US-12616158-B2 - Quality sensor, computer-implemented method of predicting inhomogeneities in milk extracted from an animal, computer program and non-volatile data carrier

US12616158B2US 12616158 B2US12616158 B2US 12616158B2US-12616158-B2

Abstract

A quality sensor that predicts a degree of inhomogeneities in milk extracted from an animal, by receiving a set of input variables reflecting at least one characteristic each of the animal, the extracted milk, and at least process during which milk was extracted from the animal, and by feeding the input variables into a trained artificial neural network in the quality sensor, which generates an estimate of a predicted degree of inhomogeneities in the milk of the animal.

Inventors

Dorota Anglart
Charlotte HALLÉN SANDGREN

Assignees

DELAVAL HOLDING AB

Dates

Publication Date: 20260505
Application Date: 20220114
Priority Date: 20210115

Claims (15)

1 . A quality sensor ( 200 ) for predicting inhomogeneities in milk extracted during an extraction process, the quality sensor ( 200 ) comprising: a processor, in communication with a memory, an input interface, and an output interface, the processor being configured to implement a trained artificial neural network, and further configured to: receive a set of input variables (D IN ) via the input interface, the input variables (D IN ) reflecting: at least one characteristic of an animal, at least one characteristic of milk extracted from the animal, and at least one characteristic of at least one process during which the milk was extracted from the animal; feed the set of input variables (D IN ) into the trained artificial neural network ( 210 ); and generate and output, from the trained artificial neural network ( 210 ) operating on the set of input variables, an estimate (D OUT ) of a predicted degree of inhomogeneities in the milk of the animal, wherein the trained artificial neural network ( 210 ) has weights that have been determined via a backpropagation training process comprised of a scoring of densities of milk clot deposits on filters ( 410 ) through which extracted milk has been passed during milk extraction processes in relation to training data (D TR-IN ) expressing the set of input variables, the scoring of the densities of milk clot deposits on the filters ( 410 ) expressing output training data (D TR-OUT ).
2 . The quality sensor ( 200 ) according to claim 1 , wherein the trained artificial neural network ( 210 ) is a feedforward neural network.
3 . The quality sensor ( 200 ) according to claim 1 , wherein the scoring of densities of milk clot deposits on the filters ( 410 ) is based on at least one of: a visual classification of the milk clot deposits by a human assessor, an automatic image classification of the milk clot deposits by a computer-implemented algorithm, an electric impedance measurement on the filter ( 410 ) when the milk clot deposits are wet, an electric impedance measurement on the filter ( 410 ) when the milk clot deposits are dry, a wet weight of the milk clot deposits, a dry weight of the milk clot deposits, and an enzyme content in the milk clot deposits.
4 . The quality sensor ( 200 ) according to claim 1 , wherein the processor is configured to receive via the input interface at least one input variable of the set of input variables (D IN ) from a milking apparatus used to extract milk from the animal.
5 . The quality sensor ( 200 ) according to claim 1 , wherein the at least one characteristic reflecting the animal comprises at least one of: at least one earlier estimate (D OUT ) generated by the trained artificial neural network ( 210 ) with respect to the animal, a number of days in milk for the animal, an interval between consecutive milk extraction processes for the animal, and a parity number for the animal.
6 . The quality sensor ( 200 ) according to claim 1 , wherein the at least one characteristic reflecting the extracted milk comprises at least one of: at least one udder quarter electric conductivity value registered during milk extraction from the animal, at least one mean udder quarter electric conductivity value for the animal determined based on milk extraction from the animal, a lactate dehydrogenase measure for the animal, a parameter expressing a color of the milk extracted from the animal, a set of parameters expressing a respective color of the milk extracted from each udder quarter of the animal, and a somatic cell count in milk extracted from the animal.
7 . The quality sensor ( 200 ) according to claim 1 , wherein the at least one characteristic reflecting the process during which the milk was extracted comprises at least one of: a duration of at least one milk extraction process for the animal, a milk yield from the animal in at least one milk extraction process, an expected milk yield from the animal in at least one milk extraction process, a set of parameters expressing a respective expected milk yield from each udder quarter of the animal in at least one milk extraction process, an expected rate of milk secretion in at least one milk extraction process for the animal, a set of parameters expressing a respective expected rate of milk secretion from each udder quarter of the animal in at least one milk extraction process, an udder quarter milk flow rate during a period of a milk extraction process for the animal, a peak milk flow rate during at least one milk extraction process for the animal, a set of parameters expressing a respective peak milk flow rate from each udder quarter of the animal in at least one milk extraction process, an average udder quarter milk flow rate during at least one milk extraction process for the animal, a set of parameters expressing a respective milk flow rate from each udder quarter of the animal during at least one milk extraction process, a time required for attaching teatcups to the animal's teats in connection with at least one milk extraction process, an indication whether the animal has kicked off at least one teatcup in connection with at least one milk extraction process, and an indication whether at least one udder quarter of the animal was not milked in at least one milk extraction process.
