Power transmission positive and negative sample fusion foreign matter identification method and system based on time sequence optimization
By constructing a foreign object recognition model based on temporal optimization and using end-to-end training with static and dynamic feature extraction layers, the problems of insufficient ability to distinguish foreign objects and imprecise capture of foreign object movement characteristics in existing technologies are solved. This achieves higher recognition accuracy and robustness, reduces false alarm rate, and improves the safety of transmission lines.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JIANGSU ELECTRIC POWER INFORMATION TECH
- Filing Date
- 2026-01-08
- Publication Date
- 2026-04-17
AI Technical Summary
Existing foreign object identification technologies for power transmission lines have shortcomings in terms of accuracy and robustness, especially in their inability to distinguish negative samples of foreign objects and their insufficient precision in capturing the motion characteristics of foreign objects, resulting in a high false alarm rate and affecting the reliability of the identification results.
A foreign object recognition model based on temporal optimization is constructed. Through end-to-end training of a static feature extraction layer, a perturbation feature extraction layer, and a classification decision layer, dynamic features are extracted using real foreign object positive samples and foreign object-like interference negative samples. Combined with a unique composite loss function design, the features of positive and negative samples are deeply fused.
It significantly improves the ability to distinguish between foreign objects and foreign object-like interference, reduces the false alarm rate, enhances the detection sensitivity of tiny foreign objects, and provides more reliable intelligent protection for the safety of power transmission lines.
Smart Images
Figure CN121884041A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of power transmission line safety assurance, and particularly relates to a method and system for identifying foreign objects in power transmission lines based on time-series optimization through the fusion of positive and negative samples. Background Technology
[0002] With the continuous expansion of the power grid, the safe and stable operation of transmission lines has become increasingly important. Among these challenges, the intrusion of foreign objects (such as plastic film, kite strings, and advertising banners) into transmission corridors is one of the main threats causing power grid short circuits, tripping, and even large-scale power outages. Therefore, automatic foreign object identification technology for transmission lines based on computer vision has become a research hotspot in the field of intelligent operation and maintenance of power systems.
[0003] Currently, deep learning-based image recognition algorithms have been widely applied in this field. The mainstream methods are mainly divided into two categories: one is single-frame target detection based on static images, that is, using trained models (such as YOLO, Faster R-CNN, etc.) to identify foreign objects in real-time acquired power transmission line images; the other is video analysis that introduces temporal information, which judges suspicious targets by comparing the differences between consecutive frames or by using tracking algorithms.
[0004] However, existing technologies still face significant bottlenecks in practical deployments, resulting in recognition accuracy, and especially robustness, failing to meet field requirements. This is mainly reflected in the following two aspects:
[0005] 1. Static image recognition methods are insufficient in distinguishing negative samples that resemble "foreign objects".
[0006] The background of power transmission lines is complex, containing numerous interfering objects that are highly similar to real foreign objects in static texture, color, and shape. Examples include distant cloud shadows, birds, leaves, and the shadows of the towers themselves. These "foreign object-like" negative samples have extremely minor feature differences from real foreign objects (positive samples) in a single image. Existing static models primarily rely on large-scale labeled datasets for training, but these datasets cannot exhaustively represent all possible negative sample scenarios. This leads to the model being highly prone to false alarms when encountering unlearned negative samples, misclassifying cloud shadows, birds, etc., as foreign objects. This results in a persistently high false alarm rate, severely impacting the reliability of the recognition results.
[0007] 2. Methods based on simple time series analysis are not precise enough in capturing the motion characteristics of foreign objects.
[0008] To reduce false alarms, some studies have introduced temporal information, using frame differencing or simple motion trajectory filtering to eliminate stationary or regularly moving interference. However, foreign objects in real-world environments (such as plastic bags tangled in wires) often exhibit complex temporal characteristics such as non-rigidity, intermittency, and wind-driven swaying. Simple frame differencing is sensitive to changes in lighting and struggles to effectively characterize these complex motion patterns; conventional tracking algorithms are prone to losing track of small targets, occluded objects, or objects that briefly disappear. More importantly, existing methods fail to systematically model and fuse the dynamic behavioral differences between positive samples (real foreign objects) and negative samples (foreign object-like interference) in the temporal dimension. They typically handle detection and tracking in isolation, lacking an end-to-end temporal optimization mechanism to deeply integrate the discriminative features of positive and negative samples within their spatiotemporal context. Consequently, they cannot achieve high-precision identification of foreign object motion characteristics and exhibit instability under complex weather and lighting conditions.
