AT527305B1 - Computer-implemented method for determining at least one physical condition of a farm animal to be examined - Google Patents

Abstract

The invention relates to a computer-implemented method for determining at least one physical condition of a livestock animal (2) to be checked at a checking time (T1), comprising determining temporally successive values of at least one physical parameter within the gastrointestinal tract (3) of the livestock animal (2) to be checked by means of a probe device (1) arranged therein, transmitting the values to an evaluation unit (12), determining the probability of the presence of at least one physical condition of the livestock animal (2) to be checked by applying a trained artificial neural network to the transmitted values, and generating a message about the presence of a physical condition of the livestock animal (2) to be checked at the checking time (T1) if the probability determined in step c) exceeds a predefined limit value, wherein information about the implementation of herd management measures and livestock (2) affected thereby is stored in the evaluation unit (12) and before determining the probability it is checked whether the The time (TH) of implementation of the herd management measure is within the review period (TR) and the farm animal (2) to be reviewed is affected by the herd management measure, whereby in the event that the time (TH) of implementation of the herd management measure is within the review period (TR) and the farm animal (2) to be reviewed is affected, those values determined in step a) which are within a predefined time interval following the time (TH) of the herd management measure are removed and in step c) the neural network is only applied to the remaining values.

Description

Description

COMPUTER-IMPLEMENTED METHOD FOR DETERMINING AT LEAST ONE PHYSICAL CONDITION OF A FARM ANIMAL TO BE EXAMINED

[0001] The invention relates to a computer-implemented method for determining at least one physical condition of a farm animal to be examined at an inspection time, comprising the following steps:

a) determining temporally successive values of at least one physical parameter within the gastrointestinal tract of the farm animal to be examined by means of a probe device arranged in the gastrointestinal tract of the farm animal to be examined, at least in a verification period prior to the time of the inspection,

b) transmitting the values, preferably at least the values within the inspection period, from the probe device to an evaluation unit arranged outside the farm animal to be inspected,

c) determining the probability of the presence of at least one physical condition of the farm animal to be checked at the time of the check by applying a trained artificial neural network to at least part of the values transmitted for the predefined check period in step b) in the evaluation unit,

d) generating a report on the existence of a physical condition of the farm animal to be checked at the time of the check if the probability determined in step c) exceeds a predefined limit.

[0002] The invention further relates to a system for determining at least one physical condition of a farm animal to be examined, comprising at least one probe device and at least one evaluation unit, a computer program product and a computer-readable data carrier.

[0003] Worldwide, there is a trend in livestock farming towards large-scale operations with hundreds, sometimes thousands of animals. At the same time, even with smaller herd sizes, there is a need for optimized management, not least due to increasing cost pressure, for example with regard to feeding and knowledge of the general physical condition of the individual animals.

[0004] For this reason, livestock farming is increasingly turning to electronic management aids to take advantage of information technology. This approach is known under the umbrella term "smart farming", and the specialization in connection with livestock farming is referred to as "livestock management". For example, DE 199 01 124 A1 describes a bolus-shaped probe that can be used for such purposes. When inserted into the gastrointestinal tract of a cow, sensors measure various physical state variables such as pressure, temperature, conductivity, pH value or ammonia content. AT 509 255 B1 of the applicant describes a similar probe unit in which measurement data can be transmitted wirelessly, with the measurement sensors being at least partially surrounded by a cylindrical protective device within an acid-resistant housing to protect them from mechanical damage. The probe unit can transmit measurement values to base stations distributed throughout a stable and can be operated or read out by a control device via these.

[0005] The data obtained in this way allow the application of modern information and communication technologies, in particular those for the analysis of large amounts of data. For example, machine learning algorithms can be trained using statistical methods in such a way that they can learn from data and make predictions or interpret new, unknown data on the basis of the training.

[0006] As training data, measured values of the above-mentioned probes or probe units

8 NN

which are additionally provided with information - so-called "labels" - on the physical conditions of the farm animals. The artificial neural networks or artificial neural networks (abbreviated to "ANN") used are now state of the art for many applications - modern architectures of this kind are, for example, convolutional artificial neural networks ("convolutional neural networks", abbreviated to "CNN").

[0007] US 2022/0310259 A1 describes a system and method for "livestock monitoring": A measuring device records biometric data - e.g. heart and respiratory rate, body weight and rectal temperature - of farm animals that, for example, gather at a feeding place in order to identify animals suffering from respiratory diseases. The recorded values are used as input data for a trained algorithm to give a user a recommendation as to whether or not the farm animal in question needs to be treated for respiratory diseases.

[0008] There is a high risk of incorrect recommendations here, as the recorded values are used as input data without context. For example, a farm animal that has just been driven to a new pasture can be identified as being ill due to increased heart and respiratory rates, although these are caused by the rapid movement and the excitement of the change of location.

[0009] The disadvantage of the known solutions is that, due to the poor data quality of the input data for such algorithms or the small amount of biometric data used, incorrect results regarding the physical condition of a farm animal can be obtained.

[0010] It is therefore an object of the invention to eliminate the above-mentioned disadvantages of the prior art and to provide a computer-implemented method, a system, a computer program product and a computer-readable data carrier which reliably allow the determination of a physical condition of a farm animal to be checked.

[0011] This object is achieved according to the invention by the method mentioned at the outset in that information at least about the time of implementation of a herd management measure and about livestock affected thereby is stored and/or can be stored in the evaluation unit and before carrying out step c) it is checked whether the time of implementation of the herd management measure is in the verification period and the livestock to be checked is affected by the herd management measure, wherein in the event that the time is in the verification period and the livestock to be checked is affected, those values determined in step a) which are within a predefined time interval following the time of the herd management measure are removed and in step c) the neural network is only applied to the remaining values.

