Single-objective real-time tracking method and system for power systems

Abstract

The application discloses a single-target real-time tracking method for a power system, comprising the following steps: acquiring a single-target real-time tracking data set for the power system; constructing a single-target real-time tracking primary model and training a single-target real-time tracking model based on a twin network and a Transformer architecture; and using the single-target real-time tracking model for the power system to complete single-target real-time tracking of the power system. The application also discloses a system for implementing the single-target real-time tracking method for the power system. The single-target real-time tracking method for the power system is realized through the twin network and the Transformer architecture, and the method not only realizes real-time tracking of a single target in the power system, but also has higher reliability and better accuracy.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of electrical automation, and particularly relates to a single-target real-time tracking method and system for a power system. BACKGROUND

[0002] With the development of economy and technology and the improvement of people's living standards, electric energy has become an essential secondary energy in people's production and life, bringing endless convenience to people's production and life. Therefore, ensuring the stable and reliable supply of electric energy has become one of the most important tasks of the power system.

[0003] Power system inspection is an important prerequisite for ensuring the normal operation of the power system. The 3D single-target tracking algorithm (SOT algorithm) has a wide range of applications in the field of power inspection, especially in the scenarios of unmanned aerial vehicle inspection and power grid equipment detection and maintenance. The SOT algorithm can continuously monitor the movement or position changes of objects over time. For example, when monitoring intruders moving along the perimeter or boundary of power facilities, the SOT algorithm can continuously monitor their positions and predict their next actions. When using unmanned aerial vehicles or ground robots for automatic inspection, the SOT algorithm can monitor and predict the path of the inspection equipment, ensuring that it operates within the predetermined route and adjusts the trajectory in real time as needed, thereby more effectively covering the target area. In addition, during infrastructure construction, maintenance or modification, the SOT algorithm can be used to monitor the movement and placement of materials, components or machinery, ensure their coordinate alignment, and predict possible deviations in advance. Finally, the SOT algorithm can also be used to track the growth of vegetation near power lines in order to arrange timely maintenance before the vegetation poses a threat to the lines.

[0004] Through feature extraction and target detection, the SOT algorithm can accurately track and predict the trajectory of power equipment. At present, most researches focus on feature analysis of point cloud data, i.e. by adding different modules to improve tracking accuracy. However, information loss may occur during feature extraction, especially during the conversion from high-dimensional features to low-dimensional features. In addition, point cloud data is sparse and noisy, especially when the target distance increases, which further exacerbates the difficulty of target detection and tracking. At the same time, when using traditional CNN algorithms to extract three-dimensional point cloud features, the fine-grained geometric details of power equipment are easily lost, which affects the detection accuracy. Statistical correlation algorithms, such as Bayesian methods, have a significant increase in computational complexity as the number of targets increases when dealing with multi-target tracking, leading to the curse of dimensionality in data association. SUMMARY

[0005] One of the purposes of the present application is to provide a single-target real-time tracking method for a power system with high reliability and good accuracy.

[0006] The second objective of this invention is to provide a system for implementing the aforementioned single-target real-time tracking method for power systems.

[0007] The single-target real-time tracking method for power systems provided by this invention includes the following steps:

[0008] S1. Obtain a single-target real-time tracking point cloud dataset for power systems;

[0009] S2. Based on Siamese networks and Transformer architecture, construct a basic model for real-time single-target tracking;

[0010] S3. Using the dataset obtained in step S1, train the primary single-target real-time tracking model constructed in step S2 to obtain the single-target real-time tracking model.

[0011] S4. Apply the single-target real-time tracking model obtained in step S3 to the power system to complete the single-target real-time tracking of the power system.

[0012] Step S2, which describes building a primary model for real-time single-target tracking based on Siamese networks and the Transformer architecture, includes the following steps:

[0013] A basic model for real-time single-target tracking is constructed based on Siamese networks and the Transformer architecture; the model's processing procedure is as follows:

[0014] A. Construct a point cloud feature extraction network based on a Siamese network, and use the constructed point cloud feature extraction network to extract the point cloud feature tensor of the search region and the point cloud feature tensor of the template region in the dataset obtained in step S1, and fuse them to obtain the final point cloud feature tensor.

[0015] B. Perform position encoding, tensor decomposition, and tensor completion operations on the final feature tensor of the point cloud obtained in step A;

[0016] C. Based on the Transformer architecture, obtain the local augmented feature tensor and the global augmented feature tensor corresponding to the final feature tensor of the point cloud;

[0017] D. The obtained local enhancement feature tensor and global enhancement feature tensor are weighted and fused to obtain the information feature tensor;

[0018] E. Perform inverse tensor decomposition and Hough transform on the obtained information feature tensor to obtain the target candidate bounding box;

[0019] F. The target candidate bounding boxes obtained in step E are processed using 3D center head, heatmap head, score head, and orientation head to obtain the final target 3D bounding box.

[0020] Step A specifically includes the following steps:

[0021] A1. For the original point cloud of the search region and the original point cloud of the template region in the point cloud dataset obtained in step S1, the farthest point is sampled through set abstraction operation to obtain the feature tensor of the search point cloud and the feature tensor of the template point cloud.

[0022] A2. Perform a segmentation operation on the feature tensor obtained in step A1 to obtain the coordinates of each point in the search region and the template region, as well as the attribute of the point; the attribute is either a foreground point or a background point.

[0023] A3. Perform a classification operation on the feature tensors obtained in step A2 to obtain the feature tensors of each foreground element category in the search region and the feature tensors of each foreground element category in the template region.

[0024] A4. Using a Siamese network, the feature tensors obtained in step A3 are fused to obtain the final feature tensor of the point cloud.

[0025] Step A1 specifically includes the following steps:

[0026] A1-1. Obtain the original point cloud data of the search area and template area. The size of the original point cloud data is 8×16384×3, where the batch size is 8, the number of points is 16384, and the dimension is 3.

[0027] A1-2. Perform feature permutation on the original point cloud data, and the output size is 8×3×16384;

[0028] A1-3. Perform farthest point sampling on the obtained point cloud data, and the output size is 8×3×4096;

[0029] A1-4. The results obtained from sampling the farthest point are grouped for sphere query to generate 4096 local regions; for each local region, two types of sphere radii, 16 and 32, are generated, and the output sizes are 8×3×4096×16 and 8×3×4096×32.

[0030] A1-5. Perform point network operations on the point cloud data contained in the region with a radius of 16 and the point cloud data contained in the region with a radius of 32 respectively;

[0031] A1-6. Perform connection, feature permutation and farthest point sampling operations on the results of the point network operation obtained in step A1-5 in sequence;

[0032] A1-7. Perform the sphere query grouping operation again on the results obtained in step A1-6 to generate 4096 local regions; for each local region, generate two types of sphere radii: 16 and 32.

[0033] A1-8. Perform point network operations on the point cloud data contained in the region with a radius of 16 and the point cloud data contained in the region with a radius of 32 respectively;

[0034] A1-9. Connect the results of the point network operation obtained in step A1-8 to obtain the final search point cloud feature tensor and template point cloud feature tensor.

[0035] Steps A1-5 specifically include the following steps:

[0036] For point cloud data encompassing a region with a radius of 16, the following steps are used to perform point network operations:

[0037] A rigid body transformation is performed on the point cloud data contained in a region with a radius of 16, and the size of the transformation matrix is ​​3×3;

[0038] The obtained output is subjected to a fully connected operation based on a multilayer perceptron network with a kernel size of 3×1, a channel number of 64, and a step size of 1.