8 . A method, performed by a computer programmed to implement a trained artificial neural network, for predicting inhomogeneities in milk that is extracted during an extraction process, the method comprising: receiving a set of input variables (D IN ) from an input interface of the computer, the input variables (D IN ) reflecting at least one characteristic of an animal, at least one characteristic of milk extracted from the animal, and at least one characteristic of at least one process during which the milk was extracted from the animal; feeding the set of input variables (D IN ) into the trained artificial neural network ( 210 ); and generating and outputting, from the trained artificial neural network ( 210 ) operating on the set of input variables, an estimate (D OUT ) of a predicted degree of inhomogeneities in the milk of the animal, wherein the trained artificial neural network ( 210 ) has weights that have been determined via a backpropagation training process comprised of a scoring of densities of milk clot deposits on filters ( 410 ) through which extracted milk has been passed during milk extraction processes in relation to training data (D TR-IN ) expressing the set of input variables, the scoring of the densities of milk clot deposits on the filters ( 410 ) expressing output training data (D TR-OUT ).
9 . The method according to claim 8 , wherein the trained artificial neural network ( 210 ) is a feedforward neural network.
10 . The method according to claim 8 , wherein the scoring of the densities of milk clot deposits on the filters ( 410 ) is based on at least one of: a visual classification of the milk clot deposits by a human assessor, an automatic image classification of the milk clot deposits by a computer-implemented algorithm, an electric impedance measurement on the filter when the milk clot deposits are wet, an electric impedance measurement on the filter when the milk clot deposits are dry, a wet weight of the milk clot deposits, a dry weight of the milk clot deposits, and an enzyme content in the milk clot deposits.
11 . The method according to claim 8 , further comprising: receiving at least one input variable of the set of input variables (D IN ) from a milking apparatus used to extract milk from the animal.
12 . The method according to claim 8 , wherein the at least one characteristic reflecting the animal comprises at least one of: at least one earlier an estimate (D OUT ) generated by the trained artificial neural network ( 210 ) with respect to the animal, a number of days in milk for the animal, an interval between consecutive milk extraction processes for the animal, and a parity number for the animal.
13 . The method according to claim 8 , wherein the at least one characteristic reflecting the extracted milk comprises at least one of: at least one udder quarter electric conductivity value registered during milk extraction from the animal, at least one mean udder quarter electric conductivity value for the animal determined based on milk extraction from the animal, a lactate dehydrogenase measure for the animal, a parameter expressing a color of the milk extracted from the animal, a set of parameters expressing a respective color of the milk extracted from each udder quarter of the animal, and a somatic cell count in milk extracted from the animal.
14 . The method according to claim 8 , wherein the at least one characteristic reflecting the process during which the milk was extracted comprises at least one of: a duration of at least one milk extraction process for the animal, a milk yield from the animal in at least one milk extraction process, an expected milk yield from the animal in at least one milk extraction process, a set of parameters expressing a respective expected milk yield from each udder quarter of the animal in at least one milk extraction process, an expected rate of milk secretion in at least one milk extraction process for the animal, a set of parameters expressing a respective expected rate of milk secretion from each udder quarter of the animal in at least one milk extraction process, an udder quarter milk flow rate during a period of a milk extraction process for the animal, a peak milk flow rate during at least one milk extraction process for the animal, a set of parameters expressing a respective peak milk flow rate from each udder quarter of the animal in at least one milk extraction process, an average udder quarter milk flow rate during at least one milk extraction process for the animal, a set of parameters expressing a respective milk flow rate from each udder quarter of the animal during at least one milk extraction process, a time required for attaching teatcups to the animal's teats in connection with at least one milk extraction process, an indication whether the animal has kicked off at least one teatcup in connection with at least one milk extraction process, and an indication whether at least one udder quarter of the animal was not milked in at least one milk extraction process.
15 . A non-volatile, non-transitory data carrier ( 316 ) having stored thereon a computer program ( 317 ) readable by a processor ( 315 ), the computer program ( 317 ) comprising software that causes the processor ( 315 ) to execute the method according to claim 8 upon execution of the computer program ( 317 ) by the processor ( 315 ).

