A method for on-orbit identification of power facilities

Abstract

The application discloses a kind of on-orbit identification methods of power facilities on satellite, belong to remote sensing image processing and identification technical field, steps include: obtaining the original image data stream generated by satellite platform collection power facility remote sensing image;Original image data stream is input into on-board lightweight identification model and generates the identification result containing position, type and initial confidence degree in real time inference;According to initial confidence degree grading processing, generate the differentiated data containing high confidence degree abnormal alarm and low confidence degree image data package to be verified and be downloaded to ground station;Receive the correction label verification result fed back by ground station based on data package to be verified;On-orbit incremental learning is carried out using verification result, and on-board lightweight identification model is updated.The application adopts the closed-loop evolution mechanism combining on-orbit identification and ground verification, can discover power anomaly in real time and download key information as needed, improve emergency response timeliness and identification precision, reduce communication bandwidth pressure.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image processing and recognition technology, and particularly relates to an on-orbit identification method for power facilities. Background Technology

[0002] Power facilities, such as transmission towers, substations, and transmission lines, are a crucial component of infrastructure. With the continuous expansion of power grids and the impact of factors such as natural disasters, traditional manual inspections and periodic maintenance methods face numerous challenges, including low efficiency, high costs, long inspection cycles, and difficulty in responding to sudden anomalies. To address these issues, combining remote sensing technology with artificial intelligence to intelligently identify power facility remote sensing images has become a new trend.

[0003] Existing intelligent identification solutions for remote sensing images of power facilities typically rely on transmitting all raw remote sensing image data acquired by satellites back to ground stations. On the ground station's servers, deep neural network models are deployed to centrally analyze and process the massive amounts of received raw data to identify the geographical location, type, and potential anomalies of power facilities. While this type of ground processing solution can achieve high identification accuracy by utilizing powerful computing resources and complex model structures, its core operating mode relies on transmitting all raw data from satellite to the ground for subsequent processing.

[0004] However, this traditional "transmit first, process later" approach has significant technical drawbacks. First, the transmission of massive amounts of raw remote sensing image data requires enormous satellite communication bandwidth, especially given limited satellite resources. This severely restricts the real-time nature of data acquisition, mission flexibility, and communication efficiency. Second, when power facility anomalies occur, the transmission of large amounts of data and centralized ground processing inevitably introduce significant latency, making it impossible to detect and respond to emergencies promptly. This is unacceptable for power facility fault early warning and repair, which require high real-time performance. Therefore, overcoming the limitations of onboard resources and ground communication bottlenecks to achieve efficient, real-time onboard intelligent identification and continuously improve identification performance has become a critical technical problem that urgently needs to be solved. Summary of the Invention

[0005] To address the aforementioned issues, the present invention aims to provide an on-orbit identification method for power facilities. This method employs a closed-loop mechanism of on-board real-time inference, confidence-level hierarchical downlink, and ground-based collaborative incremental learning. This mechanism can improve the real-time performance and accuracy of power facility anomaly identification, optimize the utilization efficiency of satellite communication resources, and enable continuous self-optimization of the model.

[0006] To achieve the above objectives, this application adopts the following technical solution: A method for on-orbit identification of power facilities, comprising the following steps: Acquire remote sensing images of power facilities in the target area in real time from the satellite platform and generate raw image data stream; The raw image data stream is input into the on-board lightweight recognition model for real-time inference, generating on-board recognition results that include the location information of the target power facility, the facility type recognition result, and the initial confidence level. The on-board identification results are graded based on the initial confidence level to generate differentiated identification result data that includes high-confidence anomaly alarm information and low-confidence image data packets to be verified. The differential identification results data are transmitted to the ground station via satellite communication link; The system receives verification results from the ground station based on the differential identification results data. The verification results are generated by the ground station after verifying the low-confidence image data packets to be verified, and include correction labels for the low-confidence image data packets to be verified. The validation results are used to perform on-orbit incremental learning on the lightweight recognition model and update the lightweight recognition model.

[0007] As a preferred embodiment of the present invention, the acquisition of real-time remote sensing images of power facilities in the target area by the satellite platform is triggered by the on-board mission scheduling unit. When the satellite flies over the power facility monitoring area or receives an emergency observation command sent from the ground, the satellite payload is automatically controlled to image the target area.

[0008] As a preferred embodiment of the present invention, the deployment process for initializing the lightweight recognition model is as follows: Obtain a historical remote sensing sample set of power facilities containing sample images labeled with normal and several abnormal states; A ground-based high-precision verification model was trained using a historical remote sensing sample set of power facilities. The ground-based high-precision verification model is a deep neural network model deployed on a ground server. Its network structure and parameter scale are not limited by spaceborne resources. It is used to perform high-precision verification and identification of low-confidence image data packets to be verified. Using knowledge distillation technology, a simplified, lightweight on-board recognition model is derived by using a high-precision ground-based verification model as the teacher model, and then deployed on the satellite platform.

