Lamb vitality identification method and system based on multi-source feature fusion

Abstract

The invention belongs to the technical field of precise and intelligent breeding, and provides a lamb vitality identification method and system based on multi-source feature fusion, and the method comprises the steps: collection of to-be-identified image data and to-be-identified physiological signal data, extraction of to-be-identified multi-source feature data, lamb vitality identification, and abnormal vitality discrimination and monitoring. According to the method, the image data and the physiological signal data of the target lamb are collected, so that the dimensionality and the quantity of the vitality identification detection data are enriched, and the identification accuracy and the robustness of the model are improved; the identification accuracy of the model is improved through the main-auxiliary cooperative lamb activity identification model; through the abnormal vitality identification model, real-time monitoring and early warning of the vitality state of the lambs are realized.

Description

Technical Field

[0001] The present invention relates to the field of precision and intelligent breeding technology, and in particular to a lamb vitality identification method and system based on multi-source feature fusion. Background Art

[0002] The output of mutton products is closely related to the survival rate of lambs, so improving the survival rate of lambs is one of the important ways to resolve this contradiction. In this process, lamb vitality is a key indicator to measure its current health status, future survival rate and future production capacity. Identifying lamb vitality can help farmers detect lamb health problems in time and take relevant measures, thereby improving lamb survival rate.

[0003] The traditional method of identifying lamb vitality mainly relies on manual observation of its apparent characteristics. Although this method is intuitive, it lacks scientific basis and it is difficult to accurately and reliably judge individual vitality. Summary of the invention

[0004] In order to overcome the shortcomings of the prior art, the purpose of the present invention is to provide a lamb vitality identification method and system based on multi-source feature fusion, which enriches the dimension and quantity of vitality identification detection data by collecting image data and physiological signal data of target lambs, and improves the recognition accuracy and robustness of the model; through the main and auxiliary collaborative lamb vitality identification model, the recognition accuracy of the model is improved; through the abnormal vitality identification model, real-time monitoring and early warning of the vitality status of lambs are realized.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] A lamb vitality identification method based on multi-source feature fusion, comprising:

[0007] Using a wearable sensor device to collect the physiological signal data to be identified of the target lamb, and using an image acquisition sensor device to obtain the image data to be identified of the target lamb;

[0008] Constructing a feature extraction fusion model, and using the feature extraction fusion model to extract features from the image data to be identified and the physiological signal data to be identified, to obtain multi-source feature data to be identified; the multi-source feature data to be identified includes: feature extraction results of physiological signal data and feature extraction results of image data;

[0009] A lamb vitality identification model with primary and secondary collaboration is constructed, and the multi-source feature data to be identified is identified using the lamb vitality identification model with primary and secondary collaboration to obtain a vitality identification result; the lamb vitality identification model with primary and secondary collaboration includes: a first lamb vitality identification sub-model and a second lamb vitality identification sub-model that are coupled to each other;

[0010] An abnormal vitality recognition model is constructed, and the abnormal vitality recognition model is used to perform encoding, decoding, reconstruction operations and abnormal vitality discrimination on the vitality recognition result and the multi-source feature data to be identified, so as to obtain an abnormal vitality discrimination result.

[0011] Preferably, a feature extraction fusion model is constructed, and the feature extraction fusion model is used to extract features from the image data to be identified and the physiological signal data to be identified to obtain multi-source feature data to be identified, including:

[0012] Constructing a feature extraction fusion model; the feature extraction fusion model includes: a physiological feature extraction path and an image feature extraction path; the physiological feature extraction path includes: a memory filter, an information injector and an output regulator; the image feature extraction path includes: a feature extraction unit, a nonlinear activation unit and a feature downsampling unit;

[0013] The physiological signal data to be identified is input into the memory filter and the information injector to obtain information candidates; the calculation formula of the information candidate is:

[0014]

[0015] in, is the candidate item of the information; t-1 is the extraction result of the previous cycle; X t is the input sequence of the current cycle with a sequence length of T; is element-wise subtraction; A weight matrix of the information candidates randomly generated by the memory filter and the information injector for the input sequence; A base coefficient of the information candidate randomly generated by the memory filter and the information injector for the input sequence;

[0016] According to the information candidate, the memory state is updated using the memory filter and the information injector; the calculation formula of the memory state is:

[0017]

[0018] Among them, S t-1 It is the memory state of the previous cycle; and ⊙ are element-wise addition and element-wise multiplication respectively; A weight matrix of the memory state randomly generated for the input sequence by the memory filter and the information injector; A reference coefficient of the memory state randomly generated for the input sequence through the memory filter and the information injector;

[0019] The memory state S that will be updated is completed t Input into the output regulator to obtain the physiological signal data feature extraction result; the calculation formula of the physiological signal data feature extraction result is:

[0020]

[0021] Among them, h t The physiological signal data feature extraction result of the current cycle; A weight matrix of the extraction results randomly generated by the output regulator for the input sequence; A reference coefficient of the extraction result randomly generated by the output regulator for the input sequence; is the tensor product;

[0022] The image data to be identified is input into the feature extraction unit to obtain a feature map; the calculation formula of the feature map is:

[0023]

[0024] in,

[0025]

[0026] Y t is the characteristic graph of the current cycle; t (i, j) is the value of the feature graph at (i, j); K H , K W and C are the sizes of the convolution kernel; W k and b k are the weight and base coefficient of the convolution kernel respectively;

[0027] Y is the image data to be identified that is randomly generated by the feature extraction unit t The weight matrix of (i,j); Y is the image data to be identified that is randomly generated by the feature extraction unit t The base coefficient of (i,j);

[0028] The feature map is input into the nonlinear activation unit for nonlinear transformation to obtain a transformation result; the calculation formula of the transformation result is:

[0029]

[0030] Among them, Z t is the transformation result of the current cycle; Z t(i, j) is the value of the transformation result at (i, j); ReLU(·) is a nonlinear activation function;

[0031] The transformation result is input into the feature downsampling unit to perform a dimensionality reduction operation to obtain a feature downsampling result; the calculation formula of the feature downsampling result is:

[0032]

[0033] in,

[0034] is the feature downsampling result; is the value of the feature downsampling result at (i, j); P H and P W are the sizes of downsampling windows;

[0035] Performing a flattening operation on the feature downsampling result to obtain the image data feature extraction result;

[0036] The physiological signal data feature extraction result and the image data feature extraction result are associated and fused to obtain the multi-source feature data to be identified.

