Single-target real-time tracking method and system for power system

Abstract

The invention discloses a single-target real-time tracking method for an electric power system. The method comprises the following steps: acquiring a single-target real-time tracking data set for the electric power system; the method comprises the following steps: constructing a single-target real-time tracking primary model based on a twin network and a Transform architecture, and training to obtain a single-target real-time tracking model; and applying the single-target real-time tracking model to the power system to complete single-target real-time tracking of the power system. The invention also discloses a system for realizing the single-target real-time tracking method for the power system. According to the invention, through the twin network and the Transform architecture, real-time tracking of a single target in the power system is realized, the reliability is higher, and the accuracy is better.

Description

Technical Field

[0001] The invention belongs to the field of electrical automation, and in particular relates to a single target real-time tracking method and system for a power system. Background Art

[0002] With the development of economy and technology and the improvement of people's living standards, electricity has become an indispensable secondary energy source in people's production and life, bringing endless convenience to people's production and life. Therefore, ensuring the stable and reliable supply of electricity 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 scenarios such as drone inspection and power grid equipment inspection and maintenance. The SOT algorithm can continuously monitor the movement or position change of objects over time. For example, when monitoring intruders moving along the perimeter or boundary of power grid facilities, the SOT algorithm can continuously monitor their position and predict their next move. When using drones or ground robots for automatic inspections, the SOT algorithm can monitor and predict the path of the inspection equipment to ensure that it operates within the predetermined route and adjust the trajectory in real time as needed, so as to cover the target area more effectively. In addition, during infrastructure construction, maintenance or renovation, the SOT algorithm can be used to monitor the movement and placement of materials, components or machinery, ensure that their coordinates are aligned, and predict possible deviations in advance. Finally, the SOT algorithm can also be used to track the growth of vegetation near power lines so that timely maintenance can be arranged 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 studies focus on the feature analysis of point cloud data, that is, improving the tracking accuracy by adding different modules. However, information loss may occur during the feature extraction process, especially the conversion process 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 aggravates the difficulty of target detection and tracking. At the same time, when the traditional CNN algorithm is used to extract three-dimensional point cloud features, it is easy to lose the fine-grained geometric details of power equipment, which in turn affects the detection accuracy. When dealing with multi-target tracking, the computational complexity of statistical association algorithms, such as the Bayesian method, increases significantly with the increase in the number of targets, resulting in the dimensionality disaster of data association. Summary of the invention

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

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

[0007] The single target real-time tracking method for a power system provided by the present invention comprises the following steps:

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

[0009] S2. Build a primary model for single target real-time tracking based on the Siamese network and Transformer architecture;

[0010] S3. Using the data set obtained in step S1, the primary single target real-time tracking model constructed in step S2 is trained to obtain a 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] The primary model for real-time tracking of a single target is constructed based on the twin network and Transformer architecture described in step S2, including the following steps:

[0013] Based on the twin network and Transformer architecture, a primary model for single target real-time tracking is constructed; the processing process of the model is as follows:

[0014] A. constructing a point cloud feature extraction network based on the twin network, and using the constructed point cloud feature extraction network to extract the search area point cloud feature tensor and the template area point cloud feature tensor in the data set obtained in step S1, and fusing 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 enhanced feature tensor and the global enhanced feature tensor corresponding to the final feature tensor of the point cloud;

[0017] D. Perform weighted fusion on the obtained local enhanced feature tensor and the global enhanced feature tensor 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 box obtained in step E is processed by the 3D center head, heat map head, score head and direction head to obtain the final target 3D bounding box.

[0020] The step A specifically comprises the following steps:

[0021] A1. For the original point cloud of the search area and the original point cloud of the template area in the point cloud data set obtained in step S1, the farthest point sampling is performed through the set abstraction operation to obtain the search point cloud feature tensor and the template point cloud feature tensor;

[0022] A2. Segment the feature tensor obtained in step A1 to obtain the coordinates of each point in the search area and the template area, as well as the attributes of the point; the attributes are foreground points or background points;

[0023] A3. Classify the feature tensor obtained in step A2 to obtain the feature tensor of each foreground point category in the search area and the feature tensor of each foreground point category in the template area.

[0024] A4. Use the twin network to fuse the feature tensors obtained in step A3 to obtain the final feature tensor of the point cloud.