[0039] The obtained output is subjected to rigid body transformation again, and the size of the transformation matrix is ​​64×64;

[0040] The obtained output is subjected to a fully connected operation based on a multilayer perceptron network with a kernel size of 3×1, a channel number of 128, and a step size of 1.

[0041] The obtained output is then subjected to a fully connected operation based on a multilayer perceptron network, with a kernel size of 3×1, a channel number of 1024, and a step size of 1.

[0042] The obtained output is subjected to max pooling, with a kernel size of 3×3 and 32 output channels;

[0043] The obtained output is then subjected to max pooling again, with a kernel size of 3×3 and an output channel of 1.

[0044] For point cloud data encompassing a region with a radius of 32, the following steps are used to perform point network operations:

[0045] A rigid body transformation is performed on the point cloud data contained in a region with a radius of 32, and the size of the transformation matrix is ​​3×3;

[0046] The obtained output is subjected to a fully connected operation based on a multilayer perceptron network with a kernel size of 3×1, a channel number of 64, and a step size of 1.

[0047] The obtained output is subjected to rigid body transformation again, and the size of the transformation matrix is ​​64×64;

[0048] The obtained output is subjected to a fully connected operation based on a multilayer perceptron network with a kernel size of 3×1, a channel number of 128, and a step size of 1.

[0049] The obtained output is then subjected to a fully connected operation based on a multilayer perceptron network, with a kernel size of 3×1, a channel number of 1024, and a step size of 1.

[0050] The obtained output is subjected to max pooling, with a kernel size of 3×3 and 64 output channels;

[0051] The obtained output is then subjected to max pooling again, with a kernel size of 3×3 and an output channel of 1.

[0052] Steps A1-7 specifically include the following steps:

[0053] For point cloud data encompassing a region with a radius of 16, the following steps are used to perform point network operations:

[0054] A rigid body transformation is performed on the point cloud data contained in a region with a radius of 16, and the size of the transformation matrix is ​​3×3;

[0055] The obtained output is subjected to a fully connected operation based on a multilayer perceptron network with a kernel size of 3×1, a channel number of 256, and a step size of 1.

[0056] The obtained output is subjected to rigid body transformation again, and the size of the transformation matrix is ​​256×256;

[0057] The obtained output is subjected to a fully connected operation based on a multilayer perceptron network with a kernel size of 3×1, a channel number of 512, and a step size of 1.

[0058] The obtained output is then subjected to a fully connected operation based on a multilayer perceptron network, with a kernel size of 3×1, a channel number of 1024, and a step size of 1.

[0059] The obtained output is subjected to max pooling, with a kernel size of 3×3 and 128 output channels;

[0060] The obtained output is then subjected to max pooling again, with a kernel size of 3×3 and an output channel of 1.

[0061] For point cloud data encompassing a region with a radius of 32, the following steps are used to perform point network operations:

[0062] A rigid body transformation is performed on the point cloud data contained in a region with a radius of 32, and the size of the transformation matrix is ​​3×3;

[0063] The obtained output is subjected to a fully connected operation based on a multilayer perceptron network with a kernel size of 3×1, a channel number of 64, and a step size of 1.

[0064] The obtained output is subjected to rigid body transformation again, and the size of the transformation matrix is ​​256×256;

[0065] The obtained output is subjected to a fully connected operation based on a multilayer perceptron network with a kernel size of 3×1, a channel number of 512, and a step size of 1.

[0066] The obtained output is then subjected to a fully connected operation based on a multilayer perceptron network, with a kernel size of 3×1, a channel number of 1024, and a step size of 1.

[0067] The obtained output is subjected to max pooling, with a kernel size of 3×3 and 128 output channels;

[0068] The obtained output is then subjected to max pooling again, with a kernel size of 3×3 and an output channel of 1.

[0069] Step A2 specifically includes the following steps:

[0070] The feature tensor obtained in step A1 is then subjected to upsampling and feature concatenation operations in sequence.

[0071] The resulting output of 8×128×1280 is then subjected to a point network operation; the point network operation specifically includes the following steps:

[0072] A rigid body transformation is performed on the point cloud data contained in a region with a radius of 16, and the size of the transformation matrix is ​​3×3;

[0073] The obtained output is subjected to a fully connected operation based on a multilayer perceptron network with a kernel size of 3×1, a channel number of 256, and a step size of 1.

[0074] The obtained output is subjected to rigid body transformation again, and the size of the transformation matrix is ​​256×256;

[0075] The obtained output is subjected to a fully connected operation based on a multilayer perceptron network with a kernel size of 3×1, a channel number of 512, and a step size of 1.

[0076] The obtained output is then subjected to a fully connected operation based on a multilayer perceptron network, with a kernel size of 3×1, a channel number of 256, and a step size of 1.

[0077] The obtained output is subjected to max pooling, with a kernel size of 3×3 and 128 output channels;

[0078] The obtained output is then subjected to max pooling again, with a kernel size of 3×3 and an output channel of 1.

[0079] For the output of the point network, for points ignored during sampling, find k known points in the neighborhood and use a computational formula. Perform upsampling; where f j (x) is a j-dimensional feature. d represents distance, f i (j) Let j be the j-dimensional feature of the i-th point;

[0080] Perform feature concatenation on the upsampled output;

[0081] The resulting output 8×16384×128+6 is then subjected to a point network operation; the point network operation specifically includes the following steps:

[0082] The obtained output is processed by a fully connected multilayer perceptron network with a kernel size of 3×1, 64 channels, and a stride of 1.

[0083] A rigid body transformation is performed on the point cloud data contained in a region with a radius of 16, and the size of the transformation matrix is ​​3×3;

[0084] The obtained output is subjected to a fully connected operation based on a multilayer perceptron network with a kernel size of 3×1, a channel number of 256, and a step size of 1.

[0085] The obtained output is subjected to rigid body transformation again, and the size of the transformation matrix is ​​256×256;

[0086] The obtained output is subjected to a fully connected operation based on a multilayer perceptron network with a kernel size of 3×1, a channel number of 512, and a step size of 1.

[0087] The obtained output is then subjected to a fully connected operation based on a multilayer perceptron network, with a kernel size of 3×1, a channel number of 1024, and a step size of 1.

[0088] The obtained output is subjected to 4 max pooling operations with a kernel size of 3×3 and 16384 output channels.

[0089] The obtained output is then subjected to max pooling again, with a kernel size of 3×1 and an output channel of 1.

[0090] The output is subjected to two convolution operations in sequence; the kernel size is 3×3 and the stride is 8.

[0091] Step A3 specifically includes the following steps:

[0092] Perform a rigid body transformation operation on the feature tensor obtained in step A2, with a transformation matrix size of 256×256;

[0093] The obtained output is subjected to a fully connected operation based on a multilayer perceptron, with a kernel size of 3×1, a channel number of 512, and a step size of 1.

[0094] The obtained output is subjected to rigid body transformation again, with a transformation matrix size of 512×512;

[0095] The obtained output is then subjected to a fully connected operation based on a multilayer perceptron, with a kernel size of 3×1, a channel count of 1024, and a step size of 1.

[0096] The output is subjected to a 4-layer fully connected operation to obtain the final classification result; the kernel size of the fully connected operation is 3×1, the number of channels is 1, and the stride is 1.

[0097] Step A4 specifically includes the following steps:

[0098] The following formula is used to continue the fusion of feature tensors:

[0099]

[0100] In the formula d + f(X) is the norm distance between two positive samples. + ) represents the features of positive samples; X + For positive samples; f(X) a ) represents the characteristics of the anchor point; X a For anchor point; d - The norm distance between two negative samples; f(X) - ) represents the features of negative samples; X - For negative samples; || ||2 is the 2-norm.