[0012] The solution according to the invention thus allows the influence that herd management measures have on the physical parameters of a farm animal to be taken into account, and incorrect statements about the probabilities of certain physical conditions of the farm animal to be checked can be prevented. As a result, incorrect, unnecessary measures are not taken on the farm animal, so that effort and costs can be saved and unnecessary animal suffering, e.g. due to over- or under-feeding, incorrect medication, etc., is prevented. Livestock farming can thus be optimized.

[0013] In particular, the physical parameters in step a) are at least one of the following parameters: temperature within the gastrointestinal tract of the farm animal; body temperature of the farm animal; pH value within the gastrointestinal tract of the farm animal; acceleration of the probe device; motility, determined from the acceleration of the probe device; activity of the farm animal, determined, among other things, from the acceleration of the probe device. The acceleration is determined in at least one spatial direction, preferably in three spatial directions of a Cartesian coordinate system or as a result of such measurements.

[0014] Preferably, the physical parameters in step a) comprise at least the parameters

8 NN

meter temperature within the gastrointestinal tract of the farm animal and motility, determined from the acceleration of the probe device, and activity of the farm animal. Other parameters can also be used, but the parameters mentioned are in any case present in a preferred variant of the invention.

[0015] In a further variant of the invention, after the values have been transmitted to the evaluation unit in step b) and before the probabilities have been determined in step c), the transmitted values are processed in a step cO0) by carrying out at least one of the following steps: - checking whether a predefined number of values was determined in the checking period and discarding all values if the number is not reached; - filling data holes in the data determined in step a), preferably by means of linear interpolation; - replacing non-numerical values in the data determined in step a), preferably with a predefined numerical value in each case.

[0016] In this way, the data quality can be significantly increased and proper execution, in particular of step c), i.e. the application of the trained artificial neural network, can be ensured. A data gap is understood here to mean that in a temporal sequence of values, no value is recorded at a time when a measurement should actually be taken. Non-numerical values are incorrect measured values or artifacts that can occur when operating a probe device.

[0017] In principle, artificial neural networks of different architectures can be used, for example "long short-term memory" networks, abbreviated to LSTM. Preferably, the trained artificial neural network in step c) is a convolutional artificial neural network or networks. The English term for this is "convolutional neural networks" with the acronym CNN. Such ANNs are well researched in the state of the art and enable correct and reproducible results.

[0018] In order to allow a simple and rapid identification of the livestock affected by herd management measures, the information stored on the evaluation unit about livestock affected by herd management measures is in the form of: 1) lists of individual livestock; or 2) lists of groups of livestock, wherein the evaluation unit further stores which livestock belong to which group; or 3) lists of organizations, wherein the evaluation unit stores which livestock and/or which groups of livestock belong to which organizations.

[0019] In this way, it is possible to quickly determine whether the farm animal to be examined was subjected to a herd management measure during the inspection period.

[0020] The predefined time interval, the values of which are removed before step c) is carried out, can preferably be specified by a user of the method or is specified depending on the herd management measure and stored on the evaluation unit. This measure makes it possible to adequately depict the fact that different herd management measures influence the physical parameters in the gastrointestinal tract of a farm animal for different lengths of time. In particular, users can also take into account farm animal, herd or organization-specific characteristics.

[0021] The above-mentioned object is also achieved by the system mentioned at the outset according to the invention in that the at least one probe device and the at least one evaluation unit are configured such that they carry out the steps of the above-mentioned method.

[0022] Furthermore, the above-mentioned object is achieved according to the invention by the computer program product mentioned at the outset in that it comprises instructions which, when the program is executed by a computer, cause the computer to carry out the steps of the above-mentioned method.

[0023] The above-mentioned object is also achieved by the computer-readable data carrier mentioned at the outset, on which the above-mentioned computer program product is stored.

[0024] The invention is explained in more detail below using a non-limiting embodiment shown in the drawings. The drawings are for illustration purposes only and therefore do not limit the invention in any way. They show

[0025] Fig. 1 shows a cow as an exemplary farm animal and the arrangement of a probe unit in its gastrointestinal tract;

[0026] Fig. 2 is a schematic representation of a probe device according to the invention and its components;

[0027] Fig. 3 is a schematic representation of a system according to the invention with several probe devices and an evaluation unit;

[0028] Fig. 4 is a flow chart of the method according to the invention;

[0029] Fig. 5 schematic representation of an embodiment of an evaluation unit according to the invention; and

[0030] Fig. 6 Flowchart for training an artificial neural network.

[0031] The solution according to the invention is explained below using an example, where the farm animal described is a cow. In principle, however, the invention is also applicable to other farm animals, in particular ruminant farm animals such as sheep, goats or wild ruminants such as red deer.

[0032] Fig. 1 shows a schematic cross-section of a cow 2. The feed ingested and - at least partially - chewed by the cow 2 reaches its gastrointestinal tract 3, for example the rumen or reticulum (also called the hood or reticulum). From the reticulum, the ingested feed can be transported further into the rumen on the one hand, and transported back into the mouth of the cow 2 for chewing on the other. A probe device 1 is inserted into the gastrointestinal tract 3 of the cow 2, with which values of physical parameters within the gastrointestinal tract 3 can be measured. From the values of one or more physical parameters within the gastrointestinal tract 3 of the cow 2, the presence of a physical condition of the cow 2 at a time of inspection can be determined using a trained artificial neural network. In particular, good results can be achieved if the probe device 1 is permanently in an end position in the reticulum. In this way, it is possible to check the cow and determine its physical condition. The cow is therefore also referred to below as the livestock animal to be checked or cow to be checked 2.