[0025] The step A1 specifically includes the following steps:

[0026] A1-1. Obtain the original point cloud data of the search area and the 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 operation on the original point cloud data, and the output size is 8×3×16384;

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

[0029] A1-4. The results obtained by sampling the farthest point are grouped for ball query to generate 4096 local areas. For each local area, two types of ball 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 area with a radius of 16 and the point cloud data contained in the area with a radius of 32 respectively;

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

[0032] A1-7. The result obtained in step A1-6 is grouped again for ball query to generate 4096 local areas; for each local area, two types of ball radii, 16 and 32, are generated;

[0033] A1-8. Perform point network operations on the point cloud data contained in the area with a radius of 16 and the point cloud data contained in the area 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] The steps A1-5 specifically include the following steps:

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

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

[0038] The obtained output is fully connected based on a multi-layer perceptron network with a kernel size of 3×1, 64 channels, 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 fully connected based on a multi-layer 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 once again fully connected based on the multi-layer 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 a maximum pooling operation with a kernel size of 3×3 and an output channel of 32;

[0043] For the obtained output, the maximum pooling operation is performed again, with a kernel size of 3×3 and an output channel of 1;

[0044] For the point cloud data contained in the area with a radius of 32, the point network operation is performed using the following steps:

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

[0046] The obtained output is fully connected based on a multi-layer perceptron network with a kernel size of 3×1, 64 channels, 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 fully connected based on a multi-layer 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 once again fully connected based on the multi-layer 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 a maximum pooling operation with a kernel size of 3×3 and an output channel of 64;

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

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

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

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

[0055] The obtained output is fully connected based on a multi-layer 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 fully connected based on a multi-layer 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 once again fully connected based on the multi-layer 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 a maximum pooling operation with a kernel size of 3×3 and an output channel of 128;

[0060] For the obtained output, the maximum pooling operation is performed again, with a kernel size of 3×3 and an output channel of 1;

[0061] For the point cloud data contained in the area with a radius of 32, the point network operation is performed using the following steps:

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

[0063] The obtained output is fully connected based on a multi-layer perceptron network with a kernel size of 3×1, 64 channels, 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 fully connected based on a multi-layer 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 once again fully connected based on the multi-layer 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 a maximum pooling operation with a kernel size of 3×3 and an output channel of 128;

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

[0069] The step A2 specifically includes the following steps:

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

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

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

[0073] The obtained output is fully connected based on a multi-layer 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 fully connected based on a multi-layer 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 once again fully connected based on the multi-layer 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 a maximum pooling operation with a kernel size of 3×3 and an output channel of 128;

[0078] For the obtained output, the maximum pooling operation is performed again, with a kernel size of 3×3 and an output channel of 1;

[0079] For the output of the point network, for the points ignored during the sampling process, find the k known points in the field and use the calculation formula Upsampling is performed; where f j (x) is the j-dimensional feature, d represents the distance, f i (j) is the j-dimensional feature of the i-th point;

[0080] Perform feature concatenation on the upsampled output;

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

[0082] For the obtained output, a fully connected multilayer perceptron network is performed with a kernel size of 3×1, 64 channels, and a step size of 1.

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

[0084] The obtained output is fully connected based on a multi-layer 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 fully connected based on a multi-layer 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 once again fully connected based on the multi-layer 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 maximum pooling operations with a kernel size of 3×3 and an output channel of 16384;

[0089] For the obtained output, the maximum pooling operation is performed again, with a kernel size of 3×1 and an output channel of 1;

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

[0091] The step A3 specifically comprises the following steps:

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

[0093] The obtained output is fully connected based on a multi-layer 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, and the transformation matrix size is 512×512;

[0095] The obtained output is again fully connected based on the multi-layer perceptron, with a kernel size of 3×1, a channel number of 1024, and a step size of 1;

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

[0097] The step A4 specifically comprises the following steps:

[0098] Use the following formula to continue the fusion of feature tensors:

[0099]

[0100] Where d + is the norm distance between two positive samples; f(X + ) is the feature of the positive sample; X + is a positive sample; f(X a ) is the feature of the anchor point; X a is the anchor point; d - is the norm distance between two negative samples; f(X - ) is the feature of negative samples; X - is a negative sample; || ||2 is the 2-norm.

[0101] The step B specifically comprises the following steps:

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

[0103] Construct hash position coding, by constructing optimization problems in high-dimensional space, based on the Lagrangian optimization problem of the nearest neighbor, determine the center point, and find neighboring points for the center point, form a local point cluster structure that uses neighboring points to represent the center point, and obtain the optimal weight coefficient in the nearest neighbor;

[0104] The optimal weight coefficient is introduced into the low-dimensional space to find the low-dimensional point cluster structure that maintains the relative position relationship between high-dimensional point clusters. The high-dimensional features are uniquely projected into the hash table of the low-dimensional space to obtain the corresponding position encoding of the high-dimensional data in the low-dimensional space.