[0101] Step B specifically includes the following steps:

[0102] Perform a hash position encoding operation on the final feature tensor of the point cloud obtained in step A:

[0103] A hash position encoding is constructed by creating an optimization problem in a high-dimensional space. Based on the nearest neighbor Lagrange optimization problem, the center point is determined, and neighboring points are found for the center point to form a local point cluster structure in which the center point is represented by the neighboring points. The optimal weight coefficient in the nearest neighbor is obtained.

[0104] By introducing the optimal weight coefficients into the low-dimensional space, we can find the low-dimensional point cluster structure that preserves the relative positional relationship between high-dimensional point clusters. We can then uniquely project the high-dimensional features into a hash table in the low-dimensional space to obtain the positional encoding of the high-dimensional data in the low-dimensional space.

[0105] Perform tensor decomposition:

[0106] Tensor decomposition is used to reduce the dimensionality of feature tensors, extract the core sub-tensors of feature tensors, and thus speed up the computation process;

[0107] Establish a combination of display patterns and latent factors; given any tensor X, the corresponding augmented tensor decomposition is represented as...

[0108]

[0109] In the formula x ijt μ is the element of tensor X at position (i,j,t); i φ represents the weighting coefficients of the elements in the core subtensor. i Let θ be the molecular matrix weight vector along the U direction. i Let η be the molecular matrix weight vector in the V direction. i g is the molecular matrix weight vector in the W direction; ik For core sub-tensor elements; u ik Let v be the molecular matrix vector along the U direction. ik Let w be the molecular matrix vector along the V direction. ik is the molecular matrix vector in the W direction; r is the rank of the tensor;

[0110] Perform tensor completion:

[0111] Given a data matrix X, find three tensors y, z, and τ to approximate X, obtaining an approximate value X. * ;

[0112] The spectrum soft thresholding of the data matrix X is used as a shrinkage process for singular values, defined as follows:

[0113]

[0114] In the formula Let U be the projection operator on the radius λ spectral norm sphere; U and V are both orthogonal matrices; X = US'V T SVD for matrix transformation of input tensor X.

[0115] Step C specifically includes the following steps:

[0116] The feature tensor and positional encoding obtained in step B are used as input to the Transformer architecture to mine the correlation between different targets in the same frame, resulting in a locally enhanced feature tensor. The processing includes the following steps:

[0117] The template point cloud feature tensor is input to the embedding layer, where the transformation matrix has a size of 8×256×256.

[0118] The obtained output is subjected to encoder stacking operation, where the kernel size is 3×1, the stacking layer is 6 layers, the number of channels is 512, and the step size is 1.

[0119] The output is subjected to a multi-head attention mechanism, which is split into 8 heads, each with a dimension of 64.

[0120] The attention weights are calculated independently for each of the eight heads, and the outputs of the eight heads are then concatenated.

[0121] The obtained output is then subjected to residual concatenation and normalization operations.

[0122] The obtained output is then processed by a feedforward neural network; the hidden layer size is 8×512×512.

[0123] The obtained output is then subjected to residual connection and normalization operations again;

[0124] Finally, the local enhancement feature tensor is obtained;

[0125] The obtained local augmented feature tensor and positional encoding are used as input to the Transformer architecture to mine the correlation between the same target in different frames, resulting in a global augmented feature tensor. The processing includes the following steps:

[0126] The local enhancement feature tensor is input to the embedding layer, where the transformation matrix has a size of 8×256×256.

[0127] The obtained output is subjected to encoder stacking operation, where the kernel size is 3×1, the stacking layer is 6 layers, the number of channels is 512, and the step size is 1.

[0128] The output is subjected to a multi-head attention mechanism, which is split into 8 heads, each with a dimension of 64.

[0129] The attention weights are calculated independently for each of the eight heads, and the outputs of the eight heads are then concatenated.

[0130] The obtained output is then subjected to residual concatenation and normalization operations.

[0131] The obtained output is then processed by a feedforward neural network; the hidden layer size is 8×512×2048.

[0132] The obtained output is then subjected to residual connection and normalization operations again;

[0133] Finally, the global enhanced feature tensor is obtained.

[0134] The training described in step S3 specifically includes the following steps:

[0135] The following formula is used as the weighted loss function L. w :

[0136]

[0137] In the formula M s a is the number of point clouds; i This represents the label of the i-th seed point; Use the smooth activation function; S represents the spatial coordinates of the point cloud; l The coordinates of the midpoint of the ground truth; δ i Let δ be the weight vector. i =MLP(f i g ), f i g For the i-th ground real feature, MLP() is the processing function of the multilayer perceptron;

[0138] The following formula is used as the training loss function L. cls :

[0139]

[0140] In the formula, ρ() is the predictor classifier; Positive samples; For negative samples; k + k is the number of positive samples. - The number of negative samples;

[0141] The following formula is used as the value difference loss function L. t :

[0142]

[0143] In the formula, softmax() is the softmax function; These are hyperparameters used to balance the loss; To represent a positive sample linear layer with a sigmoid activation function; To represent a linear layer of negative samples with a sigmoid activation function;

[0144] The following formula is used as the location-aware ranking loss function L. l :

[0145]

[0146] In the formula For paired samples; The crossover ratio (CUP) is a metric used to calculate the crossover ratio for positive samples. The crossover ratio (CRO) is used to calculate the index for negative samples.

[0147] Finally, the total training function L is obtained as follows:

[0148] L=λ1L w +λ2L cls +λ3L t +λ4L l

[0149] In the formula, λ1 is the first weight; λ2 is the second weight; λ3 is the third weight; and λ4 is the fourth weight.

[0150] This invention also provides a system for implementing the single-target real-time tracking method for power systems, comprising a data acquisition module, a model building module, a model training module, and a real-time tracking module; the data acquisition module, model building module, model training module, and real-time tracking module are connected in series; the data acquisition module acquires a single-target real-time tracking dataset for power systems and uploads the data information to the model building module; the model building module constructs a primary single-target real-time tracking model based on a Siamese network and Transformer architecture according to the received data information and uploads the data information to the model training module; the model training module trains the primary single-target real-time tracking model constructed in step S2 using the acquired dataset according to the received data information to obtain a single-target real-time tracking model and uploads the data information to the real-time tracking module; the real-time tracking module applies the obtained single-target real-time tracking model to the power system according to the received data information to complete the single-target real-time tracking of the power system.

[0151] The single-target real-time tracking method and system for power systems provided by this invention, through twin networks and Transformer architecture, not only achieves real-time tracking of single targets in power systems, but also has higher reliability and better accuracy. Attached Figure Description

[0152] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0153] Figure 2 This is a schematic diagram of the position encoding in the method of the present invention.

[0154] Figure 3 This is a schematic diagram of tensor decomposition in the method of the present invention.

[0155] Figure 4 This is a schematic diagram of tensor completion in the method of the present invention.

[0156] Figure 5 This is a schematic diagram of tensor restoration in the method of the present invention.

[0157] Figure 6This is a schematic diagram comparing the tracking effects of an embodiment of the method of the present invention.