[0037] Preferably, a lamb vitality identification model with primary and secondary collaboration is constructed, and the multi-source feature data to be identified is identified using the primary and secondary collaboration lamb vitality identification model to obtain a vitality identification result, including:

[0038] The first lamb vitality identification sub-model and the second lamb vitality identification sub-model are set; the first lamb vitality identification sub-model includes: a first auxiliary classification sub-model, a second auxiliary classification sub-model and a first main classification sub-model coupled to each other; the second lamb vitality identification sub-model includes: a third auxiliary classification sub-model, a fourth auxiliary classification sub-model and a second main classification sub-model coupled to each other;

[0039] Perform unlabeled operations and true labeled operations on the pre-collected multi-source feature data to obtain an unlabeled multi-source feature dataset and a true labeled multi-source feature dataset;

[0040] Inputting the physiological signal feature extraction result and the image data feature extraction result in the unlabeled multi-source feature data set into the first auxiliary classification sub-model and the second auxiliary classification sub-model respectively to obtain a first prediction probability score and a second prediction probability score;

[0041] Calculating a comprehensive predicted probability score based on the first predicted probability score and the second predicted probability score;

[0042] Screening and combining the preset marker set according to the comprehensive prediction probability score to obtain a prediction pseudo marker set;

[0043] Fusion of the predicted pseudo-label set and the pre-collected multi-source feature data to obtain a pseudo-label multi-source feature data set;

[0044] Inputting the pseudo-labeled multi-source feature data set and the real-labeled multi-source feature data set into the first main classification sub-model to perform matching degree calculation, thereby obtaining a matching degree calculation result;

[0045] When the matching degree calculation result is equal to 0, iteratively optimizing the first auxiliary classification sub-model and the second auxiliary classification sub-model;

[0046] When the matching degree calculation result is 1, outputting the first lamb vitality identification sub-model that has been trained;

[0047] According to the training process of the first lamb vitality identification sub-model, the third auxiliary classification sub-model, the fourth auxiliary classification sub-model and the second main classification sub-model are trained to obtain the trained second lamb vitality identification sub-model;

[0048] The first lamb vitality identification sub-model that has been trained is coupled with the second lamb vitality identification sub-model that has been trained to obtain the main-auxiliary coordinated lamb vitality identification model.

[0049] Preferably, it also includes:

[0050] Inputting the multi-source feature data to be identified into the first lamb vitality identification sub-model and the second lamb vitality identification sub-model respectively to obtain a first probability score to be identified and a second probability score to be identified;

[0051] The first probability score to be identified and the second probability score to be identified are weightedly calculated to obtain a comprehensive probability score; the calculation formula of the comprehensive probability score is:

[0052]

[0053] in, is the comprehensive probability score; is the first probability score to be identified; is the second probability score to be identified; L is the data volume of the first probability score to be identified and the second probability score to be identified; A first probability score which is the first probability score to be identified;

[0054] The maximum probability mark of the comprehensive probability score is extracted to obtain the vitality recognition result.

[0055] Preferably, the calculation formula of the comprehensive prediction probability score is:

[0056]

[0057] in, is the comprehensive prediction probability score; and are the first predicted probability score and the second predicted probability score, respectively; t is the feature extraction result of the physiological signal data; t is the feature extraction result of the image data; c is the preset tag set; u c is the center of the preset marker set; d(h t ,u c ) and d(v t ,u c ) are the distances between the physiological signal data feature extraction result and the image data feature extraction result and the center of the preset tag set; θ is the similarity coefficient; L is the number of the predicted pseudo tag set; ω is the weight coefficient; is a predicted pseudo label in the generated predicted pseudo label set.

[0058] Preferably, the calculation formula of the matching degree calculation result is:

[0059]

[0060] Among them, Ψ(D real ,D pseudo ) is the matching degree between the real labeled multi-source feature dataset and the pseudo labeled multi-source feature dataset; and δ are weight coefficients; is the real labeled multi-source feature data in the real labeled multi-source feature data set of the current cycle; is the pseudo-labeled multi-source feature data in the pseudo-labeled multi-source feature data set of the current cycle; a is the maximum number of the physiological signal data to be identified and the image data to be identified.

[0061] Preferably, an abnormal vitality recognition model is constructed, and the abnormal vitality recognition model is used to perform encoding, decoding, reconstruction operations and abnormal vitality discrimination on the vitality recognition result and the multi-source feature data to be identified, to obtain an abnormal vitality discrimination result, including:

[0062] Constructing an abnormal vitality recognition model; the abnormal vitality recognition model includes: an input layer, an encoding layer, a decoding layer and an output layer;

[0063] Using the input layer to receive the physiological signal data to be identified, the image data to be identified, and the vitality recognition result to obtain input data;

[0064] Using the encoding layer to reduce the dimension of the input data to obtain a potential representation;

[0065] Reconstructing the latent representation using the decoding layer to obtain a reconstruction result;

[0066] Perform error calculation according to the reconstruction result to obtain a reconstruction error;

[0067] Utilizing the output layer to analyze the degree of deviation of the reconstruction error, and obtaining the abnormal activity discrimination result;

[0068] When the reconstruction error is within a preset threshold range, determining that the abnormal vitality discrimination result is normal;

[0069] When the reconstruction error exceeds a preset threshold range, the abnormal activity discrimination result is determined to be abnormal, and an alarm is triggered.

[0070] Preferably, a lamb vitality identification system based on multi-source feature fusion comprises:

[0071] A biological signal acquisition unit, used to acquire the physiological signal data to be identified of the target lamb using a wearable sensor device, and to acquire the image data to be identified of the target lamb using an image acquisition sensor device;

[0072] A feature extraction and fusion unit, used for constructing a feature extraction and fusion model, and using the feature extraction and fusion model to extract features from the image data to be identified and the physiological signal data to be identified, so as to obtain multi-source feature data to be identified;

[0073] A lamb vitality identification unit is used to construct a primary and secondary collaborative lamb vitality identification model, and use the primary and secondary collaborative lamb vitality identification model to identify the multi-source feature data to be identified to obtain a vitality identification result;

[0074] The abnormal activity monitoring unit is used to construct an abnormal activity recognition model, and use the abnormal activity recognition model to perform encoding, decoding, reconstruction operations and abnormal activity discrimination on the activity recognition result and the multi-source feature data to be identified, so as to obtain an abnormal activity discrimination result.

[0075] The present invention discloses the following technical effects:

[0076] The present invention provides a method for enriching the dimension and quantity of vitality recognition detection data by collecting image data and physiological signal data of target lambs, thereby improving the recognition accuracy and robustness of the model; through a primary and auxiliary collaborative lamb vitality recognition model, the recognition accuracy of the model is improved; through an abnormal vitality recognition model, real-time monitoring and early warning of the vitality status of lambs are achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0078] Figure 1 A schematic diagram of a lamb vitality identification process based on multi-source feature fusion provided by an embodiment of the present invention;

[0079] Figure 2 A schematic diagram of the structure of a lamb vitality identification model based on primary and secondary collaboration of multi-source feature fusion provided in an embodiment of the present invention;

[0080] Figure 3 A schematic diagram of the structure of a lamb vitality identification system based on multi-source feature fusion provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0081] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0082] The purpose of the present invention is to provide a lamb vitality identification method and system based on multi-source feature fusion, which enriches the dimension and quantity of vitality identification detection data by collecting image data and physiological signal data of target lambs, and improves the recognition accuracy and robustness of the model; improves the recognition accuracy of the model through a main and auxiliary collaborative lamb vitality identification model; and realizes real-time monitoring and early warning of the vitality status of lambs through an abnormal vitality identification model.

