Wild animal positioning method based on self-supervised learning

Abstract

The invention discloses a wild animal positioning method based on self-supervised learning. The method comprises the following steps: S1, obtaining preprocessed signal data; s2, describing features of wild animal signals by the initial feature vectors; s3, constructing a self-supervised learning model for extracting animal signal features; s4, optimizing parameters of the self-supervised learning model in real time through the reward signal; s5, forming a global feature matrix; s6, optimizing parameters of the self-supervised learning model through the reward signal; s7, based on the preliminary positioning result, performing high-frequency depth acquisition on the area where the animal is located and the surrounding area; s8, the central node optimizes the structure and parameters of the self-supervised learning model through a reinforcement learning strategy module according to the deep collection data; and S9, the central node uses the optimized self-supervised learning model and deep acquisition data to accurately position the position of the animal through a signal propagation model and a cooperative computing algorithm. According to the invention, wild animals can be efficiently positioned in a large range.

Description

Technical Field

[0001] The present invention relates to the technical field of wild animal positioning, and in particular to a wild animal positioning method based on self-supervised learning. Background Art

[0002] In the existing technology, wildlife positioning mainly relies on the combination of signal transmitters and ground receiving stations, and usually uses radio signals or satellite signals for positioning. In this method, wild animals carry signal transmitters, and ground receiving stations or satellites calculate the location of animals based on the received signal strength and time delay information. Although traditional positioning methods can track wild animals to a certain extent, they face many challenges.

[0003] First of all, existing wildlife positioning technology mainly relies on the deployment of ground receiving stations or satellite receiving equipment. Ground receiving stations need to be deployed in vast wild environments. In areas with complex terrain such as mountains and forests, they face significant deployment difficulties and equipment maintenance difficulties, and their signal reception range is greatly limited. Although satellite positioning can solve the problem of ground equipment deployment, satellite signals are limited by environmental conditions. For example, dense forests make it impossible to accurately obtain the location of animals in certain scenarios. In addition, most of these methods rely on the direct propagation characteristics of the signal and lack the means to deal with multipath effects and noise interference in the signal path, resulting in low positioning accuracy.

[0004] Existing positioning systems rely on a large amount of annotated data to build effective models. Due to the diversity and dynamic changes of the wild environment, signal propagation characteristics are difficult to accurately predict and model. Therefore, a large amount of annotated data is required to train the positioning model. However, obtaining these annotated data is often costly and it is difficult to obtain accurate ground truth data in the actual environment. As a result, the existing models are not adaptable enough in the face of environmental changes and it is difficult to maintain positioning accuracy and robustness.

[0005] Existing positioning technologies usually adopt a centralized processing architecture, in which signal data is transmitted from receiving stations or satellites to a central processing center for processing and calculation. The centralized architecture not only brings about delays in signal data transmission, but also has significant bottlenecks in the system's response speed and processing capabilities when processing large-scale radio signal data. At the same time, due to the complexity of data synchronization and task allocation, the collaborative work of multiple nodes is difficult to achieve efficient data sharing and real-time collaboration, which limits the overall performance of the system in positioning.

[0006] To sum up, the existing technology has the following main shortcomings: first, the positioning system is difficult to deploy in complex environments, the positioning accuracy is limited, and it cannot effectively deal with the influence of terrain and occlusion factors; second, the positioning model that relies on a large amount of labeled data has poor adaptability and is difficult to cope with the frequent changes in signal characteristics; finally, the centralized system architecture has bottlenecks in data sharing and real-time collaboration among multiple nodes, resulting in low overall system performance and response speed, making it difficult to achieve efficient positioning and tracking of wild animals. Summary of the invention

[0007] One object of the present invention is to provide a method for locating wild animals based on self-supervised learning, which can efficiently locate wild animals over a large range.

[0008] A wild animal positioning method based on self-supervised learning according to an embodiment of the present invention comprises the following steps:

[0009] S1. Deploy signal receiving equipment on multiple drones, with each drone acting as a signal receiving end to receive signals from signal transmitters carried by wild animals within the drone's flight coverage area, and perform preliminary denoising and bandpass filtering preprocessing on the received signals to obtain preprocessed signal data;

[0010] S2, each UAV extracts frequency domain and time domain features from the preprocessed signal data to form a preliminary feature vector, which describes the characteristics of the wildlife signal;

[0011] S3. Construct a self-supervised learning model for extracting animal signal features. The self-supervised learning model accepts preliminary feature vectors as training data and generates a variety of self-supervised learning tasks, so that the self-supervised learning model can automatically discover signal features without manually labeled data.

[0012] S4. Embed the reinforcement learning strategy module as the optimization mechanism of the self-supervised learning model, monitor the prediction results of the self-supervised learning model, construct a reward function based on the animal positioning accuracy and the effectiveness of the self-supervised learning model output, and optimize the parameters of the self-supervised learning model in real time through the reward signal;

[0013] S5, each drone uploads the preliminary feature vector and the results of multiple self-supervised learning tasks to the central node, and shares the feature information among multiple drones through the data synchronization mechanism to form a global feature matrix;

[0014] S6, the central node receives the global feature matrix and analyzes and matches the feature matrix in combination with the self-supervised learning model to preliminarily locate the location of the wild animal;

[0015] S7. Based on the preliminary positioning results, each drone dynamically adjusts the flight path and signal sampling strategy to conduct high-frequency in-depth sampling of the area where the animal is located and its surrounding areas;

[0016] S8, the central node optimizes the structure and parameters of the self-supervised learning model through the reinforcement learning strategy module based on the deep collection data;

[0017] S9, the central node uses the optimized self-supervised learning model and deep acquisition data to accurately locate the animal's position through signal propagation model and collaborative computing algorithm.

