Device, method and system for detecting the behaviour of fish in order to improve the feeding process thereof

Abstract

The present invention proposes a device, method and system for detecting the behaviour of a group of fish in order to improve the feeding process thereof, based on passive acoustics of hydrodynamic noise produced by the school of fish and the use of artificial intelligence. Specifically, a method, device and smart system is proposed, which, by means of at least one acoustic sensor, collects the acoustic signal generated by the movement of the fish during feeding, processes the signal and implements models developed using artificial intelligence in order to determine the feeding status of the fish. By means of these models, the efficiency of the feeding process of the population being farmed is detected, enabling, among other advantages real-time decision-making in order to manage the feeding process and reduce the amount of wasted food, thereby optimising feed conversion rates.

Description

[0001] DEVICE, METHOD AND SYSTEM FOR DETECTING FISH BEHAVIOR TO IMPROVE THEIR FEEDING PROCESS

[0002] DESCRIPTION

[0003] TECHNICAL FIELD OF THE INVENTION

[0004] The present invention relates to the field of fish farming and more specifically, to an improved device, method and system for detecting the behavior of a group of fish with application in improving the fish feeding process, based on the capture (through passive acoustics) of the hydrodynamic noise generated by the fish in their aquatic environment and the processing of the same using machine learning.

[0005] BACKGROUND OF THE INVENTION

[0006] When farming fish for consumption, environmental and biological factors can cause stressful situations in fish, causing them to substantially modify their behavior. Recognizing these abnormal behavior patterns is essential to prevent them and promote optimal adaptation of the organism to the farming conditions, thus achieving the best possible production. One of these situations (although there may be others) in which fish can modify their behavior is during feeding.

[0007] Proper management of fish feed in aquaculture is essential for several reasons, such as:

[0008] Optimal Growth: Providing an adequate diet is important for healthy fish growth. Well-managed feeding ensures that fish receive the necessary nutrients in the correct amounts for their development. Resource Efficiency: Properly controlling feed supply helps optimize feed conversion, that is, the amount of feed needed to produce a given amount of fish flesh. Greater feed conversion efficiency contributes to a more sustainable use of feed resources.

[0009] Waste Prevention: Preventing overfeeding through proper feed supply control reduces feed waste, which can negatively affect water quality and the aquatic environment, and also helps prevent pollution and minimize environmental impact.

[0010] Product Quality: Fish diet influences the quality of meat and other fish products. Precise feed control contributes to high-quality end products.

[0011] Animal Health: Careful monitoring of feed helps prevent diet-related health problems, such as nutritional deficiencies or diseases associated with overfeeding, and contributes to maintaining the overall health of farmed fish.

[0012] Environmental Sustainability: Proper feed supply management helps minimize the release of nutrients into the aquatic environment that will not be consumed by fish, which promotes more sustainable and environmentally friendly aquaculture practices.

[0013] Cost Optimization: Proper feed management allows for efficient use of feed resources, which contributes to optimizing production costs and making aquaculture operations more profitable.

[0014] In short, proper management of the feeding process, and in particular, precise control of feed supply in aquaculture, is essential to achieving a balance between fish growth, product quality, animal health, and environmental sustainability. Therefore, numerous solutions for fish feeding management have been developed, but all of them present certain disadvantages or problems in their application, which make them less efficient and optimal than the solution proposed in the present invention.

[0015] Thus, there are various solutions based on passive acoustics (only listening) to the sounds emitted by the jaws of crustaceans, particularly shrimp, when they ingest food. In other words, they use acoustic sensors to monitor the noise emitted by farmed species while they eat (especially crustaceans like shrimp). However, these types of techniques can only be applied to species that emit characteristic sounds when eating and are not applicable to other species.

[0016] Other proposed solutions use active acoustics (sound transmission and reception) for food management, emitting sounds and then receiving and processing the echoes of these sounds as they bounce off the fish (this is especially used for salmon farming). However, this has the disadvantage that using active acoustics can interfere with fish behavior, and, in general, emitting noise in open-air fish farms can cause a potential environmental impact.

[0017] In other cases, cameras are used to visually monitor fish feeding (i.e., monitoring images when fish are eating or stopping, and / or when there are remains of the feed provided that the fish are not ingesting). However, this requires poor visibility in the water, the use of multiple cameras, and constant maintenance of their lenses to ensure the necessary image quality, which requires significant resource consumption.

[0018] In view of the challenges faced by the sector and the problems presented by existing solutions, there is a need to develop a device and method for controlling the feeding of aquatic species that do not emit sounds when eating (clicks), such as the one set forth in the present invention, which is environmentally friendly (does not generate noise emissions into the environment), which is more effective and precise, which is adaptable to almost all species and aquatic environments regardless of visibility conditions, and which allows the data generated by monitoring to be managed in real time. SUMMARY OF THE INVENTION

[0019] The present invention proposes a device, method, and system for detecting the behavior of a group of fish, with application in improving the fish feeding process, based on passive acoustics, by recording and subsequently processing the hydrodynamic noise produced by the school of fish (group of fish) and the use of artificial intelligence, and more specifically, machine learning. Hydrodynamic noise refers to the sound generated by the movement of fish and / or the interaction of water with them.