[0083] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0084] Figure 1 A schematic diagram of a lamb vitality identification process based on multi-source feature fusion provided by an embodiment of the present invention is shown in FIG. Figure 1 As shown, the present invention provides a lamb vitality identification method based on multi-source feature fusion, comprising:

[0085] Step 100: using a wearable sensor device to collect physiological signal data to be identified of the target lamb, and using an image acquisition sensor device to obtain image data to be identified of the target lamb;

[0086] Step 200: construct a feature extraction fusion model, and use the feature extraction fusion model to extract features from the image data to be identified and the physiological signal data to be identified, to obtain multi-source feature data to be identified; the multi-source feature data to be identified includes: feature extraction results of the physiological signal data and feature extraction results of the image data;

[0087] Step 300: constructing a primary-auxiliary collaborative lamb vitality identification model, and using the primary-auxiliary collaborative lamb vitality identification model to identify the multi-source feature data to be identified to obtain a vitality identification result; the primary-auxiliary collaborative lamb vitality identification model includes: a first lamb vitality identification sub-model and a second lamb vitality identification sub-model that are coupled to each other;

[0088] Step 400: construct an abnormal vitality recognition model, and use the abnormal vitality recognition model to perform encoding, decoding, reconstruction operations and abnormal vitality discrimination on the vitality recognition result and the multi-source feature data to be identified, so as to obtain the abnormal vitality discrimination result.

[0089] Furthermore, a feature extraction fusion model is constructed, and the feature extraction fusion model is used to extract features from the image data to be identified and the physiological signal data to be identified, so as to obtain multi-source feature data to be identified, including:

[0090] Constructing a feature extraction fusion model; the feature extraction fusion model includes: a physiological feature extraction path and an image feature extraction path; the physiological feature extraction path includes: a memory filter, an information injector, and an output regulator; the image feature extraction path includes: a feature extraction unit, a nonlinear activation unit, and a feature downsampling unit;

[0091] Input the physiological signal data to be identified into the memory filter and the information injector to obtain information candidates; the calculation formula of the information candidate is:

[0092]

[0093] in, is the information candidate; h t-1 is the extraction result of the previous cycle; X t is the input sequence of the current cycle with a sequence length of T; is element-wise subtraction; The weight matrix of information candidates randomly generated by the memory filter and information injector for the input sequence; The base coefficients of the information candidates randomly generated by the memory filter and the information injector for the input sequence;

[0094] According to the information candidates, the memory state is updated using the memory filter and information injector; the calculation formula of the memory state is:

[0095]

[0096] Among them, S t-1 It is the memory state of the previous cycle; and ⊙ are element-wise addition and element-wise multiplication respectively; The weight matrix of the memory state randomly generated by the memory filter and information injector for the input sequence; The base coefficients of the memory states randomly generated for the input sequence through the memory filter and information injector;

[0097] The memory state S of the end of the update t Input into the output regulator to obtain the physiological signal data feature extraction result; the calculation formula of the physiological signal data feature extraction result is:

[0098]

[0099] Among them, h t The feature extraction result of the physiological signal data of the current cycle; The weight matrix of the extracted results randomly generated by the output regulator for the input sequence; The base coefficient of the extraction result randomly generated by the output regulator for the input sequence; is the tensor product;

[0100] The image data to be identified is input into the feature extraction unit to obtain a feature map; the calculation formula of the feature map is:

[0101]

[0102] in,

[0103]

[0104] Y t is the feature map of the current cycle; Y t (i, j) is the value of the feature map at (i, j); K H , K W and C are the sizes of the convolution kernel; W k and b k are the weight and base coefficient of the convolution kernel respectively;

[0105] Y is the image data to be identified and randomly generated by the feature extraction unit t The weight matrix of (i,j); Y is the image data to be identified and randomly generated by the feature extraction unit t The base coefficient of (i,j);

[0106] The feature map is input into the nonlinear activation unit for nonlinear transformation to obtain the transformation result; the calculation formula of the transformation result is:

[0107]

[0108] Among them, Z t is the transformation result of the current cycle; Z t (i, j) is the value of the transformation result at (i, j); ReLU(·) is the nonlinear activation function;

[0109] The transformation result is input into the feature downsampling unit for dimensionality reduction operation to obtain the feature downsampling result; the calculation formula of the feature downsampling result is:

[0110]

[0111] in,

[0112] is the result of feature downsampling; is the value of the feature downsampling result at (i, j); P H and P W are the sizes of downsampling windows;

[0113] Flatten the feature downsampling result to obtain the image data feature extraction result;

[0114] The feature extraction results of physiological signal data and image data are associated and fused to obtain multi-source feature data to be identified.

[0115] Furthermore, a lamb vitality identification model with primary and secondary collaboration is constructed, and the lamb vitality identification model with primary and secondary collaboration is used to identify the multi-source feature data to be identified, and the vitality identification results are obtained, including:

[0116] A first lamb vitality identification sub-model and a second lamb vitality identification sub-model are set; the first lamb vitality identification sub-model includes: a first auxiliary classification sub-model, a second auxiliary classification sub-model and a first main classification sub-model coupled to each other; the second lamb vitality identification sub-model includes: a third auxiliary classification sub-model, a fourth auxiliary classification sub-model and a second main classification sub-model coupled to each other;

[0117] Perform unlabeled operations and true labeled operations on the pre-collected multi-source feature data to obtain an unlabeled multi-source feature dataset and a true labeled multi-source feature dataset;

[0118] Inputting the physiological signal feature extraction result and the image data feature extraction result in the unlabeled multi-source feature data set into the first auxiliary classification sub-model and the second auxiliary classification sub-model respectively to obtain a first prediction probability score and a second prediction probability score;

[0119] Calculating a comprehensive predicted probability score based on the first predicted probability score and the second predicted probability score;

[0120] The preset marker set is screened and assembled according to the comprehensive prediction probability score to obtain a prediction pseudo marker set;

[0121] The predicted pseudo-label set and the pre-collected multi-source feature data are fused to obtain a pseudo-label multi-source feature data set;

[0122] Inputting the pseudo-labeled multi-source feature data set and the real-labeled multi-source feature data set into the first main classification sub-model to perform matching degree calculation, thereby obtaining a matching degree calculation result;

[0123] When the result of the matching degree calculation is equal to 0, the first auxiliary classification sub-model and the second auxiliary classification sub-model are iteratively optimized;

[0124] When the matching degree calculation result is 1, the first lamb vitality identification sub-model that has been trained is output;

[0125] According to the training process of the first lamb vitality identification sub-model, the third auxiliary classification sub-model, the fourth auxiliary classification sub-model and the second main classification sub-model are trained to obtain a trained second lamb vitality identification sub-model;

[0126] The trained first lamb vitality identification sub-model and the trained second lamb vitality identification sub-model are coupled to obtain a main-auxiliary coordinated lamb vitality identification model.