[0105] Perform tensor decomposition:

[0106] Tensor decomposition is used to reduce the dimension of the feature tensor and extract the core sub-tensors of the feature tensor, thereby speeding up the calculation process;

[0107] Establish a combination of display modes and potential factors; given any tensor X, the corresponding augmented tensor decomposition is expressed as

[0108]

[0109] Where x ijt is the element of tensor X at position (i, j, t); μ i is the weight coefficient of the elements in the core sub-tensor; φ i is the molecular matrix weight vector in the U direction, θ i is the molecular matrix weight vector in the V direction, η i is the molecular matrix weight vector in the W direction; g ik is the core sub-tensor element; u ik is the molecular matrix vector in the U direction, v ik is the molecular matrix vector in the V direction, w 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, and get the approximate value X * ;

[0112] The spectrum soft thresholding of the data matrix X is treated as shrinking of singular values ​​and is defined as

[0113]

[0114] In the formula is the projection operator on the spectral norm ball of radius λ; U and V are both orthogonal matrices; X=US'V T is the matrixized SVD of the input tensor X.

[0115] The step C specifically comprises the following steps:

[0116] The feature tensor and position encoding obtained in step B are used as the input of the Transformer architecture to mine the association between different objects in the same frame and obtain the local enhanced feature tensor. The processing process includes the following steps:

[0117] Perform input embedding layer operation on the template point cloud feature tensor, where the size of the transformation matrix is ​​8×256×256;

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

[0119] For the obtained output, a multi-head attention mechanism is performed, and the attention mechanism is split into 8 heads, and the dimension of each head is 64;

[0120] The attention weights of the eight heads are calculated independently, and the outputs of the eight heads are concatenated;

[0121] Perform residual connection and normalization operations on the obtained output;

[0122] The obtained output is subjected to a feed-forward neural network operation; wherein the hidden layer size is 8×512×512;

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

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

[0125] The obtained local enhanced feature tensor and position encoding are used as the input of the Transformer architecture to mine the association relationship of the same target in different frames and obtain the global enhanced feature tensor; the processing process includes the following steps:

[0126] The local enhanced feature tensor is input into the embedding layer, where the size of the transformation matrix is ​​8×256×256;

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

[0128] For the obtained output, a multi-head attention mechanism is performed, and the attention mechanism is split into 8 heads, and the dimension of each head is 64;

[0129] The attention weights of the eight heads are calculated independently, and the outputs of the eight heads are concatenated;

[0130] Perform residual connection and normalization operations on the obtained output;

[0131] The obtained output is subjected to feed-forward neural network operation; wherein the hidden layer size is 8×512×2048;

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

[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 weight loss function L w :

[0136]

[0137] Where M s is the number of point clouds; a i Represents the label of the i-th seed point; is the smooth activation function; is the spatial coordinate of the point cloud; S l is the midpoint coordinate of the ground truth; δ i is the weight vector, and δ i =MLP(f i g ), f i g is the i-th ground truth 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] Where ρ() is the prediction classifier; is a positive sample; is a negative sample; k + is the number of positive samples; k - is the number of negative samples;

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

[0142]

[0143] Where softmax() is the softmax function; is a hyperparameter used to balance the loss; represents the positive sample linear layer with sigmoid activation function; represents the negative sample linear layer with 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; It is the intersection-over-union ratio calculation index of positive samples; It is the intersection-over-union ratio calculation index of negative samples;

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

[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] The present invention also provides a system for implementing the single-target real-time tracking method for an electric power system, comprising a data acquisition module, a model construction module, a model training module and a real-time tracking module; the data acquisition module, the model construction module, the model training module and the real-time tracking module are connected in series in sequence; the data acquisition module is used to acquire a single-target real-time tracking data set for an electric power system, and upload the data information to the model construction module; the model construction module is used to construct a single-target real-time tracking primary model based on a twin network and a 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 single-target real-time tracking primary model constructed in step S2 according to the received data information and the acquired data set, 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 electric power system according to the received data information, and complete the single-target real-time tracking of the electric power system.

[0151] The single target real-time tracking method and system for the power system provided by the present invention, through the twin network and Transformer architecture, not only realizes the real-time tracking of the single target in the power system, but also has higher reliability and better accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

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

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

[0154] Figure 3 Schematic diagram of tensor decomposition in the method of the present invention.

[0155] Figure 4 Schematic diagram of tensor completion in the method of the present invention.

[0156] Figure 5 Schematic diagram of tensor reduction in the method of the present invention.

[0157] Figure 6Schematic diagram for comparing tracking effects of an embodiment of the method of the present invention.

