoring.I. DEFINITION

[0040] "Aquaculture treatment system" refers to any system or infrastructure designed 25 for the handling, treating, and / or processing aquatic animals (e.g., fish). Such systems can include equipment configured for crowding, transfer, cleaning (e.g., delousing), or other husbandry or welfare-related interventions. Aquaculture treatment systems may include WSLEGAL\097387\00003\42560596v2both sea-based and land-based configurations as known in the art, and further exemplified in FIGs. 1A and IB. The systems include various equipment including net pens, rearing tanks, pipelines, wells, pumps, and the like.

[0041] Stress indicating condition or "stress condition" is a detected condition (e g., 5 environmental, physical and / or imaged) within an aquaculture treatment system that evidences elevated welfare risk or adverse operational state for aquatic animals (e.g., fish) in the treatment system. Examples include, without limitation: abnormal fish behavior patterns, clustering or density spikes, rapid dissolved oxygen (DO) decline or persistently low DO zones, abrupt temperature deviations, irregular pressure transients indicative of unstable flow or blockage, and sudden increases in acceleration or G-force consistent with mechanical shock.

[0042] Environmental data or "environmental data measurements" are sensor data measurements characterizing the aquatic medium in an aquaculture treatment system. These include chemical and thermal parameters and related water or fluid quality indicators.15 Examples include dissolved oxygen concentration, temperature, pH, conductivity, salinity, turbidity, and related water or fluid chemistry metrics.

[0043] Physical data or "physical data measurements" are sensor data measurements characterizing mechanical and / or kinematic conditions experienced within an aquaculture treatment system. Examples include static and dynamic pressure, depth, flow rate, acceleration and G-force, vibration spectra, and impact or shock events.II. GENERAL OVERVIEW

[0044] During harvesting and crowding processes, fish biomass may be gathered in large aquaculture net pens and gradually crowded into smaller enclosures, wells, or pipelines to 25 facilitate transfer, treatment, or harvesting.

[0045] A common treatment during these operations is delousing, i.e., removal of parasites such as sea lice. Traditionally, chemical agents were used for delousing; however, their use has been increasingly restricted or banned due to regulatory and environmental concerns. As a result, mechanical treatment methods are now employed to achieveWSLEGAL\097387\00003\42560596v2delousing and cleaning functions without chemical additives. Mechanical methods often involve the application of controlled water pressure or flushing systems within aquaculture treatment systems. In practice, such treatment operations may be carried out in sea-based or land-based environments.5

[0046] FIG. 1 A illustrates an example of a sea-based treatment system 100a for handling and treating aquatic animals, such as fish.

[0047] As shown, untreated fish are drawn from a sea cage, net pen, or open-water enclosure 102a through an intake conduit 104a and directed into a treatment or delousing unit 106. Within the delousing unit 106, various treatment processes (e.g., mechanical delousing, filtration, or flushing) may be performed under controlled flow and pressure conditions.

[0048] In some cases, delousing is achieved by exposing fish to pressurized waterjets or turbulent flow that physically dislodges sea lice and other external parasites without the use of chemical agents. The delousing unit 106 may include a pump room 108 configured to 15 regulate circulation rate, pressure level, and suction flow to ensure gentle handling and uniform treatment. A filtration unit 110 may also be provided to remove detached parasites and debris from the circulating water before discharge or recirculation.

[0049] After treatment, the fish are then discharged back into the sea environment, or to a discharge tank 102b (via an outtake conduit 104b).20

[0050] Various controllable system hardware may be provided within environment 100a.These include winches configured to adjust the shape and density of the intake nets 102a, thereby managing crowding levels prior to pumping or treatment. Additionally, the speed of the pumps 108 can be regulated to optimize the suction and transfer of fish into the treatment infrastructure. Oxygenation and temperature control can also be integrated (e.g., inline oxygen injectors or heat-exchange loops) to maintain acceptable dissolved oxygen (DO) and thermal conditions during transfer.

[0051] FIG. IB illustrates an example alternative land-based treatment system 100b. In this example, the treatment system includes a series of rearing tanks 112 arranged within a controlled facility. The tanks form part of, for example, a recirculation loop (e.g., RAS) in 30 which water is continuously filtered, oxygenated, and temperature-controlled for reuse. WSLEGAL\097387\00003\42560596v2

[0052] As shown, each tank 112 provides a closed or semi -recirculating environment in which water quality, temperature, salinity, and flow can be closely monitored and adjusted.

[0053] Delousing, in the land-based system 100b, may be performed using mechanical or hydraulic treatment processes integrated into the tank circulation loop. For example, fish 5 may be guided into a designated treatment chamber where low-pressure nozzles (or directed waterjets) generate controlled turbulence to dislodge sea lice or debris from the fish surface. The treated water may pass through a filtration and separation stage to remove dislodged parasites and particulates before recirculation.

[0054] Similar to the sea-based system 100a, the land-based system 100b may also 10 include various controllable mechanisms, including crowding mechanisms, water filtration control, and pump controls to manage fish density and facilitate handling or treatment operations.

[0055] Irrespective of whether a sea-based or land-based system is used, there are sensitivity problems in operating these treatment systems. For example, excessive water pressure may sometimes cause scale loss and other damage to the fish. Other issues that are found to adversely affect the quality of the fish include, low oxygen levels in the water (which can occur when a large number of fish are crowded together), fish swimming into equipment or each other in the pipelines, incorrect temperature levels, improper salinity or pH of the water, blocked pipes, etc.20

[0056] Such issues that can harm the fish are often not noted until the fish are harvested.Moreover, the actual cause of the harm is not always determined correctly. For example, incorrect oxygen levels in the water can cause the fish to swim erratically and cause other stress on the fish. Particularly, such erratic swimming may cause the fish to swim into each other or equipment in the pipelines or wells. While the corrective measure should be to adjust the oxygen levels (i.e., fixing the actual root cause), the blame may be incorrectly placed on equipment, water pressure during delousing, etc.

[0057] Currently, these issues are addressed through manual observation by humans monitoring fish in a well or aquiline net pen. Some treatment systems may also introduce one sensor at a fixed location in a well, pen or other infrastructure to assist with the visual 30 monitoring or water condition monitoring. However, such monitoring will only detect theWSLEGAL\097387\00003\42560596v2final outcome that is presented in the fish; not the actual root cause. As such, any corrective measures taken may not actually fix the problem.

[0058] In view of the foregoing, disclosed systems and methods are designed to optimize operations in aquaculture treatment systems. As provided herein, disclosed systems use 5 advanced sensors and intelligent monitoring capabilities, which provide real-time or near real time data on operational and fish behaviour.

[0059] In some examples, systems described herein also provide for a user-friendly interface that empowers aquaculture professionals to make informed decisions, ensuring optimal conditions for fish health and reducing stress during critical operations.10III. EXAMPLE FISH HANDLING AND MONITORING SYSTEM

[0060] FIG. 2 exemplifies a system 200 for automated fish handling and monitoring, according to at least one example. System 200 may be deployed in conjunction with, or in association with, an aquaculture treatment system 100 (as exemplified above).

[0061] As shown, system 200 may include one or more sensor assemblies 202a - 202n which couple, via a network 250, to server 204. System 200 may also include an aquaculture treatment control subsystem 206, also coupled to network 250. In some cases, one or more user terminals 208 are provided, and further coupled to the network 250.

[0062] Sensor assemblies 202 are deployable within the aquaculture treatment system 20 100 (e.g., FIGs. 1A and IB). As discussed below, the sensor assemblies 202 are equipped with a variety of sensors for monitoring key aquaculture system parameters in a treatment system 100.

[0063] In some examples, the sensors assemblies 202 include sensors for monitoring environmental data, such as dissolved oxygen, pH, conductivity, and temperature levels 25 within the fluid (e.g., water) of the treatment system 100.

[0064] In addition, sensor assemblies 202 may incorporate sensors for measuring physical data, such as pressure, acceleration, and G-force. It is also possible that sensor assemblies 202 incorporate image sensors (e.g., cameras, including video cameras). In at least one case, the sensor assemblies 202 also incorporate location sensors which generate WSLEGAL\097387\00003\42560596v2location data (e.g., global positioning system (GPS) or global navigation satellite system (GNSS) sensors).