[0033] Physical states such as being awake or sleeping can be identified, and the presence of heat can also be detected. A pH value that is too low can indicate, for example, an illness, especially rumen acidosis; changes in heart rate, rumen motility, rumination and movement activity allow conclusions to be drawn about the presence of milk fever or mastitis, for example. These states are also divided into classes. In other words, the term physical state in the context of the present disclosure essentially includes the status of the organism of the body of the farm animal, which can be assigned to certain, predefined classes, so that assessments, diagnoses and recommendations for action can subsequently be made, but this is not part of the present invention.

[0034] Probe devices 1 for determining the values of the physical parameters are known from the prior art, for example from earlier patent applications of the applicant such as AT 509 255 B1, AT 517 847 B1 and AT 521 597 B1. A possible embodiment is outlined in Fig. 2, which shows a schematic, partially transparent view of the probe device 1:

[0035] Within a housing 4 with a first closure element 4a and a second

8 NN

A first 5a and a second sensor element 5b are arranged in the closure element 4b. The first sensor element 5a is, for example, an acceleration sensor, the second sensor element 5b is designed as a temperature sensor. In addition to or instead of this, other sensors can also be used, e.g. those for measuring temperature, pH value, density, pressure, conductivity, sound, optical properties or oxygen, CO2, ammonia, glucose, volatile fatty acids, acetate, propionate, butyrate and lactate.

[0036] Physical parameters, the values of which are determined directly or indirectly by the sensor elements 5a, 5b, are understood to be parameters that characterize the physical conditions in the gastrointestinal tract 3 of the farm animal, e.g. temperature, pH value, acceleration values, image information and the like. In the case of acceleration, the acceleration in all three spatial directions of a Cartesian coordinate system in which the probe device 1 is located is used in particular, which is determined in the illustrated embodiment with the first sensor element 5a, the acceleration sensor, or a resulting acceleration in space.

[0037] The probe device 1 also allows the determination of values of physical parameters that are not measured directly but are calculated from the measured values of the probe device 1, such as motility, rumination, heartbeat, eating and drinking times of the farm animal 1 or quantities relating to the state and progress of the digestive tract/process, for example the duration of the contractions of a motility and/or their periodicity or frequency. Motility is the rumen or stomach activity, essentially intestinal movements of the gastrointestinal tract 3 of the farm animal 2, which are shown by corresponding contractions. In other words, motility in the context of the present disclosure refers to contractions of the gastrointestinal tract 3, which act on the probe device 1 in the form of acceleration forces. Motility as a physical parameter is therefore expressed as acceleration forces with a specific characteristic, which can be measured by the acceleration sensor of the probe device 1 due to the contractions of the gastrointestinal tract 3, which propagate, among other things, through the stomach contents. Motility is the characteristic acceleration forces of the contractions of the gastrointestinal tract 3.

[0038] From the pulse width and thus the duration of the contractions and/or their periodicity - i.e. the time that passes between two consecutive contraction phases - and/or their frequency, the rumination or chewing activity can also be determined, in particular the periods in which the farm animal 2 devotes itself to chewing the cud. In the context of the present disclosure, rumination refers to the chewing activity of the farm animal, in particular its intensity, duration and periodicity or frequency.

[0039] In addition, further behavioral information in the form of activities such as standing, slow/fast movement and lying down can be determined from the acceleration measurements, which can be summarized under the term activity of the farm animal: The activity of the farm animal 2 comprises the characteristic acceleration forces that result from physical movement of the farm animal 2 and are measurable by means of the probe device 1.

[0040] While motility, rumination and activity of the farm animal 2 are all measured by acceleration sensors, a distinction in the measurement signal is possible based on the characteristic properties of the acceleration patterns of motility, rumination and activity. In particular, these characteristic properties are the frequency range of the acceleration signals, i.e. the frequency with which acceleration peaks occur, the periodicity and the amplitudes.

[0041] Furthermore, a timer in the form of a clock generator 5c, e.g. an RTC (“Real Time Clock”), is provided in the probe device 1.

[0042] The sensor elements 5a, 5b of the probe device 1 determine temporally successive values of the physical parameters. This can be done continuously, in particular with defined sampling rates (e.g. 100 Hz), so that immediately successive measurements are taken and the determined values are stored or sent to an evaluation unit 12.

(see Fig. 3). However, it is also possible for the measurements to be carried out at regular intervals - for example every minute, every hour after certain time intervals, e.g. every few minutes, every few hours, etc., or triggered by certain events, e.g. feeding, drinking, etc. This results in time series data of values of the physical parameters, supported by the clock generator 5c. The clock generator 5c can output absolute time values and/or relative time information such as the time since the last measurement, the number of measurements carried out, etc.

[0043] Sensor elements 5a, 5b and clock generator 5c are connected to a probe control unit 6, which serves to control the probe device 1. The probe control unit 6 is designed, for example, as a correspondingly programmed microprocessor. The probe control unit 6 controls and processes the data from the sensor elements 5a, 5b and from the clock generator 5c.

[0044] A storage element 7 can be provided for storing the data, for example a memory chip, an SD card or comparable technologies known to those skilled in the art. The storage element 7 stores both measured values of the sensor elements 5a, 5b, in particular also time-coded using the clock generator 5c, and operating parameters of the probe device 1 such as radio frequency (for communication with an evaluation unit 12 described further below in connection with Fig. 3), transmission channel, system time, but also configuration parameters such as the sampling rate of the acceleration sensor, sampling rate of the conversion into the first state variable and others.