[0158] Figure 7 Schematic diagram of the functional modules of the system of the present invention. DETAILED DESCRIPTION

[0159] like Figure 1 The method flow chart of the method of the present invention is shown as follows: the single target real-time tracking method for power system disclosed in the present invention comprises the following steps:

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

[0161] S2. Based on the Siamese network and Transformer architecture, a primary model for real-time tracking of a single target is constructed. The steps include:

[0162] Based on the twin network and Transformer architecture, a primary model for single target real-time tracking is constructed; the processing process of the model is as follows:

[0163] A. Constructing a point cloud feature extraction network based on the twin network, and using the constructed point cloud feature extraction network to extract the search area point cloud feature tensor and the template area point cloud feature tensor in the data set obtained in step S1, and fusing them to obtain the final point cloud feature tensor; specifically comprising the following steps:

[0164] A1. For the original point cloud of the search area and the original point cloud of the template area in the data set obtained in step S1, the farthest point sampling is performed through the set abstraction operation to obtain the search point cloud feature tensor and the template point cloud feature tensor; specifically, the following steps are included:

[0165] A1-1. Obtain the original point cloud data of the search area and the 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 operation on the original point cloud data, and the output size is 8×3×16384;

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

[0168] A1-4. The results obtained by sampling the farthest point are grouped for ball query to generate 4096 local areas. For each local area, two types of ball 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 the area with a radius of 16 and the point cloud data contained in the area with a radius of 32 respectively; specifically, the following steps are included:

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

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

[0172] The obtained output is fully connected based on a multi-layer perceptron network with a kernel size of 3×1, 64 channels, 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 fully connected based on a multi-layer 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 once again fully connected based on the multi-layer 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 a maximum pooling operation with a kernel size of 3×3 and an output channel of 32;

[0177] For the obtained output, the maximum pooling operation is performed again, with a kernel size of 3×3 and an output channel of 1;

[0178] For the point cloud data contained in the area with a radius of 32, the point network operation is performed using the following steps:

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

[0180] The obtained output is fully connected based on a multi-layer perceptron network with a kernel size of 3×1, 64 channels, 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 fully connected based on a multi-layer 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 once again fully connected based on the multi-layer 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 a maximum pooling operation with a kernel size of 3×3 and an output channel of 64;

[0185] For the obtained output, the maximum pooling operation is performed again, with a kernel size of 3×3 and an output channel of 1;

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

[0187] A1-7. The result obtained in step A1-6 is grouped again for ball query to generate 4096 local areas; for each local area, two types of ball radii, 16 and 32, are generated; specifically, the steps include:

[0188] For the point cloud data contained in the area with a radius of 16, the point network operation is performed using the following steps:

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

[0190] The obtained output is fully connected based on a multi-layer 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 fully connected based on a multi-layer 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 once again fully connected based on the multi-layer 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 a maximum pooling operation with a kernel size of 3×3 and an output channel of 128;

[0195] For the obtained output, the maximum pooling operation is performed again, with a kernel size of 3×3 and an output channel of 1;

[0196] For the point cloud data contained in the area with a radius of 32, the point network operation is performed using the following steps:

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

[0198] The obtained output is fully connected based on a multi-layer perceptron network with a kernel size of 3×1, 64 channels, 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 fully connected based on a multi-layer 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 once again fully connected based on the multi-layer 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 a maximum pooling operation with a kernel size of 3×3 and an output channel of 128;

[0203] For the obtained output, the maximum pooling operation is performed 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 area with a radius of 16 and the point cloud data contained in the area 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. Segment the feature tensor obtained in step A1 to obtain the coordinates of each point in the search area and the template area, as well as the attributes of the point; the attributes are foreground points or background points; specifically, the following steps are included:

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

[0208] The obtained output 8×128×1280 is subjected to point network operation; the point network operation is specifically as follows:

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

[0210] The obtained output is fully connected based on a multi-layer 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 fully connected based on a multi-layer 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 once again fully connected based on the multi-layer 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 a maximum pooling operation with a kernel size of 3×3 and an output channel of 128;

[0215] For the obtained output, the maximum pooling operation is performed again, with a kernel size of 3×3 and an output channel of 1;

[0216] For the output of the point network, for the points ignored during the sampling process, find the k known points in the field and use the calculation formula Upsampling is performed; where f j (x) is the j-dimensional feature, d represents the distance. The smaller the distance, the greater the weight. i (j) is the j-dimensional feature of the i-th point;

[0217] Perform feature concatenation on the upsampled output;

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

[0219] For the obtained output, a fully connected multilayer perceptron network is performed with a kernel size of 3×1, 64 channels, and a step size of 1.