[0009] In summary, existing foreign object identification technologies for power transmission lines urgently need a new method that can effectively utilize temporal information and deeply integrate positive and negative sample features to improve the model's identification ability and robustness. Summary of the Invention
[0010] The purpose of this invention is to provide a method for identifying foreign objects in power transmission lines based on temporal optimization through the fusion of positive and negative samples. This method can effectively utilize temporal information and deeply fuse positive and negative sample features, thereby improving the model's identification ability and robustness, and providing more reliable intelligent protection for the safety of power transmission lines.
[0011] A method for foreign object identification in power transmission based on temporal optimization through positive and negative sample fusion, characterized by the following steps:
[0012] S10. Construct a training dataset using known real foreign object positive samples and foreign object-like interference negative samples;
[0013] S20. Construct a foreign object recognition model, which includes a static feature extraction layer, a perturbation feature extraction layer, and a classification decision layer; wherein, the perturbation feature extraction layer contains a perturbation decomposition sub-model;
[0014] S30. Use the training dataset to perform end-to-end training on the foreign object recognition model and optimize the model's internal parameters.
[0015] S40. Deploy the trained foreign object recognition model to the power transmission line monitoring system. For real-time video streams, identify real foreign objects based on the output of the foreign object recognition model and trigger an alarm.
[0016] Furthermore, a training dataset is constructed using known real foreign object positive samples and foreign object-like interference negative samples, specifically through the following sub-steps:
[0017] Import labeled transmission line monitoring video clips into the sample library;
[0018] For each video segment in the sample library, based on the existing frame-by-frame bounding boxes, extract the complete motion trajectory of each independent target in the entire segment;
[0019] Using the motion trajectories of the previous N frames as input samples and the labeled category of the Nth frame as the true label, multiple training samples are generated by sliding and cropping from each complete trajectory, and then organized to form a training dataset.
[0020] The foreign object recognition model is trained end-to-end using the training dataset, which consists of the following sub-steps:
[0021] The samples in the training dataset are divided into training batches, and the forward propagation process is performed sequentially.
[0022] After all samples in the current batch have completed forward propagation, calculate the total loss for that batch.
[0023] The model is updated by applying the gradients of all trainable parameters based on the total loss until it converges.
[0024] For real-time video streams, the model's output is used to identify real foreign objects and trigger alarms. This process involves the following sub-steps:
[0025] For each frame of video image, the front-end module is used to identify suspicious targets and extract center coordinates;
[0026] The motion trajectory of the suspicious target is constructed based on the identification results of the front-end module, and the motion trajectory is input into the foreign object identification model;
[0027] The output of the foreign object recognition model is evaluated. If the object is identified as a real foreign object, an alarm message is generated and triggered through the human-computer interaction page.
[0028] A power transmission positive and negative sample fusion foreign object identification system based on time-series optimization includes: a sample reprocessing module, a foreign object identification model construction and access module, a model pre-processing module, and a foreign object alarm module;
[0029] The sample reprocessing module is used to construct a training dataset using known real foreign object positive samples and foreign object-like interference negative samples;
[0030] The foreign object recognition model construction and access module is used to perform end-to-end training of the foreign object recognition model using the training dataset, and to connect the trained foreign object recognition model to the existing power transmission line monitoring system.
[0031] The model pre-processing module is used to identify suspicious targets in the real-time video stream acquired by the power transmission line monitoring system and to construct the motion trajectory of the suspicious targets based on the identification results.
[0032] The foreign object alarm module is used to input the constructed motion trajectory into the foreign object recognition model, and then identify the real foreign object based on the model output and trigger an alarm.
[0033] The beneficial effects of this invention are as follows:
[0034] It accurately separates background regular motion from target-specific vibration, extracts the most discriminative dynamic features through an adaptive learning mechanism, and combines a unique composite loss function design to significantly improve the ability to distinguish between foreign objects and foreign object-like interference; it effectively reduces the false alarm rate in complex environments, enhances the detection sensitivity of small foreign objects, and provides more reliable intelligent protection for the safety of power transmission lines. Attached Figure Description
[0035] Figure 1 This is a flowchart of a foreign object identification method based on time-series optimization for power transmission positive and negative sample fusion according to the present invention;
[0036] Figure 2 This is a schematic diagram of a power transmission positive and negative sample fusion foreign object identification system based on time-series optimization according to the present invention. Detailed Implementation
[0037] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings.