[0018] Optionally, the specific steps of S1 are as follows:

[0019] S11. Deploy signal receiving equipment on multiple drones, each of which is equipped with R i (t) signal receiving module, i∈{1, 2, ..., N} represents the i-th UAV, and t is a time variable used to mark the time point of receiving the signal;

[0020] S12. The signal transmitter carried by wild animals sends a signal S j (f, t), j∈{1, 2, ..., M} represents the jth wild animal, f is the frequency variable of the signal, indicating the frequency range used by the transmitter;

[0021] S13, each drone receives the transmission signal of wild animals in its flight coverage area S j (f, t), and the strength and phase information of the signal captured by the antenna array is expressed as the received signal data matrix:

[0022] Y i (t) = A i ·S j (f, t)+n i (t);

[0023] Among them, Y i (t) is the signal received by the i-th UAV at time t, A i is the propagation coefficient of the signal, reflecting the attenuation and interference in the signal propagation path, n i (t) is the noise interference during the receiving process;

[0024] S14, using an adaptive filtering algorithm in the signal receiving module of the drone to receive the received signal data matrix Y i (t) Perform preliminary denoising to eliminate noise interference n i (t);

[0025] S15, band-pass filtering is performed on the signal data after denoising, and the frequency range [f min, f max ], only keep the target signal S j (ft) effective frequency band, and the filtered signal is obtained

[0026] S16, the filtered signal Normalization processing is performed to eliminate the amplitude inconsistency caused by the difference in receiving strength of different drones, and the normalized signal data matrix is ​​obtained:

[0027]

[0028] Optionally, the specific steps of S2 are as follows:

[0029] S21, each UAV based on the pre-processed signal data matrix Z i (t), the spectrum representation is obtained by converting the time domain signal to the frequency domain signal through Fourier transform:

[0030]

[0031] Among them, X i (f) is the spectrum representation of the signal received by the i-th UAV, where f is the frequency variable and j is the imaginary unit;

[0032] S22, spectrum-based representation of X i (f) Extract frequency domain features, including the main frequency component f of the signal peak , signal bandwidth B and spectral energy density ε(f);

[0033] S23. Calculate the statistical characteristics of the signal in the time domain, including the mean μ(Z i ), variance σ 2 (Z i ) and autocorrelation function R(τ), where the autocorrelation function R(τ) is:

[0034]

[0035] Where τ is the time delay and T is the observation period;

[0036] S24, integrating the extracted frequency domain features and time domain features into a preliminary feature vector F i :

[0037] F i =[f peak , B, ε(f), μ(Z i ), σ 2 (Z i ), R(τ)].

[0038] Optionally, the specific steps of S3 are as follows:

[0039] S31, based on the preliminary feature vector F uploaded by each drone i Constructing a self-supervised learning model M θ , where θ is the self-supervised learning model parameter. The self-supervised learning model is based on F i As input, the latent feature representation H of the output animal signal i :

[0040] H i =M θ (F i );

[0041] Among them, H i is the potential feature representation of the animal signal corresponding to the i-th drone, which is used to capture the deep feature pattern of the animal signal;

[0042] S32. Design signal pattern prediction task, self-supervised learning model M θ According to the potential feature representation H at the current moment i (t) Predict the preliminary feature vector for the next moment

[0043]

[0044] Among them, W p and b p is the weight and bias parameter of the prediction layer, by minimizing the prediction error Optimize the self-supervised learning model parameters so that the self-supervised learning model can learn the temporal variation pattern of animal signal characteristics:

[0045]

[0046] S33, construct a contrastive learning task, using the potential feature representations of different drones or at different times for comparison; define a positive sample pair as the feature representation of the same drone at adjacent times (H i (t), H i (t+1)), the negative sample pairs are the feature representations of different drones (H i (t), H j (t′)), the contrastive learning loss function is:

[0047]

[0048] in, is the cosine similarity, τ1 is the temperature parameter;

[0049] S34, design the signal anomaly detection task, use the autoencoder structure to analyze the initial feature vector F i To refactor:

[0050]

[0051] in, and are encoder and decoder networks respectively; calculate the reconstruction error:

[0052]

[0053] When the reconstruction error exceeds the threshold δ, the corresponding signal is determined to be an abnormal signal. The self-supervised learning model parameters are optimized by minimizing the reconstruction error so that the self-supervised learning model can detect abnormal animal signal characteristics.

[0054] Optionally, the specific steps of S4 are as follows:

[0055] S41. Embed the reinforcement learning strategy module into the self-supervised learning model M θ , the reinforcement learning strategy module uses the current state s of the self-supervised learning model t is the input, s t It consists of the latent feature representation and loss function value of the self-supervised learning model, defined as:

[0056]

[0057] Among them, H i (t) represents the potential feature representation of the animal signal of the i-th UAV at time t;

[0058] S42. Define the action space A of reinforcement learning, including the update strategy of the self-supervised learning model parameter θ and the adjustment of the drone flight strategy. Action a t Defined as:

[0059] a t =[Δθ t , Δu i (t)];

[0060] Among them, Δθ t is the update amount of the self-supervised learning model parameter θ at time t, guiding the learning direction of the self-supervised learning model, Δu i (t) is the flight control command increment of the i-th UAV at time t, including the adjustment of position and speed, which is used to optimize the flight path and signal acquisition strategy of the UAV;

[0061] Δu i (t) = [Δx i (t), Δy i (t), Δz i (t), Δv i (t)];

[0062] Where Δx i(t), Δy i (t), Δz i (t) is the spatial coordinate adjustment, Δv i (t) is the speed adjustment;