[0158] Figure 7 This is a schematic diagram of the functional modules of the system of the present invention. Detailed Implementation

[0159] like Figure 1 The diagram shown is a flowchart of the method of the present invention: The single-target real-time tracking method for power systems disclosed in this invention includes the following steps:

[0160] S1. Obtain a single-target real-time tracking dataset for power systems;

[0161] S2. Based on Siamese networks and the Transformer architecture, construct a basic model for real-time single-target tracking; including the following steps:

[0162] A basic model for real-time single-target tracking is constructed based on Siamese networks and the Transformer architecture; the model's processing procedure is as follows:

[0163] A. Construct a point cloud feature extraction network based on a Siamese network, and use the constructed point cloud feature extraction network to extract the search region point cloud feature tensor and the template region point cloud feature tensor from the dataset obtained in step S1, and fuse them to obtain the final point cloud feature tensor; specifically including the following steps:

[0164] A1. For the original point clouds of the search region and the original point clouds of the template region in the dataset obtained in step S1, the farthest point is sampled through set abstraction operations to obtain the feature tensors of the search point cloud and the template point cloud; specifically, the following steps are included:

[0165] A1-1. Obtain the original point cloud data of the search area and template area. The size of the original point cloud data is 8×16384×3, where the batch size is 8, the number of points is 16384, and the dimension is 3.

[0166] A1-2. Perform feature permutation on the original point cloud data, and the output size is 8×3×16384;

[0167] A1-3. Perform farthest point sampling on the obtained point cloud data, and the output size is 8×3×4096;

[0168] A1-4. The results obtained from sampling the farthest point are grouped for sphere query to generate 4096 local regions; for each local region, two types of sphere radii, 16 and 32, are generated, and the output sizes are 8×3×4096×16 and 8×3×4096×32.

[0169] A1-5. Perform point network operations on the point cloud data contained in a region with a radius of 16 and the point cloud data contained in a region with a radius of 32, respectively; specifically including the following steps:

[0170] For point cloud data encompassing a region with a radius of 16, the following steps are used to perform point network operations:

[0171] A rigid body transformation is performed on the point cloud data contained in a region with a radius of 16, and the size of the transformation matrix is ​​3×3;

[0172] The obtained output is subjected to a fully connected operation based on a multilayer perceptron network with a kernel size of 3×1, a channel number of 64, and a step size of 1.

[0173] The obtained output is subjected to rigid body transformation again, and the size of the transformation matrix is ​​64×64;

[0174] The obtained output is subjected to a fully connected operation based on a multilayer perceptron network with a kernel size of 3×1, a channel number of 128, and a step size of 1.

[0175] The obtained output is then subjected to a fully connected operation based on a multilayer perceptron network, with a kernel size of 3×1, a channel number of 1024, and a step size of 1.

[0176] The obtained output is subjected to max pooling, with a kernel size of 3×3 and 32 output channels;

[0177] The obtained output is then subjected to max pooling again, with a kernel size of 3×3 and an output channel of 1.

[0178] For point cloud data encompassing a region with a radius of 32, the following steps are used to perform point network operations:

[0179] A rigid body transformation is performed on the point cloud data contained in a region with a radius of 32, and the size of the transformation matrix is ​​3×3;

[0180] The obtained output is subjected to a fully connected operation based on a multilayer perceptron network with a kernel size of 3×1, a channel number of 64, and a step size of 1.

[0181] The obtained output is subjected to rigid body transformation again, and the size of the transformation matrix is ​​64×64;

[0182] The obtained output is subjected to a fully connected operation based on a multilayer perceptron network with a kernel size of 3×1, a channel number of 128, and a step size of 1.

[0183] The obtained output is then subjected to a fully connected operation based on a multilayer perceptron network, with a kernel size of 3×1, a channel number of 1024, and a step size of 1.

[0184] The obtained output is subjected to max pooling, with a kernel size of 3×3 and 64 output channels;

[0185] The obtained output is then subjected to max pooling again, with a kernel size of 3×3 and an output channel of 1.

[0186] A1-6. Perform connection, feature permutation and farthest point sampling operations on the results of the point network operation obtained in step A1-5 in sequence;

[0187] A1-7. Using the results obtained in step A1-6, perform another grouping operation on the sphere query to generate 4096 local regions; for each local region, generate two types of sphere radii: 16 and 32; specifically including the following steps:

[0188] For point cloud data encompassing a region with a radius of 16, the following steps are used to perform point network operations:

[0189] A rigid body transformation is performed on the point cloud data contained in a region with a radius of 16, and the size of the transformation matrix is ​​3×3;

[0190] The obtained output is subjected to a fully connected operation based on a multilayer perceptron network with a kernel size of 3×1, a channel number of 256, and a step size of 1.

[0191] The obtained output is subjected to rigid body transformation again, and the size of the transformation matrix is ​​256×256;

[0192] The obtained output is subjected to a fully connected operation based on a multilayer perceptron network with a kernel size of 3×1, a channel number of 512, and a step size of 1.

[0193] The obtained output is then subjected to a fully connected operation based on a multilayer perceptron network, with a kernel size of 3×1, a channel number of 1024, and a step size of 1.

[0194] The obtained output is subjected to max pooling, with a kernel size of 3×3 and 128 output channels;

[0195] The obtained output is then subjected to max pooling again, with a kernel size of 3×3 and an output channel of 1.

[0196] For point cloud data encompassing a region with a radius of 32, the following steps are used to perform point network operations:

[0197] A rigid body transformation is performed on the point cloud data contained in a region with a radius of 32, and the size of the transformation matrix is ​​3×3;

[0198] The obtained output is subjected to a fully connected operation based on a multilayer perceptron network with a kernel size of 3×1, a channel number of 64, and a step size of 1.

[0199] The obtained output is subjected to rigid body transformation again, and the size of the transformation matrix is ​​256×256;

[0200] The obtained output is subjected to a fully connected operation based on a multilayer perceptron network with a kernel size of 3×1, a channel number of 512, and a step size of 1.

[0201] The obtained output is then subjected to a fully connected operation based on a multilayer perceptron network, with a kernel size of 3×1, a channel number of 1024, and a step size of 1.

[0202] The obtained output is subjected to max pooling, with a kernel size of 3×3 and 128 output channels;

[0203] The obtained output is then subjected to max pooling again, with a kernel size of 3×3 and an output channel of 1.

[0204] A1-8. Perform point network operations on the point cloud data contained in the region with a radius of 16 and the point cloud data contained in the region with a radius of 32 respectively;

[0205] A1-9. Connect the results of the point network operation obtained in step A1-8 to obtain the final search point cloud feature tensor and template point cloud feature tensor;

[0206] A2. Perform segmentation on the feature tensor obtained in step A1 to obtain the coordinates of each point in the search region and the template region, as well as the attribute of that point; the attribute is either a foreground point or a background point; specifically, this includes the following steps:

[0207] The feature tensor obtained in step A1 is then subjected to upsampling and feature concatenation operations in sequence.

[0208] The resulting output of 8×128×1280 is then subjected to a point network operation; the point network operation is as follows:

[0209] A rigid body transformation is performed on the point cloud data contained in a region with a radius of 16, and the size of the transformation matrix is ​​3×3;

[0210] The obtained output is subjected to a fully connected operation based on a multilayer perceptron network with a kernel size of 3×1, a channel number of 256, and a step size of 1.

[0211] The obtained output is subjected to rigid body transformation again, and the size of the transformation matrix is ​​256×256;

[0212] The obtained output is subjected to a fully connected operation based on a multilayer perceptron network with a kernel size of 3×1, a channel number of 512, and a step size of 1.

[0213] The obtained output is then subjected to a fully connected operation based on a multilayer perceptron network, with a kernel size of 3×1, a channel number of 256, and a step size of 1.

[0214] The obtained output is subjected to max pooling, with a kernel size of 3×3 and 128 output channels;

[0215] The obtained output is then subjected to max pooling again, with a kernel size of 3×3 and an output channel of 1.