[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0039] Example 1
[0040] like Figure 1 As shown, a method for foreign object identification based on time-series optimization and fusion of positive and negative power transmission samples includes:
[0041] Step S10: Construct a training dataset using known real foreign object positive samples and foreign object-like interference negative samples;
[0042] This step aims to transform existing, labeled sample data into a training dataset containing spatiotemporal trajectory information suitable for this invention, and specifically consists of the following sub-steps:
[0043] Step S11: Import the labeled transmission line monitoring video clips into the sample library;
[0044] The labeled transmission line monitoring video clips are derived from historical data of power companies or publicly available benchmark datasets. In each frame of the image, real foreign objects (positive samples) and foreign object interference (negative samples) have been accurately labeled with their categories and locations using bounding boxes.
[0045] Step S12: For each video segment in the sample library, extract the complete motion trajectory of each independent target in the entire segment based on the existing frame annotation boxes;
[0046] Using multi-target tracking algorithms (such as DeepSORT and ByteTrack), based on the existing bounding boxes in each frame, the complete motion trajectory of each independent target in the entire video segment is automatically associated and extracted. Each trajectory data includes: trajectory identifier, used to distinguish different targets; target category, directly inherited from the original annotation, i.e., "real foreign object" or "foreign object interference"; spatiotemporal coordinate sequence, referring to the coordinate sequence of the center point of the target in consecutive image frames; and image patch sequence, which is the sequence of target area image patches extracted from each frame image based on the bounding boxes.
[0047] Step S13: Using the motion trajectories of the previous N frames as input samples and the labeled category of the Nth frame as the true label, multiple training samples are generated by sliding and cropping from each complete trajectory, and then organized to form a training dataset.
[0048] N is a system preset value, which is 30 in this embodiment. Note that only the image patch of the Nth frame is retained as the current target image in the motion trajectory of the first N frames, that is, each input sample contains only one image patch; the generated training samples also need to be standardized preprocessed, including image patch scaling and pixel normalization operations.
[0049] Step S20: Construct a foreign object recognition model, which includes a static feature extraction layer, a perturbation feature extraction layer, and a classification decision layer; wherein, the perturbation feature extraction layer further includes a perturbation decomposition sub-model;
[0050] The static feature extraction layer consists of a convolutional neural network. Its function is to perform forward propagation and feature transformation on image patches in the input sample through the convolutional neural network, and output a high-dimensional appearance feature vector rich in semantic information. It is used to characterize the static visual attributes of a target, such as texture, shape, and color;
[0051] The perturbation feature extraction layer is the core of this invention, responsible for extracting key dynamic discriminative features, i.e., perturbation features, from the input samples using the perturbation decomposition sub-model. The internal workflow of the perturbation decomposition sub-model is as follows:
[0052] Fit the target's recent historical motion trajectory to its current underlying oscillation trajectory;
[0053] The basic oscillation trajectory is used to characterize the inherent regular motion components of the background caused by environmental wind loads, and its fitting function is... Represented as:
[0054]
[0055] Where t represents the time index in the current input sample, i is the index of the fundamental oscillation component, i takes values from 1 to M, and M is the preset total number of fundamental oscillation components, which determines the number of sinusoidal function terms used to fit the fundamental oscillation. Its value is set based on prior knowledge of the typical oscillation modes of transmission lines. Let represent the instantaneous amplitude of the i-th fundamental oscillation component at time t. It is any time point between time 0 (the start time of the input sample) and time t. This indicates that the i-th basis oscillation component is at time i. instantaneous frequency, Let represent the instantaneous phase of the i-th fundamental oscillation component at time t. , , The value is determined by a lightweight neural network. Dynamically generated based on the target's recent historical movement trajectory. The set of trainable network parameters, i.e. , Indicates the target is The historical trajectory up to time t-1 This represents the length of the historical observation window.
[0056] Calculate the function sequentially on the time series of the input samples. The return value forms a complete basic oscillation trajectory, represented as .
[0057] Calculate the difference between the input trajectory and the basic oscillation trajectory to obtain the perturbation residual vector;
[0058] The perturbation residual vector is denoted as R, which contains the key dynamic information that best distinguishes foreign objects from interference. ,in The trajectory of the target at time t, i.e., the currently input trajectory. This is the fitted basic oscillation trajectory.