[0045] Data is transmitted via a transmitting device 8, which has at least one antenna 9 - only sketched in Fig. 2 - for example to an evaluation unit 12 (Fig. 3) which is located outside the farm animal to be checked. The evaluation unit 12 can be located in the immediate vicinity of the cow 2, but can also be located far away and accessible via appropriate wired or wireless communication connections. The transmitting device 8 is advantageously designed as a transceiver device which can both send and receive data.

[0046] The energy supply of the probe device 1 is provided, for example, via an energy supply device 10, which can be designed, for example, as a battery, accumulator or capacitor (advantageously a thin-film or supercapacitor). The energy supply device 10 can also be designed to be rechargeable; for example, recharging is possible by "energy harvesting" or other methods.

[0047] In the illustrated embodiment, the components described in accordance with the above-mentioned AT 509 255 B1 are enclosed within the housing 4 by a hollow protective device 11 which surrounds at least the energy supply device 10 and which protects against mechanical impact, e.g. by biting. The protective device 11 can be made of any resistant material, for example plastic or metal.

[0048] While Fig. 2 shows a very detailed variant of a probe device 1, the probe device 1 can also be designed in a simpler manner to carry out the invention. It is sufficient if values of physical parameters in the gastrointestinal tract 3 of a farm animal 2 can be measured and transmitted to an evaluation unit 12 or if the probe device 1 has means with which these tasks can be fulfilled.

[0049] Fig. 3 shows an embodiment of a system 100 for determining the physical condition of several farm animals to be checked using several described probe devices 1 - for reasons of clarity, the farm animals in whose gastrointestinal tract 3 the probe devices 1 are arranged are not shown. The probe devices 1 communicate wirelessly with an evaluation unit 12. The aggregation of larger amounts of data and their evaluation, or the application of the trained artificial neural network to the measured values of the probe devices 1, takes place in the evaluation unit 12.

[0050] The transmission of the values from the probe device 1 to the evaluation unit 12 is preferably carried out wirelessly using a suitable protocol (e.g. LoRa, ZigBee, RFiD, WLAN,

8 NN

or others), preferably using a suitable frequency range, e.g. 300 MHz to 450 MHz, in which the permeability for radio waves in animals is particularly high, since the data of the probe devices 1 are sent from the gastrointestinal tract 3 of the farm animals.

[0051] This transmission can take place continuously, but in reality transmission is not always possible or desired because the cow 2 is not in the reception area of the evaluation unit 12 or because the data should only be sent in packets for energy saving reasons. For this reason, the memory element 7 is provided, which can be designed as RAM and/or ROM. For example, values can be stored in the random access memory ("RAM") during a measurement interval, and are then transferred from the temporary memory to an EEPROM, i.e. a read-only memory. The intermediate step with the EEPROM is particularly advantageous in order to be able to store sufficient amounts of data before the next transmission to the evaluation unit 12 is possible.

[0052] To increase the range or to reduce the necessary transmission power, in the embodiment shown, several transmitting/receiving units 13 are provided between the evaluation unit 12 and the probe devices 1, which can function as relays. This makes it possible to implement a star-shaped architecture for using the LoRa network protocol (or "LoRaWAN" - "Long Range Wide Area Network"), with the probe devices 1 representing the end devices and the transmitting/receiving units 13 being gateways that send data packets to the evaluation unit 12, e.g. a network server. The transmitting/receiving units 13, particularly in the function as relays, are optional and not absolutely necessary. A typical system 100 will consist of probe units 1 on the one hand and at least one or more evaluation units 12, whereby the communication between probe units 1 and evaluation unit 12 or evaluation units 12 takes place via transmitting/receiving units 13, which forward measured values or values from probe units 1 in their reception area to the evaluation unit 12. Usually, therefore, there is no forwarding between transmitting/receiving units 13, but from probe units 1 via an assigned transmitting/receiving unit 13 to an evaluation unit 12.

[0053] The evaluation unit 12 can, for example, be a mobile or stationary computer on which the method according to the invention and further control and evaluation routines run, or a server that can be connected via corresponding network connections, with several processors or program logics, or several such servers.

[0054] In a system 100 comprising probe unit 1 (or several probe units) and evaluation unit 12, the computer-implemented method according to the invention for determining at least one physical condition of at least one farm animal 2 to be checked at a time of checking is carried out, as shown schematically in Fig. 4 and explained below. In particular, the evaluation unit 12 determines and outputs a message about a physical condition of a farm animal 2 to be checked.

[0055] The method described is computer-implemented, i.e. a method is proposed that is carried out with or on a computer system (evaluation unit 12) or with computer support. The trained artificial neural network described below is therefore implemented in the computer system - the evaluation unit 12 - and is designed to be trainable.

[0056] In other words, a physical condition of a farm animal 2 to be checked is detected by means of deep learning by applying a trained model in the form of a trained artificial neural network - in the illustrated embodiment based on a CNN architecture, i.e. a "convolutional neural network" - to values for physical parameters in the gastrointestinal tract 3 of a farm animal 2, which are determined by a probe device 1 arranged in the gastrointestinal tract 3. Alternatively, the condition can be detected using statistical methods by applying a statistical model in the form of a trained artificial neural network to the values.

8 NN

[0057] The starting situation is a farm animal 2, here a cow, in whose gastrointestinal tract 3 there is a probe device 1 as described above. A first step, which is considered to have already been completed here and is therefore not shown, consists in inserting a probe device 1 into the gastrointestinal tract 3 of a farm animal 2 to be examined.

