CN116071783A - Sheep reproductive health early warning system and method - Google Patents

Abstract

The invention discloses a sheep reproduction health early warning system and a method, which are used for acquiring triaxial acceleration data of a ewe to obtain behavior data of the ewe, preprocessing the behavior data to obtain pre-production behavior information of the ewe, extracting characteristics of the behavior information to obtain characteristic values, identifying pre-production movement behaviors of the ewe according to the characteristic values to obtain sample behavior data, inputting the sample behavior data into a trained long-short-period memory network LSTM for training so as to classify the pre-production behaviors of the ewe to obtain classification results, carrying out sheep reproduction health early warning according to the classification results, describing the effects of a plurality of characteristics on the pre-production behavior identification of the ewe and the effects of the behavior identification, realizing accurate, stable and efficient long-distance transmission of the pre-production behavior data of the ewe, and transmitting the acquired data into a computer through wireless data for data processing, so that the accuracy of real-time processing and health early warning of the behavior data can be improved.

Description

Sheep reproductive health early warning system and method

Technical Field

The invention belongs to the technical field of livestock health management, and particularly relates to a sheep reproduction health early warning system and method.

Background

The growth and health status of animals can be reflected by their behavior, so that the monitoring data of their behavior during early warning and diagnosis of livestock diseases become the main data source. The ewes before delivery are easily affected by external environment to cause abortion, and the ewes with multiple delivery can be paralysed and other diseases due to the inadequacy of application and unknowingly supplemented, and the health of the ewes can be affected by any condition, the life of the ewes can be seriously endangered, and the sick manifestations of the ewes before delivery are mostly abnormal behaviors, unstable gait, unwilling to walk and the like. The timely and accurate detection of the disease condition of the prenatal ewe has important practical significance for improving the breeding efficiency of the ewe and improving the economic benefit of a feeder, so that the prenatal ewe disease needs to be warned very much.

Disclosure of Invention

In view of the above, the invention provides a sheep reproduction health early warning system and a method for constructing a data model corresponding to four typical behavior features occurring in a prenatal movement process of a ewe and improving the accuracy of classification results of the prenatal behavior features of the ewe, so as to solve the technical problems, and the invention is realized by adopting the following technical scheme.

In a first aspect, the present invention provides a sheep reproductive health warning system, comprising:

the data acquisition unit is used for acquiring triaxial acceleration data of the ewe to obtain behavior data of the ewe, and preprocessing the behavior data to obtain prenatal behavior information of the ewe, wherein the preprocessing comprises denoising, windowing and inclination correction;

the characteristic extraction unit is used for extracting characteristics of the behavior information to obtain characteristic values, wherein three time domain characteristics of acceleration data are selected, namely, the variances of three axial data of x, y and z, the frequency domain characteristic is the mean value of the main peak frequency of the three axial data and the first five values of the energy value of the frequency of the three axial data, and nineteen characteristic values are counted;

the behavior recognition unit is used for recognizing the prenatal exercise behaviors of the ewes according to the characteristic values to obtain sample behavior data, wherein the exercise behaviors comprise walking, standing, lying prone and planing, and the prenatal exercise behaviors of the ewes are recognized by adopting a clustering algorithm and a neural network classification algorithm;

and the early warning unit is used for inputting the sample behavior data into a trained long-short-period memory network LSTM for training so as to classify the prenatal behaviors of the ewes to obtain classification results, and carrying out sheep reproduction health early warning according to the classification results.

As a further improvement of the above technical solution, inputting the sample behavior data into a trained long-short-term memory network LSTM for training to classify the pre-partum behavior of the ewe to obtain a classification result, including:

the J epsilon {1, 2..J } is adopted as all the key feature points of the movement behaviors in a certain window, the number of windows represented by a certain movement behavior is represented by T epsilon {1, 2..T }, each LSTM finishes the data input, the analysis result of the movement behavior feature points of the same category output in the previous window and the output result of the adjacent movement nodes in the current time range are subjected to the data input, and the data are output to the next movement behavior feature representation window and the adjacent movement behavior feature points through the LSTM, wherein the corresponding expression is h _j,t ＝f(x _j,t h _j-1,t ,h _j,t-1 ) The operational procedure of LSTM is represented by a function f;

with two forgetting doors

Wherein->

Representing in space and space characteristics, forgetting gate represents whether the internal behavior feature change unit can generate data forgetting calculation in the LSTM calculation process, namely, obtaining behaviors in the last hierarchy within a certain probability rangeThe change result of the feature;

the expression of the input gate is i _j,t ＝σ(W _f *[h _j,t-1 ,h _j-1,t ,X _j,t ]+b _i ) The input gate completes the integration processing of the input data in the algorithm structure, and the input data comprises the behavior change characteristics of the ewe in the previous time range, the motion behavior characteristic points in the current time range and the position coordinates of the motion behavior occurrence points of the ewe in the current time period.