[0065] Sensor assemblies 202 may be constructed with robust materials to withstand harsh marine environments, ensuring long-lasting performance. Further, as best shown in 5 FIGs. 3A - 3B (and described below), the sensor assemblies 202 are configured such that they can sit or flow within tight spaces, such as pipelines within a treatment system. For instance, the sensor assemblies 202 may be deployed in a free-floating manner. They can also be deployed in a fixed position.

[0066] It is possible that multiple sensor assemblies 202a - 202n are deployed in a given treatment system 100. This can assist in monitoring data at various locations or depths throughout the net, pipes, wells or other system infrastructure. As described herein with respect to FIG. 3C, one or more sensor assemblies 202 may also couple together to form a “daisy chain” array.

[0067] In at least one example, a dynamic light alarm system is included for real-time or 15 near real time visualization of changes in environmental conditions, enhancing situational awareness at the edge of the net pen or land based facility. A disinfection sleeve may be provided for calibration and cleaning, ensuring optimal sensor performance and accuracy.

[0068] Continuing with reference to FIG. 2, server 204 can communicatively couple, and receive, sensor data generated by the one or more sensor assemblies 202a - 202n. In some cases, server 204 is a cloud-based server as known in the art. While only a single server is exemplified, the system may include any number of servers (e.g., interconnected servers).

[0069] In some examples, server 204 hosts a treatment system monitoring module 250. Treatment system monitoring module 250 can be stored in a memory 1110, of the server 204 (FIG. 11).25

[0070] As explained below, monitoring module 250 analyzes the sensor data generated by sensor assemblies 202. In response to the analysis, module 250 may one or more of: (i) detect stress indicator conditions (also known as "stress conditions") associated with increased welfare risk to aquatic animals (e.g., fish) within the treatment system 100 (e.g., crowding or abnormal swimming behaviour); (ii) determine at least one contributing factorWSLEGAL\097387\00003\42560596v2to the stress condition; and (iii) identify at least one system operational adjustment to mitigate or reduce the effect of the contributing factor.

[0071] Module 250 is also configured to generate various outputs. Outputs can include visual outputs that may be presented on a display interface, e.g., of user terminal 208. These 5 outputs can include indications of detected stress conditions, associated contributing factors, and recommended system adjustments. For example, the output may indicate that a stress event is being caused by fish overcrowding in a pipe, which is occurring due to low water oxygen levels in the pipe. Therefore, the recommended system adjustment is actuating a pump to release more oxygen into the water.10

[0072] In some examples, module 250 also generates control instructions that are transmitted to control subsystem 206. The control instructions are operable to adjust the control subsystem 206 in order to implement the identified system adjustments (e.g., adjust a pump) and mitigate the detected issues.

[0073] In at least one example, the module 250 can include one or more trained machine learning models which implement one or more of the above-described functions.

[0074] Still referencing FIG. 2, aquaculture treatment control subsystem 206 represents the controllable electronic hardware components of treatment system 100 (as discussed above). This subsystem encompasses devices and mechanisms that can be actuated to regulate operational aspects (e.g., environmental and / or physical) of the treatment system.20 Examples of such hardware include electronically controlled pumps / valves for adjusting water flow and oxygenation, automated winches for modifying net pen configuration, and crowding mechanisms for managing fish distribution during transfer or treatment processes (e.g., baffles, fates, barriers, curtains, etc.). These components may be operated either manually by an operator or automatically in response to control instructions generated by module 250.

[0075] User terminal 208 is any suitable user device configured to interact with the aquaculture treatment system. The user terminal may include an output interface, such as a display screen, for presenting system outputs, alerts, or recommended adjustments to an operator. Additional input interfaces, such as a keyboard, touchscreen, or mouse, may be 30 provided to facilitate user interaction with the system.WSLEGAL\097387\00003\42560596v2

[0076] Network 250 may be implemented using wired or wireless technologies. These include, but are not limited to, Ethernet, Wi-Fi, Bluetooth, or cellular networks. It is not necessary that the same network is used to connect all system components; rather, various subnetworks may be employed to facilitate independent connectivity between different 5 elements of the system.IV. EXAMPLE SENSOR ASSEMBLY

[0077] FIGs. 3A - 3B illustrate an example configuration for a sensor assembly 202, in accordance with some embodiments.

[0078] As shown, sensor assembly 202 may be shaped to travel (or be relocatable) through pipes, conduits, and other narrows spaces defining the treatment system infrastructure. For example, they may have a generally spherical shape.

[0079] Sensor assembly 202 generally includes a sensor unit 240. In some cases, the sensor unit 240 is further surrounded, and protected by, one or more protective members 15 206. In this example, the protective members 206 include soft guards which are made of polyuretane (e.g., 12 mm).

[0080] As discussed below with reference to FIG. 12, sensor unit 240 includes a processor 1202 coupled to one or more of a memory 1204, a sensor system 1206, an input / output (I / O) interface 1208, and a communication or network interface 1210.

[0081] Sensor system 1206 can incorporate one or more sensors for monitoring key parameters in a treatment system 100. These include sensors for monitoring environmental data parameters, as well as physical data parameters.

[0082] For instance, as shown in FIG. 3A, the sensor system 1206 can include one or more sensors for measuring environmental parameters, such as a temperature sensor 302, a 25 conductivity sensor 304, a pH sensor 306, a dissolved oxygen (DO) sensor 308. It may also include various sensors for monitoring physical parameters, including a pressure sensor 310, and a G-forces sensor 312 (e.g., acceleration sensor).WSLEGAL\097387\00003\42560596v2

[0083] In some cases, sensor system 1206 can further include one or more image sensors, such as cameras for capturing images or video (e.g., image frames). Sensor system 1206 can also include location-based sensors, as discussed previously.

[0084] In some embodiments, sensor measurements are used to detect conditions and 5 trigger responsive actions. For example, (i) low dissolved oxygen can trigger oxygenation or a reduction in crowding; (ii) elevated temperature may prompt slower handling or increased oxygenation; (iii) high acceleration or turbulence can indicate overcrowding and lead to flow or crowding adjustments; (iv) pressure or flow anomalies may reveal suction imbalance or obstruction; and (v) video analysis can detect abnormal fish swimming or clustering and.

[0085] The sensors, in sensor system 1206, may generate sensor data at any desired time or frequency interval, including in real time or near real time.

[0086] As best shown in FIG. 3C, one or more sensor assemblies 202 may be coupled together to form a sensor array 350.15

[0087] Sensor array 350 includes an array component 352 (e.g., a buoyancy element) and two or more sensor assemblies 202. They may be coupled together using a string, wire, or other suitable coupling medium. The array 350 may include features such as adjustable buoyancy, etc. Sensor assemblies 202 in the array 350 may be placed at various depths (or locations) in order to obtain measurements at the different depths (or locations), rather than at a single location.V. TREATMENT SYSTEM MONITRING MODULE

[0088] FIG. 11 shows an exemplary software architecture for the treatment system monitoring module 250 (FIG. 2). As shown, the module 250 may be stored or hosted in a 25 memory 1110, of server 204.

[0089] In at least one example, the monitoring module 250 uses a machine learning architecture for monitoring and data analysis of sensor data. This includes real-time or near real-time monitoring and data analysis.WSLEGAL\097387\00003\42560596v2

[0090] The monitoring and data analysis allows the system to use sensor units 240 to continuously monitor fish movement, water flow, and pressure changes. By analyzing this data, the machine learning detects early indicators of stress or inefficiencies that could escalate into larger problems. Accordingly, rather than only reacting to symptoms (e.g., fish 5 clustering in low-oxygen zones), pattern recognition and predictive insights of the machine learning model identifies the root causes, such as: poor water flow distribution, physical obstructions in pipes, and / or stress responses caused by improper net movement.

[0091] In some cases, once a root cause is identified, the system may automatically intervene to fix it. For example, the system may actuate a valve or other equipment component to: adjust water flow to optimize oxygen distribution, reduce net speed to minimize fish stress, and / or clear pressure irregularities in the pipeline.

[0092] By employing a continuous feedback loop, the machine learning model may also learn from its adjustments and continuously improves its ability to predict and resolve root causes. By learning over time, the system prevents recurring symptoms and enhances 15 operational efficiency.