[0020] The present invention proposes an intelligent method, device, and system for detecting (and monitoring) the behavior of farmed fish. This system, using at least one acoustic sensor, collects the acoustic signal generated by the movement of the fish during (and before and after) feed delivery, processes it, and implements developed artificial intelligence models to determine whether or not the fish are eating the supplied food (e.g., feed). These models detect the feeding efficiency of the farmed population, allowing a high degree of accuracy (around 80-90%) to estimate whether or not the fish are eating.

[0021] Some of the main functionalities of this technological solution are:

[0022] • Provide the producer with instant information on the likelihood of fish ingesting the feed supplied, for example, by sending light alerts conditioned on the established feeding efficiency level or threshold.

[0023] • Collection, processing, and analysis of data related to feeding strategy and fish behavior.

[0024] • Possibility of automatic connectivity with the feeder based on this information and the producer's parameters.

[0025] These functionalities allow, for example: • To assist in real-time decision-making for managing the feeding process.

[0026] • Reduce the amount of wasted feed and optimize feed conversion rates.

[0027] • Reduce the environmental impact of spills and waste from leftover food supplied and not consumed.

[0028] • Provide valuable information for the design of feeding strategies based on the thresholds established by the producer.

[0029] • Improve species knowledge by rapidly identifying low intake activity events in which fish are not eating efficiently.

[0030] • Automation of the feeding process by connecting the system to fish feeders.

[0031] In the face of the problems presented by existing devices and systems, some of the advantages of the present invention are listed below, together with some of the technical characteristics thereof that give rise to said advantages:

[0032] Passive (not active) acoustic detection of the characteristic sound associated with the hydrodynamic behavior of the entire school of fish. A distinguishing feature of the present invention is that it monitors the behavior of the fish to determine whether or not they are hungry, by listening to the hydrodynamic noise of the school of fish. Unlike other technologies that employ active acoustics (i.e., which emit sounds and subsequently receive and process the reflections of this sound in the fish), the present invention is innovative in that it does not employ any sound-emitting device, but rather solely records the sound using at least one directional hydrophone submerged in the water. Thus, the hydrodynamic noise generated by the agitation of the school of fish during feeding is recorded and analyzed to detect the moments when the fish are eating and not eating.Furthermore, the use of passive acoustic systems has a direct impact on the device's use in the aquatic environment, as it is a non-invasive device.

[0033] Detection of fish feeding responses in soundless fish species. One of the main distinctive advantages of the present invention lies in its ability to detect feeding patterns from the hydrodynamic noise generated by fish in their aquatic environment. This allows, unlike conventional sensors that focus on specific sounds emitted by the jaws of crustaceans when they ingest food (shrimp and other specific crustacean species farmed for consumption), to be based on the interpretation of the broad spectrum of noise generated by the modification of the hydrodynamic behavior of fish during feeding.Thus, another important difference of the present invention lies in its ability to address a wide diversity of species, clearly distinguishing it from others that focus exclusively on the detection of sounds emitted by shrimp or other crustaceans, and are therefore only useful for a specific type of species. While many existing solutions may overlook the diversity in aquatic environments by limiting themselves to specific species, here we present a comprehensive solution that covers a diverse range of fish species and behaviors, as it can be valid for any type of aquatic animal that modifies its behavior when faced with a food supply.

[0034] Additionally, detected information regarding the fish's response to feed can be displayed using a light signal emitted by a color-changing LED sensor (e.g., red = stop feeding, green = continue feeding, amber = transition). In the present invention, a light signal can be incorporated as a distinctive and efficient means of immediately communicating the feeding status of the fish. This feature is especially valuable in offshore aquaculture environments (far from the coast or offshore), where direct observation is difficult. The light signal facilitates a rapid response and facilitates efficient management of feed administration, optimizing fish health and contributing to sustainable aquaculture practices.

[0035] It is feeder-independent. The present invention uses a technology that allows its deployment in aquaculture nurseries independently of the particular feeding process, technology, or technique used. Thus, a significant advantage is its operation's independence from the feeding system employed. To this end, the present invention can incorporate various communication technologies, adaptable according to proximity to the control point (both in the offshore and onshore installation). This provides an interoperable technical solution independent of the feeder technology provider and the feeding process followed in the fish farm, allowing for connectivity with multiple control or feeding platforms.

[0036] Data processing. Since a passive acoustic system is used, it operates over a wide frequency range (in one exemplary embodiment, between 10 Hz and 10 kHz), enabling it to record varying degrees of movement, agitation, and behavior of the school of fish, and even other sounds not originating from the fish tank itself (nearby boats, feed hopper, etc.). Since it works with a wider frequency range than active acoustic devices (those that emit sound in a limited range of specific frequencies and collect their echoes) and existing passive acoustic devices (which record the characteristic sounds emitted by the jaws of crustaceans), data processing is multispectral and multivariate.To distinguish between different sounds, a series of adaptive filters can be applied in different frequency bands within the working spectrum, thereby obtaining specific characteristics for each of the sounds recorded during feeding. These characteristics are then introduced into the intelligent multivariate models used to implement feeding prediction models.