[0127] Specifically, it also includes:

[0128] Inputting the multi-source feature data to be identified into the first lamb vitality identification sub-model and the second lamb vitality identification sub-model respectively, to obtain a first probability score to be identified and a second probability score to be identified;

[0129] The first probability score to be identified and the second probability score to be identified are weighted to obtain a comprehensive probability score; the calculation formula for the comprehensive probability score is:

[0130]

[0131] in, is the comprehensive probability score; is the first probability score to be identified; is the second probability score to be identified; L is the data volume of the first probability score to be identified and the second probability score to be identified; A first probability score is a first probability score to be identified;

[0132] The maximum probability mark of the comprehensive probability score is extracted to obtain the vitality identification result.

[0133] Preferably, the calculation formula for the comprehensive prediction probability score is:

[0134]

[0135] in, is the comprehensive prediction probability score; and are the first predicted probability score and the second predicted probability score respectively; h t is the feature extraction result of physiological signal data; v t is the result of feature extraction of image data; c is the preset tag set; u c is the center of the preset marker set; d(h t ,u c ) and d(v t ,u c ) are the distances between the physiological signal data feature extraction results and the image data feature extraction results and the center of the preset label set; θ is the similarity coefficient; L is the number of predicted pseudo label sets; ω is the weight coefficient; is the predicted pseudo label in the generated predicted pseudo label set.

[0136] Specifically, the calculation formula for the matching degree calculation result is:

[0137]

[0138] Among them, Ψ(D real ,D pseudo ) is the matching degree between the real labeled multi-source feature dataset and the pseudo labeled multi-source feature dataset; and δ are weight coefficients; The real labeled multi-source feature data in the real labeled multi-source feature dataset of the current cycle; is the pseudo-labeled multi-source feature data in the pseudo-labeled multi-source feature data set of the current cycle; a is the maximum number of physiological signal data to be identified and image data to be identified.

[0139] Optionally, an abnormal vitality recognition model is constructed, and the abnormal vitality recognition model is used to perform encoding, decoding, reconstruction operations and abnormal vitality discrimination on the vitality recognition result and the multi-source feature data to be identified, to obtain the abnormal vitality discrimination result, including:

[0140] Construct an abnormal vitality recognition model; the abnormal vitality recognition model includes: an input layer, an encoding layer, a decoding layer and an output layer;

[0141] Using the input layer to receive the physiological signal data to be identified, the image data to be identified, and the vitality identification result to obtain input data;

[0142] Use the encoding layer to reduce the dimension of the input data to obtain the potential representation;

[0143] The latent representation is reconstructed using the decoding layer to obtain a reconstructed result;

[0144] Perform error calculation based on the reconstruction result to obtain the reconstruction error;

[0145] The output layer is used to analyze the deviation degree of reconstruction error and obtain the abnormal activity discrimination result;

[0146] When the reconstruction error is within the preset threshold, the abnormal vitality discrimination result is determined to be normal;

[0147] When the reconstruction error exceeds the preset threshold range, the abnormal vitality discrimination result is determined to be abnormal and an alarm is triggered.

[0148] Furthermore, the abnormal activity recognition model performs encoding and decoding reconstruction operations on the two types of data, and compares the result with the normal data. If the error is greater than a threshold, it is abnormal activity.

[0149] Furthermore, the number of physiological signal data and image data is consistent and one-to-one corresponding; the physiological signal data and image data of the t sequence are feature extracted and combined to obtain the feature data of the t sequence, which are marked to obtain the marked data of the t sequence.

[0150] Figure 2 The overall framework diagram of the lamb vitality identification method based on multi-source feature fusion provided in this embodiment; Figure 2 As shown, this embodiment provides a lamb vitality identification method based on multi-source feature fusion, including: a feature extraction fusion model, a main-auxiliary collaborative lamb vitality identification model and an abnormal vitality identification model working in coordination with each other.

[0151] Specifically, a method for constructing a feature extraction fusion model includes the following processes:

[0152] Step 1: Obtaining the vitality-related data of the target lamb in real time in a dynamic environment, including but not limited to: collecting the lamb's physiological signal data, such as heart rate, body temperature and movement data, through wearable sensor devices, which can reveal its internal physiological state; obtaining the lamb's image data through image acquisition sensor devices, which can reflect its external activities, behavioral performance and appearance characteristics;

[0153] Step 2: Input the target lamb physiological signal data acquired by the wearable sensor device in real time and the target lamb image data acquired by the image acquisition sensor device into different feature extraction paths respectively, and finally complete the feature fusion by associating the one-dimensional feature data output by each path; specifically:

[0154] Physiological signal data feature extraction path:

[0155] The physiological signal data feature extraction path is composed of mutually coupled units, including: memory filter M t 、Information Injector I t and the output regulator R t ; Input sequence X of length T t =[x1,x2,…x T ] Input memory filter M respectively t and Information Injector I t , get information candidates Combined with Memory Filter M t 、Information Injector I t and information candidates Update memory state S t ; Combined with memory state S t and the output regulator R t Output extraction result h t ;

[0156] Information candidates The formula is as follows:

[0157]

[0158] Among them, h t-1 Indicates the extraction result of the previous cycle (optional, set the initial h0 = 1); X t represents the t-th input sequence of length T, t∈[1,a]; Represents element-level subtraction and obtains the combination of subtraction operation dimensions; Represents the input sequence X t After the memory filter M t and Information Injector I t Randomly generated information candidates The weight matrix of Represents the input sequence X t After the memory filter M t and Information Injector I t Randomly generated information candidates The base coefficient of

[0159] Memory status S tThe formula is as follows:

[0160]

[0161] Among them, S t-1 Indicates the memory state of the previous cycle (optional, set initial S0=1); and ⊙ represent element-level addition and element-level multiplication, respectively, and obtain the addition operation dimension combination and the multiplication operation dimension combination; Represents the input sequence X t After the memory filter M t and Information Injector I t The randomly generated memory state S t The weight matrix of Represents the input sequence X t After the memory filter M t and Information Injector I t The randomly generated memory state S t The base coefficient of

[0162] Extraction results t The formula is as follows:

[0163]

[0164] in, Represents the input sequence X t Through the output regulator R t The randomly generated extraction result h t The weight matrix of Represents the input sequence X t Through the output regulator R t The randomly generated extraction result h t The base coefficient of Represents tensor product and obtains the combination of concatenation operation dimensions;

[0165] Image data feature extraction path:

[0166] The image data feature extraction path is composed of mutually coupled units, including: feature extraction unit F t , nonlinear activation unit N t and feature downsampling unit D t ; A piece of image data P with a height × width × number of channels H × W × C t (t∈[1,a]) is sequentially input into the feature extraction unit F t , output feature map Y t ; Nonlinear activation unit N t For feature map Y t Perform nonlinear transformation and output the transformation result Z t; Transform the result Z t Input feature downsampling unit D t Perform dimensionality reduction operations and finally output feature downsampling results The feature downsampling result Perform a flattening operation and output the extracted result v t ;