[0059] Perform a fractional Fourier transform on the magnitude sequence of the perturbation residual vector, then calculate the energy in the adaptive sensitive subband based on the obtained fractional domain, and output it as the extracted perturbation feature;
[0060] First, the magnitude sequence is calculated based on the perturbation residual vector R; then, a fractional Fourier transform is performed on this magnitude sequence to obtain the fractional Fourier transform of the sequence. Representation on the corresponding fractional domain ; then in the fractional domain In the calculation of adaptive sensitive subbands The energy within is extracted and output as the perturbation feature, represented as... ,Right now Where u is a fractional-order field variable, and its physical meaning in the fractional Fourier transform varies with the order. The change transitions between the time domain and the frequency domain. For the upper and lower bounds of the adaptive sensitive sub-band;
[0061] It should be noted that Together they form a trainable set of parameters. It is automatically optimized using gradient descent during the end-to-end training of the model.
[0062] The classification decision layer is used to extract the apparent feature vectors output by the static feature extraction layer. , and the perturbation feature scalar output by the perturbation feature extraction layer The data is then spliced together, activated by a fully connected layer and a Softmax function, and the output is a probability distribution containing "real foreign objects" and "foreign object-like interference".
[0063] Step S30: Use the training dataset to perform end-to-end training on the foreign object recognition model and optimize the model's internal parameters;
[0064] The foreign object recognition model undergoes end-to-end deep training to optimize all trainable parameters within the model, including those of the static feature extraction layer, perturbation feature extraction layer, and classification decision layer. This process is divided into the following sub-steps:
[0065] Step S31: Divide the samples in the training dataset into training batches and execute the forward propagation process sequentially;
[0066] The training dataset is randomly divided into multiple fixed-size training batches. For each sample in the training batch, the following forward propagation process is performed sequentially:
[0067] Apparent feature vectors are calculated through a static feature extraction layer. ;
[0068] The perturbation feature scalar is calculated through the perturbation feature extraction layer. ;
[0069] Will and After concatenation, the data is input into the classification decision layer to obtain the predicted probability distribution. .
[0070] Step S32: After all samples in the current batch have completed forward propagation, calculate the total loss for that batch;
[0071] To optimize classification accuracy, enhance the model's discriminative ability on temporal dynamic features, and ensure the complementarity of static appearance features and dynamic perturbation features, the total loss function designed for the foreign object recognition model in this invention is expressed as follows:
[0072] ,in It is the total loss. , The preset regularization coefficient;
[0073] Let b be the cross-entropy loss for the b-th input sample, where b ranges from 1 to B, and B is the total number of samples in the current training batch. The calculation formula is:
[0074] , This represents the true probability that the output is the c-th label for the b-th input sample. Let c be the probability that the model predicts the b-th input sample as the c-th label, where c takes values from 1 to C, and C is the total number of category labels.
[0075] The dynamic feature separation loss for the current batch is calculated using the following formula:
[0076] , , These represent the perturbation features of all positive and negative samples in the current batch, respectively. mean , These represent the perturbation features of all positive and negative samples in the current batch, respectively. variance, This is the separation threshold (preset hyperparameter). It is the variance suppression coefficient (preset hyperparameter);
[0077] The feature orthogonality loss for the current batch is calculated using the following formula:
[0078] , Represents the apparent features of the b-th input sample. The j-th element (component) in. This represents the perturbation feature of the b-th input sample. Indicates all samples in the current batch The mean of the j-th element, This represents the mean of the perturbation features of all samples in the current batch, where j ranges from 1 to D, and D is the apparent feature. The loss term minimizes the linear correlation between the static feature dimensions and the dynamic feature dimensions in the batch data, forcing the model to learn complementary and non-redundant information representations.
[0079] Step S33: Update the parameters of all trainable parameters of the model based on the gradient of the total loss until the model converges;
[0080] The total loss is calculated using backpropagation. Gradients are applied to all trainable parameters of the model, and gradient descent optimization algorithms (such as the Adam optimizer) are used to update the parameters; the total loss from different batches is reused. Model parameters were tuned, and an early stopping method was introduced to avoid overfitting.
[0081] Step S40: Deploy the trained foreign object recognition model to the power transmission line monitoring system. For real-time video streams, identify real foreign objects based on the output of the model and trigger an alarm.