As a further improvement of the technical scheme, the expression for updating the movement behavior characteristic state of the ewe is that

Wherein the current state of the athletic performance characteristic in the LSTM is represented by C _j,t To show that the two forgetting doors can jointly realize the control of the movement behavior characteristic state;

the output gate is calculated by the following relation o _j,t ＝σ(W _o *[h _j,t-1 ,h _j-1,t ,X _j,t ]+b _o ) The output gate can realize the processing of the data to be output, and the output gate is represented by the following relation

After the correlation operation on the LSTM node is completed, the result is output to the output layer and the next time space iteration.

As a further improvement of the technical scheme, the identification of the prenatal exercise behavior of the ewes according to the characteristic values to obtain sample behavior data comprises the following steps:

preset training sample set d= { (x) _i ,y _i ) I=1, 2..n }, where x is _i ∈R ⁿ Represents the ith feature vector, R ⁿ Representing sample space, y _i E { -1,1} represents two classes of class labels, n represents the lumped number of samples, and the hyperplane can be classified by a classification function f (x) =ω ^T x+b, where ω represents the hyperplane direction parameter, b represents the dividing hyperplay parameter, T represents the transpose of the vector, when f (x) =0X is the point on the hyperplane;

maximizing the geometric interval of all support vectors to obtain a maximum interval classifier, namely, a basic expression of SVM, as follows

The expression obtained by using Lagrangian multiplier method and converting is

Wherein alpha is _i Representing the lagrangian multiplier added by the ith constraint condition, alpha represents a set of n lagrangian multipliers, and L (ω, b, a) represents the resulting lagrangian function;

solving the minimum value of L (omega, b, alpha) to omega, b, then solving the maximum value to alpha, and finally converting into

Where i, j denote the numbers of any two samples.

As a further improvement of the technical scheme, the three-axis acceleration data is mapped into a high-dimensional space by adopting a Gaussian kernel function, so that the three-axis acceleration data becomes linearly separable, and the Gaussian kernel function has the expression of

Wherein the method comprises the steps of ||x _i -x _j And I represents the mode of any two input vectors, sigma represents the nuclear parameter of a Gaussian kernel function, then three classifiers are built by combining drinking, feeding and ruminating in pairs by adopting a one-to-one method, finally a test sample is sequentially input into the three classifiers and voted, and the result output by the classifier with the largest number of votes is used as a final behavior classification result.

As a further improvement of the technical scheme, the feature extraction of the behavior information to obtain the feature value comprises the following steps:

the expression for extracting the time domain characteristic root mean square value from the sensor signal of the ewe behavior recognition is

Wherein N represents the number of behavior recognition samples, V _i A sensor acceleration value representing the instant i;

training a prediction model of acceleration coordinate difference value and corresponding RMS of behavior recognized by the ewe based on LSTM, calculating various parameters through a front-back time sequence relation of the acceleration coordinate recognized by the actual ewe behavior mapped with the RMS value in a training set and the predicted coordinate difference value, comprehensively considering estimated values and observed values of two adjacent moments, realizing updating of state variables, and estimating current requirements;

after the LSTM model is trained, the acceleration coordinate difference value can be predicted by introducing a prediction set RMS, the coordinate value is subjected to reverse-push prediction, and finally the visual prediction of the behavior acceleration data of the ewe is realized.

As a further improvement of the above solution, the prediction of the acceleration coordinate difference may be achieved by introducing a prediction set RMS, including:

input model trained acceleration signal eigenvalue RMS _i X, Y and coordinate value x in Z direction _i 、y _i And z _i ，i＝1,2...n；

The output includes process excitation noise covariance Q, observation noise covariance R, and parameter state transition matrix A, and M _x 、M _y M is as follows _z A three-axis coordinate difference prediction model;

based on the original signal, respectively for the original coordinate value x _i 、y _i And z _i Calculating the time domain feature RMS of acceleration _i Calculation is performed, i=1, 2..n;

in the front and rear time, the expression for calculating the coordinate value is

i=1, 2..n, the construction of a deep LSTM network is completed, in which the input and target output are respectively represented by RMS _i Δx _i And Deltay _i Acting as, in two coordinate directions, by correlation trainingImplementing construction of predictive models, i.e. Mx (RMS) _i ) And My (RMS) _i )；

The calculated parameters A, Q and R are respectively

Then

Wherein DeltaX _i ＝[Δx ₁ ,Δx ₂ ...Δx _i ]，ΔY _i ＝[Δy ₁ ,Δy ₂ ...Δy _i ]，ΔZ _i ＝[Δz ₁ ,Δz ₂ ...Δz _i ]，F _BiLinear Representing a binary linear regression function for calculating the correlation coefficients and residuals of the x, y, z coordinates of the previous moment and the x, y, z coordinates of the next moment as

Wherein a is ₁ 、a ₂ 、a ₃ 、b ₁ 、b ₂ 、b ₃ 、c ₁ 、c ₂ And c ₃ Representing the correlation coefficient, u _i And v _i Residual error, thereby determining->

And a= cov (β _x ,β _y ,β _z ) Two LSTM models are used to build a nonlinear observation model, namely

Wherein μ' _i 、v′ _i And omega' _i The residual is represented, and the observation noise R is obtained.