Description

CROSS-REFERENCE TO RELATED APPLICATIONS This application is the U.S. national phase of International Application No. PCT/SE2022/050044 filed Jan. 14, 2022, which designated the U.S. and claims priority to SE 2150028-5 filed Jan. 15, 2021, the entire contents of each of which are hereby incorporated by reference. TECHNICAL FIELD The present invention relates generally to milk quality monitoring. Especially, the invention relates to a quality sensor for predicting inhomogeneities in the milk extracted from an animal and a corresponding computer-implemented method. The invention also relates to a computer program and a non-volatile data carrier storing such a computer program. BACKGROUND For health and hygienic reasons it is important that a dairy farmer has accurate and reliable information about the milk being produced at his/her facilities. Moreover, various regulatory frameworks demand that the milk placed on the consumer market meets certain quality standards. For example, milk extracted from animals with udder inflammation cannot be used in food production. However, there are few methods for generating predictions of generally accepted indicators of udder inflammation and poor milk quality, such as somatic cell count (SCC) or changes in milk homogeneity. Traditionally, in manual milking, the milker pre-strips each quarter of the animal's udder and inspects the stripped milk before attaching the milking unit. During the inspection, the milker looks for deviations in the milk quality, such as clots, color changes, or other abnormalities. Thereby, the milker may identify sick cows and prevent abnormal milk from ending up in the bulk tank. In addition, regular milk samplings are typically performed in which the levels of somatic cells (i.e. white blood cells) are measured to monitor the udder health as well as milk quality. Of course, the above visual inspection cannot be made in automatic milking systems, such as milking robots, where the cows are milked on a voluntary basis. Here, some of said quality aspects may be measured by sensors, for example electric conductivity and color. However, it is challenging to automatically detect other kinds of deviations, such as milk clots, the presence of which is positively correlated with mastitis (udder inflammation). EP 1 264 537 discloses a method and a device for selecting milk. Here, within a predetermined period of time, a flow of milk is guided into a measuring chamber with at least one filter and a detector unit. After the predetermined period of time, the flow of milk is guided past the measuring chamber in a bypass line. After the predetermined period of time, a surface of the filter is detected. Depending on the evaluation result of the detection, the milk flow is either directed to a collecting container for usable milk or discarded. EP 1 273 224 describes a solution wherein, for the selection of milk, a predetermined milk volume of a milk flow is introduced into a measuring chamber with at least one detector unit. At least one area of the bottom surface of the measuring chamber is then detected, the detection is evaluated and, depending on the evaluation result, the milk flow either being directed to the collecting container for usable milk or being discarded. Thus, the known solutions basically mimic the milker's visual inspection described above. This strategy has proven to be both time consuming and unreliable. More important, the milk quality from a particular animal cannot be predicted before the milk extraction has been initiated. Consequently, there is a risk that substandard milk enters the bulk tank even if the detector signals a quality-related problem. SUMMARY The object of the present invention is therefore to offer a solution that solves the above problem and enables efficient and reliable prediction of inhomogeneities, for instance in the form of so-called clots in the milk to be extracted from a particular animal, or in the milk that is currently being extracted from this animal. According to one aspect of the invention, the object is achieved by a quality sensor for predicting inhomogeneities in milk extracted from an animal. The quality sensor is configured to receive a set of input variables reflecting at least one characteristic of the animal as such; at least one characteristic of the milk extracted—in the present milking session, in at least one earlier milking session, or both; and at least one characteristic of at least one process during which milk was extracted from the animal—i.e. the present milking session, at least one earlier milking session, or both. The quality sensor is further configured to feed the set of input variables into a trained artificial neural network. In response thereto, the trained artificial neural network is configured to generate an estimate of a predicted degree of inhomogeneities in the milk. The above quality sensor is advantageous because it truly enables prediction of a milk quality to be expected i