[0093] In more detail, as exemplified in FIG. 11, the module 250 may include: (i) a data collection layer 1121, (ii) a data processing and communication layer 1122; (iii) a machine learning (ML) engine and processing layer 1123, (iv) a decision-making and control layer 1124, (v) a visualization and monitoring layer 1125, and (vi) a feedback and optimization layer 1126. Other layers may be added to the architecture.

[0094] (a.) Data Collection Laver

[0095] Data collection layer 1121 allows the system to continuously collect data from sensor assemblies 202. This includes data collected anytime fish are moved (not only during crowding), including general handling, transportation and other movements.25

[0096] (b.) Data Processing and Communication Laver

[0097] Data processing and communication layer 1122 may include edge computing and data integration. Edge computing allows for preprocessing of large volumes of sensor and video data close to the source to reduce latency. Data integration allows for the combinationWSLEGAL\097387\00003\42560596v2of acceleration data, camera inputs, environmental data, and other sensor data to build a comprehensive real-time picture of fish movement and conditions.

[0098] (c.) Machine Learning (ML) Engine and Processing Laver

[0099] Machine learning (ML) processor layer 1123 is configured for identifying and 5 fixing root causes of issues. The ML processing layer 1123 may include and implement a number of submodules, including: computer vision submodule, physical sensor data analysis submodule, environmental sensor data analysis submodule, a root cause analysis submodule (e.g., to determine contributing factors to detected stress indicator conditions), predictive analytics submodules, and a behavioral simulation submodule.

[0100] Computer vision submodule analyzes video feed of monitored fish, generated by a sensor assembly 202. For example, video feeds can be analyzed to identify fish density, stress signals, and abnormal movement patterns.

[0101] Physical sensor data analysis submodule may process movement and / or physical sensor data, generated by a sensor assembly 202. For example, this includes processing 15 acceleration data. The sensor data is processed to determine various movement anomalies in monitored fish, including: sudden movement spikes (stress reactions), and / or irregular slowdowns or erratic patterns indicating discomfort.

[0102] Environmental sensor data analysis submodule may monitor for various environmental data parameters, including pressure changes and water flow inside pipes to detect blockages, uneven flow, or stress zones.

[0103] Root cause analysis submodule is configured for detecting the underlying cause (i.e., root causes analysis) of a detected anomaly, based on the various analyzed video and / or sensor data. Examples of detected underlying causes include determining: pressure irregularities from obstructions in pipes, poor water flow leading to uneven oxygen 25 distribution, and excessive net speed causing stress. In some cases, based on the detected root cause, the system may then proactively intervene to fix the cause. As used herein, the root cause is also referenced herein as the "contributing factor" to an identified stress condition.WSLEGAL\097387\00003\42560596v2

[0104] Predictive analytics submodule may involve the use of ML models (e.g., LSTM, Deep Q-Leaming) to predict: future movement bottlenecks based on environmental conditions and fish responses, and / or optimal adjustments to minimize stress and improve efficiency.5

[0105] Behavioral simulation submodule may simulate fish responses to changes in net positioning, water flow, or handling practices to recommend and / or acuate best actions.

[0106] (d.) Decision-Making and Control Laver

[0107] In the decision-making and control layer 1124, the system acts on machine learning insights to address root causes, to determine suggested operational system adjustments. That is, layer 1124 can determine various adjustments to apply to the treatment control subsystem 206 (FIG. 2) to mitigate, address, or reduce the identified root causes (i.e., contributing factors).

[0108] For example, layer 1124 may determine net control (to automatically adjust net speed and positioning to minimize stress), water flow optimization (to ensure smooth, even 15 flow to avoid clustering and oxygen-deprived zones), and / or pipe management (to detect and resolve pressure anomalies or obstructions in real-time).

[0109] In some cases, control layer 1124 can generate suggested system adjustments, which are output to an operator. In other cases, it may generate control instructions transmitted to the control subsystem 206, to automatically effect system adjustments.

[0110] (e.) Visualization and Monitoring Laver

[0111] Visualization and monitoring layer 1125 generates various visual outputs based on collected data and analysis. For instance, this includes generating real-time dashboards (to provide clear visibility into fish movement, environmental conditions, and pipeline health), and / or alerts and insights (to highlight root causes and recommended actions rather 25 than symptoms).

[0112] For example, instead of simply alerting “Fish clustering detected,” the system may identify and report “Oxygen distribution imbalance detected due to water flow irregularity.” Alerts may also be generated for operators to flag persistent root causes that require manual interventions.WSLEGAL\097387\00003\42560596v2

[0113] In some cases, these outputs are produced on a display interface of user terminal 208.

[0114] (f.) Feedback and Optimization Laver

[0115] Feedback and optimization layer 1126 is used by the system to continuously learn 5 by integrating: (i) feedback from manual overrides, and (ii) data from sensor assemblies 202. Over time, the machine learning layer 1123 becomes better at predicting and solving root causes, reducing recurring stress events and inefficiencies.

[0116] It should be noted that traditional systems react to symptoms, such as: high fish mortality, sudden clustering behavior, and pipe blockages or irregular pressure readings.10 These symptoms are effects of an underlying issue. Reacting to them can reduce immediate risks, but does not solve the root problem.

[0117] By identifying and fixing root causes (not just symptoms), the system helps: (i) ensure smoother fish movements and better welfare, (ii) reduce mortality rates caused by stress or handling inefficiencies, (iii) improves operational efficiency, lowering long-term costs, and (iv) provides actionable insights to prevent issues before they escalate.VI. EXAMPLE METHOD(S)

[0118] The following is a description of various exemplary methods for operating the fish handling and monitoring system.

[0119] In some cases, the disclosed methods are stored as computer-executable 20 instructions on a memory (e.g., memory 1110 of server 204), and / or executed or executable by at least one processor (e.g., processor 1104 of server 204).

[0120] To that end, the disclosed methods may be implemented using various software routines forming part of the treatment system monitoring module 250 (e.g., software routines forming part of one or more submodules shown in FIG. 11).25

[0121] (i.) Method for Fish Handling and Monitoring

[0122] FIG. 4A shows an example method 400a for automated fish handling and monitoring. Method 400a may be performed in real time or near real time.WSLEGAL\097387\00003\42560596v2

[0123] At 402a, sensor data is obtained from one or more sensor assemblies 202 deployed within atreatment system 100. For example, the sensor assemblies 202 may be fixed or free floating within various tanks and pipes, described above. The sensor data may be obtained during high-stress events that are more likely to affect fish welfare (e.g., crowding, 5 treatment, and transfer events).

[0124] The data acquisition, at 402a, may occur in real time, near real time, or at predetermined intervals. In other cases, the data is obtained after the fact (e.g. , from memory or database 1112 in FIG. 11).

[0125] Sensor data can include one or more data streams (e.g., sensor data types). This includes one or both of measured environmental data (e.g., dissolved oxygen, temperature, pH, and conductivity), as well as measured physical data (e.g., pressure, acceleration, and G-force). Sensor data may further include image and / or video data, which can be used to provide visual monitoring of fish behavior and movement.

[0126] In at least one example, sensor data is time-stamped to ensure accurate 15 synchronization between the one or more data streams.

[0127] At 404a, the acquired sensor data may undergo a series of pre-processing steps. For example, image and / or video frames may be resized to standardized dimensions (e.g., 640x640 or 416x416 pixels). Further, pixel values may be normalized to facilitate consistent input for the model. Median filtering may also be applied to the environmental 20 and physical sensor readings to smooth out noise and enhance the reliability of the input data.

[0128] In some cases, the preprocessing involves time synching the environmental and / or physical sensor data streams to each other, as well with the corresponding image frames. Particularly, this enables association between environmental and physical conditions and imaged fish behavior.

[0129] At 406a, the sensor data is further processed to determine one or more derivative sensor features. Derivative sensor features may be computed from environmental and / or physical sensor measurements to capture, e.g., (i) dynamics such as rates of change of sensor values over time; and (ii) frequency-related data (e.g., using Fast Fourier 30 Transforms).WSLEGAL\097387\00003\42560596v2

[0130] For example, the determination at 406a may include determining pressure rate changes from measured pressure data, extracting vibration frequencies from acceleration or vibration signals, and computing dissolved oxygen rate of change from dissolved oxygen measurements. The rate of change may be computed relative to a previously obtained value 5 (e.g., the immediately preceding sample) or over another defined sampling interval.