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

[0221] The obtained output is fully connected based on a multi-layer 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 fully connected based on a multi-layer 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 once again fully connected based on the multi-layer 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 maximum pooling operations with a kernel size of 3×3 and an output channel of 16384;

[0226] For the obtained output, the maximum pooling operation is performed again, with a kernel size of 3×1 and an output channel of 1;

[0227] The obtained output is convolved twice in sequence; the size of the convolution kernel is 3×3 and the step length is 8;

[0228] A3. Classify the feature tensor obtained in step A2 to obtain a feature tensor of each foreground point category in the search area and a feature tensor of each foreground point category in the template area; specifically comprising the following steps:

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

[0230] The obtained output is fully connected based on a multi-layer 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, and the transformation matrix size is 512×512;

[0232] The obtained output is again fully connected based on the multi-layer perceptron, with a kernel size of 3×1, a channel number of 1024, and a step size of 1;

[0233] The obtained output is subjected to a 4-layer full connection operation to obtain the final classification result; the kernel size of the full connection operation is 3×1, the number of channels is 1, and the step size is 1;

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

[0235] Use the following formula to continue the fusion of feature tensors:

[0236]

[0237] Where d + is the norm distance between two positive samples; f(X + ) is the feature of the positive sample; X + is a positive sample; f(X a ) is the feature of the anchor point; X a is the anchor point; d - is the norm distance between two negative samples; f(X - ) is the feature of negative samples; X - is a negative sample; || ||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, the steps include:

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

[0240] Construct hash position coding, by constructing optimization problems in high-dimensional space, based on the nearest neighbor Lagrangian optimization problem, determine the center point, and find nearby points for the center point, forming a local point cluster structure (such as Figure 2 (b)), and obtain the optimal weight coefficient among the nearest neighbors;

[0241] The optimal weight coefficient is introduced into the low-dimensional space to find the low-dimensional point cluster structure that maintains the relative position relationship between high-dimensional point clusters, and the high-dimensional features are uniquely projected into the hash table of the low-dimensional space (such as Figure 2 (a) shows), and obtains the corresponding position encoding of high-dimensional data in low-dimensional space;

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

[0243]

[0244]

[0245] The weight coefficient w i for Where 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 weight coefficient w, let Y i (k) ∈R k×d is a matrix composed of its k neighboring points, then the optimization problem of uniquely projecting n-dimensional samples to d-dimensional is expressed as

[0247]

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

[0249] Perform tensor decomposition:

[0250] Tensor decomposition is used to reduce the dimension of the feature tensor and extract the core sub-tensors of the feature tensor, thereby speeding up the calculation process;

[0251] Build a combination of display patterns and potential factors, such as Figure 3 As shown in the figure, by augmenting the tensor decomposition, the decomposition of the high-order tensor eigenvalues ​​of the point cloud is completed; given any tensor X, the corresponding augmented tensor decomposition is expressed as

[0252]

[0253] Where x ijt is the element of tensor X at position (i, j, t); μ i is the weight coefficient of the elements in the core sub-tensor; φ i is the molecular matrix weight vector in the U direction, θ i is the molecular matrix weight vector in the V direction, η i is the molecular matrix weight vector in the W direction; g ik is the core sub-tensor element; u ik is the molecular matrix vector in the U direction, v ik is the molecular matrix vector in the V direction, w 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 tensor decomposition and retention of core sub-tensors based on high-dimensional features of point clouds has data loss. By constructing a tensor completion optimization problem based on the alternating direction multiplier algorithm to supplement and correct it, more point cloud features and related relationship information can be mined, such as Figure 4 As shown;

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

[0257] The spectrum soft thresholding of the data matrix X is treated as shrinking of singular values ​​and is defined as

[0258]

[0259] In the formula is the projection operator on the spectral norm ball of radius λ; U, V are orthogonal matrices respectively; X=US'V T is the matrixed SVD of the input tensor X;

[0260] In specific implementation, given any tensor X and observation vector y, the k-type expansion matrix of X is X k , vectorized to x; take the auxiliary matrix Z of the same size as the k-type expansion of tensor Xk , vectorized to z k , then there exists an observation operator Ω: and regularization constant λ, so that the trace norm optimization problem of the spectral soft threshold operator

[0261]

[0262] xT k x=z k

[0263] The optimality condition of

[0264]

[0265] in is the regularization term; || || F is the spectral norm; || || * is the linear sum operation of the matrix singular values;

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

[0267] The feature tensor and position encoding obtained in step B are used as the input of the Transformer architecture to mine the association between different objects in the same frame and obtain the local enhanced feature tensor. The processing process includes the following steps:

[0268] Perform input embedding layer operation on the template point cloud feature tensor, where the size of the transformation matrix is ​​8×256×256;

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

[0270] For the obtained output, a multi-head attention mechanism is performed, and the attention mechanism is split into 8 heads, and the dimension of each head is 64;

[0271] The attention weights of the eight heads are calculated independently, and the outputs of the eight heads are concatenated;

[0272] Perform residual connection and normalization operations on the obtained output;

[0273] The obtained output is subjected to feed-forward neural network operation; wherein the hidden layer size is 8×512×512;

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

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

[0276] The obtained local enhanced feature tensor and position encoding are used as the input of the Transformer architecture to mine the association relationship of the same target in different frames and obtain the global enhanced feature tensor; the processing process includes the following steps:

[0277] The local enhanced feature tensor is input into the embedding layer, where the size of the transformation matrix is ​​8×256×256;

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

[0279] For the obtained output, a multi-head attention mechanism is performed, and the attention mechanism is split into 8 heads, and the dimension of each head is 64;

[0280] The attention weights of the eight heads are calculated independently, and the outputs of the eight heads are concatenated;

[0281] Perform residual connection and normalization operations on the obtained output;

[0282] The obtained output is subjected to feedforward neural network operation; wherein the hidden layer size is 8×512×2048;

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

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

[0285] D. Perform weighted fusion on the obtained local enhanced feature tensor and the global enhanced feature tensor to obtain the information feature tensor;

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

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

[0288] S3. Using the data set obtained in step S1, the primary single target real-time tracking model constructed in step S2 is trained to obtain a single target real-time tracking model;

[0289] In specific implementation, the training comprises the following steps:

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

[0291]

[0292] Where M s is the number of point clouds; a i Represents the label of the i-th seed point; is the smooth activation function; is the spatial coordinate of the point cloud; S l is the midpoint coordinate of the ground truth; δ i is the weight vector, and δ i =MLP(f i g ), f i g is the i-th ground truth 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] Where ρ() is the prediction classifier; is a positive sample; is a negative sample; k + is the number of positive samples; k - is the number of negative samples;

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

[0297]

[0298] Where softmax() is the softmax function; is a hyperparameter used to balance the loss; represents the positive sample linear layer with sigmoid activation function; represents the negative sample linear layer with 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; It is the intersection-over-union ratio calculation index of positive samples; It is the intersection-over-union ratio calculation index of negative samples;

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

[0303] L=λ1L w+λ2L cls +λ3L t +λ4L l

[0304] Where λ1 is the first weight; λ2 is the second weight; λ3 is the third weight; 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 is an example to illustrate the effect of the method of the present invention:

[0307] like Figure 6 FIG. 1 is a schematic diagram showing a comparison of tracking effects of the method of the present invention, wherein PTHTNet represents the method of the present invention;

[0308] The method of the present invention and the existing target tracking algorithm are tested in a test data set (preferably KITTI or unSences data set), and the obtained tracking effect comparison diagram is shown in FIG. Figure 6 As shown;

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

[0310] like Figure 7As shown is a schematic diagram of the functional modules of the system of the present invention: the system disclosed in the present invention for implementing the single-target real-time tracking method for the power system includes a data acquisition module, a model construction module, a model training module and a real-time tracking module; the data acquisition module, the model construction module, the model training module and the real-time tracking module are connected in series in sequence; the data acquisition module is used to acquire a single-target real-time tracking data set for the power system, and upload the data information to the model construction module; the model construction module is used to construct a single-target real-time tracking primary model based on the twin network and the 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 single-target real-time tracking primary model constructed in step S2 according to the received data information using the acquired data set, 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, and complete the single-target real-time tracking of the power system.

Claims

1. A single target real-time tracking method for a power system, comprising the following steps: S1. Acquire a single target real-time tracking dataset for power system; S2. Build a primary model for single target real-time tracking based on the Siamese network and Transformer architecture; S3. Using the data set obtained in step S1, the primary single target real-time tracking model constructed in step S2 is trained to obtain a 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 system according to claim 1 is characterized in that The primary model for real-time tracking of a single target is constructed based on the twin network and Transformer architecture described in step S2, including the following steps: Based on the twin network and Transformer architecture, a primary model for single target real-time tracking is constructed; the processing process of the model is as follows: A. constructing a point cloud feature extraction network based on the twin network, and using the constructed point cloud feature extraction network to extract the search area point cloud feature tensor and the template area point cloud feature tensor in the data set obtained in step S1, and fusing 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 enhanced feature tensor and the global enhanced feature tensor corresponding to the final feature tensor of the point cloud; D. Perform weighted fusion on the obtained local enhanced feature tensor and the global enhanced feature tensor 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 box obtained in step E is processed by the 3D center head, heat map head, score head and direction head to obtain the final target 3D bounding box.

3. The single target real-time tracking method for power system according to claim 2 is characterized in that The step A specifically comprises the following steps: A1. For the original point cloud of the search area and the original point cloud of the template area in the data set obtained in step S1, the farthest point sampling is performed through the set abstraction operation to obtain the search point cloud feature tensor and the template point cloud feature tensor; A2. Segment the feature tensor obtained in step A1 to obtain the coordinates of each point in the search area and the template area, as well as the attributes of the point; the attributes are foreground points or background points; A3. Classify the feature tensor obtained in step A2 to obtain the feature tensor of each foreground point category in the search area and the feature tensor of each foreground point category in the template area. A4. Use the twin network to fuse the feature tensors obtained in step A3 to obtain the final feature tensor of the point cloud.