[0063] S43. Construct reward function R t To evaluate the comprehensive performance of the self-supervised learning model and the drone swarm, the reward function integrating positioning accuracy, self-supervised learning model performance and resource consumption is defined as:

[0064]

[0065] Among them, ∈(t) is the wildlife positioning error at time t, α, β, γ are weight coefficients, and E i (t) is the energy consumption of the i-th UAV at time t, C i (t) is its communication resource consumption, c energy and c comm is the cost coefficient of unit energy and communication resources;

[0066] S44, using deep reinforcement learning algorithm to optimize strategy and maximize the cumulative reward G:

[0067]

[0068] Among them, γ∈[0,1] is the discount factor, using the strategy network π φ (a t |s t ) parameterizes the strategy, φ is the policy network parameter, and the policy gradient method is used to update φ and the self-supervised learning model parameter θ:

[0069]

[0070]

[0071] Where η is the learning rate, They are the gradients of the cumulative reward to the policy network parameters and the self-supervised learning model parameters, and the parameters are updated by the gradient ascent method.

[0072] Optionally, the specific steps of S5 are as follows:

[0073] S51, each UAV extracts the initial feature vector The execution results of the self-supervised learning task are uploaded to the central node;

[0074] S52. The central node defines a global feature matrix G(t) based on the uploaded feature information and task results. The global feature matrix G(t) records the feature vectors and potential features of all drones:

[0075]

[0076] Optionally, the specific steps of S6 are as follows:

[0077] S61, the central node receives and analyzes the global feature matrix G(t), and the central node combines the self-supervised learning model M θ Analyze and pattern match the global feature matrix G(t) to calculate each potential feature representation H i (t) Similarity with the reference feature:

[0078]

[0079] Among them, H ref is the reference feature representation in historical data, sim is the cosine similarity function, which is used to evaluate the matching degree of signal patterns;

[0080] S63, clustering analysis is performed on the feature vectors according to the matching results, the location of the wild animals is located, and the cluster center C of each group of feature vectors is defined. k (t):

[0081]

[0082] in, is the set of feature vectors in the kth cluster, C k (t) is the center point of the cluster, corresponding to the location of wild animals;

[0083] S64. Use the reinforcement learning strategy module to evaluate the accuracy of the positioning result in real time and define the positioning error ∈(t):

[0084]

[0085] Among them, pi(t) is the actual position, To predict the position, the error ∈(t) is used as the input of the reward function to optimize the parameters θ of the self-supervised learning model;

[0086] S65. Use the policy gradient method to optimize the self-supervised learning model parameters θ according to the value of the reward function, and optimize the self-supervised learning model to match and locate animal signal features.

[0087] Optionally, the specific steps of S9 are as follows:

[0088] S91, the central node uses the optimized self-supervised learning model Receive the depth data and upload the feature vector F i With latent feature representation Integrate to generate the overall data matrix D(t):

[0089]

[0090] in, represents the potential feature representation of the i-th drone after the optimization model;

[0091] S92. Calculate the path of potential features based on the signal propagation model, and define the signal propagation model equation as:

[0092] P i (t) = p0 + v i ·(t-t0)+n i (t);

[0093] Among them, P i (t) represents the signal propagation path recorded by the i-th UAV at time t, p0 is the initial position of the signal source, v i is the signal propagation velocity vector, n i (t) is random noise interference;

[0094] S93, the central node performs time synchronization processing on the signal paths of all drones and uses a triangulation algorithm to accurately locate the position of the animal. At time t, the signal propagation paths recorded by drones i, j, and k are P respectively. i (t), P j (t), P k (t), the position estimate is calculated by triangulation

[0095]

[0096] in, is the estimated position of the animal, × represents the vector cross product, which is used to calculate the intersection of the signal propagation planes enclosed by the three drones;

[0097] S94, introduce the Kalman filter algorithm to update the estimated position in real time, and the precise position after Kalman filter optimization The central node packages the positioning results and distributes them to all drones via wireless communication links;

[0098] S95, each drone receives accurate location data Adjust flight paths and optimize signal collection strategies so that the system can continuously track and collect real-time location information of wildlife.

[0099] The beneficial effects of the present invention are:

[0100] (1) The present invention adopts a model optimization strategy that combines self-supervised learning with reinforcement learning. It realizes adaptive optimization of the model to address the adaptability problem of traditional positioning technology in dynamic environments. It automatically extracts signal features through the self-supervised learning model to reduce the dependence on large-scale labeled data, and optimizes model parameters in real time according to positioning errors in the reinforcement learning module, so that the system can still work efficiently in the face of complex and changeable field environments. Compared with traditional methods that rely on static models, the present invention realizes dynamic updating of model parameters, improves the adaptability and robustness of the positioning system, and significantly improves the positioning accuracy in complex environments.

[0101] (2) The present invention innovatively adopts a distributed architecture for signal collection and collaborative computing, and realizes efficient multi-node collaboration through data synchronization and global feature matrix sharing between various UAVs. The signal data and self-supervision task results of each UAV are integrated through a central node and analyzed and located using a collaborative computing algorithm. Compared with the traditional centralized architecture, the distributed design of the present invention reduces the delay of signal transmission and improves the response speed and processing power of the system through multi-node collaborative computing, so that the system can efficiently locate wild animals over a large range.