As a further improvement of the technical scheme, the method for acquiring the triaxial acceleration data of the ewe to obtain behavior data of the ewe, and preprocessing the behavior data to obtain the prenatal behavior information of the ewe comprises the following steps:

triaxial acceleration of ewes during static periodAcquiring the degree data, calculating a trigonometric function value sin alpha sin theta cos alpha cos theta of alpha theta according to space geometric knowledge, and presetting three axis measurement values of an acceleration sensor to be alpha when in rest _1x 、α _1y And alpha _1z Then

Wherein A is _x 、A _y And A _z Representing the actual triaxial acceleration value, A _x Indicating the acceleration value of the horizontal forward direction gravity acceleration direction A _y Indicating the acceleration value in the left-right direction, A _z Representing the acceleration value in the vertical direction, and the measured values of the acceleration sensor in three directions are respectively alpha _x 、α _y And alpha _z Can be calculated from the tilt angle and the measured value

And respectively calculating acceleration values in the horizontal direction, the front-rear direction and the vertical direction to obtain the acceleration data in the three corrected directions.

As a further improvement of the technical scheme, the feature dimension reduction optimization recognition process comprises the following steps:

presetting the high-dimensional matrix as { X } ₁ ,X ₂ ...X _N Computing the average vector of the high-dimensional matrix

The expression is

The covariance matrix Var (X) of the high-dimensional matrix X is calculated by using the average vector as the expression

To determine the eigenvector U of the covariance matrix Var (X) _i And a characteristic value lambda _i N eigenvalues are arranged in descending order, namely lambda ₁ >λ ₂ >λ ₃ ...>λ _k >λ _k+1 >...>λ _k Presetting lambda ₁ -λ _k Corresponding toThe k eigenvectors of (a) constitute a principal component matrix, and the variation matrix is U= { U ₁ ,U ₂ ,U ₃ ...U _k Then the expression of principal component matrix Y is y=u ^T X, the size of k is adjusted to adjust the dimension of the principal component matrix, i.e., to adjust the reduced dimension.

In a second aspect, the invention also provides a sheep reproduction health early warning method, which comprises the following steps:

acquiring triaxial acceleration data of a ewe to obtain behavior data of the ewe, and preprocessing the behavior data to obtain prenatal behavior information of the ewe, wherein the preprocessing comprises denoising, windowing and inclination correction;

extracting characteristics of the behavior information to obtain characteristic values, wherein three time domain characteristics of acceleration data, namely variances of three axial data of x, y and z, are selected, and frequency domain characteristics are taken to be the mean value of the main peak frequency of the three axial data and the first five values of the energy value of the frequency of the three axial data, and nineteen characteristic values are totally selected;

identifying the prenatal exercise behaviors of the ewes according to the characteristic values to obtain sample behavior data, wherein the exercise behaviors comprise walking, standing, lying prone and planing, and identifying the prenatal exercise behaviors of the ewes by adopting a clustering algorithm and a neural network classification algorithm;

and inputting the sample behavior data into a trained long-short-term memory network LSTM for training so as to classify the prenatal behaviors of the ewes to obtain classification results, and carrying out sheep reproduction health early warning according to the classification results.

The invention provides a sheep reproduction health early warning system and method, which are characterized in that behavior data of a ewe are obtained by collecting triaxial acceleration data of the ewe, the behavior data are preprocessed to obtain behavior information before the ewe is produced, characteristic values are obtained by characteristic extraction of the behavior information, the behavior data before the ewe is produced are identified according to the characteristic values to obtain sample behavior data, the sample behavior data are input into a trained long-short-period memory network LSTM for training to classify the behavior before the ewe to obtain classification results, sheep reproduction health early warning is carried out according to the classification results, characteristic extraction is carried out on the denoised acceleration data, and a plurality of acceleration characteristics representing the behavior of the ewe such as variance, mean value, extremum and the like are introduced and analyzed, so that the effects of the characteristics on the behavior recognition before the ewe can be described, the behavior recognition effect is realized, the accurate, stable and efficient long-distance transmission of the behavior data before the ewe is realized, the collected data are transmitted into a computer through wireless data for data processing, and the real-time processing and the health early warning of the behavior data can be improved.

Drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings that are needed in the embodiments will be briefly described below, it being understood that the following drawings only illustrate some embodiments of the present invention and therefore should not be considered as limiting the scope, and other related drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.

FIG. 1 is a block diagram of a sheep reproductive health warning system provided by the invention;

fig. 2 is a flow chart of the sheep reproduction health early warning method provided by the invention.

Detailed Description

Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein like or similar reference numerals refer to like or similar elements or elements having like or similar functions throughout. The embodiments described below by referring to the drawings are illustrative only and are not to be construed as limiting the invention.