[0058] 20: In a step a), the determination of temporally successive values of at least one physical parameter within the gastrointestinal tract 3 of the farm animal 2 to be tested is carried out by the probe device 1 arranged in the gastrointestinal tract 3 of the farm animal 2 at least in a testing period TR prior to the testing time T+. In other words, values of physical parameters prevalent in the gastrointestinal tract 3 are recorded. The values can be determined continuously or at fixed or varying time intervals. Examples of the physical parameters are described above. As already explained, these can be directly measured physical parameters as well as parameters that can be determined or calculated from the measured values, e.g. motility, activity, rumination, acceleration. In any case, a time series of values of physical parameters is obtained that can cover a period of minutes, hours, days, weeks or months to years. At least part of this time series is then the checking period TR, the values of which must be available at a checking time T+; in order to determine a physical condition. The checking period TR can be firmly defined and typically amount to a number of days, in particular seven days, but can be freely selected by a user of the method or system 100 according to the invention. The checking period TR or its duration, regardless of whether it is fixed or selected by a user, is stored on the evaluation unit 12.

[0059] 30: In a step b), the values, preferably at least the values in the checking period TR, are transmitted from the probe device 1 to the evaluation unit 12 arranged outside the livestock 2 to be checked. As already explained above, this transmission can take place continuously, i.e. the values are sent immediately after they have been determined, but it is also possible to aggregate the values in the memory 7 of the probe device 1 and then send them in packets, i.e. in the form of time series. The values can be transmitted directly to the evaluation unit 12 via a transmission path, but in variants the transmission from the livestock 2 can take place wirelessly, the further transmission path to the evaluation unit 12 can then take place via wireless and wired transmission paths. In any case, the values are transmitted from the probe device 1 to the evaluation unit 12, directly or via one or more stations. In the illustrated embodiment, the transmitted values include values for the temperature within the gastrointestinal tract 3 of the farm animal 2, motility, determined from the acceleration of the probe device 1, and activity of the farm animal.

[0060] 40: In a step c), the probability of the presence of at least one physical condition of the farm animal 2 to be checked at the check time T+ is determined by applying a trained artificial neural network to at least some of the values transmitted in step b) in the evaluation unit 12. The trained neural network is therefore applied in the evaluation unit 12 to a time series of values of physical parameters which in any case covers the check period TR. The trained artificial neural network can be a suitable network from the state of the art which the expert selects accordingly. The trained artificial neural network is preferably a folding artificial neural network or a "convolutional neural network" (CNN). Details of the trained neural network are explained in detail below.

[0061] For example, in order to detect a physical condition at the checking period T+, values from the checking period TR are used, where the period is, for example, seven days and the time of commencement is designated as To.

[0062] Therefore, TR = [To, T4], where To is: To= T1- 7 days.

[0063] In order to cover all aspects that may influence the physical condition, the solution according to the invention allows for the consideration of herd management measures that were carried out on the farm animal 2 during the inspection period TR.

[0064] Herd management measures are actions that are carried out directly on the farm animal 2 or on a group of farm animals 2 and have an impact on physical parameters in the gastrointestinal tract 3 of the farm animal 2 and are therefore measurable by the probe device 1. In other words, herd management measures are external influences that act on the farm animal 2 at a fixed point in time or within a defined period of time. Examples of herd management measures are the administration of vaccinations, the performance of hoof care or other prophylactic measures carried out on farm animal 2, group changes, i.e. when farm animal 2 moves from one group or herd to another group or herd, grazing (i.e. the movement from a stable situation to a pasture or an outdoor location), feed changes (if, for example, a feed composition is changed or feed is taken from a different silo on a farm), storms, heat or cold waves, heat stress in demanding climatic conditions or the reaction to this, e.g. when ventilation is activated in a stable and has a cooling effect on individual farm animals 2 or entire groups or herds.

[0065] Herd management measures influence physical parameters or can cause physical parameters that are very similar to certain physical conditions, although they do not correspond to them - this can lead to step c) not running properly and can lead to incorrect results regarding the existence of a physical condition of the farm animal 2.

[0066] For example, a vaccination may cause a fever reaction resulting in elevated temperatures in the gastrointestinal tract 3, which could be identified as an inflammatory condition, although the farm animal 2 is actually in good physical condition.

[0067] Hoof trimming may cause increased stress in farm animal 2, resulting in increased activity, which could be identified as an abnormal physical condition and understood as the presence of milk fever, although farm animal 2 is actually in good physical condition.

[0068] The relocation of livestock, i.e. when a livestock 2 or a group of livestock 2 are driven to a new pasture or similar, is also referred to as a herd management measure and has an influence on the physical parameters measured in the gastrointestinal tract 3 of a livestock 2.

[0069] Herd management measures carried out during the review period TR can therefore lead to the artificial neural network calculating incorrect probabilities and determining an incorrect physical condition of the farm animal 2. This is often the case in the state of the art.

[0070] The inventive solution is now suitable for preventing these previous disadvantages. For this purpose, information about at least the time TH of the implementation of a herd management measure and information about farm animals 2 affected by it is stored or can be stored in the evaluation unit 12. In other words, the evaluation unit contains data about when a herd management measure was carried out and which farm animals 2 were affected by it. Conveniently, further details are also stored, e.g. which herd management measure was carried out and how long it may have an effect on physical parameters in the gastrointestinal tract 3 of the farm animal 2.