[0037] The machine learning models are trained on a database of the school's soundscape, which the proposed system itself helps construct. Currently, there is no database, documentation, or scientific literature on the sound emitted by the hydrodynamic behavior of schools of fish. A unique feature of this invention is the methodology used to train it, as the proposed system itself collects sound data using passive acoustics and preprocesses the audio recordings so that, at a later stage, the database that will be used to train the algorithm can be constructed.To develop this invention, various sound recording campaigns were conducted at different aquaculture farms (in Spain, Greece, and Norway) with different species (seabream, corbina, sea bass, and salmon) with the aim of creating a comprehensive database with which to validate the intelligent models in each of these scenarios. Thus, a differentiating detail of this invention is that it not only applies artificial intelligence techniques, including machine learning, but also implements and validates them based on its own unique repository of sounds recorded during the feeding process. In this way, it starts from a neural network trained and validated in these feeding scenarios and species that adapts to each of the particular conditions of the hatcheries.

[0038] Control via smartphones, tablets, PCs, and many other types of electronic devices (Multiplatform). The present invention can provide real-time transmission of alarms and information across various multiplatform devices, facilitating remote monitoring and management in aquaculture environments and ensuring universal data accessibility. Thus, it is compatible with smartphones, tablets, computers, and generally any type of electronic device capable of communication.

[0039] Furthermore, integration with modern communication technologies enhances the use of predictive analytics and advanced management strategies, thereby improving efficiency and sustainability in aquaculture.

[0040] In a first aspect, the present invention proposes a device for detecting the behavior of a group of fish with application in improving the fish feeding process, in which the device comprises:

[0041] - At least one acoustic sensor configured to capture in real time the hydrodynamic noise of the group of fish during a certain time interval (for example, during the process of supplying food to the fish, optionally also including a period before and / or after the supply of food); - At least one electronic processor configured to: a) Process the hydrodynamic noise data captured by the at least one acoustic sensor; b) Determine the agitation state of the group of fish in said certain time interval, applying a previously trained machine learning model to the captured hydrodynamic noise data once processed;c) Indicate the feeding status of the group of fish corresponding to the determined agitation status by means of a light indicator located on the device itself (for example, by means of lights of different colors) and / or send information (for example about the determined agitation status or about the appetite status of the fish corresponding to the determined agitation status) to one or more external devices by means of a wired or wireless communication network.;

[0042] Where the processing of step a) may comprise: a1) Filtering the captured noise data with a broadband filter; a2) Filtering the data resulting from step a1) with a series of N band-pass filters centered on different frequencies, obtaining N groups of filtered noise data, each one coming from a different band-pass filter (where N is a design parameter); a3) Obtaining the sound pressure values ​​in the time domain corresponding to each of the N groups of filtered noise data;

[0043] The at least one processor may be further configured to communicate (send a message) with a feeding device of the group of fish via the wired or wireless communication network, to control (stop or resume) the supply of feed by the feeding device to the group of fish, based on the determined agitation state.

[0044] In one embodiment the frequency band of the broadband filter is between 20 Hz and 20 kHz.

[0045] In one embodiment the bandpass filters (which may be adaptive) are centered on third octave bands and may be of second order infinite impulse response.

[0046] The machine learning model may be of the Random Forest type. In one embodiment, hydrodynamic noise data collected by the at least one acoustic sensor (and processed as in step a)) corresponding to previous time intervals (a certain time interval) and / or hydrodynamic noise data previously recorded during the feeding process of other groups of fish stored in a database are used to train the machine learning model.

[0047] The external devices with which the device communicates can be user terminals of the operators or administrators of the aquaculture infrastructure (farm) where the device is located.

[0048] For communication with the one or more external devices, the device may have a communication module that may be enabled to connect to a wired or wireless communication network with the one or more external devices.

[0049] In a second aspect, the present invention proposes a method for detecting the behavior of a group of fish with application in improving the feeding process of the group of fish, wherein the method comprises the following steps performed by at least one electronic processor of an electronic device, where the method comprises: a) Processing hydrodynamic noise data of the group of fish captured by at least one acoustic sensor of the device. Wherein said processing may comprise: a1) Filtering the captured noise data with a broadband filter of a predetermined bandwidth; a2) Filtering the data resulting from step a1) with a series of N band-pass filters centered on different frequencies, obtaining N groups of filtered noise data, each one originating from a different band-pass filter; a3) Obtaining the sound pressure values ​​in the time domain corresponding to each one of the N groups of filtered noise data;b) Determine the agitation state of the group of fish by applying a previously trained machine learning model to the captured hydrodynamic noise data once processed (in the case that steps a1-a3 are carried out, it would be applied to the N sound pressure values ​​obtained); c) Indicate the feeding state of the group of fish corresponding to the determined agitation state by means of a light indicator located on the device and / or send information (for example about the determined agitation state or about the appetite state of the fish corresponding to the determined agitation state) to one or more external devices by means of a wireless or wired communication network.;

[0050] In a third aspect, the present invention proposes a system for detecting the behavior of a group of fish with application in improving their feeding process, comprising:

[0051] - At least one device for detecting behavior as described above;

[0052] - A gateway device configured to:

[0053] Receiving, from the at least one device for detecting behavior, information about the determined agitation state and / or the appetite state of the fish corresponding to the agitation state determined by the at least one device and / or information about the hydrodynamic noise of the group of fish;

[0054] Communicate with one or more devices external to the system (such as user terminals of operators or administrators of the aquaculture infrastructure where the device is located) to, among other things, send information received from the at least one device;

[0055] Remotely accessing at least one device to modify its operating parameters and to control its on / off operation (e.g., following commands received from user terminals with which it communicates); where the communication between the gateway device and the at least one device is carried out via a wired or wireless communication network that may be the same or different network than the one used by the gateway device to communicate with external devices.

[0056] Finally, a computer program is presented comprising executable instructions for implementing the described method, when executed on a computer, a digital signal processor, an application-specific integrated circuit, a microprocessor, a microcontroller, or any other form of programmable hardware. These instructions may be stored on a digital data storage medium.

[0057] For a more complete understanding of these and other aspects of the invention, its objects and advantages, reference may be made to the following specification and the accompanying figures.

[0058] DESCRIPTION OF THE FIGURES

[0059] To complement the description being made and in order to help better understand the characteristics of the invention, in accordance with some preferred examples of practical embodiments thereof, a set of drawings is attached as an integral part of this description, in which, for illustrative and non-limiting purposes, the following has been represented:

[0060] Figure 1 shows schematically the main body of the box device or “box” of the system proposed according to an embodiment of the present invention.

[0061] Figure 2 shows schematically a hydrophone of the box or “box” of the system proposed according to an embodiment of the present invention.

[0062] Figure 3 shows schematically an aquaculture cage where the proposed system is installed according to an embodiment of the invention.

[0063] DETAILED EXPLANATION OF AN EMBODIMENT OF THE INVENTION

[0064] The present invention proposes a device, method and system for detecting the behavior of a group of fish, especially in aquaculture farms, with application in improving the fish feeding process, by accurately estimating when a group of fish (school) is feeding (and when not). To this end, the proposed solution monitors the hydrodynamic noise produced by the school of fish (during the feeding process) and uses an appropriately trained machine learning model to determine whether or not the collected noise corresponds to a significant difference in the fish's agitation situation (state). To collect said noise from the movement of the fish (or more generally speaking of aquatic animals in general) and their interaction with the water (hydrodynamic noise), one (or more than one) acoustic sensor, preferably a hydrophone, is used.A hydrophone is a sound-to-electricity transducer (microphone) designed to be used immersed in water or another liquid, i.e., it picks up sound propagating through the water (or other liquid). Hydrophones can be directional or omnidirectional and typically operate over a wide frequency range (in one embodiment, from 10 Hz to 10 kHz; this is just a limiting example; they can operate in any other sound frequency range). Any hydrophone (durable enough to withstand the harsh operating environment) can be used, for example, an analog one. In one embodiment, for greater installation convenience, the hydrophone can feature a "plug-and-play" system.

[0065] In one embodiment, it can be said that, generally speaking, the proposed system may consist of the following differentiated parts (also called devices or units): The device called “Box” (box in Spanish) and, optionally, the device called “Gateway” (gateway in Spanish).

[0066] - The “Box” device (1) is the part that performs the data acquisition and, normally, also the data processing, to determine the fish’s agitation status and thus estimate whether they are feeding or not (in other words, it estimates the fish’s appetite). It is usually installed directly above the enclosure containing the group of fish whose feeding is to be monitored; for example, it can be installed in the infrastructure of aquaculture cages.

[0067] It consists of a main body (which is the box itself, shown schematically in figure 1) that is connected (usually by a cable) to one or more hydrophones (shown schematically in figure 2).

[0068] In one embodiment, it incorporates a light indicator (11) that indicates, in real time, whether the fish are eating (corresponding to a specific degree of agitation of the school of fish, high) or if they have stopped eating (corresponding to a different degree of agitation of the school of fish, low). To do this, it can emit a green light or a red light respectively (this is just an example and any other color combination can be used). In one embodiment, it also indicates if there is a change (transition) in the feeding rhythm (for example, by means of an amber light). The light indicator can be an LED type lighting equipment or any other type.

[0069] In one embodiment this indication of the power status can be made by means of an audible indicator.

[0070] The device is usually completely autonomous and contains a power supply battery. In one embodiment, said battery is recharged by a solar or photovoltaic panel (12). In addition, for fixing it to the structure of the enclosure where the fish are kept, the box may have anchoring or fixing means on one of its sides, for example, anchoring or fixing to the railing of the aquaculture cage.

[0071] In one embodiment, the box acts directly on the automatic feeders to control the supply of food to the fish, such that the feeding level is controlled based on the feeding status estimated by the device. That is, it determines the agitation status of the fish, from which it determines whether the fish are eating or not, and based at least on that determination, it acts on the fish feeders. For example, if it determines that the fish have stopped eating, it acts on the feeders to close them and thus not provide more food to the fish, avoiding food waste and, therefore, saving resources. To do this, the box must communicate with one or more feeders either via a cable or via wireless communication (WiFi or any other type).