[0216] For the output of the point network, for points ignored during sampling, find k known points in the neighborhood and use a computational formula. Perform upsampling; where f j (x) is a j-dimensional feature. d represents the distance; the smaller the distance, the larger the weight. i (j) Let j be the j-dimensional feature of the i-th point;

[0217] Perform feature concatenation on the upsampled output;

[0218] The resulting output 8×16384×128+6 is then subjected to a point network operation; the point network operation is as follows:

[0219] The obtained output is processed by a fully connected multilayer perceptron network with a kernel size of 3×1, 64 channels, and a stride of 1.

[0220] A rigid body transformation is performed on the point cloud data contained in a region with a radius of 16, and the size of the transformation matrix is ​​3×3;

[0221] The obtained output is subjected to a fully connected operation based on a multilayer perceptron network with a kernel size of 3×1, a channel number of 256, and a step size of 1.

[0222] The obtained output is subjected to rigid body transformation again, and the size of the transformation matrix is ​​256×256;

[0223] The obtained output is subjected to a fully connected operation based on a multilayer perceptron network with a kernel size of 3×1, a channel number of 512, and a step size of 1.

[0224] The obtained output is then subjected to a fully connected operation based on a multilayer perceptron network, with a kernel size of 3×1, a channel number of 1024, and a step size of 1.

[0225] The obtained output is subjected to 4 max pooling operations with a kernel size of 3×3 and 16384 output channels.

[0226] The obtained output is then subjected to max pooling again, with a kernel size of 3×1 and an output channel of 1.

[0227] The output is then subjected to two convolution operations in sequence; the kernel size is 3×3 and the stride is 8.

[0228] A3. Perform a classification operation on the feature tensors obtained in step A2 to obtain the feature tensors of each foreground point category in the search region and the feature tensors of each foreground point category in the template region; specifically, this includes the following steps:

[0229] Perform a rigid body transformation operation on the feature tensor obtained in step A2, with a transformation matrix size of 256×256;

[0230] The obtained output is subjected to a fully connected operation based on a multilayer perceptron, with a kernel size of 3×1, a channel number of 512, and a step size of 1.

[0231] The obtained output is subjected to rigid body transformation again, with a transformation matrix size of 512×512;

[0232] The obtained output is then subjected to a fully connected operation based on a multilayer perceptron, with a kernel size of 3×1, a channel count of 1024, and a step size of 1.

[0233] The output is subjected to a 4-layer fully connected operation to obtain the final classification result; the kernel size of the fully connected operation is 3×1, the number of channels is 1, and the stride is 1.

[0234] A4. Using a Siamese network, the feature tensors obtained in step A3 are fused to obtain the final feature tensor of the point cloud; specifically, the following steps are included:

[0235] The following formula is used to continue the fusion of feature tensors:

[0236]

[0237] In the formula d + f(X) is the norm distance between two positive samples. + ) represents the features of positive samples; X + For positive samples; f(X) a ) represents the characteristics of the anchor point; X a For anchor point; d - The norm distance between two negative samples; f(X) - ) represents the features of negative samples; X - For negative samples; || ||2 is the 2-norm.

[0238] B. Perform position encoding, tensor decomposition, and tensor completion operations on the final feature tensor of the point cloud obtained in step A; specifically, this includes the following steps:

[0239] Perform a hash position encoding operation on the final feature tensor of the point cloud obtained in step A:

[0240] Constructing a hash positional encoding involves creating an optimization problem in a high-dimensional space, based on a nearest-neighbor Lagrangian optimization problem, to determine the center point and find neighboring points for that center point, forming a local point cluster structure where the center point is represented by neighboring points (e.g., ...). Figure 2 (b) is shown), and the optimal weight coefficients in the nearest neighbors are obtained;

[0241] By introducing the optimal weight coefficients into the low-dimensional space, a low-dimensional point cluster structure that preserves the relative positional relationships between high-dimensional point clusters is found. The high-dimensional features are then uniquely projected onto a hash table in the low-dimensional space (e.g., ...). Figure 2 (a) shows the positional encoding of the high-dimensional data in the low-dimensional space;

[0242] In practice, given m n-dimensional samples {x1, x2, ..., x...} m}∈R n and its corresponding k nearest neighbors, x i The linear relationship between the k nearest neighbors can be expressed as

[0243]

[0244]

[0245] Then the weighting coefficient w i for Z i =(x i -x j ) T (x i -x j );

[0246] Given an n-dimensional sample {x1, x2, ..., x} m}∈R m×n The corresponding low-dimensional space coordinates {y1,y2,...,y m}∈R m×d And the weighting coefficient w, let Y i (k) ∈R k×d Given a matrix composed of its k nearest neighbors, the optimization problem of uniquely projecting an n-dimensional sample onto a d-dimensional matrix is ​​expressed as:

[0247]

[0248] There exists a mapping Y satisfying Y = {y2, y3, ..., y...} d+1}, where y i For matrix (IW) T The eigenvectors corresponding to the eigenvalues ​​of (IW);

[0249] Perform tensor decomposition:

[0250] Tensor decomposition is used to reduce the dimensionality of feature tensors, extract the core sub-tensors of feature tensors, and thus speed up the computation process;

[0251] Establish a combination of display patterns and underlying factors, such as Figure 3 As shown, augmented tensor decomposition is used to decompose the eigenvalues ​​of higher-order tensors in the point cloud. Given any tensor X, the corresponding augmented tensor decomposition is expressed as follows:

[0252]

[0253] In the formula x ijt μ is the element of tensor X at position (i,j,t); i φ represents the weighting coefficients of the elements in the core subtensor. i Let θ be the molecular matrix weight vector along the U direction. i Let η be the molecular matrix weight vector in the V direction. i g is the molecular matrix weight vector in the W direction; ik For core sub-tensor elements; u ik Let v be the molecular matrix vector along the U direction. ik Let w be the molecular matrix vector along the V direction. ik is the molecular matrix vector in the W direction; r is the rank of the tensor;

[0254] Perform tensor completion:

[0255] The process of preserving core subtensors in tensor decomposition based on high-dimensional features of point clouds suffers from data loss. This is addressed by constructing a tensor completion optimization problem based on an alternating direction multiplier algorithm to supplement and correct this loss, thereby uncovering more feature and relational information from the point cloud. Figure 4 As shown;

[0256] Given a data matrix X, find three tensors y, z, and τ to approximate X, obtaining an approximate value X. * ;

[0257] The spectrum soft thresholding of the data matrix X is used as a shrinkage process for singular values, defined as follows:

[0258]

[0259] In the formula Let U be the projection operator on the radius λ spectral norm sphere; U and V are orthogonal matrices; X = US'V T It is a matrix-based SVD of the input tensor X;

[0260] In practice, given any tensor X and observation vector y, the k-type expansion matrix of X is X k Vectorize it to x; take an auxiliary matrix Z with the same size as the k-type expansion of the tensor X.k Vectorized into z k Then there exists an observation operator Ω: And the regularization constant λ, which makes the trace norm optimization problem of the spectral soft thresholding operator...

[0261]

[0262] stX k x = z k

[0263] The optimality condition can be expressed as

[0264]

[0265] in For regularization terms; || || F For spectral norm; || || * Linear sum operations on the singular values ​​of a matrix;

[0266] C. Based on the Transformer architecture, obtain the local augmented feature tensor and the global augmented feature tensor corresponding to the final feature tensor of the point cloud; specifically including the following steps:

[0267] The feature tensor and positional encoding obtained in step B are used as input to the Transformer architecture to mine the correlation between different targets in the same frame, resulting in a locally enhanced feature tensor. The processing includes the following steps:

[0268] The template point cloud feature tensor is input to the embedding layer, where the transformation matrix has a size of 8×256×256.

[0269] The obtained output is subjected to encoder stacking operation, where the kernel size is 3×1, the stacking layer is 6 layers, the number of channels is 512, and the step size is 1.