[0071] Before determining the probabilities in step c) - 40 - a checking step 99 is then carried out in which it is checked whether the time TH of the implementation of the herd management measure lies within the U-checking period TR and whether the

8 NN

Livestock 2 was or is affected by the herd management measure. If both apply, i.e. if the livestock 2 to be checked was exposed to a herd management measure in the checking period TR, those values that lie within a predefined time interval following the time TH of the herd management measure are removed from the time series of values of physical parameters from the checking period TR, and the artificial neural network is only applied to the remaining values of the time series. In other words, the artificial neural network is only applied to values of the time series that were not affected by the herd management measure.

[0072] The values of all physical parameters can be removed, but in one variant only values of certain, previously defined parameters can be removed.

[0073] The length of the time interval following the time TH of the herd management measure, in which the measured values are removed from the time series used for step c), can, for example, be specified by a user of the method on an ad hoc basis and stored on the evaluation unit 12, or a length of the time interval is stored on the evaluation unit 12 depending on the type of herd management measure.

[0074] In other words, the evaluation unit 12 basically applies the trained artificial neural network to all values of a time series of a farm animal 2 to be checked from a checking period TR.

[0075] Before determining the probabilities, however, according to the invention, a check step 99 checks whether the time TH of a herd management measure lies in the check period TR and whether the farm animal 2 to be checked was affected by the herd management measure.

[0076] The information stored on the evaluation unit 12 on farm animals 2 affected by the herd management measure can be present in various forms: as lists of individual farm animals 2, where for each farm animal 2 it is stored when which herd management measure was applied; as a list of groups of farm animals 2, where herd management measures for groups of farm animals 2 are stored on the evaluation unit 12, wherein the evaluation unit 12 also stores which farm animals 2 belong to which group; lists of organizations, wherein the evaluation unit 12 stores which farm animals 2 and/or which groups of farm animals 2 belong to which organizations.

[0077] If the two conditions mentioned above apply, those values that were recorded in a time interval following the time TH of the herd management measure are removed from the time series of the verification period TR. The time interval can be one day, for example. This is the case, for example, if the herd management measure is hoof care and experience has shown that the treated farm animal 2 shows increased stress and thus increased activity for around 24 hours due to the hoof care.

[0078] In step c), the artificial neural network is applied to the remaining values, i.e. the values of the original time series of the verification period TR minus the values from the time interval following the time TH of the herd management measure.

[0079] Thus, the values of the time series which are distorted or potentially distorted by the herd management measure are no longer in the time series from which the probabilities for the presence of a physical condition are determined, and the risk of an erroneous result is eliminated.

[0080] Of course, the described procedure can also be carried out several times if several herd management measures were carried out during the review period TR.

[0081] The result of step c) - 40 - is thus a probability for a physical condition for which the artificial neural network was trained. If it was trained for several physical conditions, the probabilities result for each physical condition defined during training (see description below in connection with Fig. 6). The probability of a physical condition corresponds to the probability that the values used in the time series, which cover, for example, a review period TR of seven days, correspond to a physical condition - if necessary minus the values following the time TH of a herd management measure in a time interval.

[0082] 40': In an optional step c0), which is carried out after the values have been transmitted to the evaluation unit 12 - 30 - and before the probabilities - 40 - are determined, the values transmitted in step b) - 30 - are processed by carrying out at least one of the following steps: - checking whether a predefined number of values was determined in the verification period and discarding all values if the number is not reached; - filling data holes in the data determined in step a), preferably by means of linear interpolation; - replacing non-numerical values in the data determined in step a), preferably with a predefined numerical value in each case.

[0083] This ensures that the determination of the probabilities can be carried out correctly. In Fig. 4, step 40' is shown before the checking step 99 - in a variant not shown, step 40° can also be carried out after the checking step 99 has been carried out.

[0084] 50: In a step d), a report on the presence of a physical condition of the farm animal 2 to be checked at the time of checking T+4 is generated in the evaluation unit 12 if the probability determined in step c) - 40 - exceeds a predefined limit value. The predefined limit values for the probabilities of different physical conditions are stored on the evaluation unit 12. The limit values can generally be specified, defined by users or determined from measured values using algorithms. It is also possible for existing, specified limit values on the evaluation unit 12 to be changed by users, e.g. if specifics of farm animals 2 or groups or herds are to be taken into account.

[0085] Alternatively or in addition to generating the message, the message can be output on a display unit not shown in the figures, which is, for example, a screen that is part of the evaluation unit 12 or is connected to the evaluation unit 12 by wire or wirelessly. For example, the display unit can be a mobile phone of a user of the method or system according to the invention.

[0086] In addition or as an alternative to generating the message and/or outputting the message on a display unit, a measure can be triggered if a physical condition of the farm animal 2 to be checked is present, for example opening a gate, raising the temperature of a stable building, increasing the flow rate of a drinking system, dispensing additional or alternative feed or administering nutritional supplements or medication, in each case by means of suitable actuators.

[0087] Step a) - 20 - is carried out in the probe device 1, while step 99, step c) - 40 -, step d) - 50 - and the optional step cO) - 40' - are carried out in the evaluation unit 12, which is shown in Fig. 4 by dotted boxes.

[0088] An exemplary structure of the evaluation unit 12 is shown in Fig. 6: [0089] A first program logic 60 is assigned the time series with values of the physical parameters

meter for the review period TR and, if applicable, meta-data of the farm animal 2 such as age, number of calvings or number of lactations, days in milk (“days-in-milk”) of the current lactation, as well as information on the herd management measures, so that in the first program logic the optional step cO) - 40' - and the review step 99 take place.