[0072] In order to better capture the noise of fish movement, in one embodiment the microphone is not located together with the main body of the box but is closer to the school of fish, for example submerged inside the cage. The microphone is connected to the main body, normally by a cable (21). In the embodiment shown in Figures 1 and 2, it is assumed that there is only one hydrophone, but to better capture the noise there may be more than one hydrophone (especially if the cage is large), each connected to the main body.

[0073] The cable may incorporate a weight (22) so that the hydrophone can be submerged more easily in the cage. And it also usually has a protective casing (23) to avoid its rapid deterioration and protect it from various attacks. The acoustic data collected by the hydrophone(s) are stored and processed to determine the feeding status of the fish. This determination is normally carried out in one or more electronic processors of the same “Box” device, although in one embodiment, the device can send the data to an external device so that the determination can be made there.

[0074] In one embodiment, the acoustic data collected is stored in single-channel audio files. These files may be 2 seconds in length and have a sampling frequency of 48,000 samples per second. This means that the data is stored on the device in 2-second audio files, each containing approximately 96,000 samples, with virtually no data loss between samples (this is only a non-limiting example, and in other embodiments, the files may be shorter or longer in length and use a higher or lower sampling frequency). Each of these audio files is immediately processed by the device to determine the fish's behavior.

[0075] Audio captured during a specific time interval (2 seconds in the example in the previous paragraph) is processed before being used as input to the machine learning model. This audio processing is performed in different phases to extract relevant features for the subsequent implementation of artificial intelligence techniques. In one exemplary embodiment:

[0076] First, data preprocessing or cleaning is performed using a broadband filter, for example between 20 Hz (Hertz) and 20 kHz (Kilohertz), to narrow the working frequency range. Optionally, an anomalous value detector can be applied to the resulting sound pressure level, based on statistics from values ​​collected in other similar installations or values ​​that have been collected in the system itself since its commissioning. These recorded anomalous values ​​are discarded for this purpose but can be used for other operations such as monitoring the installations.

[0077] Secondly, a series of N (adaptive) filters can be applied in different frequency bands within the working spectrum, thereby obtaining specific characteristics of each of the sounds recorded during the feeding. Thus, in one embodiment, a series of band-pass filters, which can be adaptive, are applied, centered on the standardized third-octave bands to divide the frequency spectrum into narrower specific frequency bands (in the previous case, where data is collected in the band between 20 Hz and 20 kHz, a total of 31 frequency bands will be obtained, i.e., N=31). These filters can be infinite impulse response (IIR) filters, for example, of second order, so that they are capable of quickly adapting to the different acoustic signal shapes recorded.

[0078] Thirdly, for each of these frequency bands, the sound pressure levels obtained in the time domain are obtained using the usual expression 20 log10(p_rms / p_ref), where p_rms is the mean square value of the pressure of each third octave band, and p_ref the reference pressure value for measurements in water (1 pPa). With this, there are (31) sound pressure levels for each of the audios being recorded. Artificial intelligence techniques will be applied to these sound pressure levels to determine, in real time, the feeding state (or more specifically, the agitation state) of the group of fish.

[0079] In one embodiment, the Box sends the collected data to an external device (e.g., the Gateway, an external server, or user terminals) for storage, analysis, or post-processing. This sent data may include the fish's appetite status (eating or not eating) determined (estimated) by the device at any given time, and may also include other types of data, such as the sound directly recorded by the hydrophone at any given time.

[0080] - The "Gateway" device can be located onboard the vessel or on a feeding platform (or any other arrangement). It receives information from the Box and can also be accessed remotely to modify relevant Box parameters or to control its on / off.

[0081] Although, for the sake of simplicity, we refer to a box and a gateway, a single gateway can receive and control more than one box; for example, if the aquaculture complex consists of several cages or cage infrastructures, each one may have a box, all reporting to the same gateway (or several of them). In one embodiment, the gateway creates a WiFi network to which the boxes automatically connect, allowing data to be received and stored locally on the gateway or directly in the cloud. The data received from the box devices can be sent by the gateway to user terminals of aquaculture system operators or other interested users.Through this Wi-Fi network, each box can be accessed to control its on / off operation (or to modify relevant box parameters), either directly through commands entered by a user directly into the Gateway (using the corresponding interface) or through a user terminal belonging to the operator or operators controlling the system. User terminals can be mobile phones, tablets, PCs, or any electronic device that can be connected to a communications network.

[0082] In one embodiment, communication between the Box and the Gateway or with the user terminals can be done through mobile telephone networks (2G, 3G, 4G, 5G or any other type) or generally speaking through any other known communication network.

[0083] In one embodiment, communication between the "Box" device and external devices is performed directly from the "Box" (without going through a Gateway) via a wireless communications network of any type. In one embodiment, this is done through a WiFi connection that the device itself automatically deploys when it has power. Thus, from any user terminal (such as a mobile phone), an authorized user can connect to said WiFi network and view, in real time, the agitation status of the school supplied by the device, as well as its history (if stored on the device). Communication between the device and the outside world (whether an external user terminal or the Gateway) is performed through a communication module located on the device itself.