[0167] It is worth noting that there is a one-to-one correspondence between physiological signal data and image data features, that is, the input sequence X t =[x1,x2,…x T ] and image data P t Correspondingly, the extraction result h t With v t Match each other; the physiological signal data feature extraction path and the image data feature extraction path must be synchronized in the time dimension to ensure the smooth completion of the feature fusion operation;

[0168] Feature map Y t The formula is as follows:

[0169]

[0170] Among them, Y t (i,j) represents the feature map Y t The value at (i, j); K H , K W and C represent the size of the convolution kernel; W k and b k Represent the weight and base coefficient of the convolution kernel respectively; Represents image data P t After the feature extraction unit F t The randomly generated Y t The weight matrix of (i,j); Represents image data P t After the feature extraction unit F t The randomly generated Y t The base coefficient of (i,j);

[0171] Transformation result Z t The formula is as follows:

[0172]

[0173] Among them, Z t (i,j) represents the transformation result Z t The value at (i, j); ReLU(·) represents a nonlinear activation function,

[0174] Feature downsampling results The formula is as follows:

[0175]

[0176] Among them, P H and P W Indicates the size of the downsampling window;

[0177] Extract result v t The formula is as follows:

[0178]

[0179] The extraction results h output by each path t and v t After the association operation, feature fusion is completed and multi-source feature data is output t∈[1,a].

[0180] Furthermore, the main-auxiliary collaborative lamb vitality identification model is composed of a mutually coupled first lamb vitality identification sub-model T1 and a second lamb vitality identification sub-model T2; the first lamb vitality identification sub-model T1 is composed of a mutually coupled first auxiliary classification sub-model β1, a second auxiliary classification sub-model β2 and a first main classification sub-model α1; the second lamb vitality identification sub-model T2 is composed of a mutually coupled third auxiliary classification sub-model β3, a fourth auxiliary classification sub-model β4 and a second main classification sub-model α2;

[0181] Specifically, a method for constructing a lamb vitality identification model with primary and secondary collaboration includes the following steps:

[0182] Step 1: perform unlabeled operation and real labeling operation on the multi-source feature data stream obtained by the feature extraction fusion model, and input them into the first feature space O1 and the second feature space O2 respectively; the feature space is similar to a data set constructed by all the multi-source feature data extracted by the feature extraction fusion model; the feature space can receive the multi-source feature data generated by the feature extraction fusion model; the feature space can transfer the multi-source feature data or labels to the auxiliary classification sub-model and the main classification sub-model; the unlabeled operation, that is, no operation is performed on the multi-source feature data extracted by the feature extraction fusion model, and it is transferred to the first feature space O1; the real labeling operation, that is, the multi-source feature data extracted by the feature extraction fusion model is manually labeled, and one multi-source feature data corresponds to one vitality label; vitality labels include but are not limited to: stress, malaise and excitement;

[0183] Preferably, in order to facilitate the following description, the multi-source feature data of the unlabeled operation is referred to as the unlabeled multi-source feature data γ t ; The multi-source feature data of the true labeling operation is called the true labeling multi-source feature data Among them, yt Represents multi-source feature data γ t Multiple corresponding true labels; all unlabeled multi-source feature data γ t The constructed set is called the unlabeled multi-source feature data set γ; the set constructed by all real labels is called the real label set Y; all real label multi-source feature data The constructed collection is called the real labeled multi-source feature dataset

[0184] Step 2: Input the unlabeled data of the first feature space O1 into the first auxiliary classification sub-model β1 and the second auxiliary classification sub-model β2 constructed based on the Pseudo-Labeling idea, and output pseudo labels in combination with the comprehensive probability distribution method;

[0185] Specifically, the unlabeled multi-source feature data of the first feature space O1 The first auxiliary classification sub-model β1 and the second auxiliary classification sub-model β2 are input respectively; wherein the first auxiliary classification sub-model β1 receives the extraction result h t , and output the predicted probability score The second auxiliary classification sub-model β2 receives the extraction result v t , and output the predicted probability score Based on the above predicted probability distribution and Get the comprehensive prediction probability score p final ;

[0186] Comprehensive prediction probability score p final The formula is as follows:

[0187]

[0188] in, and are the predicted probability distributions output by the first auxiliary classification sub-model β1 and the second auxiliary classification sub-model β2 respectively; d(h t ,u c ) and d(v t ,u c ) represent the input vector h t and v t With c marking set center u c The distance between them; θ represents the similarity coefficient; L represents the number of pseudo-label sets; ω represents the weight coefficient; represents the generated pseudo-label;

[0189] Further, in this embodiment, the pseudo-marker set of lamb vitality includes but is not limited to pseudo-markers such as "stress", "laziness" or "excitement"; based on the first auxiliary classification sub-model β1 and the second auxiliary classification sub-model β2, a comprehensive prediction probability score is output; each pseudo-marker corresponds to a prediction probability score, which reflects the model's confidence in the pseudo-marker; the prediction probability score ranges from 0 to 1, and the closer it is to 1, the more confident the model is that the input data belongs to the pseudo-marker;

[0190] The pseudo-labels generated sequentially are aggregated to construct a pseudo-label set Tag Collection The formula is as follows:

[0191]

[0192] The pseudo-label set With multi-source feature data Combine and construct pseudo-labeled multi-source feature data All pseudo-labeled multi-source feature data The constructed collection is called pseudo-labeled multi-source feature dataset

[0193] Step 3: The real labeled multi-source feature dataset D real and pseudo-labeled multi-source feature dataset D pseudo Input the first main classification sub-model α1 to calculate the matching degree Ψ(D real ,D pseudo );

[0194] Matching degree Ψ(D real ,D pseudo ) is as follows:

[0195]

[0196] in, and δ represent weight coefficients; when Ψ(D real ,D pseudo )=1, indicating that the real labeled multi-source feature dataset D real and pseudo-labeled multi-source feature dataset D pseudo There is a good match between them; when Ψ(D real ,D pseudo )=0, indicating that the real labeled multi-source feature dataset D real and pseudo-labeled multi-source feature dataset D pseudo There is no good match between them, so go back to step 2 to retrain the model;

[0197] The composition architecture and composition mode of the third auxiliary classification sub-model β3, the fourth auxiliary classification sub-model β4 and the second main classification sub-model α2 are the same as those of the first auxiliary classification sub-model β1, the second auxiliary classification sub-model β2 and the first main classification sub-model α1;

[0198] When Ψ(D real ,D pseudo )=1, indicating that the real labeled multi-source feature dataset D real and pseudo-labeled multi-source feature dataset D pseudo There is a good match between them; at this time, after processing in steps 1 to 3, a model capable of identifying unlabeled multi-source feature data is obtained;