[0090] The appropriately cleaned time series is then fed to a second program logic 70 on which the trained artificial neural network runs, so that step c) - 40 - can take place.

[0091] The probability for a physical condition or the probabilities for several physical conditions resulting from the model are then fed to a third program logic 80, which generates and, if necessary, outputs a message when a predefined limit value for a physical condition is exceeded.

[0092] The trained artificial neural network used in step c) - 40 - and its training are described in more detail below.

[0093] Basically, the physical conditions are determined by KNN or artificial intelligence models, which are trained with the help of annotated training data sets for the most important physical conditions of a farm animal 1. This process is shown schematically in Fig. 6. The physical parameters used for this include in particular those that can be measured with the probe device 1 (e.g. temperature, motility, activity, ...) or calculated from it (e.g. rumination events, water intake). This also makes it possible to describe the so-called hood-rumen motor skills, i.e. the sequence of movements of the hood and the rumen. These special, defined and regular contraction sequences serve to mix and re-sort the feed and are an important part of the process of processing the feed into components that can be used by a cow 2. Other animal-specific properties such as lactation number, time of year, etc. can also be taken into account.

[0094] The starting point for the creation or training of the model in the form of an artificial neural network is a central database which is located in the evaluation unit 12 or is connected to it.

[0095] This database consists, on the one hand, of time series of values for physical parameters for a large number of farm animals 2, in particular containing values for the physical parameters temperature measured in the gastrointestinal tract 3, in particular in the forestomach system, activity and motility.

[0096] On the other hand, the database contains labels, i.e. data records on known physical conditions of the farm animal 2 with associated details, for example illnesses with medical treatments, times of heat and the like, which consist of at least a time of the physical condition, an identification of the affected farm animal 2 and, if applicable, a note on the illness, medical treatment or heat. In addition, so-called metadata of the farm animals 2 can be included, for example the organization or group to which a farm animal 2 is assigned, the number and date of calvings, the date of insertion of the probe device 1 and the time of the last heat.

[0097] The first step 90 of training the model consists in loading the time series and labels from the database.

[0098] In a second step 91, the labels are prepared. To do this, at least one or more of the following measures are carried out: - Checking to ensure that there is sufficient time series data for the farm animal 1 identified in the data set at the time identified in the data set. Several days around the identified time are preferred. If there is not enough time series data, the data set is discarded; - Loading time series data for the farm animal 1 identified in the data set around the time identified in the data set. The scope of the time series data therefore corresponds to

at a predefined time interval, e.g. seven days;

- Filling data gaps in the time series data caused by missing data points, provided that these do not exceed a maximum duration of several hours. Linear interpolation is used as the method for filling. If it is not possible to fill the data gaps or the data gap is too long, the data set is discarded;

- Replacing non-numeric measurement data caused by measurement errors. Affected data is replaced by a predefined numeric value.

[0099] In a third step 92, the loaded and cleaned labels are divided into different categories (“classes”) based on the note in the data set. The classes correspond to the physical conditions that the model enables to be estimated.

[00100] In the case that, for example, a physical condition corresponding to mastitis should be detected, possible classes are "mastitis" and "non-mastitis". Alternatively, several classes can be defined, such as "severe mastitis", "mild mastitis", "heat", "pneumonia", "livestock sleeping", "livestock walking" and "other", whereby the first two classes can be used for positive mastitis detection.

[00101] In the following fourth step 93, training takes place. For this purpose, the data set resulting from the previous steps is divided into training and validation data sets and used to train the artificial neural network. A "convolutional neural network" is used here. Hyperparameter tuning can be used in accordance with the ML software library used.

[00102] In a final step 94, the trained model is validated. To do this, it is subjected to several plausibility checks in which the accuracy of the model is compared with known physical conditions. The resulting model, called the "general model", is selected for the classification of real-time data if its accuracy is satisfactory and is used as a trained artificial neural network for operation in the method described above.

[00103] In an optional step 95, which is shown in Fig. 6 with a dashed line, the general model can be refined by refining it for a specific livestock species, a specific group of livestock 2, a business, a customer group, or a specific organization using "transfer learning". To do this, the third step 92 is repeated with a selection of data sets representative of the customer group/business/organization, and the data set obtained is used in the fourth step 93 to further train the general model, which is referred to as "transfer learning." In contrast to the general model, this results in a "special model" which, when applied to the customer group/livestock species/organization/etc., has a higher accuracy than the general model.

[00104] In the sequence of the method according to the invention described above with reference to Fig. 4, the general model as well as the specific model can be used depending on the application.

[00105] In addition to the described training method, a variant using statistical methods can also be used.

[00106] According to the above statements, the invention therefore relates to a computer-implemented method for determining a physical condition of a farm animal and a system 100 configured for this purpose, which comprises at least one probe device 1 and at least one evaluation unit 12. Accordingly, the invention also relates to a computer program product (not shown in detail in the figures), comprising commands which, when the program is executed by a computer, cause the computer to carry out the steps of the method described above, as well as a computer-readable data carrier on which this computer program product is stored, and a data carrier signal which transmits this computer program product.

[00107] A possible application of the described method of determining the physical condition of a farm animal 2 is, in a subsequent step, which is not part of the

protected invention is to make a diagnosis of a possible disease based on the condition and to suggest a response to it.

[00108] The detection of animal diseases at an early stage and the early and correct response to them is of great importance in livestock farming and reduces the frequency of disease and increases the degree of success of treatments while at the same time shortening the duration of treatment. As a further consequence, the need for the use of medication, especially antibiotics, is reduced.