[0270] The output is subjected to a multi-head attention mechanism, which is split into 8 heads, each with a dimension of 64.

[0271] The attention weights are calculated independently for each of the eight heads, and the outputs of the eight heads are then concatenated.

[0272] The obtained output is then subjected to residual concatenation and normalization operations.

[0273] The obtained output is then processed by a feedforward neural network; the hidden layer size is 8×512×512.

[0274] The obtained output is then subjected to residual connection and normalization operations again;

[0275] Finally, the local enhancement feature tensor is obtained;

[0276] The obtained local augmented feature tensor and positional encoding are used as input to the Transformer architecture to mine the correlation between the same target in different frames, resulting in a global augmented feature tensor. The processing includes the following steps:

[0277] The local enhancement feature tensor is input to the embedding layer, where the transformation matrix has a size of 8×256×256.

[0278] The obtained output is subjected to encoder stacking operation, where the kernel size is 3×1, the stacking layer is 6 layers, the number of channels is 512, and the step size is 1.

[0279] The output is subjected to a multi-head attention mechanism, which is split into 8 heads, each with a dimension of 64.

[0280] The attention weights are calculated independently for each of the eight heads, and the outputs of the eight heads are then concatenated.

[0281] The obtained output is then subjected to residual concatenation and normalization operations.

[0282] The obtained output is then processed by a feedforward neural network; the hidden layer size is 8×512×2048.

[0283] The obtained output is then subjected to residual connection and normalization operations again;

[0284] Finally, the global enhanced feature tensor is obtained;

[0285] D. The obtained local enhancement feature tensor and global enhancement feature tensor are weighted and fused to obtain the information feature tensor;

[0286] E. Perform inverse tensor decomposition on the obtained information feature tensor (e.g.) Figure 5 (as shown) and Hough transform to obtain the target candidate bounding box;

[0287] F. The target candidate bounding boxes obtained in step E are processed using 3D center head, heat map head, score head and direction head to obtain the final target 3D bounding box;

[0288] S3. Using the dataset obtained in step S1, train the primary single-target real-time tracking model constructed in step S2 to obtain the single-target real-time tracking model.

[0289] In practice, the training includes the following steps:

[0290] The following formula is used as the weighted loss function L. w :

[0291]

[0292] In the formula M s a is the number of point clouds; i This represents the label of the i-th seed point; Use the smooth activation function; S represents the spatial coordinates of the point cloud; l The coordinates of the midpoint of the ground truth; δ i Let δ be the weight vector. i =MLP(f i g ), f i g For the i-th ground real feature, MLP() is the processing function of the multilayer perceptron;

[0293] The following formula is used as the training loss function L. cls :

[0294]

[0295] In the formula, ρ() is the predictor classifier; Positive samples; For negative samples; k + k is the number of positive samples. - The number of negative samples;

[0296] The following formula is used as the value difference loss function L. t :

[0297]

[0298] In the formula, softmax() is the softmax function; These are hyperparameters used to balance the loss; To represent a positive sample linear layer with a sigmoid activation function; To represent a linear layer of negative samples with a sigmoid activation function;

[0299] The following formula is used as the location-aware ranking loss function L. l :

[0300]

[0301] In the formula For paired samples; The crossover ratio (CUP) is a metric used to calculate the crossover ratio for positive samples. The crossover ratio (CRO) is used to calculate the index for negative samples.

[0302] Finally, the total training function L is obtained as follows:

[0303] L=λ1L w+λ2L cls +λ3L t +λ4L l

[0304] In the formula, λ1 is the first weight; λ2 is the second weight; λ3 is the third weight; and λ4 is the fourth weight.

[0305] S4. Apply the single-target real-time tracking model obtained in step S3 to the power system to complete the single-target real-time tracking of the power system.

[0306] The following example illustrates the implementation effect of the method of the present invention:

[0307] like Figure 6 The diagram shows a comparison of the tracking effects of the method of the present invention, where PTHTNet represents the method of the present invention;

[0308] The method of this invention was tested against existing target tracking algorithms on a test dataset (preferably the KITTI or unSenses dataset), and the resulting tracking performance comparison diagram is shown below. Figure 6 As shown;

[0309] pass Figure 6 As you can see, we randomly selected sequence data from the KITTI dataset to better demonstrate the effectiveness and practicality of our algorithm through visualization. From Figure 6 The following conclusions can be drawn: 1) In terms of bounding box information, the method of this invention provides the most accurate information. 2) The method of this invention exhibits good stability under different last-shift conditions. 3) The method of this invention achieves the best tracking performance among all compared algorithms. This is because our algorithm uses tensor decomposition and tensor completion to extract latent information from the point cloud and mines the correlations between point clouds through global and local transformer architectures to obtain better performance. This method achieves accurate localization performance in long-term motion and appearance matching.

[0310] like Figure 7The diagram shows the functional modules of the system of the present invention: The system for implementing the single-target real-time tracking method for power systems disclosed in this invention includes a data acquisition module, a model building module, a model training module, and a real-time tracking module; the data acquisition module, model building module, model training module, and real-time tracking module are connected in series; the data acquisition module is used to acquire a single-target real-time tracking dataset for power systems and upload the data information to the model building module; the model building module is used to construct a primary single-target real-time tracking model based on a Siamese network and Transformer architecture according to the received data information and upload the data information to the model training module; the model training module is used to train the primary single-target real-time tracking model constructed in step S2 using the acquired dataset according to the received data information to obtain a single-target real-time tracking model and upload the data information to the real-time tracking module; the real-time tracking module is used to apply the obtained single-target real-time tracking model to the power system according to the received data information to complete the single-target real-time tracking of the power system.

Claims

1. A method for real-time tracking of a single target in a power system, comprising the following steps: S1. Obtain a single-target real-time tracking dataset for power systems; S2. Based on Siamese networks and the Transformer architecture, construct a basic model for real-time single-target tracking; including the following steps: A basic model for real-time single-target tracking is constructed based on Siamese networks and the Transformer architecture; the model's processing procedure is as follows: A. Construct a point cloud feature extraction network based on a Siamese network, and use the constructed point cloud feature extraction network to extract the point cloud feature tensor of the search region and the point cloud feature tensor of the template region in the dataset obtained in step S1, and fuse them to obtain the final point cloud feature tensor. B. Perform position encoding, tensor decomposition, and tensor completion operations on the final feature tensor of the point cloud obtained in step A; C. Based on the Transformer architecture, obtain the local augmented feature tensor and the global augmented feature tensor corresponding to the final feature tensor of the point cloud; D. The obtained local enhancement feature tensor and global enhancement feature tensor are weighted and fused to obtain the information feature tensor; E. Perform inverse tensor decomposition and Hough transform on the obtained information feature tensor to obtain the target candidate bounding box; F. The target candidate bounding boxes obtained in step E are processed using 3D center head, heat map head, score head and direction head to obtain the final target 3D bounding box; S3. Using the dataset obtained in step S1, train the primary single-target real-time tracking model constructed in step S2 to obtain the single-target real-time tracking model. S4. Apply the single-target real-time tracking model obtained in step S3 to the power system to complete the single-target real-time tracking of the power system.

2. The single-target real-time tracking method for power systems according to claim 1, characterized in that... Step A specifically includes the following steps: A1. For the original point cloud of the search region and the original point cloud of the template region in the dataset obtained in step S1, the farthest point is sampled through set abstraction operation to obtain the feature tensor of the search point cloud and the feature tensor of the template point cloud. A2. Perform a segmentation operation on the feature tensor obtained in step A1 to obtain the coordinates of each point in the search region and the template region, as well as the attribute of the point; the attribute is either a foreground point or a background point. A3. Perform a classification operation on the feature tensors obtained in step A2 to obtain the feature tensors of each foreground point category in the search region and the feature tensors of each foreground point category in the template region; A4. Using a Siamese network, the feature tensors obtained in step A3 are fused to obtain the final feature tensor of the point cloud.