[0102] (3) The present invention realizes high-precision positioning of the animal's position by combining the signal propagation model with the triangulation algorithm. The depth acquisition data is processed by the self-supervised learning model and combined with the signal propagation model to infer the propagation path and position of the animal. In addition, the Kalman filter algorithm is introduced to perform real-time correction of the estimated position, which further improves the positioning accuracy and the robustness of the system, thereby realizing continuous tracking and dynamic monitoring of the animal's position. BRIEF DESCRIPTION OF THE DRAWINGS

[0103] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0104] Figure 1 This is a flow chart of a wild animal positioning method based on self-supervised learning proposed by the present invention. DETAILED DESCRIPTION

[0105] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0106] refer to Figure 1 , a wildlife positioning method based on self-supervised learning, comprising the following steps:

[0107] S1. Deploy signal receiving equipment on multiple drones, with each drone acting as a signal receiving end to receive signals from signal transmitters carried by wild animals within the drone's flight coverage area, and perform preliminary denoising and bandpass filtering preprocessing on the received signals to obtain preprocessed signal data;

[0108] S2, each UAV extracts frequency domain and time domain features from the preprocessed signal data to form a preliminary feature vector, which describes the characteristics of the wildlife signal;

[0109] S3. Construct a self-supervised learning model for extracting animal signal features. The self-supervised learning model accepts preliminary feature vectors as training data and generates a variety of self-supervised learning tasks, so that the self-supervised learning model can automatically discover signal features without manually labeled data.

[0110] S4. Embed the reinforcement learning strategy module as the optimization mechanism of the self-supervised learning model, monitor the prediction results of the self-supervised learning model, construct a reward function based on the animal positioning accuracy and the effectiveness of the self-supervised learning model output, and optimize the parameters of the self-supervised learning model in real time through the reward signal;

[0111] S5, each drone uploads the preliminary feature vector and the results of multiple self-supervised learning tasks to the central node, and shares the feature information among multiple drones through the data synchronization mechanism to form a global feature matrix;

[0112] S6, the central node receives the global feature matrix and analyzes and matches the feature matrix in combination with the self-supervised learning model to preliminarily locate the location of the wild animal;

[0113] S7. Based on the preliminary positioning results, each drone dynamically adjusts the flight path and signal sampling strategy to conduct high-frequency in-depth sampling of the area where the animal is located and its surrounding areas;

[0114] S8, the central node optimizes the structure and parameters of the self-supervised learning model through the reinforcement learning strategy module based on the deep collection data;

[0115] S9, the central node uses the optimized self-supervised learning model and deep acquisition data to accurately locate the animal's position through signal propagation model and collaborative computing algorithm.

[0116] In this implementation, the specific steps of S1 are as follows:

[0117] S11. Deploy signal receiving equipment on multiple drones, each of which is equipped with R i (t) signal receiving module, i∈{1, 2, ..., N} represents the i-th UAV, and t is a time variable used to mark the time point of receiving the signal;

[0118] S12. The signal transmitter carried by wild animals sends a signal S j (f, t), j∈{1, 2, ..., M} represents the jth wild animal, f is the frequency variable of the signal, indicating the frequency range used by the transmitter;

[0119] S13, each drone receives the transmission signal of wild animals in its flight coverage area S j (ft), and the signal strength and phase information captured by the antenna array is represented as the received signal data matrix:

[0120] Y i (t) = A i ·S j (f, t)+n i (t);

[0121] Among them, Y i (t) is the signal received by the i-th UAV at time t, A i is the propagation coefficient of the signal, reflecting the attenuation and interference in the signal propagation path, n i (t) is the noise interference during the receiving process;

[0122] S14, using an adaptive filtering algorithm in the signal receiving module of the drone to receive the received signal data matrix Y i (t) Perform preliminary denoising to eliminate noise interference n i (t);

[0123] S15, band-pass filtering is performed on the signal data after denoising, and the frequency range [f min , f max ], only keep the target signal S j The effective frequency band of (f, t) is used to obtain the filtered signal

[0124] S16, the filtered signal Normalization processing is performed to eliminate the amplitude inconsistency caused by the difference in receiving strength of different drones, and the normalized signal data matrix is ​​obtained:

[0125]

[0126] In this implementation, the specific steps of S2 are as follows:

[0127] S21, each UAV based on the pre-processed signal data matrix Z i (t), the spectrum representation is obtained by converting the time domain signal to the frequency domain signal through Fourier transform:

[0128]

[0129] Among them, X i (f) is the spectrum representation of the signal received by the i-th UAV, where f is the frequency variable and j is the imaginary unit;

[0130] S22, spectrum-based representation of X i (f) Extract frequency domain features, including the main frequency component f of the signal peak , signal bandwidth B and spectral energy density ε(f);

[0131] S23. Calculate the statistical characteristics of the signal in the time domain, including the mean μ(Z i ), variance σ 2 (Z i ) and autocorrelation function R(τ), where the autocorrelation function R(τ) is:

[0132]

[0133] Where τ is the time delay and T is the observation period;

[0134] S24, integrating the extracted frequency domain features and time domain features into a preliminary feature vector F i :

[0135] F i =[f peak , B, ε(f), μ(Z i ), σ 2 (Z i ), R(τ)].