[0131] At 408a, the acquired sensor data and / or derived sensor features are input into, and processed by, a trained machine learning model.

[0132] The model analyzes the data to generate outputs that include: (i) identifying any stress indicator conditions (or abnormal conditions) affecting or increasing risk of fish welfare; (ii) determining at least one contributing factor associated with (e.g., causing) the identified condition (e.g., a root cause); and (iii) identifying at least one system operational adjustment (e.g., to control subsystem 206) to mitigate the contributing factor.

[0133] In at least one example, the trained model evaluates the sensor data against one or more predefined stress indicator conditions, for which the model has been trained to 15 identify. Non-limiting examples of stress-indicator conditions may include:• Fish clustering in a pipe segment or tank zone• Abnormal fish swimming patterns (e.g., frantic movement, surface gasping)• Sudden spikes in G-force or vibration indicating mechanical shock• Rapid drops in dissolved oxygen (DO) or persistent low DO zones• Abrupt temperature changes beyond acceptable thresholds• Irregular pressure transients suggesting flow instability or blockage• Elevated turbidity indicating debris or treatment byproduct accumulation• Salinity or pH deviations outside configured ranges

[0134] Several example scenarios are provided below to demonstrate how the model 25 generates outputs by identifying (i) stress indicator conditions, (ii) associated contributing factors, and (iii) corresponding system adjustments:WSLEGAL\097387\00003\42560596v2• Example 1 - Uneven Oxygen Distribution Causing Fish Clustering in Pipeline:Based on time-synchronized sensor and video data collected during a fish transfer operation, the trained model can identify a welfare risk condition (i.e., stress condition) characterized by elevated fish clustering within a specific pipeline 5 segment. This may be evidenced by computer-vision fish density estimates from captured image frames.The model can further analyze time-synchronized sensor streams to attribute the clustering to contributing factors comprising, for example, uneven dissolved oxygen distribution caused by suboptimal flow in that segment. This attribution is supported 10 by, for instance: dissolved oxygen sensors reporting localized DO deficits relative to adjacent segments; flow-meter readings indicating reduced or uneven velocity profiles; and pressure sensors showing transients or gradients consistent with hydraulic imbalance or partial obstruction.In response, the system generates a recommended adjustment to optimize water flow, such as actuating pumps or adjusting valves of control subsystem 206 to increase circulation and balance oxygen levels.• Example 2 - Increased Fish Agitation Due to Excessive Pump Speed:The model analyzes acceleration data(e.g., tri-axial accelerometer-derived G-force) and pressure measurements acquired from the sensor assemblies 202. Based on 20 these streams, the model detects a stress incident condition characterized by a sequence of irregular peaks in G-force and concurrent pressure transients.The model further correlates these signals with time-synchronized video frames that evidence erratic swimming patterns and elevated fish density in the affected segment.25 Using these correlated inputs, the model attributes the condition to excessive pump speed, which is known to increase local water velocity and pressure fluctuations, thereby elevating mechanical stress on the fish. In response, the system may generate a recommended adjustment to reduce pump speed to reduce water velocity.WSLEGAL\097387\00003\42560596v2• Example 3 - Increased Fish Agitation Due to Temperature Variation (e.g., Drop or Increase)The trained model processes environmental sensor data together with time- synchronized video imagery during a routine delousing operation.5 From the video frames, the model detects a deviation in fish behavior, such as increased agitation and rapid swimming, indicative of elevated stress. Concurrent analysis of temperature sensor readings reveals a sudden variation (e.g., drop or increase) in water temperature that is temporally aligned with the observed behavioral change.Based on this correlation, the model identifies the underlying contributing factor as an acute temperature decrease. In response, the system may generate a recommended adjustment to stabilize thermal conditions, such as activating heating elements.

[0135] In some examples, the trained model applied at 408a is a system-specific model.15 That is, it is a machine learning model that has been specifically trained on training data generated within the given aquaculture treatment system 100 (or any portion of it). In other examples, the model is a generic model that is generally applicable to aquaculture treatment systems, or specific types of treatment systems (e.g., land-based v. sea-based).

[0136] At 410a, the system can generate one or more outputs. The outputs can include: (i) the detected stress indicator conditions; (ii) the at least one contributing factor causing the condition (e.g., root cause); and (ii) the at least one system adjustment to mitigate, reduce, or address the contributing factor.

[0137] The outputs can be generated in any suitable form. In some examples, the outputs are visual outputs generated on a display interface, such as that of user terminal 208. 25

[0138] Additionally, or alternatively, to act 410a: at 412a, the treatment control subsystem 206 (FIG. 2) is operated and adjusted, based on the output generated from act 410a.

[0139] For example, at 412a, the system can translate the determined system adjustments, into data control instructions transmitted to the various elements of control subsystem 206WSLEGAL\097387\00003\42560596v2(FIG. 2). For instance, the system can generate and transmit control set points for different controllable hardware elements, of the control subsystem 206.

[0140] By way of example, the system may actuate pumps or adjust valves to increase circulation and balance dissolved oxygen levels within a pipeline segment exhibiting fish 5 clustering. In other cases, the system may reduce pump speed to mitigate irregular G-force peaks and pressure transients correlated with erratic swimming behavior. The system may also activate heating elements or adjust inflow mixing ratios to stabilize water temperature following a detected thermal drop aligned with increased fish agitation.

[0141] More generally, control instructions may be transmitted for both land-based and sea-based systems. In land-based systems, the instructions can adjust crowding mechanisms and pump controls. In sea-based settings, the instructions can modulate automated winches to control net shape and density, and regulate pump speeds for suction into well boats or treatment infrastructure.

[0142] In some examples, at 412a, the system can: (i) determine the control adjustments 15 suggested by the model (e.g., increase water circulation); (ii) identify the specific controllable system hardware, in control subsystem 206, required to implement the adjustments. For example, the system may store a mapping (e.g., in memory) of the different controllable hardware and their intended functions, and may accordingly select the hardware able to implement the required changes; and (iii) generate and transmit control 20 instructions to the selected c hardware to effect the recommended operational changes.

[0143] The trained model can also generate an output, at 408a-410a, specifying the magnitude of the required operational adjustment. This includes quantitative setpoints or deltas for controllable elements (e.g., pump speed, valve position, oxygenation rate, or heating power), and / or timing parameters (e.g., ramp rates, duration) to ensure stable transitions. Alternatively or in addition, the system can apply a feedback-driven sensing and response loop to iteratively refine the adjustment based on subsequent sensor measurements.

[0144] In at least one example, the model is trained to generate one or more output scores at 410a, for any given stress-indicator condition. The output scores indicate how closely 30 current environmental and physical conditions, and / or observed fish behavior, align withWSLEGAL\097387\00003\42560596v2configured criteria for normal operating conditions. In some cases, the scores quantify deviation from a learned baseline, indicating how far current conditions diverge from nominal operating ranges for the treatment system.

[0145] Examples of scores can include: (i) a crowd readiness score, which quantifies how 5 closely current environmental, physical, and behavior indicators align with configured criteria for initiating or continuing crowding; (ii) a welfare risk score, reflecting the likelihood of compromised fish welfare under current conditions, (iii) a flow stability score, indicating hydraulic steadiness in transfer lines, and (iv) an oxygen distribution uniformity score representing spatial balance of dissolved oxygen.

[0146] In some cases, when a given score surpasses a predefined threshold, this may trigger the model to: identify the contributing factor and / or operate the control subsystem 206 to mitigate the issue. For instance, when the crowding readiness score exceeds a predetermined threshold, at 412a, the system triggers the crowding actuator via the controller.15

[0147] (ii.) Example Method for Operating Control Subsystem.

[0148] FIG. 4B shows an example method 400b for operating the control subsystem 206. In some examples, method 400b is performed during act 412a, of method 400a (FIG. 4A).

[0149] At 402b, the system can determine the location of the sensor assembly associated with the detected abnormal condition (e.g., out-of-threshold score for a given condition). This may be performed in several ways. In some implementations, a sensor assembly is fixed at a known position; the system can therefore reference a stored mapping of the treatment system that specifies the predefined locations of each sensor assembly. The mapping may be stored in memory, such as on a memory of server 204.