[0084] The deployment and start-up of the proposed system involves a series of initial phases, which are described below according to an embodiment of the present invention:

[0085] 1. Installation Phase: As indicated above, the proposed device (box) is installed in (or near) the enclosure or infrastructure that houses the group of fish whose feeding is to be controlled; for example, it can be installed in the infrastructure of the aquaculture cages where the fish are kept. In one embodiment, there will be one box per cage, but multiple cages (each with its own acoustic sensor connected to the box) can also be controlled using the same box.

[0086] Figure 3 shows, as an example, an aquaculture cage (31) in which the proposed device is installed.

[0087] The installation must consider not only the arrangement of the device's box but also that of the submerged acoustic sensor(s). In an exemplary embodiment, each box will be placed on top of the cage stanchion attached to the cage railing for system stability. A cable will extend from the box, at the end of which will be the corresponding acoustic sensor.

[0088] In the example in Figure 3, the device (box, 32) is located on one of the railings of the cage in question and the acoustic sensor (33) is introduced into the cage (at a depth of, for example, between 5 and 10 meters, this depth will depend on the specific species). For greater security of the installation and to prevent the acoustic sensor (hydrophone) from being moved by currents or by the fish themselves, deteriorating the quality of the collected sound, it is usually fixed to some part of the cage infrastructure. Thus, for example, it can be introduced tied to the net cloth or it can also be hung from the turret (34) of the anti-bird net (which many aquaculture cages usually have) towards the bottom of the cage.

[0089] If the device is powered by a solar panel, it should preferably be placed in the part of the cage where the solar panel receives direct sunlight. This will optimize the charging and operation of the internal battery and maximize the device's energy efficiency.

[0090] Regarding the installation of the acoustic sensor, it should preferably be placed at the optimal point (or one of the optimal points) for receiving noise generated by the fish in the cage. To achieve this, tests can be conducted during installation to verify and / or validate the point in the cage where fish-generated noise reception is best (or at least sufficiently good), and this will be the sensor's placement point.

[0091] 2. Training phase:

[0092] The application of artificial intelligence techniques (in this case, machine learning) involves two clearly differentiated main stages. First, the device requires training and validation time in the context under study. In this stage, a training dataset is used to fine-tune and optimize the machine learning model for the specific situation / context in which it will be used. After the model has been trained and evaluated, establishing the best learning model, it is deployed on the device in production mode, where it can be used to make predictions on new, unlabeled acoustic data. In this second stage, the model is used to make inferences based on the input data provided; in this case, the model outputs the agitation state (which corresponds to the feeding state) that has classified each of the audios it records.

[0093] In other words, before a machine learning model can be used to assess whether a given hydrodynamic fish noise corresponds to a given agitation level from which an estimate of fish feeding status can be derived, such a model must be trained to define which features of the collected noise indicate a high or low agitation situation or, in other words, it must be built on a database of the school's soundscape that corresponds to different fish agitation situations.

[0094] This training phase must be long enough to correctly label the different agitation stages. In other words, to correctly establish a correspondence between the hydrodynamic noise of the group of fish and the agitation state corresponding to that noise. In the production context under study, the training and validation phase is estimated to last approximately 30 days (this is only an estimate, and depending on the specific case, the duration may be longer or shorter).During this training period, the device records the audios at each of the fish's feeding times, the data is processed and filtered, and an operations center (a center where system experts are located, for example, the Naval Technology Center) or the system administrator is in charge of evaluating, based on this data, which is the artificial intelligence model (or more specifically, the value of the parameters of the artificial intelligence model are determined) that is best able to classify the events of interest (high agitation, low agitation, change in agitation that would correspond to the states of eating, not eating, and change of state).

[0095] In the specific implementation explained above, the audios would last 2 seconds and would be filtered in 31 frequency bands, so there would be 31 sound characteristics in each 2-second audio (this is just an example, and in other implementations the duration of the audio and the number of bands could be different).

[0096] This task is a classification task, as it determines which class (high agitation, low agitation, change in agitation, which corresponds to a specific feeding state: eating, not eating, and transition, respectively) the recorded noises belong to (based on the training that has been performed). Therefore, a machine learning classification model can be used for this task.

[0097] In one embodiment, machine learning algorithms (models) based on Random Forest techniques are used for classification, preferably composed of 9 trees with their corresponding branches (this is just an example, and in other embodiments, another number of trees may be used). It has been proven that Random Forest models are apparently the best supervised learning models when it comes to classifying these events based on the characteristics of the audios, since these models, compared to others, seem to entail a reduction in overfitting, greater robustness against outliers or noise, versatility, scalability, and good performance in this particular type of applications. This is a non-limiting example, and in other embodiments, other machine learning methods such as support vector machines (SVMs) or any other may be used.

[0098] This modification of the model parameters can not only be carried out prior to the use of the proposed device (training phase) but, during its use, the models can continue to be updated (for example, if it is found that there has been a false positive, this can be used to update and improve the model used).