Referring to fig. 1, the invention provides a sheep reproduction health early warning system, comprising:

the data acquisition unit is used for acquiring triaxial acceleration data of the ewe to obtain behavior data of the ewe, and preprocessing the behavior data to obtain prenatal behavior information of the ewe, wherein the preprocessing comprises denoising, windowing and inclination correction;

the characteristic extraction unit is used for extracting characteristics of the behavior information to obtain characteristic values, wherein three time domain characteristics of acceleration data are selected, namely, the variances of three axial data of x, y and z, the frequency domain characteristic is the mean value of the main peak frequency of the three axial data and the first five values of the energy value of the frequency of the three axial data, and nineteen characteristic values are counted;

the behavior recognition unit is used for recognizing the prenatal exercise behaviors of the ewes according to the characteristic values to obtain sample behavior data, wherein the exercise behaviors comprise walking, standing, lying prone and planing, and the prenatal exercise behaviors of the ewes are recognized by adopting a clustering algorithm and a neural network classification algorithm;

and the early warning unit is used for inputting the sample behavior data into a trained long-short-period memory network LSTM for training so as to classify the prenatal behaviors of the ewes to obtain classification results, and carrying out sheep reproduction health early warning according to the classification results.

In this embodiment, inputting the sample behavior data into a trained long-short-term memory network LSTM for training to classify the pre-partum behavior of the ewe to obtain a classification result, including: the J epsilon {1, 2..J } is adopted as all the key feature points of the movement behaviors in a certain window, the number of windows represented by a certain movement behavior is represented by T epsilon {1, 2..T }, each LSTM finishes the data input, the analysis result of the movement behavior feature points of the same category output in the previous window and the output result of the adjacent movement nodes in the current time range are subjected to the data input, and the data are output to the next movement behavior feature representation window and the adjacent movement behavior feature points through the LSTM, wherein the corresponding expression is h _j,t ＝f(x _j,t h _j-1,t ,h _j,t-1 ) The operational procedure of LSTM is represented by a function f; with two forgetting doors

Wherein->

Representing in space and spatial characteristics, forgetting gate representing whether or not data loss occurs in the internal behavior feature change unit in LSTM calculation processForgetting to calculate, namely obtaining a change result of the behavior characteristic in the last level within a certain probability range; the expression of the input gate is i _j,t ＝σ(W _f *[h _j,t-1 ,h _j-1,t ,X _j,t ]+b _i ) The input gate completes the integration processing of the input data in the algorithm structure, and the input data comprises the behavior change characteristics of the ewe in the previous time range, the motion behavior characteristic points in the current time range and the position coordinates of the motion behavior occurrence points of the ewe in the current time period.

It should be noted that, the expression for updating the athletic performance characteristic state of the ewe is that

Wherein the current state of the athletic performance characteristic in the LSTM is represented by C _j,t To show that the two forgetting doors can jointly realize the control of the movement behavior characteristic state; the output gate is calculated by the following relation o _j,t ＝σ(W _o *[h _j,t-1 ,h _j-1,t ,X _j,t ]+b _o ) The output gate can realize the processing of the data to be output, and the output gate is expressed by the following relation>

After the correlation operation on the LSTM node is completed, the result is output to the output layer and the next time space iteration.

It is understood that the motion information of the ewes is collected through the nine-axis attitude sensor in the data collection part, and the data is uploaded to the cloud server through Bluetooth and WIFI. After data acquisition, denoising the acquired attitude data by adopting a wavelet threshold method. And introducing the motion characteristics of the ewes represented by multiple dimensions such as acceleration, angular speed and the like in the behavior characteristic extraction part to complete the gesture analysis of the ewe motion process. Before classifying the pre-natal behaviors of the ewes, the sensor signals of the movement behaviors of the ewes are collected in a concentrated mode, and preprocessing of behavior feature extraction is carried out. The behavior of lying in the ewe is approximately reflected in sleep and rest, and the behavior is in a lying-prone position, and the breastbone and breasts of the ewe are in contact with the ground. The ewes mostly maintain a static or slightly swaying posture in standing behaviors, and four limbs of the ewes are contacted with the ground to maintain the body standing. The body balance was maintained by two legs at all time points when half-strides were produced in the ewe walking behavior, which was a pile-up behavior, gauge, and slow. The ewe developed this action by planing or scraping the ground through the front hoof. In the preprocessing process of the collected data of the sensors, the extraction of the behavior characteristics is very important for the subsequent behavior classification process of the ewes, the prenatal movement behavior characteristics of the ewes cannot be directly represented after the preprocessing operation is carried out on the original signal data, and the characteristic values of the behaviors of lying me, standing, walking and spectral planing of the ewes are required to be extracted and analyzed. The collected original sensor data information has a plurality of gravity acceleration components and a plurality of sharp noise interferences, and the noise reduction processing is carried out on the collected original acceleration data, so that relatively pure noise-interference-free acceleration data can be obtained. The characteristics of the denoised acceleration data are extracted, a plurality of acceleration characteristics representing the behavior of the ewe, such as variance, mean value, extremum difference and the like, are introduced and analyzed, the effects of the characteristics on the behavior recognition before the delivery of the ewe are described, the accurate, stable and efficient long-distance transmission of the behavior data before the delivery of the ewe is realized, the acquired data are transmitted into a computer through wireless data for data processing, and the accuracy of real-time processing and health early warning of the behavior data can be improved.