[0199] Step 4: Obtain new unlabeled multi-source feature data The new unlabeled multi-source feature data The first lamb vitality identification sub-model T1 and the second lamb vitality identification sub-model T2 are input respectively; the probability scores of the output results of the first lamb vitality identification sub-model T1 and the second lamb vitality identification sub-model T2 are respectively and By calculating the weights and combining them with the model output, we get a comprehensive probability score

[0200] Comprehensive probability score The formula is as follows:

[0201]

[0202] Based on the comprehensive probability score Get the mark with the highest probability The vitality recognition result as the final output; the maximum probability mark The formula is as follows:

[0203]

[0204] Among them, arg max means taking the maximum value;

[0205] In this embodiment, the lamb vitality recognition result is represented by a label; the category label output by the vitality recognition sub-model is the most likely lamb vitality category it considers; each category label corresponds to a probability score, which reflects the model's confidence in the category label; the probability score ranges from 0 to 1, and the closer it is to 1, the more confident the model is that the input data belongs to the category; specifically:

[0206] The vitality identification sub-model calculates a probability score for each possible vitality category and uses these scores as a measure of the model's confidence in each category; for example, the model's output may be 0.72 for "stress", 0.18 for "depression", and 0.10 for "excitement", which means that the model believes that the input data most likely corresponds to the "stress" category with a confidence level of 72%; users can evaluate the model's prediction results based on these probability scores and can set a confidence threshold to decide whether to accept the model's prediction results; for example, if the set threshold is 0.7 and the probability of "stress" is 0.72, which exceeds the threshold, then "depression" can be used as the final classification result; if the probability score of any category does not exceed the set threshold, the user can decide whether to accept the model's prediction or request further data analysis.

[0207] Furthermore, a method for constructing an abnormal vitality recognition model includes the following process:

[0208] 1) The hierarchical structure of the abnormal vitality recognition model includes an input layer, an encoding layer, a decoding layer, and an output layer connected in sequence; the model is specifically used to monitor the abnormal vitality state of lambs, and is designed to recognize and distinguish abnormal situations by learning and reconstructing normal vitality patterns;

[0209] 2) At the input layer of the model, the received data includes two main parts: one is the multi-source feature data to be identified extracted by the data feature extraction path; the other is the output results of the vitality recognition model, which are the model's predicted labels for the lamb vitality category and their corresponding probability scores; these input data are integrated and provided to the autoencoder as the basis for its learning and processing;

[0210] 3) After entering the encoding layer, the model will perform efficient feature extraction and compression on the input data, converting it into a low-dimensional latent representation; this process helps the model capture the key patterns and features in the input data while reducing redundant information; the output of the encoder, i.e., the latent representation, can effectively summarize normal vitality patterns;

[0211] 4) Next, the decoding layer attempts to reconstruct the input data based on the latent representation, converting it back into a high-dimensional space. Through this process, the model reconstructs a result that is close to the original input data. The key to this reconstruction process is that if the input data conforms to the normal vitality pattern, the decoder can reconstruct the original input well; however, if there are anomalies in the input data, the reconstruction result is often significantly different from the original data.

[0212] 5) Finally, at the output layer, the model generates monitoring results of abnormal vitality of lambs; by analyzing the reconstruction error of the decoder, the model can determine whether the input data deviates from the normal pattern; a larger reconstruction error usually indicates that the input data may represent an abnormal vitality state; therefore, the output layer will give a result indicating abnormal vitality for further monitoring and intervention. This abnormal vitality recognition model can not only capture subtle changes in vitality, but also improve the detection accuracy of potential abnormalities by combining the output results of the vitality recognition sub-model; this comprehensive multi-level structure enables the model to effectively screen out individuals with possible abnormalities when processing complex physiological data, thereby providing an important reference for the health management of lambs.

[0213] Optionally, the construction code of the lamb abnormal vitality recognition model is as follows:

[0214] class AnomalyVitalityAutoencoder:

[0215] def__init__(self,input_dim,hidden_dim):

[0216] # Initialize the hidden layers of the encoder and decoder

[0217] self.encoder=DenseLayer(input_dim,hidden_dim)

[0218] self.decoder=DenseLayer(hidden_dim,input_dim)

[0219] def forward(self,input_data):

[0220] #Forward propagation: from input data to hidden space and then reconstruct output

[0221] latent_space=self.encoder(input_data)

[0222] return self.decoder(latent_space)

[0223] def compute_reconstruction_error(self,input_data,reconstructed_data):#Calculate the reconstruction error to determine whether it is abnormal

[0224] return mean_squared_error(input_data reconstructed_data)

[0225] def predict_anomaly(self,reconstruction_error,threshold):

[0226] #Judge whether it is abnormal based on reconstruction error and set threshold

[0227] return "abnormal" if reconstruction_error>threshold else "normal" classDenseLayer:

[0228] def__init__(self,input_dim,output_dim):

[0229] # Initialize weights and biases

[0230] self.weights=initialize_weights(input_dim,output_dim)

[0231] self.bias=initialize_bias(output_dim)

[0232] def__call__(self, input_data):

[0233] #Forward propagation: calculate the output after the activation function

[0234] return activation_function(input_data@self.weights+self.bias)

[0235] if__name__=="__main__":

[0236] #Set the dimension of input data and hidden layer dimension

[0237] input_dim = the dimension of the input data

[0238] hidden_dim = Dimension of the latent space

[0239] #Create an abnormal vitality recognition model

[0240] model=AnomalyVitalityAutoencoder(input_dim,hidden_dim)

[0241] #Get input data and perform forward propagation

[0242] input_data = get_input_data()

[0243] reconstructed_data=model.forward(input_data)

[0244] # Calculate reconstruction error

[0245] reconstruction_error=model.compute_reconstruction_error(input_data, reconstructed_data)

[0246] #Set the anomaly detection threshold and judge the result

[0247] anomaly_threshold=set_threshold()

[0248] result=model.predict_anomaly(reconstruction_error,anomaly_threshold)

[0249] # Output monitoring results

[0250] print(f"Monitoring results:{result}")

[0251] refer to Figure 3 , a lamb vitality identification system based on multi-source feature fusion includes a biological signal acquisition unit, a feature extraction and fusion unit, a lamb vitality identification unit and an abnormal vitality monitoring unit; each unit cooperates with each other to gradually complete the whole process operation from data collection to vitality identification and abnormal monitoring;

[0252] A biological signal acquisition unit, used to collect physiological signal data of the target lamb using a wearable sensor device, and to obtain image data of the target lamb using an image acquisition sensor device;

[0253] A feature extraction and fusion unit is used to construct a feature extraction and fusion model, and use the feature extraction and fusion model to extract features from image data and physiological signal data to obtain multi-source feature data;

[0254] A lamb vitality identification unit is used to construct a primary and secondary collaborative lamb vitality identification model, and use the primary and secondary collaborative lamb vitality identification model to identify unlabeled multi-source feature data to obtain vitality identification results;

[0255] The abnormal activity monitoring unit is used to construct an abnormal activity recognition model, and use the abnormal activity recognition model to perform encoding and decoding reconstruction operations and abnormal activity discrimination on the lamb to obtain an abnormal activity discrimination result.