[0099] In one embodiment, the classification of the data in this training phase is performed by one or more of the following options (presented as non-limiting examples and in other embodiments other techniques or a combination thereof may be used):

[0100] Passive training: The collected acoustic data (once processed) is sent to an operations center and / or feed platform in real time (this can be done directly from the device itself or through the Gateway). If cameras (surface and / or underwater) are present, these images are also sent to the operations center and / or feed platform in real time. This data will be analyzed at the operations center (e.g., the Naval Technology Center) to facilitate classification.

[0101] Assisted training: Acoustic and visual camera data are uploaded to the cloud by an operator (e.g., via USB) every day. The operator monitors each cage (e.g., using a mobile app) and notes when the fish are eating and / or not eating. This determines which parameter values ​​of the machine learning model, based on the input data, will most accurately match the model's output (agitation state) to the actual state of the fish.

[0102] 3. Validation phase:

[0103] Once the parameter values ​​for the artificial intelligence model to be used have been determined in the training phase, the model's correctness is verified by testing its suitability. This verification phase (also called validation) aims to evaluate the model's performance and ensure it can generalize well to unseen or future data.

[0104] The validation process is carried out, to the extent possible, during a complete feeding cycle. During this cycle, the agitation stage of the group of fish is estimated from real hydrodynamic noise samples and using the machine learning model obtained after the training phase, and the accuracy of this estimate is verified (for example, visually).

[0105] For model validation, sound recordings of different species from other aquaculture facilities can also be used, creating a database of hydrodynamic sound recordings of groups of fish corresponding to their different feeding states.

[0106] During the validation phase, you can adjust the model's parameters (and hyperparameters) and optionally evaluate its performance. A properly validated model is more reliable and can make more accurate decisions when presented with new data in the production or deployment phase.

[0107] 4. Production phase (also called implementation phase):

[0108] Once the best learning model has been established in the training (and validation) phase, it is deployed to the device in production mode (also called deployment mode). That is, this model is used to determine the agitation state (output data) that it has classified for each of the audios it records, suitably processed (input data). This agitation state output (from which, as we have said, the feeding status of the group of fish can be estimated) can be sent, on the one hand, to the device's communication module (for sending to other devices directly or through the Gateway) and, on the other hand, to the indicator light (if applicable) to visually indicate the feeding status of the fish.

[0109] Specifically, the model's input would be the audio collected by the acoustic sensor (hydrophone) after the corresponding processing / filtering (if applicable). In the specific implementation explained above, filtering would be performed in 31 frequency bands, resulting in 31 sound characteristics for each 2-second audio recording (this is just an example; in other implementations, the audio duration and number of bands could be different).

[0110] In other words, at this stage of production, after the training period, the device will be able to perform its function, which is to estimate the fish's behavior (and alert the operator and / or the feeding system and / or other external devices) while the group of fish is being fed, indicating whether they are eating, displaying a lack of appetite toward the feed supplied, or have stopped eating at a given time. In this case, the operator or the feeding system can be notified (for example, by means of an LED light or by sending a message) to interrupt the feed supply.

[0111] During this production phase (which can be said to be the device's normal operating phase), maintenance must be performed. In one embodiment, such maintenance is based on the following routine tasks (this is only a non-limiting example, and the training may include only some of these tasks, all of them, or other tasks):

[0112] Visual inspection: The condition of the hydrophone, cabling, mooring system, and solar panel is checked for integrity and cleanliness. This work is usually performed by divers, as some of these components are submerged.

[0113] Hardware: Hydrophone cleaning: Every time divers check the cage or extract casualties (this is usually done, for example, weekly). Solar panel cleaning: Every time it becomes dirty, for example, due to bird droppings (this is usually done, for example, weekly). Battery replacement: Since the battery is recharged by the solar panel, replacing it would simply be a lifespan procedure (this is usually done, for example, annually or even bi-annually).

[0114] In summary, the present invention proposes a device, method, and system for detecting the behavior of a group of fish, based on passive acoustics of the hydrodynamic noise produced by the school of fish and the use of artificial intelligence. Specifically, an intelligent method, device, and system for controlling fish feeding is proposed that, through at least one acoustic sensor, collects the acoustic signal generated by the movement of the fish during feeding, processes it, and implements developed artificial intelligence models to determine the agitation state and estimate the feeding status of the fish. Thanks to these models, the efficiency of the feeding process of the cultured population can be improved, which allows, among other advantages, to make real-time decisions for managing the feeding process and reduce the amount of wasted feed, optimizing feed conversion rates.

[0115] Many of the embodiments presented here refer to the recording and processing of the hydrodynamic noise produced by a group of fish to determine their state of agitation and estimate their corresponding feeding state, but this is only an example, since the procedure and system proposed could be applied to any other method of determining the behavior of a group of fish that can be estimated from the hydrodynamic noise produced (detection of diseases, abnormal behavior, stress levels, etc.).

[0116] In this text, the term "comprises" and its derivatives (such as "comprising", etc.) should not be understood in an exclusive sense, that is, these terms should not be interpreted as excluding the possibility that what is described and defined may include more elements, stages, etc.