Optionally, identifying the prenatal exercise behavior of the ewe according to the feature value to obtain sample behavior data, including:

preset training sample set d= { (x) _i ,y _i ) I=1, 2..n }, where x is _i ∈R ⁿ Represents the ith feature vector, R ⁿ Representing sample space, y _i E { -1,1} represents two classes of class labels, n represents the lumped number of samples, and the hyperplane can be classified by a classification function f (x) =ω ^T x+b, where ω represents the hyperplane direction parameter, b represents the partitioning hyperplay parameter, T represents the transpose of the vector, when f (x) =0When x is a point on the hyperplane;

maximizing the geometric interval of all support vectors to obtain a maximum interval classifier, namely, a basic expression of SVM, as follows

The expression obtained by using Lagrangian multiplier method and converting is

Wherein alpha is _i Representing the lagrangian multiplier added by the ith constraint condition, alpha represents a set of n lagrangian multipliers, and L (ω, b, a) represents the resulting lagrangian function;

solving the minimum value of L (omega, b, alpha) to omega, b, then solving the maximum value to alpha, and finally converting into

Where i, j denote the numbers of any two samples.

In this embodiment, the three-axis acceleration data is mapped into the high-dimensional space by using a gaussian kernel function, so that the three-axis acceleration data becomes linearly separable, and the gaussian kernel function has the expression of

Wherein the method comprises the steps of ||x _i -x _j And I represents the mode of any two input vectors, sigma represents the nuclear parameter of a Gaussian kernel function, then three classifiers are built by combining drinking, feeding and ruminating in pairs by adopting a one-to-one method, finally a test sample is sequentially input into the three classifiers and voted, and the result output by the classifier with the largest number of votes is used as a final behavior classification result.

It should be noted that, performing feature extraction on the behavior information to obtain a feature value includes: the expression for extracting the time domain characteristic root mean square value from the sensor signal of the ewe behavior recognition is

Wherein N represents the number of behavior recognition samples, V _i A sensor acceleration value representing the instant i; training a prediction model of acceleration coordinate difference value and corresponding RMS of behavior recognized by the ewe based on LSTM, calculating various parameters through a front-back time sequence relation of the acceleration coordinate recognized by the actual ewe behavior mapped with the RMS value in a training set and the predicted coordinate difference value, comprehensively considering estimated values and observed values of two adjacent moments, realizing updating of state variables, and estimating current requirements; after the LSTM model is trained, the acceleration coordinate difference value can be predicted by introducing a prediction set RMS, the coordinate value is reversely predicted, the visual prediction of the behavior acceleration data of the ewe is finally realized, and the accuracy of the sheep reproduction health early warning and the stability of the system work are improved.

Alternatively, the prediction of the acceleration coordinate difference may be achieved by the introduction of a prediction set RMS, comprising:

input model trained acceleration signal eigenvalue RMS _i X, Y and coordinate value x in Z direction _i 、y _i And z _i ，i＝1,2...n；

The output includes process excitation noise covariance Q, observation noise covariance R, and parameter state transition matrix A, and M _x 、M _y M is as follows _z A three-axis coordinate difference prediction model;

based on the original signal, respectively for the original coordinate value x _i 、y _i And z _i Calculating the time domain feature RMS of acceleration _i Calculation is performed, i=1, 2..n;

in the front and rear time, the expression for calculating the coordinate value is

i=1, 2..n, the construction of a deep LSTM network is completed, in which the input and target output are respectively represented by RMS _i Δx _i And Deltay _i Acting as a model for constructing prediction by correlation training in two coordinate directions, i.e. Mx (RMS) _i ) And My (RMS) _i )；

The calculated parameters A, Q and R are respectively

Then

Wherein DeltaX _i ＝[Δx ₁ ,Δx ₂ ...Δx _i ]，ΔY _i ＝[Δy ₁ ,Δy ₂ ...Δy _i ]，ΔZ _i ＝[Δz ₁ ,Δz ₂ ...Δz _i ]，F _BiLinear Representing a binary linear regression function for calculating the correlation coefficients and residuals of the x, y, z coordinates of the previous moment and the x, y, z coordinates of the next moment as

Wherein a is ₁ 、a ₂ 、a ₃ 、b ₁ 、b ₂ 、b ₃ 、c ₁ 、c ₂ And c ₃ Representing the correlation coefficient, u _i And v _i Residual error, thereby determining->

And a= cov (β _x ,β _y ,β _z ) Two LSTM models are used to build a nonlinear observation model, namely

Wherein μ' _i 、v′ _i And omega' _i The residual is represented, and the observation noise R is obtained.

In this embodiment, collecting triaxial acceleration data of a ewe to obtain behavior data of the ewe, and preprocessing the behavior data to obtain prenatal behavior information of the ewe, including: collecting triaxial acceleration data of a ewe when the ewe is stationary, calculating trigonometric function values sin alpha sin theta cos alpha cos theta of alpha and theta according to space geometric knowledge, and presetting three axis measurement values of an acceleration sensor when the ewe is stationary as alpha _1x 、α _1y And alpha _1z Then

Wherein A is _x 、A _y And A _z Representing the actual triaxial acceleration value, A _x Indicating the acceleration value of the horizontal forward direction gravity acceleration direction A _y Indicating the acceleration value in the left-right direction, A _z Representing the acceleration value in the vertical direction, and the measured values of the acceleration sensor in three directions are respectively alpha _x 、α _y And alpha _z From the tilt angle and the measured value, it is calculated +.>

And respectively calculating acceleration values in the horizontal direction, the front-rear direction and the vertical direction to obtain the acceleration data in the three corrected directions.