[0150] In other implementations, the system determines a relative location within 25 treatment system 100 from telemetry. For example, with reference to FIG. 1A, the system may infer that a sensor assembly is within conduits 104 between intake and outtake areas 102a and 102b based on characteristic physical signals. Elevated or transient pressure readings and increased G-force / acceleration signatures can indicate suction into, transit through, or egress from the conduits. By monitoring for threshold changes or patterns inWSLEGAL\097387\00003\42560596v2these physical measurements, the system can detect entry into and exit from specific regions and thereby localize the sensor assembly within the treatment system.

[0151] In still other cases, the sensor assemblies incorporate location sensors which generate location data. The location data is used to map the location of sensor assemblies, 5 such as by mapping the location data to the predefined referencing mapping

[0152] In cases where there are multiple sensor assemblies, it is possible that each sensor assembly transmits an identifier along with the sensor data (e.g., a unique ID). This allows the system to associate the sensor data with a given sensor assembly location.

[0153] At 404b, the system identifies one or more controllable components of control subsystem 206 that are proximate to the localized sensor assembly. In some implementations, the system references stored mapping data that specifies the positions of controllable components within the treatment system. Using the sensor assembly’s location determined at 402b, the system selects relevant components — such as pumps, valves, or oxygenation units — within a defined proximity. This selection enables the system to target 15 the specific controllable components most suitable to address the identified contributing factor and remediate the detected condition.

[0154] In some cases, once these components are identified, they may be generally output to a system operator to control (i.e., act 410a of method 400a)._Otherwise, at 406b, the system can operate these components as discussed previously with respect to act 412a (method 400a).

[0155] (iii.) Example Method for Training Model.

[0156] FIG. 4C shows an example method 400c for training a machine learning model for fish handling and monitoring.

[0157] At 402c, training sensor data is obtained from one or more sensor assemblies 202 25 deployed within a treatment system 100. Training sensor data can include one or more of environmental data, physical data, and image data (e.g., video data). Data acquisition may occur during both routine, as well as high-stress events such as crowding, transfer, and treatment.WSLEGAL\097387\00003\42560596v2

[0158] In some implementations, as mentioned previously, the training is applied to generate system-specific models. Accordingly, the training sensor data is obtained from the example system 100, for which the model is being trained for. In other cases, the training is applied to generate a generic model applicable to any treatment system, or specific types 5 of treatment systems (e.g., land v sea based). As such, the training data can again be acquired in the relevant environment.

[0159] At 404c, in some examples, reference data is concurrently collected during the operational event to serve as ground-truth for model training. This reference data may include, for instance, visual inspection footage, operator logs, documented mortality outcomes, and behavioral indicators of fish stress. More broadly, the reference data can include information about identified stress conditions and observed contributing factors, as well as potentially applied system changes (e.g., by an operator).

[0160] In some cases, the training sensor data and reference data are time-stamped to enable temporal alignment and cross-referencing between datasets.15

[0161] At 406c, the acquired sensor data may undergo pre-processing. As noted above, pre-processing steps may include resizing video frames to standardized dimensions (for example, 640x640 or 416x416 pixels), normalizing pixel values, and time-synchronizing sensor streams with corresponding video frames. Median filtering may also be applied to environmental sensor readings to reduce noise and enhance the reliability of the input data.20

[0162] At 408c, one or more derivative sensor features are determined from the pre- processed data. As also noted above, these can include determining various rates of changes for different sensor data types, as well as determining frequency values.

[0163] At 410c, a training dataset is generated, which may include one or more of the sensor data (e.g., preprocessed sensor data), derivative sensor features, as well as reference data. The training dataset may include time-synchronized sensor streams and corresponding video frames. The training data can also include the location of the various sensor subsystems at each training instance, which can be derived as discussed previously.

[0164] Data augmentation techniques can also be used to enhance the training dataset, and improve model robustness. These include random flips, rotations, noise injection, and 30 lighting variation applied to image frames.WSLEGAL\097387\00003\42560596v2

[0165] It is possible that, using the reference data - key observed events (e.g., initiation of crowding, optimal velocity for fish transfer, periods of overcrowding, and recovery phases), may be labeled and segmented within the sensor dataset to facilitate supervised learning. Further, image frames can be annotated with labels such as fish count, density, 5 behavioral stress cues, and region-of-interest boundaries to support object detection and activity recognition tasks.

[0166] In at least one example, the models are trained on training data collected across diverse operational conditions, including routine steady-state operation and periods of instability. Periods of instability can include various instabilities causing elevated stressindicating conditions, including instabilities in environmental or physical states (e.g. uneven oxygen distribution, excessive high or low temperatures, out of range pump / valve operating states). This allows training the model to determine a baseline state for the system, and to determine and predict deviations from baseline.

[0167] For training data that reflects unstable operation, the dataset may also include the 15 corresponding operational adjustments applied to rectify those states (e.g., commands issued to control subsystem 206 by an operator). Where multiple control subsystems are present, the dataset can specify which particular controllable elements were adjusted in each scenario. The dataset may further include a unique identifier for each controllable element and / or a mapping of their locations within the system. This enables the model to learn 20 associations between detected conditions and specific controllable elements.

[0168] As mentioned, the training process may enable the model to learn baseline operational states for each system 100 to generate system-specific models. These baseline states may be derived from routine operational training data, reflecting the unique characteristics of each system environment. The model uses this baseline to understand what constitutes nominal operation, which serves as a reference point for identifying deviations.

[0169] Building on this foundation, the model may be trained to generate scores that represent deviations from the learned baseline, and for different stress-indicating conditions. In at least one example, these scores provide normalized, bounded indicators 30 (e.g., values between 0 and 1) that summarize various aspects of system performance.WSLEGAL\097387\00003\42560596v2Examples of such scores include a crowding readiness score, a welfare risk score, a flow stability score, and an oxygen distribution uniformity score. The model learns to compute these scores by analyzing single or fused features from environmental, physical, and video training data streams acquired under both stable and / or unstable conditions. By learning 5 how to combine these multimodal inputs, the model can generate scores that accurately reflect the state of the system under different conditions.

[0170] In some implementations, the model may also leam thresholds that represent out- of-bound deviations from the baseline. These thresholds are not necessarily fixed or globally defined; instead, they may be learned on a per-system basis using z- 10 standardization. For a given sensor channel, or group of sensor channels (x), the model calculates baseline statistics (mean ( / z) and standard deviation (cr) during routine operation. Standardized values ((z)) are then derived using the formula (z = ^-^). The model learns to identify out-of-range conditions by recognizing when these standardized deviations exceed configurable z-band limits. This approach allows thresholds to dynamically adapt to the unique operational characteristics of each system, ensuring that deviations are assessed in the context of the specific environment.

[0171] As mentioned, the process of learning baseline normal operation, deviations, and out-of-bound thresholds can be applied on a per system basis (e.g., a system-specific model), on a per system type basis (e.g., land v sea based), or simply generically for 20 aquaculture treatment systems. At 412c, the machine learning model is trained using the training dataset.

[0172] Training may be implemented using supervised learning, where labels derived from reference data (e.g., crowding state, stress indicators, event onset / offset) are used to optimize a classification or regression objective. In some embodiments, semi-supervised learning is employed to leverage large volumes of unlabeled sensor streams alongside a smaller labeled subset, using techniques such as consistency regularization or pseudolabeling. Unsupervised or self-supervised pretraining may also be utilized to leam representations from unlabelled time-series and video data (e.g., contrastive learning across synchronized sensor / video modalities, masked signal modeling), followed by fine-tuning 30 on task-specific labels.WSLEGAL\097387\00003\42560596v2

[0173] Model optimization can include sequence models (e.g., LSTM / GRU, Temporal Convolutional Networks), transformer-based architectures for multimodal fusion, and convolutional backbones for image frames. Loss functions may include cross-entropy for discrete event detection, mean squared error for continuous severity or readiness scores, 5 focal loss for class imbalance, and multi-task losses to jointly optimize detection of abnormal conditions and estimation of causal parameters. Class imbalance can be addressed via re-sampling, cost-sensitive weighting, or hard-negative mining.