[0009] As a preferred embodiment of the present invention, the on-board lightweight recognition model is a pruned and optimized convolutional neural network model, whose model size and computational complexity are adapted to the memory and computing power constraints of the on-board processor.

[0010] As a preferred embodiment of the present invention, the classification processing of the on-board identification results based on the initial confidence level includes: Obtain a first confidence threshold and a second confidence threshold to distinguish between high-confidence and low-confidence recognition results, wherein the first confidence threshold is higher than the second confidence threshold; For satellite identification results with an initial confidence level higher than the first confidence threshold, their location information and facility type identification results are extracted and packaged into high-confidence anomaly alarm information; The original image regions corresponding to the on-board recognition results with initial confidence levels between the first and second confidence thresholds are cropped to generate low-confidence image data packets to be verified.

[0011] As a preferred embodiment of the present invention, transmitting the differential identification result data to the ground station via a satellite communication link includes: High-confidence anomaly alarm information is transmitted in real time via the satellite's real-time telemetry channel; Low-confidence image data packets to be verified are downloaded non-real-time via the satellite's broadband data transmission channel.

[0012] As a preferred embodiment of the present invention, the verification results fed back by the ground station are specifically as follows: The ground station receives low-confidence image data packets and inputs them into a high-precision ground verification model for fine identification to obtain corrected labels. The low-confidence image data to be verified and its corresponding correction labels are packaged together to generate a mini incremental training sample package as the verification result.

[0013] As a preferred embodiment of the present invention, the on-orbit incremental learning of the on-board lightweight identification model using the verification results includes: Satellite receives mini-incremental training sample packets; On the onboard processor, the current lightweight recognition model is trained using micro-incremental training sample packages, thereby improving the lightweight recognition model's ability to distinguish scenes related to low-confidence image data packets. Update the on-board lightweight recognition model with the new parameters after training.

[0014] The present invention has the following advantages: This invention achieves continuous improvement in the accuracy of a lightweight on-board identification model for complex and abnormal scenarios through an in-orbit incremental learning mechanism. This mechanism enables the model to continuously adapt to new operating conditions, environmental interference, and previously unseen fault types without frequent manual intervention, thereby extending the model's effectiveness and service life, reducing operation and maintenance costs, and improving the intelligence level of power facility monitoring systems.

[0015] This invention effectively optimizes the utilization efficiency of satellite communication resources. By classifying the confidence levels of onboard identification results, high-confidence anomaly alarms can be instantly transmitted via real-time telemetry channels with minimal data volume, ensuring rapid response in emergencies. Lower-confidence image data packets, which have relatively larger data volumes, are transmitted in batches via non-real-time broadband data transmission channels, avoiding unnecessary real-time bandwidth consumption and allowing limited satellite communication bandwidth to be allocated on demand, maximizing its effectiveness.

[0016] This invention realizes an intelligent collaborative working mode between onboard and ground systems. The lightweight onboard identification model focuses on initial screening and rapid decision-making, utilizing limited onboard resources for preliminary identification. Meanwhile, the ground system leverages its powerful computing capabilities and human expert resources to perform high-precision verification and annotation of low-confidence data that the onboard system cannot accurately assess, feeding back its expertise as corrective labels to the onboard model. This closed-loop learning mode of space-ground collaboration leverages the respective strengths of both the onboard and ground systems, improving the overall accuracy and reliability of identification and ensuring comprehensive monitoring capabilities for power facilities. Attached Figure Description

[0017] Figure 1 This is a schematic flowchart of the method of the present invention; Figure 2 This is a statistical heatmap of the abnormal status identification results of power facilities in Embodiment 1 of the present invention; Figure 3 This is a graph showing the performance change of the model iterative evolution in Embodiment 1 of the present invention. Detailed Implementation

[0018] The embodiments of the present invention will be further described below with reference to the accompanying drawings: Example 1: As Figure 1 As shown, a method for on-orbit identification of power facilities includes the following steps: Acquire remote sensing images of power facilities in the target area in real time from the satellite platform and generate raw image data stream; The raw image data stream is input into the on-board lightweight recognition model for real-time inference, generating on-board recognition results that include the location information of the target power facility, the facility type recognition result, and the initial confidence level. The on-board identification results are graded based on the initial confidence level to generate differentiated identification result data that includes high-confidence anomaly alarm information and low-confidence image data packets to be verified. The differential identification results data are transmitted to the ground station via satellite communication link; The system receives verification results from the ground station based on the differential identification results data. The verification results are generated by the ground station after verifying the low-confidence image data packets to be verified, and include correction labels for the low-confidence image data packets to be verified. The validation results are used to perform on-orbit incremental learning on the lightweight recognition model and update the lightweight recognition model.