It should be noted that, the feature dimension reduction optimizing and identifying process includes: presetting the high-dimensional matrix as { X } ₁ ,X ₂ ...X _N Computing the average vector of the high-dimensional matrix

The expression is +.>

The expression of covariance matrix Var (X) for calculating high-dimensional matrix X using average vector is +.>

To determine the eigenvector U of the covariance matrix Var (X) _i And a characteristic value lambda _i N eigenvalues are arranged in descending order, namely lambda ₁ >λ ₂ >λ ₃ ...>λ _k >λ _k+1 >...>λ _N Presetting lambda ₁ -λ _k The corresponding k eigenvectors form a principal component matrix, and the formed change matrix is U= { U ₁ ,U ₂ ,U ₃ ...U _k Then the expression of principal component matrix Y is y=u ^T And X, the size of k is adjusted to adjust the dimension of the principal component matrix, namely the reduced dimension, so that the working efficiency of data processing is improved.

Referring to fig. 2, the invention also provides a sheep reproduction health early warning method, which comprises the following steps:

s1: acquiring triaxial acceleration data of a ewe to obtain behavior data of the ewe, and preprocessing the behavior data to obtain prenatal behavior information of the ewe, wherein the preprocessing comprises denoising, windowing and inclination correction;

s2: extracting characteristics of the behavior information to obtain characteristic values, wherein three time domain characteristics of acceleration data, namely variances of three axial data of x, y and z, are selected, and frequency domain characteristics are taken to be the mean value of the main peak frequency of the three axial data and the first five values of the energy value of the frequency of the three axial data, and nineteen characteristic values are totally selected;

s3: identifying the prenatal exercise behaviors of the ewes according to the characteristic values to obtain sample behavior data, wherein the exercise behaviors comprise walking, standing, lying prone and planing, and identifying the prenatal exercise behaviors of the ewes by adopting a clustering algorithm and a neural network classification algorithm;

s4: and inputting the sample behavior data into a trained long-short-term memory network LSTM for training so as to classify the prenatal behaviors of the ewes to obtain classification results, and carrying out sheep reproduction health early warning according to the classification results.

In this embodiment, the neural network has a strong learning capability under the condition of distinguishing nonlinear separable, and is mainly characterized by the capability of autonomous learning of complex mapping, and can extract a group of nonlinear feature sets for recognition and classification. The neural network classifier has good fault tolerance to the processing process of external signals, has strong self-adaptive capacity, is easy to calculate in parallel and program calculation of software and hardware, and is very suitable for identifying the motion behavior based on the acceleration sensor. The identification of the prenatal exercise behavior of the ewe is respectively four behaviors of walking, standing, lying prone and planing, and because the same exercise mode is adopted between different behaviors, such as standing and lying prone, the ewe can be set to be static although the states of the two behaviors are different. The data sampling rate and the insensitivity of the data errors are adopted, the behavior of the female sheep such as pre-partum walking, standing, lying and planing is identified by combining a neural network classification algorithm, namely, the static behavior of the female sheep is classified by the clustering algorithm, the lying behavior of the female sheep is accurately distinguished and classified from the other three behaviors, and the three behaviors such as pre-partum standing, walking and planing are further identified by adopting the neural network classification algorithm on the basis. The generalization capability of the algorithm is improved, and the recognition rate of the algorithm to the prenatal exercise behaviors of the ewes is also improved.

Any particular values in all examples shown and described herein are to be construed as merely illustrative and not a limitation, and thus other examples of exemplary embodiments may have different values.

It should be noted that: like reference numerals and letters denote like items in the following figures, and thus once an item is defined in one figure, no further definition or explanation thereof is necessary in the following figures.

The above examples merely represent a few embodiments of the present invention, which are described in more detail and are not to be construed as limiting the scope of the present invention. It should be noted that it will be apparent to those skilled in the art that several variations and modifications can be made without departing from the spirit of the invention, which are all within the scope of the invention.

Claims (10)

1. Sheep reproduction health early warning system, characterized by, include:

the data acquisition unit is used for acquiring triaxial acceleration data of the ewe to obtain behavior data of the ewe, and preprocessing the behavior data to obtain prenatal behavior information of the ewe, wherein the preprocessing comprises denoising, windowing and inclination correction;

the characteristic extraction unit is used for extracting characteristics of the behavior information to obtain characteristic values, wherein three time domain characteristics of acceleration data are selected, namely, the variances of three axial data of x, y and z, the frequency domain characteristic is the mean value of the main peak frequency of the three axial data and the first five values of the energy value of the frequency of the three axial data, and nineteen characteristic values are counted;

the behavior recognition unit is used for recognizing the prenatal exercise behaviors of the ewes according to the characteristic values to obtain sample behavior data, wherein the exercise behaviors comprise walking, standing, lying prone and planing, and the prenatal exercise behaviors of the ewes are recognized by adopting a clustering algorithm and a neural network classification algorithm;

and the early warning unit is used for inputting the sample behavior data into a trained long-short-period memory network LSTM for training so as to classify the prenatal behaviors of the ewes to obtain classification results, and carrying out sheep reproduction health early warning according to the classification results.