[0117] Having sufficiently described the nature of the invention, as well as the manner of its implementation in practice, it must be noted that its different parts may be manufactured in a variety of materials, sizes and shapes, and that variations that practice advises may also be introduced into its constitution or procedure, as long as they do not alter the fundamental principle of the present invention.

[0118] The description and drawings merely illustrate the principles of the invention. Therefore, it should be appreciated that those skilled in the art will conceive various arrangements which, although not explicitly described or shown herein, nevertheless represent the principles of the invention and are included within its scope. Furthermore, all examples described herein are provided primarily for pedagogical purposes to assist the reader in understanding the principles of the invention and the concepts contributed by the inventor(s) to improve the art, and are to be considered as not limiting with respect to such specifically described examples and conditions. Furthermore, everything stated herein relating to the principles, aspects and embodiments of the invention, as well as the specific examples thereof, encompasses equivalents thereof.Although the present invention has been described with reference to specific embodiments, it should be understood by those skilled in the art that the foregoing and various other changes, omissions and additions in form and detail thereof may be made without departing from the scope of the invention as defined by the following claims.

Claims

CLAIMS 1. Device for detecting the behavior of a group of fish to improve their feeding process, in which the device comprises: - At least one acoustic sensor configured to capture in real time the hydrodynamic noise of the group of fish during a certain time interval; - At least one electronic processor configured to: a) Process the hydrodynamic noise data captured by the at least one acoustic sensor during the certain time interval; b) Determine the agitation state of the group of fish in said certain time interval, applying a previously trained machine learning model to the processed data resulting from a), c) Indicate the feeding state of the group of fish corresponding to the determined agitation state, by means of a light indicator located on the device itself and / or send information about the determined agitation state or about the feeding state of the group of fish corresponding to the determined agitation state, to one or more external devices by means of a communication network.

2. Device according to claim 1, wherein the processing of step a) comprises: a1) Filtering the captured noise data with a broadband filter; a2) Filtering the data resulting from step a1) with a series of N band-pass filters centered on different frequencies, obtaining N groups of filtered noise data, each one coming from a different band-pass filter; a3) Obtaining the sound pressure values ​​in the time domain corresponding to each of the N groups of filtered noise data 3. Device according to claim 2, wherein the bandpass filters are centered on third octave bands.

4. Device according to any of the preceding claims 2 to 3, wherein the bandpass filters are second-order infinite impulse response filters.

5. Device according to any of the preceding claims 2 to 4, wherein in step a1), after broadband filtering, an anomalous value detector is applied to the resulting sound pressure values, based on statistics of values ​​collected in other similar aquaculture facilities and / or values ​​that have been collected in the device itself since its start-up.

6. Device according to any of the preceding claims, wherein the machine learning model is of the Random Forests type.

7. Device according to any of the preceding claims, wherein the at least one processor is further configured to communicate with a feeding device of the group of fish through the communication network, to control the supply of feed by the feeding device to the group of fish, based on the determined agitation state.

8. Device according to any of the preceding claims, further comprising a solar panel for recharging a power battery of the device.

9. A device according to any of the preceding claims, wherein the agitation state is one of the following: high, low, or transitioning, which correspond to the feeding, eating, not eating, or transitioning states, respectively.

10. Device according to any of the preceding claims, further comprising means for fixing to an aquaculture cage where the group of fish is located.

11. Device according to any of the preceding claims, wherein the at least one electronic processor is located in a main body of the device that is connected to the at least one acoustic sensor by a cable.

12. Device according to any of the preceding claims, wherein hydrodynamic noise data collected by the at least one acoustic sensor and processed, corresponding to previous time intervals and / or hydrodynamic noise data, are used for training the machine learning model. previously recorded during the feeding process of other groups of fish stored in a database.

13. A method for detecting the behavior of a group of fish to improve their feeding process, wherein the method comprises the following steps performed by at least one electronic processor of an electronic device, where the method comprises: a) Processing hydrodynamic noise data of the group of fish captured by at least one acoustic sensor of the device during a certain time interval; b) Determining the agitation state of the group of fish, applying a previously trained machine learning model to the processed data obtained in the previous step;c) Indicate the feeding status of the group of fish corresponding to the determined agitation status by means of a light indicator located on the device and / or send information about the determined agitation status or about the feeding status of the group of fish corresponding to the determined agitation status to one or more external devices by means of a communication network.; 14. System for detecting the behavior of a group of fish to improve their feeding process, which includes: - At least one device for detecting the behavior of the group of fish according to any of claims 1-12; - A gateway device configured to: Receive, from the at least one device for detecting the behavior of the group of fish, information about the determined agitation state and / or about the appetite state of the fish corresponding to the agitation state determined by the at least one device and / or information about the hydrodynamic noise of the group of fish; Communicate with one or more devices external to the system to send them information received from at least one device; Remotely access at least one device to modify its operating parameters and to control its on / off; where communication between the gateway device and the at least one device is performed via a wired or wireless communication network.

15. A computer program comprising computer-executable instructions for implementing the method according to claim 13, when executed on a computer, a digital signal processor, an application-specific integrated circuit, a microprocessor, a microcontroller, or any other form of programmable hardware.

Images