3. The single-target real-time tracking method for power systems according to claim 2, characterized in that... Step A1 specifically includes the following steps: A1-1. Obtain the original point cloud data of the search area and template area. The size of the original point cloud data is 8×16384×3, where the batch size is 8, the number of points is 16384, and the dimension is 3. A1-2. Perform feature permutation on the original point cloud data, and the output size is 8×3×16384; A1-3. Perform farthest point sampling on the obtained point cloud data, and the output size is 8×3×4096; A1-4. The results obtained from sampling the farthest point are grouped using sphere query to generate 4096 local regions; for each local region, two types of sphere radii, 16 and 32, are generated, resulting in output sizes of 8×3×4096×16 and 8×3×4096×32. A1-5. Perform point network operations on the point cloud data contained in the region with a radius of 16 and the point cloud data contained in the region with a radius of 32 respectively; A1-6. Perform connection, feature permutation and farthest point sampling operations on the results of the point network operation obtained in step A1-5 in sequence; A1-7. Perform the sphere query grouping operation again on the results obtained in step A1-6 to generate 4096 local regions; for each local region, generate two types of sphere radii: 16 and 32. A1-8. Perform point network operations on the point cloud data contained in the region with a radius of 16 and the point cloud data contained in the region with a radius of 32 respectively; A1-9. Connect the results of the point network operation obtained in step A1-8 to obtain the final search point cloud feature tensor and template point cloud feature tensor.

4. The single-target real-time tracking method for power systems according to claim 3, characterized in that... Steps A1-5 specifically include the following steps: For point cloud data encompassing a region with a radius of 16, the following steps are used to perform point network operations: A rigid body transformation is performed on the point cloud data contained in a region with a radius of 16, and the size of the transformation matrix is ​​3×3; The obtained output is subjected to a fully connected operation based on a multilayer perceptron network with a kernel size of 3×1, a channel number of 64, and a step size of 1. The obtained output is subjected to rigid body transformation again, and the size of the transformation matrix is ​​64×64; The obtained output is subjected to a fully connected operation based on a multilayer perceptron network with a kernel size of 3×1, a channel number of 128, and a step size of 1. The obtained output is then subjected to a fully connected operation based on a multilayer perceptron network, with a kernel size of 3×1, a channel number of 1024, and a step size of 1. The obtained output is subjected to max pooling, with a kernel size of 3×3 and 32 output channels; The obtained output is then subjected to max pooling again, with a kernel size of 3×3 and an output channel of 1. For point cloud data encompassing a region with a radius of 32, the following steps are used to perform point network operations: A rigid body transformation is performed on the point cloud data contained in a region with a radius of 32, and the size of the transformation matrix is ​​3×3; The obtained output is subjected to a fully connected operation based on a multilayer perceptron network with a kernel size of 3×1, a channel number of 64, and a step size of 1. The obtained output is subjected to rigid body transformation again, and the size of the transformation matrix is ​​64×64; The obtained output is subjected to a fully connected operation based on a multilayer perceptron network with a kernel size of 3×1, a channel number of 128, and a step size of 1. The obtained output is then subjected to a fully connected operation based on a multilayer perceptron network, with a kernel size of 3×1, a channel number of 1024, and a step size of 1. The obtained output is subjected to max pooling, with a kernel size of 3×3 and 64 output channels; The obtained output is then subjected to max pooling again, with a kernel size of 3×3 and an output channel of 1. Steps A1-7 specifically include the following steps: For point cloud data encompassing a region with a radius of 16, the following steps are used to perform point network operations: A rigid body transformation is performed on the point cloud data contained in a region with a radius of 16, and the size of the transformation matrix is ​​3×3; The obtained output is subjected to a fully connected operation based on a multilayer perceptron network with a kernel size of 3×1, a channel number of 256, and a step size of 1. The obtained output is subjected to rigid body transformation again, and the size of the transformation matrix is ​​256×256; The obtained output is subjected to a fully connected operation based on a multilayer perceptron network with a kernel size of 3×1, a channel number of 512, and a step size of 1. The obtained output is then subjected to a fully connected operation based on a multilayer perceptron network, with a kernel size of 3×1, a channel number of 1024, and a step size of 1. The obtained output is subjected to max pooling, with a kernel size of 3×3 and 128 output channels; The obtained output is then subjected to max pooling again, with a kernel size of 3×3 and an output channel of 1. For point cloud data encompassing a region with a radius of 32, the following steps are used to perform point network operations: A rigid body transformation is performed on the point cloud data contained in a region with a radius of 32, and the size of the transformation matrix is ​​3×3; The obtained output is subjected to a fully connected operation based on a multilayer perceptron network with a kernel size of 3×1, a channel number of 64, and a step size of 1. The obtained output is subjected to rigid body transformation again, and the size of the transformation matrix is ​​256×256; The obtained output is subjected to a fully connected operation based on a multilayer perceptron network with a kernel size of 3×1, a channel number of 512, and a step size of 1. The obtained output is then subjected to a fully connected operation based on a multilayer perceptron network, with a kernel size of 3×1, a channel number of 1024, and a step size of 1. The obtained output is subjected to max pooling, with a kernel size of 3×3 and 128 output channels; The obtained output is then subjected to max pooling again, with a kernel size of 3×3 and an output channel of 1.

5. The single-target real-time tracking method for power systems according to claim 4, characterized in that... Step A2 specifically includes the following steps: The feature tensor obtained in step A1 is then subjected to upsampling and feature concatenation operations in sequence. The resulting output of 8×128×1280 is then subjected to a point network operation; the point network operation specifically includes the following steps: A rigid body transformation is performed on the point cloud data contained in a region with a radius of 16, and the size of the transformation matrix is ​​3×3; The obtained output is subjected to a fully connected operation based on a multilayer perceptron network with a kernel size of 3×1, a channel number of 256, and a step size of 1. The obtained output is subjected to rigid body transformation again, and the size of the transformation matrix is ​​256×256; The obtained output is subjected to a fully connected operation based on a multilayer perceptron network with a kernel size of 3×1, a channel number of 512, and a step size of 1. The obtained output is then subjected to a fully connected operation based on a multilayer perceptron network, with a kernel size of 3×1, a channel number of 256, and a step size of 1. The obtained output is subjected to max pooling, with a kernel size of 3×3 and 128 output channels; The obtained output is then subjected to max pooling again, with a kernel size of 3×3 and an output channel of 1. For the output of the point network, for points ignored during sampling, find k known points in the neighborhood and use a computational formula. Perform upsampling; where f j (x) is a j-dimensional feature. d represents distance, f i (j) Let j be the j-dimensional feature of the i-th point; Perform feature concatenation on the upsampled output; The resulting output 8×16384×128+6 is then subjected to a point network operation; the point network operation specifically includes the following steps: The obtained output is processed by a fully connected multilayer perceptron network with a kernel size of 3×1, 64 channels, and a stride of 1. A rigid body transformation is performed on the point cloud data contained in a region with a radius of 16, and the size of the transformation matrix is ​​3×3; The obtained output is subjected to a fully connected operation based on a multilayer perceptron network with a kernel size of 3×1, a channel number of 256, and a step size of 1. The obtained output is subjected to rigid body transformation again, and the size of the transformation matrix is ​​256×256; The obtained output is subjected to a fully connected operation based on a multilayer perceptron network with a kernel size of 3×1, a channel number of 512, and a step size of 1. The obtained output is then subjected to a fully connected operation based on a multilayer perceptron network, with a kernel size of 3×1, a channel number of 1024, and a step size of 1. The obtained output is subjected to 4 max pooling operations with a kernel size of 3×3 and 16384 output channels. The obtained output is then subjected to max pooling again, with a kernel size of 3×1 and an output channel of 1. The output is then subjected to two convolution operations in sequence; the kernel size is 3×3 and the stride is 8. Step A3 specifically includes the following steps: Perform a rigid body transformation operation on the feature tensor obtained in step A2, with a transformation matrix size of 256×256; The obtained output is subjected to a fully connected operation based on a multilayer perceptron, with a kernel size of 3×1, a channel number of 512, and a step size of 1. The obtained output is subjected to rigid body transformation again, with a transformation matrix size of 512×512; The obtained output is then subjected to a fully connected operation based on a multilayer perceptron, with a kernel size of 3×1, a channel count of 1024, and a step size of 1. The output is subjected to a 4-layer fully connected operation to obtain the final classification result; the kernel size of the fully connected operation is 3×1, the number of channels is 1, and the stride is 1. Step A4 specifically includes the following steps: The following formula is used to continue the fusion of feature tensors: In the formula d + f(X) is the norm distance between two positive samples. + ) represents the features of positive samples; X + For positive samples; f(X) a ) represents the characteristics of the anchor point; X a For anchor point; d - The norm distance between two negative samples; f(X) - ) represents the features of negative samples; X - For negative samples; || ||2 is the 2-norm.