[0136] In this implementation, the specific steps of S3 are as follows:

[0137] S31, based on the preliminary feature vector F uploaded by each drone i Constructing a self-supervised learning model M θ , where θ is the self-supervised learning model parameter. The self-supervised learning model is based on F i As input, the latent feature representation H of the output animal signal i :

[0138] H i =M θ (Fi);

[0139] Among them, H i is the potential feature representation of the animal signal corresponding to the i-th drone, which is used to capture the deep feature pattern of the animal signal;

[0140] S32. Design signal pattern prediction task, self-supervised learning model M θ According to the potential feature representation H at the current moment i(t) Predict the preliminary feature vector for the next moment

[0141]

[0142] Among them, W p and b p is the weight and bias parameter of the prediction layer, by minimizing the prediction error Optimize the self-supervised learning model parameters so that the self-supervised learning model can learn the temporal variation pattern of animal signal characteristics:

[0143]

[0144] S33, construct a contrastive learning task, using the potential feature representations of different drones or at different times for comparison; define a positive sample pair as the feature representation of the same drone at adjacent times (H i (t), H i (t+1)), the negative sample pairs are the feature representations of different drones (H i (t), H j (t′)), the contrastive learning loss function is:

[0145]

[0146] in, is the cosine similarity, τ1 is the temperature parameter;

[0147] S34, design the signal anomaly detection task, use the autoencoder structure to analyze the initial feature vector F i To refactor:

[0148]

[0149] in, and are encoder and decoder networks respectively; calculate the reconstruction error:

[0150]

[0151] When the reconstruction error exceeds the threshold δ, the corresponding signal is determined to be an abnormal signal. The self-supervised learning model parameters are optimized by minimizing the reconstruction error so that the self-supervised learning model can detect abnormal animal signal characteristics.

[0152] In this implementation manner, the specific steps of S4 are as follows:

[0153] S41. Embed the reinforcement learning strategy module into the self-supervised learning model M θ , the reinforcement learning strategy module uses the current state s of the self-supervised learning model t is the input, s tIt consists of the latent feature representation and loss function value of the self-supervised learning model, defined as:

[0154]

[0155] Among them, H i (t) represents the potential feature representation of the animal signal of the i-th UAV at time t;

[0156] S42. Define the action space A of reinforcement learning, including the update strategy of the self-supervised learning model parameter θ and the adjustment of the drone flight strategy. Action a t Defined as:

[0157] a t =[Δθ t , Δu i (t)];

[0158] Among them, Δθ t is the update amount of the self-supervised learning model parameter θ at time t, guiding the learning direction of the self-supervised learning model, Δu i (t) is the flight control command increment of the i-th UAV at time t, including the adjustment of position and speed, which is used to optimize the flight path and signal acquisition strategy of the UAV;

[0159] Δu i (t) = [Δx i (t), Δy i (t), Δz i (t), Δv i (t)];

[0160] Where Δx i (t), Δy i (t), Δz i (t) is the spatial coordinate adjustment, Δv i (t) is the speed adjustment;

[0161] S43. Construct reward function R t To evaluate the comprehensive performance of the self-supervised learning model and the drone swarm, the reward function integrating positioning accuracy, self-supervised learning model performance and resource consumption is defined as:

[0162]

[0163] Among them, ∈(t) is the wildlife positioning error at time t, α, β, γ are weight coefficients, and E i (t) is the energy consumption of the i-th UAV at time t, Ci(t) is its communication resource consumption, c energy and c comm is the cost coefficient of unit energy and communication resources;

[0164] S44, using deep reinforcement learning algorithm to optimize strategy and maximize the cumulative reward G:

[0165]

[0166] Among them, γ∈[0,1] is the discount factor, using the strategy network π φ (a t |s t ) parameterizes the strategy, φ is the policy network parameter, and the policy gradient method is used to update φ and the self-supervised learning model parameter θ:

[0167]

[0168]

[0169] Where η is the learning rate, They are the gradients of the cumulative reward to the policy network parameters and the self-supervised learning model parameters, and the parameters are updated by the gradient ascent method.

[0170] In this implementation, the specific steps of S5 are as follows:

[0171] S51, each UAV extracts the initial feature vector The execution results of the self-supervised learning task are uploaded to the central node;

[0172] S52. The central node defines a global feature matrix G(t) based on the uploaded feature information and task results. The global feature matrix G(t) records the feature vectors and potential features of all drones:

[0173]

[0174] In this implementation manner, the specific steps of S6 are as follows:

[0175] S61, the central node receives and analyzes the global feature matrix G(t), and the central node combines the self-supervised learning model M θ Analyze and pattern match the global feature matrix G(t) to calculate each potential feature representation H i (t) Similarity with the reference feature:

[0176]

[0177] Among them, H ref is the reference feature representation in historical data, sim is the cosine similarity function, which is used to evaluate the matching degree of signal patterns;

[0178] S63, clustering analysis is performed on the feature vectors according to the matching results, the location of the wild animals is located, and the cluster center C of each group of feature vectors is defined. k (t):

[0179]

[0180] in, is the set of feature vectors in the kth cluster, C k (t) is the center point of the cluster, corresponding to the location of wild animals;

[0181] S64. Use the reinforcement learning strategy module to evaluate the accuracy of the positioning result in real time and define the positioning error ∈(t):

[0182]

[0183] Among them, pi(t) is the actual position, To predict the position, the error ∈(t) is used as the input of the reward function to optimize the parameters θ of the self-supervised learning model;

[0184] S65. Use the policy gradient method to optimize the self-supervised learning model parameters θ according to the value of the reward function, and optimize the self-supervised learning model to match and locate animal signal features.

[0185] In this implementation manner, the specific steps of S9 are as follows:

[0186] S91, the central node uses the optimized self-supervised learning model Receive the depth data and upload the feature vector F i With latent feature representation Integrate to generate the overall data matrix D(t):

[0187]

[0188] in, represents the potential feature representation of the i-th drone after the optimization model;

[0189] S92. Calculate the path of potential features based on the signal propagation model, and define the signal propagation model equation as:

[0190] P i (t)=p0+V i ·(t-t0)+n i (t);

[0191] Among them, P i (t) represents the signal propagation path recorded by the i-th UAV at time t, p0 is the initial position of the signal source, v i is the signal propagation velocity vector, ni (t) is random noise interference;

[0192] S93, the central node performs time synchronization processing on the signal paths of all drones and uses a triangulation algorithm to accurately locate the position of the animal. At time t, the signal propagation paths recorded by drones i, j, and k are P respectively. i (t), P j (t), P k (t), the position estimate is calculated by triangulation

[0193]

[0194] in, is the estimated position of the animal, × represents the vector cross product, which is used to calculate the intersection of the signal propagation planes enclosed by the three drones;

[0195] S94, introduce the Kalman filter algorithm to update the estimated position in real time, and the precise position after Kalman filter optimization The central node packages the positioning results and distributes them to all drones via wireless communication links;

[0196] S95, each drone receives accurate location data Adjust flight paths and optimize signal collection strategies so that the system can continuously track and collect real-time location information of wildlife.