4. The single target real-time tracking method for power system according to claim 3 is characterized in that The step A1 specifically includes the following steps: A1-1. Obtain the original point cloud data of the search area and the 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 operation on the original point cloud data, and the output size is 8×3×16384; A1-3. Perform the farthest point sampling on the obtained point cloud data, and the output size is 8×3×4096; A1-4. The results obtained by sampling the farthest point are grouped for ball query to generate 4096 local areas. For each local area, two types of ball radii, 16 and 32, are generated, and the output sizes are 8×3×4096×16 and 8×3×4096×32. A1-5. Perform point network operations on the point cloud data contained in the area with a radius of 16 and the point cloud data contained in the area with a radius of 32 respectively; A1-6. Perform connection, feature replacement and farthest point sampling operations on the results of the point network operation obtained in step A1-5; A1-7. The result obtained in step A1-6 is grouped again for ball query to generate 4096 local areas; for each local area, two types of ball radii, 16 and 32, are generated; A1-8. Perform point network operations on the point cloud data contained in the area with a radius of 16 and the point cloud data contained in the area 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.

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

6. The single target real-time tracking method for power system according to claim 5 is characterized in that The step A2 specifically includes the following steps: The feature tensor obtained in step A1 is subjected to upsampling and feature concatenation operations in sequence; The obtained output 8×128×1280 is subjected to a point network operation; the point network operation specifically includes the following steps: Perform rigid body transformation on the point cloud data contained in the area with a radius of 16, and the size of the transformation matrix is ​​3×3; The obtained output is fully connected based on a multi-layer 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 fully connected based on a multi-layer perceptron network with a kernel size of 3×1, a channel number of 512, and a step size of 1. The obtained output is once again fully connected based on the multi-layer 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 a maximum pooling operation with a kernel size of 3×3 and an output channel of 128; For the obtained output, the maximum pooling operation is performed again, with a kernel size of 3×3 and an output channel of 1; For the output of the point network, for the points ignored during the sampling process, find the k known points in the field and use the calculation formula Upsampling is performed; where f j (x) is the j-dimensional feature, d represents the distance, is the j-dimensional feature of the i-th point; Perform feature concatenation on the upsampled output; The obtained output 8×16384×128+6 is subjected to a point network operation; the point network operation specifically includes the following steps: For the obtained output, a fully connected multilayer perceptron network is performed with a kernel size of 3×1, 64 channels, and a step size of 1. Perform rigid body transformation on the point cloud data contained in the area with a radius of 16, and the size of the transformation matrix is ​​3×3; The obtained output is fully connected based on a multi-layer 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 fully connected based on a multi-layer perceptron network with a kernel size of 3×1, a channel number of 512, and a step size of 1. The obtained output is once again fully connected based on the multi-layer 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 maximum pooling operations with a kernel size of 3×3 and an output channel of 16384; For the obtained output, the maximum pooling operation is performed again, with a kernel size of 3×1 and an output channel of 1; The obtained output is convolved twice in sequence; the size of the convolution kernel is 3×3 and the step length is 8; The step A3 specifically comprises the following steps: Perform a rigid body transformation operation on the feature tensor obtained in step A2, and the transformation matrix size is 256×256; The obtained output is fully connected based on a multi-layer 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, and the transformation matrix size is 512×512; The obtained output is again fully connected based on the multi-layer perceptron, with a kernel size of 3×1, a channel number of 1024, and a step size of 1; The obtained output is subjected to a 4-layer full connection operation to obtain the final classification result; the kernel size of the full connection operation is 3×1, the number of channels is 1, and the step size is 1; The step A4 specifically comprises the following steps: Use the following formula to continue the fusion of feature tensors: Where d + is the norm distance between two positive samples; f(X + ) is the feature of the positive sample; X + is a positive sample; f(X a ) is the feature of the anchor point; X a is the anchor point; d - is the norm distance between two negative samples; f(X - ) is the feature of negative samples; X - is a negative sample; || ||2 is the 2-norm.