[00109] If the physical condition of the farm animal 2 indicates a disease, a treatment is proposed depending on the measured parameters and values (recommendation for action). This treatment may vary depending on the country, the schemes chosen by the farm, etc.

[00110] This proposed treatment, for example administration of a special or supplementary feed, addition of a special feed or a medication, is carried out either manually (by a person) or automatically, e.g. via a concentrate feeding station or a feeding robot.

[00111] The treatment carried out is confirmed by observing the physical condition of the farm animal 2, on the basis of which further steps are then taken to reach a diagnosis, or if necessary adapted and suggested/changed. A change is initiated by a change/non-change in the physical condition and a subsequent diagnosis. The new treatment plan is selected based on this calculated result and on the therapy/treatment carried out so far.

[00112] If a normal physical condition is measured again from which no disease is diagnosed anymore, the treatment plan can be shortened or discontinued under certain circumstances.

[00113] Treatment plans can change depending on how the physical condition develops (i.e., the treatment plans take history into account). Accordingly, a dosage of a medication can be adjusted or the administration of a medication can be stopped completely. It can also be recognized that the treatment is not sufficient and another medication should be given or a change to a different treatment plan must be made (“Standard Operation Procedure” - SOP). A treatment plan can recommend the administration of medication or be based on a different therapy.

[00114] The invention thus allows the physical condition of a farm animal 2 to be examined to be recognized by applying a trained artificial neural network to values for physical parameters in the gastrointestinal tract 3 of a farm animal 2, which are determined by a probe device 1 arranged in the gastrointestinal tract 3. The invention does not allow a diagnosis of an illness - as shown above, this requires additional steps that are not part of the method or system according to the invention.

Claims (10)

patent claims 1. Computer-implemented method for determining at least one physical condition of a farm animal (2) to be inspected at an inspection time (T +4), comprising the following steps: a) determining temporally successive values of at least one physical parameter within the gastrointestinal tract (3) of the farm animal (2) to be checked by means of a probe device (1) which is arranged in the gastrointestinal tract (3) of the farm animal (2) to be checked, at least in a checking period (TR) prior to the checking time (T+), ” b) transmitting the values, preferably at least the values within the inspection period (TR), from the probe device (1) to an evaluation unit (12) arranged outside the farm animal (2) to be inspected, c) determining the probability of the presence of at least one physical condition of the farm animal (2) to be checked at the time of checking (T+;) by applying a trained artificial neural network to at least part of the values transmitted in step b) in the evaluation unit (12), d) generating a report on the existence of a physical condition of the farm animal (2) to be checked at the time of the check (T+) if the probability determined in step Cc) exceeds a predefined limit, characterized in that in the evaluation unit (12) information at least about the time (TH) of implementation of a herd management measure and about farm animals (2) affected thereby is stored and/or can be stored and before step c) is carried out it is checked whether the time (TH) of implementation of the herd management measure is within the inspection period (TR) and the farm animal (2) to be checked is affected by the herd management measure, wherein in the event that the time (TH) of implementation of the herd management measure is within the inspection period (TR) and the farm animal (2) to be checked is affected, those values determined in step a) which are within a predefined time interval following the time (TH) of the herd management measure are removed and in step c) the neural network is only applied to the remaining values. 2. Method according to claim 1, characterized in that the physical parameters in step a) are at least one of the following parameters: temperature within the gastrointestinal tract (3) of the farm animal (2); body temperature of the farm animal (2); pH value within the gastrointestinal tract (3) of the farm animal (2); acceleration of the probe device (1); motility, determined from the acceleration of the probe device (1); activity of the farm animal (2), determined, among other things, from the acceleration of the probe device (1). 3. Method according to claim 1 or 2, characterized in that the physical parameters in step a) comprise at least the parameters temperature within the gastrointestinal tract (3) of the farm animal (2) and motility, determined from the acceleration of the probe device (1) and activity of the farm animal (2). 4. Method according to one of claims 1 to 3, characterized in that after transmitting the values to the evaluation unit (12) in step b) and before determining the probabilities in step c), in a step c0) the transmitted values are processed by carrying out at least one of the following steps: - Check whether a predefined number of values was determined in the verification period (TR) and discard all values if the number is not reached; - filling data holes in the data determined in step a), preferably by means of linear interpolation; - Replacing non-numerical values in the data determined in step a), preferably by a predefined numerical value. 5. Method according to one of claims 1 to 4, characterized in that the trained artificial neural network in step c) is a convolutional artificial neural network. 6. Method according to one of claims 1 to 5, characterized in that the information stored on the evaluation unit (12) about livestock (2) affected by herd management measures is in the form of: 1) lists of individual farm animals (2); or 2) lists of groups of farm animals (2), whereby the evaluation unit also contains information on which farm animals (2) belong to which group; or 3) Lists of organisations, whereby the evaluation unit (12) stores which farm animals (2) and/or which groups of farm animals (2) belong to which organisations. 7. Method according to one of claims 1 to 6, characterized in that the predefined time interval, the values of which are removed before carrying out step c), can be specified by a user of the method or is specified depending on the herd management measure and is stored on the evaluation unit (12). 8. System for determining at least one physical condition of a farm animal (2) to be checked, comprising at least one probe device (1) and at least one evaluation unit (12) which are configured to carry out the steps of the method according to one of claims 1 to 7. 9. A computer program product comprising instructions which, when the program is executed by a computer, cause the computer to carry out the steps of the method according to any one of claims 1 to 7. 10. A computer-readable data carrier on which the computer program product according to claim 9 is stored. Here 2 sheets of drawings