[0174] Validation may be performed using hold-out sets, k-fold cross-validation, or timebased splits to respect temporal dependencies. Hyperparameter tuning can be conducted via grid search, Bayesian optimization, or population-based training. Regularization may include weight decay, dropout, early stopping, and data augmentation across modalities (e.g., temporal jitter, noise injection for sensors, brightness / rotation for imagery). Deployment-oriented techniques such as model pruning, quantization, or knowledge distillation may be applied to meet edge-compute constraints while preserving accuracy.15

[0175] While the foregoing generally references training a single model, it is understood that the described functions may be implemented by multiple models that operate independently or in combination. Accordingly, any reference herein to training or applying “the model” is to be construed as encompassing one or more models configured to perform the recited functions.VII. EXAMPLE DASHBOARD GRAPHICAL USER INTERFACE (GUIs)

[0176] The system may involve software as a mobile application. In some embodiments, a data dashboard is implemented by the visualization and monitoring layer 1125 (FIG. 11)

[0177] FIGs. 5 A - 5D illustrate, in screenshots, an example of a data dashboard 500, 525, 25 550, 575, in accordance with some embodiments. These may be displayed on a display interface of a user terminal 208, for example.

[0178] Sensor readings (see FIG. 5A) may be sent from the sensor assemblies 202 to the data dashboard via wireless communication (e.g., Bluetooth, Wi-Fi, etc.). The data dashboard may be navigated to view data (see FIG. 5B) from each sensor assembly 202WSLEGAL\097387\00003\42560596v2currently communicatively connected or previously communicatively connected to the data dashboard.

[0179] Users may monitor sensor real-time and / or historical data readings from the data dashboard. In some embodiments, areas of the data dashboard may be zoomed in and back 5 out via use of a mouse to select an area or section of the data and / or an auto scale button feature.

[0180] Once sensor data has been logged, graphs (e.g., one or more data plots on a graph) may be generated (see FIG. 5C) and exported (see FIG. 5D), for example, to a .pdf file.VIII. EXAMPLE APPLICATIONS

[0181] The following are a number of non-limiting examples of use applications for disclosed embodiments.

[0182] Example 1 - Fish Handling and Treatment Operations

[0183] It has been reported that aquaculture salmon operations can experience significant 15 losses in fish biomass due to preventable mortality, resulting in substantial financial impacts for producers each year.

[0184] The disclosed monitoring and handling fish system may be used to enable fish farmers to receive immediate feedback and data on how to adjust their fish handling systems during mission critical processes.

[0185] For example, operations are digitized, allowing fish farmers to make better management decisions or allowing automated decisions to be made in real-time on the fish biomass. Management and fish health experts may use real-time and historical data to improve infrastructure, welfare and fish health practices on a daily basis. For example, sensors may measure dissolved oxygen, pH, temperature, acceleration / shock, conductivity 25 and pressure of the water in which the fish are in.

[0186] In some embodiments, the sensor assemblies 202 increase understanding of the equipment and internal transfers of the fish through pipes, pumps and tanks. The system may document the full transfer of the fish, wherever the fish is taken, linking all operationsWSLEGAL\097387\00003\42560596v2together. The system increases fish welfare during crowding operations and reducing risk of stress and mortality. The system documents and calibrates the delousing-system, before fish is sent through. The system documents physical impact and oxygen levels during operations and transfers, thereby reducing the chance of developing melanized focal 5 changes.

[0187] Example 2 - Return on Investment (ROD of Automated Monitoring

[0188] In one evaluated application, the total financial loss for an operation in one net pen for a sea-based system was estimated to be:21,750 fish * 3.5kg per fish * CAD $11.56 per kg = $880,00510 This figure only represents the economic loss from immediate mortality the day after crowding, and does not take into account additional losses related to stress and reduction in fish appetite and consumables (liquid oxygen).

[0189] FIG. 6 illustrates an example of a graph 600 monitoring fish welfare throughout the treatment operation. Each value represents a depth profile. If an event is detected and 15 an action taken, stress on the fish can be reduced (e.g., by 60%). Mortality can therefore be reduced. For example, if mortality rates are reduced by 40%, then total savings using the system can be calculated to be at least:21,750 fish * 3.5kg per fish * CAD $11.56 per kg * 40% = $352,002.

[0190] Example 3 - Operational Benefits of Sensor Assemblies in Aquaculture 20 Treatment Systems

[0191] Sensor assemblies, as described herein, are adaptable for use in a variety of aquaculture environments, including land-based facilities, well boats, harvesting operations, and integrated treatment equipment.

[0192] These sensor assemblies enable the identification of systemic issues during fish 25 transfers and facility operations, such as detecting faults in pumps or inefficiencies in equipment design. By providing actionable data, the sensor units facilitate immediate improvements in operational processes, resulting in measurable cost savings.WSLEGAL\097387\00003\42560596v2

[0193] In certain implementations, modifications informed by sensor unit data have led to significant reductions in operational expenses and enhanced overall system performance.

[0194] Example 4 - Generating Fish Health Index

[0195] The disclosed machine learning enabled analytics may be used with performance 5 of equipment and production methods in real-time. This allows for the generation of an industry-wide fish health index, i.e., allowing farmers to know which assets, technologies or handling methods affect the survivability and quality of their biomass.

[0196] Example 5 - Fish Clustering Root Cause Identification

[0197] Consider a case scenario where the symptom presented is that fish cluster in one area show signs of stress. The root cause may be uneven water flow or poor oxygen distribution. A traditional system may add more oxygen in response to the stress. However, disclosed examples may detect and fix the root cause by automatically optimizing water flow, thereby ensuring oxygen is evenly distributed, and preventing stress before it occurs.

[0198] Accordingly, by addressing the root cause, the machine learning reduces the need 15 for reactive interventions, improves fish welfare and survival rates, and optimizes operational efficiency and reduces long-term costs. This proactive problem-solving sets the present system apart, enabling sustainable and efficient fish management.

[0199] Example 6 - Optimizing Fish-Carrying Pipelines

[0200] A test of the sensor assemblies and system along a pipeline was performed to map and optimize the fish-carrying pipes leading into a harvest plant. The three-day testing involved using two sensor assemblies to collect data and physically optimize the pipeline based on the collected data. The aim was to improve the pipeline and reduce quality loss during harvesting.

[0201] During testing, the sensor assemblies uncovered a zip-tie case where the ties 25 surrounding the device had cut through the pump, causing probable shell loss and quality issues. A sharp edge was found on the inside of one pump, which was likely the cause of the issue.

[0202] Example 7 - Reducing Peak Acceleration Forces in Fish-Carrying PipelineWSLEGAL\097387\00003\42560596v2

[0203] The data collected from a test performed with the sensor assembly 202 was divided into two analyses based on collected acceleration data:• Max Peak Acceleration (G-forces); and• All acceleration data (G-forces) throughout the line (50 sec).5

[0204] FIGs. 7A and 7B illustrate results for Test 1 and Test 3. Test 1 found the temperature to be 10.0°C, dissolved oxygen to be 9.7mg / L and peak acceleration to be 1.00g. Test 3 found the temperature to be 14.8°C, dissolved oxygen to be 8.0mg / L and peak acceleration to be 1.02g.

[0205] The first analysis showed that after installing a pipeline with the same dimensions throughout and reducing the number of 90 degree bends, the Max Peak (G-forces) decreased by 30%. The average G-forces for Test 1, Test 2, and Test 3 were 22.25 G, 20.64 G, and 16.31 G, respectively. The result was 1 / 3 less total G-force after line correction.

[0206] Example 8 - Smolt Transfer

[0207] Data collected from sensor assembly 202 during smolt transfer operations enabled 15 the generation of a comprehensive historical record of the entire transfer process, including detailed monitoring of fish welfare.

[0208] The sensor assembly 202 provides coverage of operational blind spots within pipes and pumps, facilitating the detection of physical forces experienced by the fish, such as G-force peaks when passing through equipment like centrifugal pumps. The sensor data, including acceleration and environmental parameters, allows for precise identification of stress events and operational inefficiencies. By analyzing the collected data, operators are able to identify areas for process improvement and implement optimizations, resulting in enhanced fish welfare, reduced risk of quality loss, and improved overall operational efficiency.25

[0209] FIG. 8 illustrates, in a screen shot, an example of G-force readings through a smolt transfer. FIG. 9 illustrates, in a graph, an example of results for sensor unit data through a smolt transfer.