7. The single target real-time tracking method for power system according to claim 6 is characterized in that The step B specifically comprises the following steps: Perform hash position encoding operation on the final feature tensor of the point cloud obtained in step A: Construct hash position coding, by constructing optimization problems in high-dimensional space, based on the Lagrangian optimization problem of the nearest neighbor, determine the center point, and find neighboring points for the center point, form a local point cluster structure that uses neighboring points to represent the center point, and obtain the optimal weight coefficient in the nearest neighbor; The optimal weight coefficient is introduced into the low-dimensional space to find the low-dimensional point cluster structure that maintains the relative position relationship between high-dimensional point clusters. The high-dimensional features are uniquely projected into the hash table of the low-dimensional space to obtain the corresponding position encoding of the high-dimensional data in the low-dimensional space. Perform tensor decomposition: Tensor decomposition is used to reduce the dimension of the feature tensor and extract the core sub-tensors of the feature tensor, thereby speeding up the calculation process; Establish a combination of display modes and potential factors; given any tensor X, the corresponding augmented tensor decomposition is expressed as Where x ijt is the element of tensor X at position (i, j, t); μ i is the weight coefficient of the elements in the core sub-tensor; φ i is the molecular matrix weight vector in the U direction, θ i is the molecular matrix weight vector in the V direction, η i is the molecular matrix weight vector in the W direction; g ik is the core sub-tensor element; u ik is the molecular matrix vector in the U direction, v ik is the molecular matrix vector in the V direction, w 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, and get the approximate value X * ; The spectrum soft thresholding of the data matrix X is treated as shrinking of singular values ​​and is defined as In the formula is the projection operator on the spectral norm ball of radius λ; U and V are both orthogonal matrices; X=US'V T is the matrixized SVD of the input tensor X.

8. The single target real-time tracking method for power system according to claim 7 is characterized in that The step C specifically comprises the following steps: The feature tensor and position encoding obtained in step B are used as the input of the Transformer architecture to mine the association between different objects in the same frame and obtain the local enhanced feature tensor. The processing process includes the following steps: Perform input embedding layer operation on the template point cloud feature tensor, where the size of the transformation matrix is ​​8×256×256; The obtained output is subjected to an encoder stacking operation, where the kernel size is 3×1, the stacking layer is 6 layers, the number of channels is 512, and the stride is 1; For the obtained output, a multi-head attention mechanism is performed, and the attention mechanism is split into 8 heads, and the dimension of each head is 64; The attention weights of the eight heads are calculated independently, and the outputs of the eight heads are concatenated; Perform residual connection and normalization operations on the obtained output; The obtained output is subjected to feed-forward neural network operation; wherein the hidden layer size is 8×512×512; The obtained output is again subjected to residual connection and normalization operations; Finally, the local enhanced feature tensor is obtained; The obtained local enhanced feature tensor and position encoding are used as the input of the Transformer architecture to mine the association relationship of the same target in different frames and obtain the global enhanced feature tensor; the processing process includes the following steps: The local enhanced feature tensor is input into the embedding layer, where the size of the transformation matrix is ​​8×256×256; The obtained output is subjected to an encoder stacking operation, where the kernel size is 3×1, the stacking layer is 6 layers, the number of channels is 512, and the stride is 1; For the obtained output, a multi-head attention mechanism is performed, and the attention mechanism is split into 8 heads, and the dimension of each head is 64; The attention weights of the eight heads are calculated independently, and the outputs of the eight heads are concatenated; Perform residual connection and normalization operations on the obtained output; The obtained output is subjected to feedforward neural network operation; wherein the hidden layer size is 8×512×2048; The obtained output is again subjected to residual connection and normalization operations; Finally, the global enhanced feature tensor is obtained.

9. The single target real-time tracking method for power system according to claim 8, characterized in that The training described in step S3 specifically includes the following steps: The following formula is used as the weight loss function L w : Where M s is the number of point clouds; a i Represents the label of the i-th seed point; is the smooth activation function; is the spatial coordinate of the point cloud; S l is the midpoint coordinate of the ground truth; δ i is the weight vector, and is the i-th ground truth feature, MLP() is the processing function of the multilayer perceptron; The following formula is used as the training loss function L cls : Where ρ() is the prediction classifier; is a positive sample; is a negative sample; k + is the number of positive samples; k - is the number of negative samples; The following formula is used as the value difference loss function L t : Where softmax() is the softmax function; is a hyperparameter used to balance the loss; represents the positive sample linear layer with sigmoid activation function; represents the negative sample linear layer with sigmoid activation function; The following formula is used as the location-aware ranking loss function L l : In the formula For paired samples; It is the intersection-over-union ratio calculation index of positive samples; It is the intersection-over-union ratio calculation index of negative samples; Finally, the total training function L is obtained as 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.

10. A system for implementing the single target real-time tracking method for a power system according to any one of claims 1 to 9, 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, the model building module, the model training module and the real-time tracking module are connected in series in sequence; the data acquisition module is used to acquire a single-target real-time tracking data set for the power system, and upload the data information to the model building module; The model building module is used to build a single-target real-time tracking primary model based on the received data information, the twin network and the Transformer architecture, and upload the data information to the model training module; The model training module is used to train the single target real-time tracking primary model constructed in step S2 according to the received data information and the acquired data set, 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 according to the received data information, so as to complete the single-target real-time tracking of the power system.

Images