6. The single-target real-time tracking method for power systems according to claim 5, characterized in that... Step B specifically includes the following steps: Perform a hash position encoding operation on the final feature tensor of the point cloud obtained in step A: A hash position encoding is constructed by creating an optimization problem in a high-dimensional space. Based on the nearest neighbor Lagrange optimization problem, the center point is determined, and neighboring points are found for the center point to form a local point cluster structure in which the center point is represented by the neighboring points. The optimal weight coefficient in the nearest neighbor is obtained. By introducing the optimal weight coefficients into the low-dimensional space, we can find the low-dimensional point cluster structure that preserves the relative positional relationship between high-dimensional point clusters. We can then uniquely project the high-dimensional features into a hash table in the low-dimensional space to obtain the positional encoding of the high-dimensional data in the low-dimensional space. Perform tensor decomposition: Tensor decomposition is used to reduce the dimensionality of feature tensors, extract the core sub-tensors of feature tensors, and thus speed up the computation process; Establish a combination of display patterns and latent factors; given any tensor X, the corresponding augmented tensor decomposition is represented as... In the formula x ijt μ is the element of tensor X at position (i,j,t); i φ represents the weighting coefficients of the elements in the core subtensor. i Let θ be the molecular matrix weight vector along the U direction. i Let η be the molecular matrix weight vector in the V direction. i g is the molecular matrix weight vector in the W direction; ik For core sub-tensor elements; u ik Let v be the molecular matrix vector along the U direction. ik Let w be the molecular matrix vector along the V direction. ik is the molecular matrix vector in the W direction; r is the rank of the tensor; Perform tensor completion: Given a data matrix X, find three tensors y, z, and τ to approximate X, obtaining an approximate value X. * ; The spectrum soft thresholding of the data matrix X is used as a shrinkage process for singular values, defined as follows: In the formula Let U be the projection operator on the radius λ spectral norm sphere; U and V are both orthogonal matrices; X = US'V T It is a matrix-based SVD of the input tensor X.

7. The single-target real-time tracking method for power systems according to claim 6, characterized in that... Step C specifically includes the following steps: The feature tensor and positional encoding obtained in step B are used as input to the Transformer architecture to mine the correlation between different targets in the same frame, resulting in a locally enhanced feature tensor. The processing includes the following steps: The template point cloud feature tensor is input to the embedding layer, where the transformation matrix has a size of 8×256×256. The obtained output is subjected to encoder stacking operation, where the kernel size is 3×1, the stacking layer is 6 layers, the number of channels is 512, and the step size is 1. The output is subjected to a multi-head attention mechanism, which is split into 8 heads, each with a dimension of 64. The attention weights are calculated independently for each of the eight heads, and the outputs of the eight heads are then concatenated. The obtained output is then subjected to residual concatenation and normalization operations. The obtained output is then processed by a feedforward neural network; the hidden layer size is 8×512×512. The obtained output is then subjected to residual connection and normalization operations again; Finally, the local enhancement feature tensor is obtained; The obtained local augmented feature tensor and positional encoding are used as input to the Transformer architecture to mine the correlation between the same target in different frames, resulting in a global augmented feature tensor. The processing includes the following steps: The local enhancement feature tensor is input to the embedding layer, where the transformation matrix has a size of 8×256×256. The obtained output is subjected to encoder stacking operation, where the kernel size is 3×1, the stacking layer is 6 layers, the number of channels is 512, and the step size is 1. The output is subjected to a multi-head attention mechanism, which is split into 8 heads, each with a dimension of 64. The attention weights are calculated independently for each of the eight heads, and the outputs of the eight heads are then concatenated. The obtained output is then subjected to residual concatenation and normalization operations. The obtained output is then processed by a feedforward neural network; the hidden layer size is 8×512×2048. The obtained output is then subjected to residual connection and normalization operations again; Finally, the global enhanced feature tensor is obtained.

8. The single-target real-time tracking method for power systems according to claim 7, characterized in that... The training described in step S3 specifically includes the following steps: The following formula is used as the weighted loss function L. w : In the formula M s a is the number of point clouds; i This represents the label of the i-th seed point; Use the smooth activation function; S represents the spatial coordinates of the point cloud; l The coordinates of the midpoint of the ground truth; δ i Let δ be the weight vector. i =MLP(f i g ), f i g For the i-th ground real feature, MLP() is the processing function of the multilayer perceptron; The following formula is used as the training loss function L. cls : In the formula, ρ() is the predictor classifier; Positive samples; For negative samples; k + k is the number of positive samples. - The number of negative samples; The following formula is used as the value difference loss function L. t : In the formula, softmax() is the softmax function; These are hyperparameters used to balance the loss; To represent a positive sample linear layer with a sigmoid activation function; To represent a linear layer of negative samples with a sigmoid activation function; The following formula is used as the location-aware ranking loss function L. l : In the formula For paired samples; The crossover ratio (CUP) is a metric used to calculate the crossover ratio for positive samples. The crossover ratio (CRO) is used to calculate the index for negative samples. Finally, the total training function L is obtained as follows: L=λ1L w +λ2L cls +λ3L t +λ4L l In the formula, λ1 is the first weight; λ2 is the second weight; λ3 is the third weight; and λ4 is the fourth weight.

9. A system for implementing the single-target real-time tracking method for power systems according to any one of claims 1 to 8, characterized in that... It includes a data acquisition module, a model building module, a model training module, and a real-time tracking module; the data acquisition module, model building module, model training module, and real-time tracking module are connected in series; the data acquisition module is used to acquire a single-target real-time tracking dataset for the power system and upload the data information to the model building module; The model building module is used to build a primary model for real-time tracking of a single target based on the received data information, using the Siamese network and Transformer architecture, and upload the data information to the model training module. The model training module is used to train the primary single-target real-time tracking model constructed in step S2 based on the received data information and the acquired dataset, so as to obtain the single-target real-time tracking model and upload the data information to the real-time tracking module. The real-time tracking module is used to apply the obtained single-target real-time tracking model to the power system based on the received data information, thereby completing the single-target real-time tracking of the power system.

Images