[0082] The best object recognition model obtained from training is deployed to the analysis server of the power transmission line monitoring system, and the following processing steps are performed on the real-time video stream:
[0083] Step S41: For each frame of video image, use the front-end module to identify suspicious targets and extract center coordinates;
[0084] The front-end module continuously receives real-time video streams transmitted from camera devices at monitoring points along the power transmission line, decodes the video streams, and processes them frame by frame. For each frame of video image, a general target detection module (such as YOLO, SSD, etc.) is used to identify all suspicious moving targets in the image and obtain their bounding boxes. Subsequently, a multi-target tracking algorithm (such as DeepSORT) is used to assign a unique ID to each target and correlate them across frames to obtain its center coordinate sequence.
[0085] Step S42: Construct the motion trajectory of the suspicious target based on the recognition results of the front-end module, and input the motion trajectory into the foreign object recognition model;
[0086] For each suspicious target identified by the front-end module, its image patch in the current frame and the center point coordinates in the previous N frames are extracted to form a complete motion trajectory, which is the input sample of the foreign object recognition model. The constructed input sample is then sent to the foreign object recognition model.
[0087] Step S43: Determine the output of the foreign object recognition model. If it is determined to be a real foreign object, generate an alarm message and trigger the alarm through the human-computer interaction page.
[0088] The foreign object identification model outputs a probability distribution of whether the suspicious target belongs to "real foreign object" or "foreign object interference" based on the input sample. If the probability value of "real foreign object" output by the model is greater than a preset threshold, the target is determined to be a real foreign object; otherwise, it is ignored. Once determined to be a real foreign object, the system immediately performs the following operations:
[0089] Generate alarm information: The alarm information should include at least the camera location where the foreign object appeared, the timestamp, and the location of the foreign object in the image;
[0090] Trigger alarms: The monitoring system’s human-computer interaction interface will notify maintenance personnel through highlighted icons, sounds, etc., and alarm information can also be sent to the mobile terminals of relevant responsible persons via SMS, email, etc.
[0091] Recording and storage: The system automatically saves video clips and images of foreign objects before and after the alarm is triggered for subsequent accident tracing and model optimization.
[0092] Example 2
[0093] like Figure 2 As shown, a power transmission positive and negative sample fusion foreign object identification system based on time-series optimization includes: a sample reprocessing module 21, a foreign object identification model construction and access module 22, a model pre-processing module 23, and a foreign object alarm module 24.
[0094] Sample reprocessing module 21 is used to construct a training dataset using known real foreign object positive samples and foreign object interference negative samples;
[0095] Foreign object recognition model construction and access module 22 is used to perform end-to-end training of the foreign object recognition model using the training dataset, and to connect the trained foreign object recognition model to the existing power transmission line monitoring system.
[0096] The model pre-module 23 is used to identify suspicious targets in the real-time video stream acquired by the power transmission line monitoring system and to construct the motion trajectory of the suspicious targets based on the identification results.
[0097] The foreign object alarm module 24 is used to input the constructed motion trajectory into the foreign object recognition model, and then identify the real foreign object based on the model output and trigger an alarm.
[0098] Corresponding to the above embodiments, the present invention provides a computer storage medium, including: at least one memory and at least one processor;
[0099] The memory is used to store one or more program instructions;
[0100] A processor for running one or more program instructions to execute a timing-optimized method for identifying foreign objects in power transmission positive and negative sample fusion.
[0101] Corresponding to the above embodiments, this embodiment of the invention provides a computer-readable storage medium containing one or more program instructions, which are executed by a processor to provide a timing-optimized method for identifying foreign objects by fusing positive and negative power transmission samples.
[0102] The embodiments disclosed in this invention provide a computer-readable storage medium storing computer program instructions. When the computer program instructions are executed on a computer, the computer performs the aforementioned method for identifying foreign objects by fusing positive and negative power transmission samples based on time-series optimization.
[0103] In this embodiment of the invention, the processor can be an integrated circuit chip with signal processing capabilities. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.
[0104] The various methods, steps, and logic diagrams disclosed in the embodiments of this invention can be implemented or executed. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this invention can be directly implemented by a hardware decoding processor, or implemented by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The processor reads information from the storage medium and, in conjunction with its hardware, completes the steps of the above methods.
[0105] The storage medium can be memory, such as volatile memory or non-volatile memory, or may include both volatile and non-volatile memory.
[0106] Among them, non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory.
[0107] Volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous linked dynamic random access memory (Synchlink DRAM, SLDRAM), and direct memory bus RAM (DRRAM).