[0197] Embodiment 1:

[0198] In order to verify the feasibility and effectiveness of the present invention, site A was selected as the experimental scene, and the golden monkey group living in the area was tracked and located. As highly active and frequently migrating wild animals, golden monkeys place high demands on the real-time and accuracy of the tracking system. A 30-day test comparison was conducted between the traditional satellite positioning system and the distributed UAV signal acquisition system of the present invention to evaluate the positioning accuracy, response speed and adaptability of the model.

[0199] In the test scenario, three golden monkeys in the group carried small signal transmitters with a signal frequency set within the 900MHz range and a transmission frequency interval of 1 second. The transmitter had a built-in accelerometer to monitor the movement status of the golden monkeys. Ten drones were deployed in the test area, each equipped with a signal receiving module and a self-supervised learning model. A central node was also set up to aggregate and process the data uploaded by the drones.

[0200] After the test started, each drone cruised within its designated flight area and received signals from golden monkeys in real time. At a certain moment during the experiment, the system received a golden monkey signal with severe attenuation, which usually means that the golden monkey has entered a dense forest area. The traditional positioning system cannot accurately track it at this time because the signal reflection and attenuation in the dense forest seriously affect the satellite positioning accuracy. The system of the present invention uses multiple drones to work together to upload the surrounding signal data collected by other drones to the central node, and combines the signal propagation model for calculation and compensation, and successfully predicts and tracks the movement path of the golden monkey.

[0201] On the seventh day of the test, the signal strength of a golden monkey dropped sharply to -80dBm at 09:45, which meant that it entered a dense forest area where signals were difficult to penetrate. The positioning error of the traditional system reached 350 meters, while the drone of the system of the present invention quickly adjusted the flight path, performed high-frequency data collection, and analyzed the signal characteristics of the surrounding nodes through a self-supervised model to infer its possible position. The system used triangulation to estimate the position of the golden monkey at 09:52, with the error reduced to 45 meters. Subsequently, the Kalman filter was used to make real-time corrections to the position, and the position of the golden monkey was accurately determined at 09:55, with the positioning error controlled within 15 meters.

[0202] To further verify the system's adaptability, the system automatically optimized the model parameters several times during the test. On the 14th day, the migration route of the golden monkeys changed, and they entered the undulating mountainous area from the open river channels and forest windows. The positioning accuracy of the traditional system dropped rapidly when the signal propagation model was not updated in time. However, this system monitored the positioning error through the reinforcement learning module and adaptively adjusted the model parameters. It took only 20 minutes to complete the parameter update, allowing the system to adapt to the new environment again and stabilize the positioning error within 10 meters.

[0203] Table 1 Experimental test results

[0204]

[0205] In order to further demonstrate the training effect of the self-supervised learning model, signal data from the 10th to the 20th day were sampled as the training set. The training set included the movement patterns of golden monkeys in different environments and the interference characteristics during signal propagation. During the training process, the model optimized the matching ability of signal features through comparative learning tasks and predicted the movement trajectory of golden monkeys in advance through signal pattern prediction tasks. The training results of the model showed that within a 10-day training cycle, the prediction accuracy of the self-supervised model increased by 12% and the anomaly detection ability increased by 15%.

[0206] This experiment shows that the method of the present invention has significant advantages in complex wild environments, especially in dealing with signal interference and environmental changes. The dynamic optimization system of the distributed UAV collaboration and self-supervision model can efficiently track the location of wild animals and maintain positioning accuracy. In addition, the present invention reduces the dependence on large-scale labeled data, greatly reduces the cost of data acquisition while improving model adaptability. The test results confirm that the method of the present invention is superior to traditional methods in positioning accuracy, response speed and model adaptability, providing reliable technical support for real-time monitoring and protection of wild animals.

[0207] The present invention adopts a model optimization strategy that combines self-supervised learning with reinforcement learning, and realizes adaptive optimization of the model to address the adaptability problem of traditional positioning technology in dynamic environments. The self-supervised learning model automatically extracts signal features to reduce dependence on large-scale labeled data, and optimizes model parameters in real time according to positioning errors in the reinforcement learning module, so that the system can still work efficiently in the face of complex and changeable field environments. Compared with traditional methods that rely on static models, the present invention realizes dynamic updating of model parameters, improves the adaptability and robustness of the positioning system, and significantly improves positioning accuracy in complex environments.

[0208] The present invention innovatively adopts a distributed architecture for signal collection and collaborative computing, and realizes efficient multi-node collaboration through data synchronization and global feature matrix sharing between drones. The signal data and self-supervision task results of each drone are integrated through a central node and analyzed and located using a collaborative computing algorithm. Compared with the traditional centralized architecture, the distributed design of the present invention reduces the delay in signal transmission and improves the system's response speed and processing capabilities through multi-node collaborative computing, enabling the system to efficiently locate wild animals over a large area.