2. The sheep reproductive health warning system of claim 1, wherein inputting the sample behavioral data into a trained long-short term memory network LSTM for training to classify the pre-natal behaviors of the ewe to obtain classification results, comprises:

the J epsilon {1, 2..J } is adopted as all the key feature points of the movement behaviors in a certain window, the number of windows represented by a certain movement behavior is represented by T epsilon {1, 2..T }, each LSTM finishes the data input, the analysis result of the movement behavior feature points of the same category output in the previous window and the output result of the adjacent movement nodes in the current time range are subjected to the data input, and the data are output to the next movement behavior feature representation window and the adjacent movement behavior feature points through the LSTM, wherein the corresponding expression is h _j,t ＝f(x _j,t h _j-1,t ,h _j,t-1 ) The operational procedure of LSTM is represented by a function f;

with two forgetting doors

Wherein->

The forgetting gate indicates whether the internal behavior feature change unit can generate data forgetting calculation in the LSTM calculation process in the space and space characteristics, namely, the change result generated by the behavior feature obtained in the last layer in a certain probability range;

the expression of the input gate is i _j,t ＝σ(W _f *[h _j,t-1 ,h _j-1,t ,X _j,t ]+b _i ) The input gate completes the integration processing of the input data in the algorithm structure, and the input data comprises the behavior change characteristics of the ewe in the previous time range, the motion behavior characteristic points in the current time range and the position coordinates of the motion behavior occurrence points of the ewe in the current time period.

3. The sheep reproductive health warning system of claim 2, further comprising:

the expression for updating the motion behavior characteristic state of the ewe is

Wherein the current state of the athletic performance characteristic in the LSTM is represented by C _j,t To show that the two forgetting doors can jointly realize the control of the movement behavior characteristic state;

the output gate is calculated by the following relation o _j,t ＝σ(W _o *[h _j,t-1 ,h _j-1,t ,X _j,t ]+b _o ) The output gate can realize the processing of the data to be output, and the output gate is represented by the following relation

After the correlation operation on the LSTM node is completed, the result is output to the output layer and the next time space iteration. />

4. The sheep reproduction health warning system according to claim 1, wherein the identifying the prenatal exercise behavior of the ewe according to the characteristic values to obtain sample behavior data comprises:

preset training sample set d= { (x) _i ,y _i ) I=1, 2..n }, where x is _i ∈R ⁿ Represents the ith feature vector, R ⁿ Representing sample space, y _i E { -1,1} represents two classes of class labels, n represents the lumped number of samples, and the hyperplane can be classified by a classification function f (x) =ω ^T x+b, where ω represents a hyperplane direction parameter, b represents a dividing hyperplane direction parameter, T represents a transpose of the vector, and when f (x) =0, x is a point on the hyperplane;

maximizing the geometric interval of all support vectors to obtain a maximum interval classifier, namely, a basic expression of SVM, as follows

The expression obtained by using Lagrangian multiplier method and converting is

Wherein alpha is _i Representing the lagrangian multiplier added by the ith constraint condition, alpha represents a set of n lagrangian multipliers, and L (ω, b, a) represents the resulting lagrangian function;

solving the minimum value of L (omega, b, alpha) to omega, b, then solving the maximum value to alpha, and finally converting into

Where i, j denote the numbers of any two samples.

5. The sheep reproductive health warning system of claim 4, wherein the three-axis acceleration data is mapped into a high-dimensional space using a gaussian kernel function that is expressed as

Wherein the method comprises the steps of ||x _i -x _j I represents the modulus of any two input vectors, σ represents the gaussian kernelAnd (3) the nuclear parameters of the function are combined in pairs by adopting a one-to-one method for three behaviors of drinking water, feeding and rumination to build three classifiers altogether, finally, the test sample is sequentially input into the three classifiers and votes are carried out, and the result output by the classifier with the largest number of votes is taken as the final behavior classification result.

6. The sheep reproduction health warning system according to claim 1, wherein the feature extraction of the behavior information to obtain a feature value includes:

the expression for extracting the time domain characteristic root mean square value from the sensor signal of the ewe behavior recognition is

Wherein N represents the number of behavior recognition samples, V _i A sensor acceleration value representing the instant i;

training a prediction model of acceleration coordinate difference value and corresponding RMS of behavior recognized by the ewe based on LSTM, calculating various parameters through a front-back time sequence relation of the acceleration coordinate recognized by the actual ewe behavior mapped with the RMS value in a training set and the predicted coordinate difference value, comprehensively considering estimated values and observed values of two adjacent moments, realizing updating of state variables, and estimating current requirements;

after the LSTM model is trained, the acceleration coordinate difference value can be predicted by introducing a prediction set RMS, the coordinate value is subjected to reverse-push prediction, and finally the visual prediction of the behavior acceleration data of the ewe is realized.