[0210] Example 9 - Calibration of Flushing SystemWSLEGAL\097387\00003\42560596v2

[0211] Sensor assemblies 202 and the associated system may be utilized to calibrate flushing systems employed in aquaculture operations. Flushing systems, which use water jets to move or treat fish, can be harmful if not properly calibrated, potentially causing excessive force or pressure that negatively impacts fish welfare. By measuring the force 5 and pressure of the water jets, the sensor assemblies 202 provide actionable data that enables precise calibration of the flushing system. This ensures that the system operates in a manner that is gentler on the fish, thereby reducing the risk of injury and improving overall operational outcomes.

[0212] FIG. 10A illustrates an example of results for a worst-case scenario for G-force readings through a delousing system. FIG. 10B illustrates an example of results for a calibrated G-force readings through the delousing system.IX. EXAMPLE MACHINE LEARNING MODEL

[0213] The following section provides an exemplary description of the training and 15 deployment of a machine learning model deployed at act 408a (FIG. 4A).

[0214] (i.) Machine Learning Engine

[0215] The machine learning engine utilized in the system comprises a custom computer vision model based on YOLOvl2 and DeepSort. These models are employed for object detection tasks, such as identifying fish presence, assessing density, and detecting escape attempts (e.g., from image data outputs of image sensors).

[0216] For feature extraction, the model leverages either a ResNet-50 or CSPDarknet53 backbone.

[0217] The system also implements multimodal data fusion, integrating video data with environmental sensor readings (e.g., dissolved oxygen, temperature, and turbidity) as well 25 as physical sensor inputs (e.g. accelerometer, pressure, and flow measurements). This comprehensive fusion of visual and sensor data enables robust and accurate analysis of fish behavior and environmental conditions.

[0218] (ii.) Training MethodologyWSLEGAL\097387\00003\42560596v2

[0219] The training methodology (method 400b in FIG. 4B) utilizes a training dataset composed of annotated video frames collected from multiple tanks and cages. These frames are labeled with key indicators such as fish count, stress signs (including erratic swimming) and crowding state.5

[0220] The pre-processing steps include resizing video frames to either 640x640 or 416x416 pixels and normalizing pixel values. Sensor streams are time-synchronized with the corresponding video frames to ensure accurate data alignment. Additionally, median filtering is applied to smooth out noisy environmental readings, enhancing the quality of the input data for subsequent analysis.

[0221] To improve the robustness and generalizability of the model, data augmentation techniques are applied to the training video frames, including random flips, rotations, noise injection, and lighting variation.

[0222] The training configuration for the model involved running between 100 and 300 epochs, depending on the convergence of the model. The batch size was set between 16 and 15 32, and the learning rate is established at 0.001 with cosine annealing applied to optimize training progression. The model utilizes either the AdamW or SGD optimizer to further enhance learning efficiency and performance.

[0223] For validation, 20% of the data was reserved to assess model performance. Additionally, early stopping and checkpointing techniques are employed during training to prevent overfitting and to ensure optimal model selection.

[0224] (iii.) Applying Trained Model

[0225] During model application (act 408a in FIG. 4A), sensor data streams were again subject to preprocessing. As noted above, the pre-processing steps include resizing video frames to either 640x640 or 416x416 pixels and normalizing pixel values. Physical and 25 environmental sensor streams are time-synchronized with the corresponding video frames to ensure accurate data alignment. Additionally, median filtering is applied to smooth out noisy environmental readings, enhancing the quality of the input data for subsequent analysis.WSLEGAL\097387\00003\42560596v2

[0226] The image frame and video streams were then fed into the trained model to generate outputs.

[0227] The post-processing steps involve applying non-max suppression to remove duplicate detections and fusing the model output with real-time environmental sensor 5 thresholds.

[0228] The deployment setup includes an edge computing device, such as an NVIDIA Jetson Xavier NX or AGX Orin, which is positioned dockside or on a barge. The trained model is deployed within a Docker container, allowing for an auto-updating pipeline. For remote data communications and override capability, the system utilizes MQTT or 10 LoRaWAN protocols.

[0229] (iv.) Performance Metrics of Trained Model

[0230] The following section provides an overview of the performance metrics used to evaluate the trained model:><< < <>WSLEGAL\097387\00003\42560596v2Table 1 - Evaluation of Trained ModelX. EXAMPLE CONFIGURATION FOR SERVER

[0231] FIG. 11 illustrates, in a schematic diagram, an example server 204 (FIG. 2), in 5 accordance with some embodiments.

[0232] The server 204 may include at least one processor 1104 coupled to a memory 1110 (transitory and / or non-transitory) storing machine executable instructions to configure the at least one processor 1104 to receive data (from e.g., sensor units 202 or sensor arrays 350). The at least one processor 1104 can receive a trained neural network and / or can train a neural network using a machine learning engine 1123. The at least one processor 1104 can also execute instructions in memory 1110 to implement aspects of processes described herein (e.g., methods 400a-400b).

[0233] The at least one processor 1104 can be, for example, any type of general-purpose microprocessor or microcontroller, a digital signal processing (DSP) processor, an 15 integrated circuit, a field programmable gate array (FPGA), a reconfigurable processor, or any combination thereof.

[0234] Memory 1110 can include database(s) 1112 and persistent storage 1114. Memory 1110 may include a suitable combination of any type of computer memory that is located either internally or externally such as, for example, random-access memory (RAM), readonly memory (ROM), compact disc read-only memory (CDROM), electro-optical memory, magneto-optical memory, erasable programmable read-only memory (EPROM), and electrically-erasable programmable read-only memory (EEPROM), Ferroelectric RAM (FRAM) or the like.

[0235] In some examples, memory stores computer executable instructions which are 25 executed by the at least one processor (e.g., methods 400a and 400b).

[0236] Databases 1112 may be configured to store information associated with or created by the server 204. Databases 1110 and / or persistent storage 1114 may be provided using various types of storage technologies, such as solid-state drives, hard disk drives, flashWSLEGAL\097387\00003\42560596v2memory, and may be stored in various formats, such as relational databases, non-relational databases, flat files, spreadsheets, extended markup files, etc.

[0237] The server 204 can also include an I / O Unit 1102 and a communication interface 1106, coupled to processor 1104.5

[0238] The I / O unit 1102 can enable the server 204 to interconnect with one or more input devices, such as a keyboard, mouse, camera, touch screen and a microphone, and / or with one or more output devices such as a display screen and a speaker.

[0239] Communication interface 1106 can enable the server 204 to communicate with other components (via network 250), to exchange data with other components, to access 10 and connect to network resources, to serve applications, and perform other computing applications by connecting to a network (or multiple networks) capable of carrying data including the Internet, Ethernet, plain old telephone service (POTS) line, public switch telephone network (PSTN), integrated services digital network (ISDN), digital subscriber line (DSL), coaxial cable, fiber optics, satellite, mobile, wireless (e.g. Wi-Fi, WiMAX), SS7 signaling network, fixed line, local area network, wide area network, and others, including any combination of these.XI. EXAMPLE HARDWARE CONFIGURATION FOR SENSOR ASSEMBLY20

[0240] FIG. 12 is a schematic diagram of a sensor unit 240 (FIG. 3 A). As depicted, the sensor unit 240 includes at least one processor 1202 coupled to a memory 1204, and one or more of a sensor system 1206, at least one I / O interface 1208, and at least one network or communication interface 1210.

[0241] Processor 1202 may be an Intel or AMD x86 or x64, PowerPC, ARM processor, 25 or the like. Memory 1204 may include a suitable combination of computer memory that is located either internally or externally such as, for example, random-access memory (RAM), read-only memory (ROM), compact disc read-only memory (CDROM). It may be a transitory or non-transitory memory.WSLEGAL\097387\00003\42560596v2

[0242] Sensor system 1206 can include multiple sensor types configured to capture environmental and physical parameters relevant to fish handling and monitoring.

[0243] Environmental sensors can include dissolved oxygen, temperature, pH, conductivity, and turbidity sensors. Physical sensors can include accelerometers, impact 5 (high g) sensors, pressure / depth sensors, and flow meters.

[0244] In some examples, sensor system 1206 can also include one or more cameras (RGB and / or IR) acquire image frames or videos.

[0245] Each I / O interface 1206 enables sensor assembly 240 to interconnect with one or more input devices, such as a keyboard, mouse, camera, touch screen and a microphone, or with one or more output devices such as a display screen and a speaker.