[0209] The present invention realizes high-precision positioning of the animal's position by combining the signal propagation model with the triangulation algorithm. The depth acquisition data is processed by the self-supervised learning model and then combined with the signal propagation model to infer the propagation path and position of the animal. In addition, the estimated position is corrected in real time by introducing the Kalman filtering algorithm, which further improves the positioning accuracy and the robustness of the system, thereby realizing continuous tracking and dynamic monitoring of the animal's position.

[0210] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A wildlife positioning method based on self-supervised learning, characterized in that: The steps include: S1. Deploy signal receiving equipment on multiple drones, with each drone acting as a signal receiving end to receive signals from signal transmitters carried by wild animals within the drone's flight coverage area, and perform preliminary denoising and bandpass filtering preprocessing on the received signals to obtain preprocessed signal data; S2, each UAV extracts frequency domain and time domain features from the preprocessed signal data to form a preliminary feature vector, which describes the characteristics of the wildlife signal; S3. Construct a self-supervised learning model for extracting animal signal features. The self-supervised learning model accepts preliminary feature vectors as training data and generates a variety of self-supervised learning tasks, so that the self-supervised learning model can automatically discover signal features without manually labeled data. S4. Embed the reinforcement learning strategy module as the optimization mechanism of the self-supervised learning model, monitor the prediction results of the self-supervised learning model, construct a reward function based on the animal positioning accuracy and the effectiveness of the self-supervised learning model output, and optimize the parameters of the self-supervised learning model in real time through the reward signal; S5, each drone uploads the preliminary feature vector and the results of multiple self-supervised learning tasks to the central node, and shares the feature information among multiple drones through the data synchronization mechanism to form a global feature matrix; S6, the central node receives the global feature matrix and analyzes and matches the feature matrix in combination with the self-supervised learning model to preliminarily locate the location of the wild animal; S7. Based on the preliminary positioning results, each drone dynamically adjusts the flight path and signal sampling strategy to conduct high-frequency in-depth sampling of the area where the animal is located and its surrounding areas; S8, the central node optimizes the structure and parameters of the self-supervised learning model through the reinforcement learning strategy module based on the deep collection data; S9, the central node uses the optimized self-supervised learning model and deep acquisition data to accurately locate the animal's position through signal propagation model and collaborative computing algorithm.

2. A method for locating wild animals based on self-supervised learning according to claim 1, characterized in that: The specific steps of S1 are as follows: S11. Deploy signal receiving equipment on multiple drones. Each drone is equipped with a Ri(t) signal receiving module, where i∈{1, 2, ..., N} represents the i-th drone and t is a time variable used to mark the time point of receiving the signal. S12. The signal transmitter carried by wild animals sends a signal S j (f, t), j∈{1, 2, ..., M} represents the jth wild animal, f is the frequency variable of the signal, indicating the frequency range used by the transmitter; S13, each drone receives the transmission signal of wild animals in its flight coverage area S j (f, t), and the strength and phase information of the signal captured by the antenna array is expressed as the received signal data matrix: Y i (t)=A i ·S j (f,t)+n i (t); Among them, Yi(t) is the signal received by the i-th UAV at time t, Ai is the propagation coefficient of the signal, which reflects the attenuation and interference in the signal propagation path, and n i (t) is the noise interference during the receiving process; S14, using adaptive filtering algorithm in the signal receiving module of the UAV to perform preliminary denoising on the received signal data matrix Yi(t) to eliminate noise interference n i (t); S15, band-pass filtering is performed on the signal data after denoising, and the frequency range [f min , f max ], only keep the target signal S j (ft) effective frequency band, and the filtered signal is obtained S16, the filtered signal Normalization processing is performed to eliminate the amplitude inconsistency caused by the difference in receiving strength of different drones, and the normalized signal data matrix is ​​obtained:

3. A method for locating wild animals based on self-supervised learning according to claim 1, characterized in that: The specific steps of S2 are as follows: S21, each UAV based on the pre-processed signal data matrix Z i (t), the spectrum representation is obtained by converting the time domain signal to the frequency domain signal through Fourier transform: Among them, X i (f) is the spectrum representation of the signal received by the i-th UAV, where f is the frequency variable and j is the imaginary unit; S22, spectrum-based representation of X i (f) Extract frequency domain features, including the main frequency component f of the signal peak , signal bandwidth B and spectral energy density ε(f); S23. Calculate the statistical characteristics of the signal in the time domain, including the mean μ(Z i ), variance σ 2 (Z i ) and autocorrelation function R(τ), where the autocorrelation function R(τ) is: Where τ is the time delay and T is the observation period; S24, integrating the extracted frequency domain features and time domain features into a preliminary feature vector F i : F i =[f peak ,B,ε(f),μ(Z i ),s 2 (Z i ), R(τ)].

4. The method for locating wild animals based on self-supervised learning according to claim 1, characterized in that: The specific steps of S3 are as follows: S31, based on the preliminary feature vector F uploaded by each drone i Constructing a self-supervised learning model M θ , where θ is the self-supervised learning model parameter. The self-supervised learning model is based on F i As input, the latent feature representation H of the output animal signal i : H i =M θ (Be); Among them, H i is the potential feature representation of the animal signal corresponding to the i-th drone, which is used to capture the deep feature pattern of the animal signal; S32. Design signal pattern prediction task, self-supervised learning model M θ According to the potential feature representation H at the current moment i (t) Predict the preliminary feature vector for the next moment Among them, W p and b p is the weight and bias parameter of the prediction layer, by minimizing the prediction error Optimize the self-supervised learning model parameters so that the self-supervised learning model can learn the temporal variation pattern of animal signal characteristics: S33, construct a contrastive learning task, use the potential feature representations of different drones or at different times for comparison; define the positive sample pair as the feature representation of the same drone at adjacent times (H i (t), H i (t+1)), the negative sample pairs are the feature representations of different drones (H i (t), H j (t′)), the contrastive learning loss function is: in, is the cosine similarity, τ1 is the temperature parameter; S34, design the signal anomaly detection task, use the autoencoder structure to analyze the initial feature vector F i To refactor: in, and are encoder and decoder networks respectively; calculate the reconstruction error: When the reconstruction error exceeds the threshold δ, the corresponding signal is determined to be an abnormal signal. The self-supervised learning model parameters are optimized by minimizing the reconstruction error so that the self-supervised learning model can detect abnormal animal signal characteristics.