7. The sheep reproduction health warning system of claim 6, wherein the prediction of the acceleration coordinate difference is achieved by the introduction of a prediction set RMS, comprising:

input model trained acceleration signal eigenvalue RMS _i X, Y and coordinate value x in Z direction _i 、y _i And z _i ，i＝1,2...n；

The output includes process excitation noise covarianceQ, observed noise covariance R, and parameter state transition matrix A, and also includes M _x 、M _y M is as follows _z A three-axis coordinate difference prediction model;

based on the original signal, respectively for the original coordinate value x _i 、y _i And z _i Calculating the time domain feature RMS of acceleration _i Calculation is performed, i=1, 2..n;

in the front and rear time, the expression for calculating the coordinate value is

The construction of a deep LSTM network is completed, in which the input and the target output are respectively controlled by RMS _i Δx _i And Deltay _i Acting as a model for constructing prediction by correlation training in two coordinate directions, i.e. Mx (RMS) _i ) And My (RMS) _i )；

The calculated parameters A, Q and R are respectively

Then

Wherein DeltaX _i ＝[Δx _1, Δx ₂ ...Δx _i ]，ΔY _i ＝[Δy ₁ ,Δy ₂ ...Δy _i ]，ΔZ _i ＝[Δz ₁ ,Δz ₂ ...Δz _i ]，F _BiLinear Representing a binary linear regression function for calculating the correlation coefficients and residuals of the x, y, z coordinates of the previous moment and the x, y, z coordinates of the next moment as

Wherein a is ₁ 、a ₂ 、a ₃ 、b ₁ 、b ₂ 、b ₃ 、c ₁ 、c ₂ And c ₃ Representing the correlation coefficient, u _i And v _i Residual error, thereby determining->

And a= cov (β _x ,β _y ,β _z ) Two LSTM models are used to build a nonlinear observation model, namely

Wherein μ' _i 、v′ _i And omega' _i The residual is represented, and the observation noise R is obtained.

8. The sheep reproductive health warning system of claim 1, wherein collecting triaxial acceleration data of a ewe to obtain behavioral data of the ewe, and preprocessing the behavioral data to obtain pre-natal behavioral information of the ewe, comprises:

collecting triaxial acceleration data of a ewe when the ewe is stationary, calculating trigonometric function values sin alpha sin theta cos alpha cos theta of alpha and theta according to space geometric knowledge, and presetting three axis measurement values of an acceleration sensor when the ewe is stationary as alpha _1x 、α _1y And alpha _1z Then

Wherein A is _x 、A _y And A _z Representing the actual triaxial acceleration value, A _x Indicating the acceleration value of the horizontal forward direction gravity acceleration direction A _y Indicating the acceleration value in the left-right direction, A _z Representing the acceleration value in the vertical direction, and the measured values of the acceleration sensor in three directions are respectively alpha _x 、α _y And alpha _z Can be calculated from the tilt angle and the measured value

Respectively calculate the horizontal direction, the front-back directionAnd the acceleration value in the vertical direction, namely obtaining the acceleration data in the three corrected directions.

9. The sheep reproductive health warning system of claim 8, wherein employing a feature dimension reduction optimization recognition process comprises:

presetting the high-dimensional matrix as { X } ₁ ,X ₂ ...X _N Computing the average vector of the high-dimensional matrix

The expression is

The covariance matrix Var (X) of the high-dimensional matrix X is calculated by using the average vector as the expression

To determine the eigenvector U of the covariance matrix Var (X) _i And a characteristic value lambda _i N eigenvalues are arranged in descending order, namely lambda ₁ >λ ₂ >λ ₃ ...>λ _k >λ _k+1 >...>λ _N Presetting lambda ₁ -λ _k The corresponding k eigenvectors form a principal component matrix, and the formed change matrix is U= { U ₁ ,U ₂ ,U ₃ ...U _k Then the expression of principal component matrix Y is y=u ^T X, the size of k is adjusted to adjust the dimension of the principal component matrix, i.e., to adjust the reduced dimension.

10. A method for sheep reproductive health warning according to any one of claims 1 to 9, comprising the steps of:

acquiring triaxial acceleration data of a ewe to obtain behavior data of the ewe, and preprocessing the behavior data to obtain prenatal behavior information of the ewe, wherein the preprocessing comprises denoising, windowing and inclination correction;

extracting characteristics of the behavior information to obtain characteristic values, wherein three time domain characteristics of acceleration data, namely variances of three axial data of x, y and z, are selected, and frequency domain characteristics are taken to be the mean value of the main peak frequency of the three axial data and the first five values of the energy value of the frequency of the three axial data, and nineteen characteristic values are totally selected;

identifying the prenatal exercise behaviors of the ewes according to the characteristic values to obtain sample behavior data, wherein the exercise behaviors comprise walking, standing, lying prone and planing, and identifying the prenatal exercise behaviors of the ewes by adopting a clustering algorithm and a neural network classification algorithm;

and inputting the sample behavior data into a trained long-short-term memory network LSTM for training so as to classify the prenatal behaviors of the ewes to obtain classification results, and carrying out sheep reproduction health early warning according to the classification results.

Images