[0246] Each network interface 1210 enables sensor unit 240 to communicate with other components, to exchange data with other components, to access and connect to network resources, to serve applications, and perform other computing applications by connecting to a network (or multiple networks) capable of carrying data including the Internet, 15 Ethernet, plain old telephone service (POTS) line, public switch telephone network (PSTN), integrated services digital network (ISDN), digital subscriber line (DSL), coaxial cable, fiber optics, satellite, mobile, wireless (e.g. Wi-Fi, WiMAX), SS7 signaling network, fixed line, local area network, wide area network, and others.XII. INTERPRETATION

[0247] The foregoing discussion provides example embodiments of the inventive subject matter. Although each embodiment may represent a single combination of inventive elements, the inventive subject matter is considered to include all possible combinations of the disclosed elements. Thus, if one embodiment comprises elements A, B, and C, and a 25 second embodiment comprises elements B and D, then the inventive subject matter is also considered to include other remaining combinations of A, B, C, or D, even if not explicitly disclosed.

[0248] The embodiments of the devices, systems and methods described herein may be implemented in a combination of both hardware and software. These embodiments may beWSLEGAL\097387\00003\42560596v2implemented on programmable computers, each computer including at least one processor, a data storage system (including volatile memory or non-volatile memory or other data storage elements or a combination thereof), and at least one communication interface.

[0249] Program code is applied to input data to perform the functions described herein 5 and to generate output information. The output information is applied to one or more output devices. In some embodiments, the communication interface may be a network communication interface. In embodiments in which elements may be combined, the communication interface may be a software communication interface, such as those for inter-process communication. In still other embodiments, there may be a combination of communication interfaces implemented as hardware, software, and combination thereof.

[0250] Throughout the foregoing discussion, numerous references will be made regarding servers, services, interfaces, portals, platforms, or other systems formed from computing devices. It should be appreciated that the use of such terms is deemed to represent one or more computing devices having at least one processor configured to 15 execute software instructions stored on a computer readable tangible, non-transitory medium. For example, a server can include one or more computers operating as a web server, database server, or other type of computer server in a manner to fulfill described roles, responsibilities, or functions.

[0251] The technical solution of embodiments may be in the form of a software product.20 The software product may be stored in a non-volatile or non-transitory storage medium, which can be a compact disk read-only memory (CD-ROM), a USB flash disk, or a removable hard disk. The software product includes a number of instructions that enable a computer device (personal computer, server, or network device) to execute the methods provided by the embodiments.

[0252] The embodiments described herein are implemented by physical computer hardware, including computing devices, servers, receivers, transmitters, processors, memory, displays, and networks. The embodiments described herein provide useful physical machines and particularly configured computer hardware arrangements.

[0253] Although the embodiments have been described in detail, it should be understood 30 that various changes, substitutions and alterations can be made herein.WSLEGAL\097387\00003\42560596v2

[0254] Moreover, the scope of the present application is not intended to be limited to the particular embodiments of the process, machine, manufacture, composition of matter, means, methods and steps described in the specification.

[0255] As can be understood, the examples described above and illustrated are intended 5 to be exemplary only.WSLEGAL\097387\00003\42560596v2

Claims

WHAT IS CLAIMED IS:

1. A computer-implemented method for monitoring and handling fish in an aquaculture treatment system, the method comprising:receiving at least one sensor data stream from at least one sensor assembly deployed in the treatment system,wherein the at least one sensor data stream comprises one or more of measured environmental data, measured physical data, and image data; andapplying a trained machine learning model to the at least one sensor data stream to generate one or more outputs comprising: (i) an identified stress condition associated with an increased risk to fish welfare in the treatment system; (ii) at least one contributing factor to the condition; and (iii) an operational adjustment for a controllable device of the treatment system to mitigate the at least one contributing factor2. The method of claim 1, wherein the method is applied in real-time or near real-time.

3. The method of any one of claims 1 or 2, wherein the sensor data streams are generated during high stress events comprising crowding, delousing, or transfer of fish in the treatment system.

4. The method of any one of claims 1 to 3, further comprising time-synchronizing the sensor data streams, and applying the trained model to the time-synchronized streams.

5. The method of any one of claims 1 to 4, wherein,the environmental data measurements comprise one or more of dissolved oxygen, temperature, pH, conductivity, and turbidity measurements, andthe physical data measurements comprise one or more of pressure, acceleration, G- force, depth, and flow rate measurements.

6. The method of any one of claims 1 to 5, wherein the stress conditions comprise one or more of:fish clustering in a pipe segment or tank zone;WSLEGAL\097387\00003\42560596v2abnormal fish swimming paterns;spikes in G-force or vibration indicating mechanical shock;rapid drops in dissolved oxygen (DO) or persistent low DO zones; abrupt temperature changes beyond acceptable thresholds;irregular pressure transients suggesting flow instability or blockage; elevated turbidity indicating debris or treatment byproduct accumulation; and salinity or pH deviations outside configured ranges.

7. The method of any one of claims 1 to 6, further comprising generating control instructions, for a control subsystem of the treatment system, to implement the operational system adjustment,wherein the control instructions comprise quantitative setpoints and / or timing parameters for controllable elements of the control subsystem.

8. The method of claim 7, wherein the control instructions comprise setpoints for one or more of pump speed, valve position, oxygenation rate, heating power, ramp rates, and winch setings for adjusting net pen shape or density.

9. The method of any one of claims 1 to 8, wherein prior to applying the trained model, the method comprises:processing the sensor data streams to generate derivative sensor features; andapplying the trained model to the derivative sensor features.

10. The method of any one of claims 1 to 9, wherein the sensor assemblies are floating or fixed within the treatment system.

11. A system for monitoring and handling fish in an aquaculture treatment system, the system comprising:one or more sensor assemblies deployed in the aquaculture treatment system, each sensor assembly configured to generate sensor data streams comprising one or more of measured environmental data, measured physical data, and image data;at least one processor; anda memory storing computer-executable instructions, which when executed by the at least one processor, configure the at least one processor to implement the method comprising:receiving the sensor data streams; andapplying a trained machine learning model to the sensor data streams to generate one or more outputs comprising: (i) an identified stress condition associated with an increased risk to fish welfare in the treatment system; (ii) at least one contributing factor to the condition; and (iii) an operational adjustment to a controllable device of the treatment system to address the at least one contributing factor.

12. The system of claim 11, wherein the method is applied in real-time or near real-time.

13. The system of any one of claims 11 or 12, wherein the method further comprises timesynchronizing the sensor data streams, and applying the trained model to the time- synchronized streams.

14. The system of any one of claims 11 to 13, wherein,the environmental data measurements comprise one or more of dissolved oxygen, temperature, pH, conductivity, and turbidity measurements, andthe physical data measurements comprise one or more of pressure, acceleration, G- force, depth, and flow rate measurements.

15. The system of any one of claims 11 to 14, wherein the stress conditions comprise one or more of:fish clustering in a pipe segment or tank zone;abnormal fish swimming patterns;spikes in G-force or vibration indicating mechanical shock;rapid drops in dissolved oxygen (DO) or persistent low DO zones; abrupt temperature changes beyond acceptable thresholds;irregular pressure transients suggesting flow instability or blockage; elevated turbidity indicating debris or treatment byproduct accumulation; and salinity or pH deviations outside configured ranges.

16. The system of any one of claims 11 to 15, further comprising generating and transmitting control instructions, for a control subsystem of the treatment system, to implement the operational system adjustment,wherein the control instructions comprise quantitative setpoints and timing parameters for controllable elements of the control subsystem.

17. The system of claim 16, wherein the control instructions comprise setpoints for one or more of pump speed, valve position, oxygenation rate, heating power, ramp rates or durations, and winch settings for adjusting net pen shape or density.

18. The system of any one of claims 11 to 17, wherein prior to applying the trained model, the method comprising:processing the sensor data streams to generate derivative sensor data comprising determined rates of change for environmental and physical measurements; andapplying the trained model to the derivative sensor data.

19. The system of any one of claims 11 to 18, wherein the sensor assemblies are floating or fixed within the treatment system.

20. The system of any one of claims 11 to 19, wherein the sensor assemblies are arranged in a sensor array located at various depths within the treatment system.

Images