[0108] The storage media described in the embodiments of the present invention are intended to include, but are not limited to, these and any other suitable types of memory.
[0109] Those skilled in the art will recognize that, in one or more of the examples above, the functions described in this invention can be implemented using a combination of hardware and software. When applied as software, the corresponding functions can be stored in a computer-readable medium or transmitted as one or more instructions or code on a computer-readable medium. Computer-readable media include computer storage media and communication media, wherein communication media include any medium that facilitates the transmission of computer programs from one place to another. Storage media can be any available medium that can be accessed by a general-purpose or special-purpose computer.
[0110] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solution of the present invention should be included within the scope of protection of the present invention.
Claims
1. A foreign matter recognition method based on timing optimization of positive and negative power transmission samples, characterized by, Includes the following steps: S10. Construct a training dataset using known real foreign object positive samples and foreign object-like interference negative samples; S20. Construct a foreign object recognition model, which includes a static feature extraction layer, a perturbation feature extraction layer, and a classification decision layer; wherein, the perturbation feature extraction layer contains a perturbation decomposition sub-model; S30. Use the training dataset to perform end-to-end training on the foreign object recognition model and optimize the model's internal parameters. S40. Deploy the trained foreign object recognition model to the power transmission line monitoring system. For real-time video streams, identify real foreign objects based on the output of the foreign object recognition model and trigger an alarm.
2. The method for foreign object identification based on time-series optimization of positive and negative power transmission samples as described in claim 1, characterized in that, In step S10, the specific steps are as follows: Import labeled transmission line monitoring video clips into the sample library; For each video segment in the sample library, based on the existing frame-by-frame bounding boxes, extract the complete motion trajectory of each independent target in the entire segment; Using the motion trajectories of the previous N frames as input samples and the labeled category of the Nth frame as the true label, multiple training samples are generated by sliding and cropping from each complete trajectory, and then organized to form a training dataset.
3. The method for foreign object identification based on time-series optimization of positive and negative power transmission samples as described in claim 1, characterized in that, In step S20, the static feature extraction layer is composed of a convolutional neural network. For image patches in the input sample, the convolutional neural network performs forward propagation and feature transformation to output a high-dimensional appearance feature vector rich in semantic information.
4. The method for foreign object identification based on time-series optimization of positive and negative power transmission samples as described in claim 1, characterized in that, The perturbation feature extraction layer is used to extract key dynamic discriminative features, i.e., perturbation features, from the input samples using the perturbation decomposition sub-model; the internal workflow of the perturbation decomposition sub-model is as follows: 1) Fit the target's recent historical motion trajectory to its current basic oscillation trajectory; The basic oscillation trajectory is used to characterize the inherent regular motion components of the background caused by environmental wind loads, and its fitting function is... Represented as: ； Where t represents the time index in the current input sample, i is the index of the fundamental oscillation component, i takes values from 1 to M, and M is the preset total number of fundamental oscillation components, which determines the number of sinusoidal function terms used to fit the fundamental oscillation. Its value is set based on prior knowledge of the typical oscillation modes of transmission lines. Let represent the instantaneous amplitude of the i-th fundamental oscillation component at time t. It is any time point between the start time 0 and time t of the input sample. This indicates that the i-th basis oscillation component is at time i. instantaneous frequency, Let represent the instantaneous phase of the i-th fundamental oscillation component at time t. , , The value is determined by a lightweight neural network. Dynamically generated based on the target's recent historical movement trajectory. The set of trainable network parameters, i.e. , Indicates the target is The historical trajectory up to time t-1 This refers to the length of the historical observation window; Calculate the function sequentially on the time series of the input samples. The return value forms a complete basic oscillation trajectory, represented as ; 2) Calculate the difference between the input trajectory and the basic oscillation trajectory to obtain the perturbation residual vector; The perturbation residual vector is denoted as R, which contains the key dynamic information that best distinguishes foreign objects from interference. ,in The trajectory of the target at time t, i.e., the currently input trajectory. The fitted basic oscillation trajectory; 3) Perform a fractional Fourier transform on the magnitude sequence of the perturbation residual vector, then calculate the energy within the adaptive sensitive sub-band based on the obtained fractional domain, and output it as the extracted perturbation feature, as follows: First, the magnitude sequence is calculated based on the perturbation residual vector R; then, a fractional Fourier transform is performed on this magnitude sequence to obtain the fractional Fourier transform of the sequence. Representation over the corresponding fractional domain ; then in the fractional domain In the calculation of adaptive sensitive subbands The energy within is extracted and output as the perturbation feature, represented as... ,Right now Where u is a fractional-order field variable, and its physical meaning in the fractional Fourier transform varies with the order. The change transitions between the time domain and the frequency domain. For the upper and lower bounds of the adaptive sensitive sub-band; Together they form a trainable set of parameters. It is automatically optimized using gradient descent during the end-to-end training of the model.