[0256] Optionally, an electronic device includes a memory and a processor, and when the computer instructions stored in the memory are executed by the processor, the methods in embodiments one to three can be implemented; the processor can be a central processing unit (CPU) or other types of processors, such as a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, as well as discrete logic or hardware components, etc.; a general-purpose processor can be a microprocessor or other conventional processor; the methods in embodiments one to three can be directly embodied as being executed by the processor, or can be executed by a combination of hardware and software modules in the processor.

[0257] The beneficial effects of the present invention are as follows:

[0258] (1) The present invention introduces a mutually coupled feature extraction fusion model, a main and auxiliary collaborative lamb vitality identification model, and an abnormal vitality identification model, and combines software and hardware to effectively monitor lamb vitality in real time and distinguish between normal and abnormal vitality; the real-time monitoring results have high accuracy and stability, which solves the problem of difficulty in automatically identifying and monitoring lamb vitality in the prior art, and can accurately identify abnormal vitality and perform early risk control in a dynamic environment.

[0259] (2) The method and system provided by the present invention can simultaneously process biological data from different sources, such as physiological signals and image data, and extract important features through their respective dedicated paths; the model can construct a multi-source feature data stream, comprehensively consider the appearance and physiological status information of the lamb, and thus achieve more accurate and comprehensive vitality recognition. This multi-source data fusion method improves the generalization ability and robustness of the model, can better adapt to various complex practical application scenarios, and improves the accuracy and stability of vitality recognition.

[0260] (3) The method and system provided by the present invention adopt a collaborative lamb vitality identification model that uses multiple auxiliary classification sub-models and a main classification sub-model to work together, and outputs the final prediction result through a comprehensive probability distribution method; this multi-model collaborative mechanism can make full use of the advantages of different models, further improve the accuracy and robustness of the prediction, and ensure that the final prediction result is more reliable.

[0261] (4) The method and system provided by the present invention, the main and auxiliary collaborative lamb vitality identification model combines unlabeled data with labeled data through a joint training strategy of pseudo-labels and real labels, and gradually adjusts the model parameters by calculating the matching degree between the pseudo-label prediction and the real label prediction, so that the model can more accurately predict the label of the unlabeled data, thereby enhancing the learning ability of the model when the labeled data is insufficient.

[0262] (5) The method and system provided by the present invention, the main and auxiliary collaborative lamb vitality identification model outputs the probability score of each category through probability distribution calculation, and takes the category with the highest confidence as the final prediction result; the user can set the threshold according to the confidence score and flexibly adjust the decision criteria of the model, thereby ensuring that the prediction results are highly reliable and controllable.

[0263] (6) The hierarchical structure design of the abnormal vitality recognition model of the method and system provided by the present invention enables it to monitor the vitality status of the lamb in real time and dynamically adjust the judgment according to the changes in the input data; by analyzing the reconstruction error of the decoder, the model can promptly detect potential abnormal situations and provide early warning.

[0264] (7) In the method and system provided by the present invention, the abnormal vitality recognition model not only utilizes its own learning ability, but also combines the output results of the vitality recognition model to further improve the detection accuracy of abnormal states. This combination enables the model to more comprehensively and accurately screen out individuals that may be abnormal when processing complex data.

[0265] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0266] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A lamb vitality identification method based on multi-source feature fusion, characterized in that: include: Using a wearable sensor device to collect the physiological signal data to be identified of the target lamb, and using an image acquisition sensor device to obtain the image data to be identified of the target lamb; Constructing a feature extraction fusion model, and using the feature extraction fusion model to extract features from the image data to be identified and the physiological signal data to be identified, to obtain multi-source feature data to be identified; The multi-source feature data to be identified includes: physiological signal data feature extraction results and image data feature extraction results; A lamb vitality identification model with primary and secondary collaboration is constructed, and the multi-source feature data to be identified is identified using the lamb vitality identification model with primary and secondary collaboration to obtain a vitality identification result; the lamb vitality identification model with primary and secondary collaboration includes: a first lamb vitality identification sub-model and a second lamb vitality identification sub-model that are coupled to each other; An abnormal vitality recognition model is constructed, and the abnormal vitality recognition model is used to perform encoding, decoding, reconstruction operations and abnormal vitality discrimination on the vitality recognition result and the multi-source feature data to be identified, so as to obtain an abnormal vitality discrimination result.

2. A lamb vitality identification method based on multi-source feature fusion according to claim 1, characterized in that: Constructing a feature extraction fusion model, and using the feature extraction fusion model to extract features from the image data to be identified and the physiological signal data to be identified, to obtain multi-source feature data to be identified, including: Constructing a feature extraction fusion model; the feature extraction fusion model includes: a physiological feature extraction path and an image feature extraction path; the physiological feature extraction path includes: a memory filter, an information injector and an output regulator; the image feature extraction path includes: a feature extraction unit, a nonlinear activation unit and a feature downsampling unit; The physiological signal data to be identified is input into the memory filter and the information injector to obtain information candidates; the calculation formula of the information candidate is: in, is the candidate item of the information; t-1 is the extraction result of the previous cycle; X t is the input sequence of the current cycle with a sequence length of T; is element-wise subtraction; A weight matrix of the information candidates randomly generated by the memory filter and the information injector for the input sequence; A base coefficient of the information candidate randomly generated by the memory filter and the information injector for the input sequence; According to the information candidate, the memory state is updated using the memory filter and the information injector; the calculation formula of the memory state is: Among them, S t-1 It is the memory state of the previous cycle; and ⊙ are element-wise addition and element-wise multiplication respectively; A weight matrix of the memory state randomly generated for the input sequence by the memory filter and the information injector; A reference coefficient of the memory state randomly generated for the input sequence through the memory filter and the information injector; The memory state S that will be updated is completed t Input into the output regulator to obtain the physiological signal data feature extraction result; the calculation formula of the physiological signal data feature extraction result is: Among them, h t The physiological signal data feature extraction result of the current cycle; A weight matrix of the extraction results randomly generated by the output regulator for the input sequence; A reference coefficient of the extraction result randomly generated by the output regulator for the input sequence; is the tensor product; The image data to be identified is input into the feature extraction unit to obtain a feature map; the calculation formula of the feature map is: in, Y t is the characteristic graph of the current cycle; t (i, j) is the value of the feature graph at (i, j); K H , K W and C are the sizes of the convolution kernel; W k and b k are the weight and base coefficient of the convolution kernel respectively; Y is the image data to be identified that is randomly generated by the feature extraction unit t The weight matrix of (i,j); Y is the image data to be identified that is randomly generated by the feature extraction unit t The base coefficient of (i,j); The feature map is input into the nonlinear activation unit for nonlinear transformation to obtain a transformation result; the calculation formula of the transformation result is: Among them, Z t is the transformation result of the current cycle; Z t (i, j) is the value of the transformation result at (i, j); ReLU(·) is a nonlinear activation function; The transformation result is input into the feature downsampling unit to perform a dimensionality reduction operation to obtain a feature downsampling result; the calculation formula of the feature downsampling result is: in, is the feature downsampling result; is the value of the feature downsampling result at (i, j); P H and P W are the sizes of downsampling windows; Performing a flattening operation on the feature downsampling result to obtain the image data feature extraction result; The physiological signal data feature extraction result and the image data feature extraction result are associated and fused to obtain the multi-source feature data to be identified.