5. The method for locating wild animals based on self-supervised learning according to claim 1, characterized in that: The specific steps of S4 are as follows: S41. Embed the reinforcement learning strategy module into the self-supervised learning model M θ , the reinforcement learning strategy module uses the current state s of the self-supervised learning model t is the input, s t It consists of the latent feature representation and loss function value of the self-supervised learning model, defined as: Among them, H i (t) represents the potential feature representation of the animal signal of the i-th UAV at time t; S42. Define the action space A of reinforcement learning, including the update strategy of the self-supervised learning model parameter θ and the adjustment of the drone flight strategy. Action a t Defined as: a t =[Δθ t ,I don't i (t)]; Among them, Δθ t is the update amount of the self-supervised learning model parameter θ at time t, guiding the learning direction of the self-supervised learning model, Δu i (t) is the flight control command increment of the i-th UAV at time t, including the adjustment of position and speed, which is used to optimize the flight path and signal acquisition strategy of the UAV; Δu i (t)=[Δx i (t),Δy i (t),Δz i (t),Δv i (t)]; Where Δx i (t), Δy i (t), Δz i (t) is the spatial coordinate adjustment, Δv i (t) is the speed adjustment; S43. Construct reward function R t To evaluate the comprehensive performance of the self-supervised learning model and the drone swarm, the reward function integrating positioning accuracy, self-supervised learning model performance and resource consumption is defined as: Among them, ∈(t) is the wildlife positioning error at time t, α, β, γ are weight coefficients, and E i (t) is the energy consumption of the i-th UAV at time t, C i (t) is its communication resource consumption, c energy and c comm is the cost coefficient of unit energy and communication resources; S44, using deep reinforcement learning algorithm to optimize strategy and maximize the cumulative reward G: Among them, γ∈[0,1] is the discount factor, using the strategy network π φ (a t |s t ) parameterizes the strategy, φ is the policy network parameter, and the policy gradient method is used to update φ and the self-supervised learning model parameter θ: Where η is the learning rate, They are the gradients of the cumulative reward to the policy network parameters and the self-supervised learning model parameters, and the parameters are updated by the gradient ascent method.

6. The method for locating wild animals based on self-supervised learning according to claim 1, characterized in that: The specific steps of S5 are as follows: S51, each UAV extracts the initial feature vector The execution results of the self-supervised learning task are uploaded to the central node; S52. The central node defines a global feature matrix G(t) based on the uploaded feature information and task results. The global feature matrix G(t) records the feature vectors and potential features of all drones:

7. The method for locating wild animals based on self-supervised learning according to claim 1, characterized in that: The specific steps of S6 are as follows: S61, the central node receives and parses the global feature matrix G(t), and the central node combines the self-supervised learning model M θ Analyze and pattern match the global feature matrix G(t) to calculate each potential feature representation H i (t) Similarity with the reference feature: Among them, H ref is the reference feature representation in the historical data, and sim is the cosine similarity function, which is used to evaluate the matching degree of the signal pattern; S63, clustering analysis is performed on the feature vectors according to the matching results, the location of the wild animals is located, and the cluster center C of each group of feature vectors is defined. k (t): in, is the set of feature vectors in the kth cluster, C k (t) is the center point of the cluster, corresponding to the location of wild animals; S64. Use the reinforcement learning strategy module to evaluate the accuracy of the positioning result in real time and define the positioning error ∈(t): Among them, pi(t) is the actual position, To predict the position, the error ∈(t) is used as the input of the reward function to optimize the parameters θ of the self-supervised learning model; S65. Use the policy gradient method to optimize the self-supervised learning model parameters θ according to the value of the reward function, and optimize the self-supervised learning model to match and locate animal signal features.

8. The method for locating wild animals based on self-supervised learning according to claim 1, characterized in that: The specific steps of S9 are as follows: S91, the central node uses the optimized self-supervised learning model Receive the depth data and upload the feature vector F i With latent feature representation Integrate to generate the overall data matrix D(t): in, represents the potential feature representation of the i-th drone after the optimization model; S92. Calculate the path of potential features based on the signal propagation model, and define the signal propagation model equation as: P i (t)=p0+v i ·(t-t0)+n i (t); Among them, P i (t) represents the signal propagation path recorded by the i-th UAV at time t, p0 is the initial position of the signal source, v i is the signal propagation velocity vector, n i (t) is random noise interference; S93, the central node performs time synchronization processing on the signal paths of all drones and uses a triangulation algorithm to accurately locate the position of the animal. At time t, the signal propagation paths recorded by drones i, j, and k are P respectively. i (t), P j (t), P k (t), the position estimate is calculated by triangulation in, is the estimated position of the animal, × represents the vector cross product, which is used to calculate the intersection of the signal propagation planes enclosed by the three drones; S94, introduce the Kalman filter algorithm to update the estimated position in real time, and the precise position after Kalman filter optimization The central node packages the positioning results and distributes them to all drones via wireless communication links; S95, each drone receives accurate location data Adjust flight paths and optimize signal collection strategies so that the system can continuously track and collect real-time location information of wildlife.

Images