5. The method for foreign object identification based on time-series optimization of positive and negative power transmission samples as described in claim 1, characterized in that, In step S30, the foreign object recognition model is trained end-to-end using the training dataset, which is specifically divided into the following sub-steps: The samples in the training dataset are divided into training batches, and the forward propagation process is performed sequentially. After all samples in the current batch have completed forward propagation, calculate the total loss for that batch. The model is updated by applying the gradients of all trainable parameters based on the total loss until it converges.
6. The method for foreign object identification based on time-series optimization of positive and negative power transmission samples as described in claim 5, characterized in that, After all samples in the current batch have completed forward propagation, the total loss for that batch is calculated as follows: To optimize classification accuracy, enhance the model's discriminative ability on temporal dynamic features, and ensure the complementarity of static appearance features and dynamic perturbation features, the total loss function of the foreign object recognition model is expressed as: ,in It is the total loss. , The preset regularization coefficient; Let b be the cross-entropy loss for the b-th input sample, where b ranges from 1 to B, and B is the total number of samples in the current training batch. The calculation formula is: , This represents the true probability that the output is the c-th label for the b-th input sample. Let c be the probability that the model predicts the b-th input sample as the c-th label, where c takes values from 1 to C, and C is the total number of category labels. The dynamic feature separation loss for the current batch is calculated using the following formula: , , These represent the perturbation features of all positive and negative samples in the current batch, respectively. mean , These represent the perturbation features of all positive and negative samples in the current batch, respectively. variance, The resolution threshold, It is the variance suppression coefficient; The feature orthogonality loss for the current batch is calculated using the following formula: , Represents the apparent features of the b-th input sample. The j-th element in This represents the perturbation feature of the b-th input sample. Indicates all samples in the current batch The mean of the j-th element, This represents the mean of the perturbation features of all samples in the current batch, where j ranges from 1 to D, and D is the apparent feature. Dimensions.
7. The method for foreign object identification based on time-series optimization of positive and negative power transmission samples as described in claim 1, characterized in that, In step S40, for the real-time video stream, the model's output is used to identify real foreign objects and trigger an alarm. This process is divided into the following sub-steps: For each frame of video image, the front-end module is used to identify suspicious targets and extract center coordinates; The motion trajectory of the suspicious target is constructed based on the identification results of the front-end module, and the motion trajectory is input into the foreign object identification model; The output of the foreign object recognition model is evaluated. If the object is identified as a real foreign object, an alarm message is generated and triggered through the human-computer interaction page.
8. The method for foreign object identification based on time-series optimization of positive and negative power transmission samples as described in claim 7, characterized in that, The evaluation of the output of the foreign object recognition model involves the following sub-steps: The foreign object identification model outputs the probability distribution of whether the suspicious target is a real foreign object or a foreign object-like interference based on the input sample; If the probability value of a real foreign object output by the model is greater than the preset threshold, the target is determined to be a real foreign object; otherwise, it is ignored. Once a foreign object is identified as a real foreign object, the system immediately performs operations such as generating alarm information, triggering alarms, and recording and storing data.
9. A system for implementing the time-series optimization-based transmission positive and negative sample fusion foreign object identification method as described in claim 1, characterized in that, include: Sample reprocessing module, foreign object identification model construction and access module, model pre-processing module, foreign object alarm module; The sample reprocessing module is used to construct a training dataset using known real foreign object positive samples and foreign object-like interference negative samples; The foreign object recognition model construction and access module is used to perform end-to-end training of the foreign object recognition model using the training dataset, and to connect the trained foreign object recognition model to the existing power transmission line monitoring system. The model pre-processing module is used to identify suspicious targets in the real-time video stream acquired by the power transmission line monitoring system and to construct the motion trajectory of the suspicious targets based on the identification results. The foreign object alarm module is used to input the constructed motion trajectory into the foreign object recognition model, and then identify the real foreign object based on the model output and trigger an alarm.