3. The lamb vitality identification method based on multi-source feature fusion according to claim 1 is characterized in that: Constructing a lamb vitality identification model with primary and secondary collaboration, and using the lamb vitality identification model with primary and secondary collaboration to identify the multi-source feature data to be identified, to obtain a vitality identification result, including: The first lamb vitality identification sub-model and the second lamb vitality identification sub-model are set; the first lamb vitality identification sub-model includes: a first auxiliary classification sub-model, a second auxiliary classification sub-model and a first main classification sub-model coupled to each other; the second lamb vitality identification sub-model includes: a third auxiliary classification sub-model, a fourth auxiliary classification sub-model and a second main classification sub-model coupled to each other; Perform unlabeled operations and true labeled operations on the pre-collected multi-source feature data to obtain an unlabeled multi-source feature dataset and a true labeled multi-source feature dataset; Inputting the physiological signal feature extraction result and the image data feature extraction result in the unlabeled multi-source feature data set into the first auxiliary classification sub-model and the second auxiliary classification sub-model respectively to obtain a first prediction probability score and a second prediction probability score; Calculating a comprehensive predicted probability score based on the first predicted probability score and the second predicted probability score; Screening and combining the preset marker set according to the comprehensive prediction probability score to obtain a prediction pseudo marker set; Fusion of the predicted pseudo-label set and the pre-collected multi-source feature data to obtain a pseudo-label multi-source feature data set; Inputting the pseudo-labeled multi-source feature data set and the real-labeled multi-source feature data set into the first main classification sub-model to perform matching degree calculation, thereby obtaining a matching degree calculation result; When the matching degree calculation result is equal to 0, iteratively optimizing the first auxiliary classification sub-model and the second auxiliary classification sub-model; When the matching degree calculation result is 1, outputting the first lamb vitality identification sub-model that has been trained; According to the training process of the first lamb vitality identification sub-model, the third auxiliary classification sub-model, the fourth auxiliary classification sub-model and the second main classification sub-model are trained to obtain the trained second lamb vitality identification sub-model; The first lamb vitality identification sub-model that has been trained is coupled with the second lamb vitality identification sub-model that has been trained to obtain the main-auxiliary coordinated lamb vitality identification model.

4. The lamb vitality identification method based on multi-source feature fusion according to claim 3 is characterized in that: Also includes: Inputting the multi-source feature data to be identified into the first lamb vitality identification sub-model and the second lamb vitality identification sub-model respectively to obtain a first probability score to be identified and a second probability score to be identified; The first probability score to be identified and the second probability score to be identified are weightedly calculated to obtain a comprehensive probability score; the calculation formula of the comprehensive probability score is: in, is the comprehensive probability score; is the first probability score to be identified; is the second probability score to be identified; L is the data volume of the first probability score to be identified and the second probability score to be identified; A first probability score which is the first probability score to be identified; The maximum probability mark of the comprehensive probability score is extracted to obtain the vitality recognition result.

5. The lamb vitality identification method based on multi-source feature fusion according to claim 3 is characterized in that: The calculation formula of the comprehensive prediction probability score is: in, is the comprehensive prediction probability score; and are the first predicted probability score and the second predicted probability score respectively; t is the physiological signal data feature extraction result; t is the feature extraction result of the image data; c is the preset tag set; u c is the center of the preset marker set; d(h t ,u c ) and d(v t ,u c ) are the distances between the physiological signal data feature extraction result and the image data feature extraction result and the center of the preset tag set; θ is the similarity coefficient; L is the number of the predicted pseudo tag set; ω is the weight coefficient; is a predicted pseudo label in the generated predicted pseudo label set.

6. The lamb vitality identification method based on multi-source feature fusion according to claim 3 is characterized in that: The calculation formula of the matching degree calculation result is: Among them, Ψ(D real ,D pseudo ) is the matching degree between the real labeled multi-source feature data set and the pseudo labeled multi-source feature data set; θ and δ are both weight coefficients; is the real labeled multi-source feature data in the real labeled multi-source feature data set of the current cycle; is the pseudo-labeled multi-source feature data in the pseudo-labeled multi-source feature data set of the current cycle; a is the maximum number of the physiological signal data to be identified and the image data to be identified.

7. The lamb vitality identification method based on multi-source feature fusion according to claim 1 is characterized in that: Constructing an abnormal vitality recognition model, and using the abnormal vitality recognition model to perform encoding, decoding, reconstruction operations and abnormal vitality discrimination on the vitality recognition result and the multi-source feature data to be identified, to obtain an abnormal vitality discrimination result, including: Constructing an abnormal vitality recognition model; the abnormal vitality recognition model includes: an input layer, an encoding layer, a decoding layer and an output layer; Using the input layer to receive the physiological signal data to be identified, the image data to be identified, and the vitality recognition result to obtain input data; Using the encoding layer to reduce the dimension of the input data to obtain a potential representation; Reconstructing the latent representation using the decoding layer to obtain a reconstruction result; Perform error calculation according to the reconstruction result to obtain a reconstruction error; Utilizing the output layer to analyze the degree of deviation of the reconstruction error, and obtaining the abnormal activity discrimination result; When the reconstruction error is within a preset threshold range, determining that the abnormal vitality discrimination result is normal; When the reconstruction error exceeds a preset threshold range, the abnormal activity discrimination result is determined to be abnormal, and an alarm is triggered.

8. A lamb vitality identification system based on multi-source feature fusion, characterized in that: The lamb vitality identification method based on multi-source feature fusion as described in claim 1, the system comprises: A biological signal acquisition unit, used to acquire the physiological signal data to be identified of the target lamb using a wearable sensor device, and to acquire the image data to be identified of the target lamb using an image acquisition sensor device; A feature extraction and fusion unit, used for constructing a feature extraction and fusion model, and using the feature extraction and fusion model to extract features from the image data to be identified and the physiological signal data to be identified, so as to obtain multi-source feature data to be identified; A lamb vitality identification unit is used to construct a primary and secondary collaborative lamb vitality identification model, and use the primary and secondary collaborative lamb vitality identification model to identify the multi-source feature data to be identified to obtain a vitality identification result; The abnormal activity monitoring unit is used to construct an abnormal activity recognition model, and use the abnormal activity recognition model to perform encoding, decoding, reconstruction operations and abnormal activity discrimination on the activity recognition result and the multi-source feature data to be identified, so as to obtain an abnormal activity discrimination result.

Images