[0019] This embodiment combines the rapid response capability of onboard real-time processing, the accuracy of high-precision verification at ground stations, and the continuous optimization mechanism of on-orbit incremental learning to construct an intelligent identification closed-loop method for power facility monitoring: First, the satellite platform uses its onboard lightweight recognition model to perform preliminary intelligent analysis on real-time acquired remote sensing imagery, quickly identifying power facilities and their potential anomalies, and outputting recognition results with confidence levels. Subsequently, based on these confidence levels, the results are finely graded. High-confidence information indicating a high probability of anomalies is immediately alerted, while low-confidence image samples where the model's assessment is uncertain are transmitted back to the ground. On the ground, low-confidence images are validated using a high-precision verification model with unlimited resources or by human experts, generating accurate correction labels. Finally, these samples with correction labels are transmitted back to the satellite for on-orbit updates to the lightweight recognition model. This entire process forms an iterative learning feedback loop, enabling the onboard model to continuously adapt to new scenarios and challenges and evolve itself through collaborative work with the ground.

[0020] The acquisition of real-time remote sensing images of power facilities in the target area by the satellite platform is triggered by the onboard mission scheduling unit. When the satellite flies over the power facility monitoring area or receives an emergency observation command sent from the ground, it automatically controls the satellite payload to image the target area.

[0021] Based on a pre-set flight plan or unexpected ground demands, the satellite's optical remote sensing payload is precisely controlled to image the power facility monitoring area, acquiring high-quality real-time remote sensing image data as input for the onboard intelligent identification system. This process is driven by the onboard mission scheduling unit and responds according to different triggering conditions.

[0022] It is necessary to ensure that the satellite platform can proactively identify and respond to two key imaging triggering conditions: autonomously planned regional observation tasks and emergency observation requests from the ground. The onboard mission scheduling unit is one of the core intelligent components of the satellite platform, responsible for managing and coordinating various tasks on the satellite, including attitude control, payload operation, and data transmission.

[0023] The onboard mission scheduling unit has a built-in or receives and updates a comprehensive mission list or schedule. The first trigger condition is an autonomous judgment based on the satellite's orbit: when the onboard navigation and positioning system (GNSS receiver and star sensor) calculates that the satellite's real-time position overlaps spatially with a predefined power facility monitoring area, the onboard mission scheduling unit will trigger an imaging mission. The second trigger condition is an emergency observation command sent from the ground: when the ground control center determines, based on an emergency event such as a natural disaster or a report of suspicious activity, that immediate observation of a specific power facility area is necessary, it will send an emergency observation command containing the geographical coordinates of the target area and observation parameters via the satellite's telemetry and command link.

[0024] The onboard mission scheduling unit receives, parses, and verifies the legality and priority of the instruction, and then triggers the imaging mission. After identifying valid imaging trigger conditions, it is necessary to precisely control the satellite platform and optical remote sensing payload to perform high-resolution remote sensing imaging of the target area and generate raw image data streams.

[0025] Once the onboard mission scheduling unit is activated under one of the two triggering conditions mentioned above, it first sends a command to the satellite attitude control system (ACS) to adjust the satellite's attitude, ensuring that the field of view (FoV) of the optical remote sensing payload is precisely aligned with the target power facility area. The attitude adjustment accuracy is typically controlled between 0.01 degrees and 0.001 degrees to ensure the target is centered in the image. Subsequently, the scheduling unit activates and configures the remote sensing payload, setting appropriate imaging parameters, such as selecting panchromatic or multispectral mode, exposure time, and gain.

[0026] For example, for applications requiring precise identification of power facility faults, a panchromatic imaging mode with a resolution better than 1 meter is often selected. Once the payload is configured, the scheduling unit sends a command to the camera command unit to initiate the imaging sequence. The optical remote sensing payload begins acquiring images of the Earth's surface in the target area. During image acquisition, the raw light signal is focused by the optical system and converted into an electrical signal. After quantization by an analog-to-digital converter (ADC), digital image data is generated. This raw, uncompressed, or only initially lossless compressed digital image data is packaged according to a predetermined format to form a raw image data stream. This data stream is then sent to the onboard data processor for buffering.

[0027] One application example is that the onboard mission scheduling unit calculates the satellite orbit in real time through the onboard navigation and positioning system. When it determines that the satellite's real-time position has entered the geofence of the pre-defined power transmission line corridor defined by latitude and longitude, or when it receives an emergency observation command containing the target coordinates sent by the ground control center, it immediately issues a scheduling command to the satellite attitude control system to accurately align the field of view of the optical remote sensing payload with the target area with an attitude adjustment accuracy of 0.005 degrees.

[0028] Subsequently, the payload was configured to initiate the acquisition sequence in a 0.5-meter resolution panchromatic imaging mode. The optical payload captured the raw light signal through a CMOS detector and generated a digital image after quantization by an ADC. If the size of a single raw image is... Given pixels and a bit depth of 12 bits, the uncompressed data size of the original image is calculated as follows: The JPEG-LS algorithm is used for initial lossless compression. If the compression ratio is 2:1, the size of a single image in the final generated original image data stream will be [size missing]. The data stream is then fed into the onboard processor cache.

[0029] The deployment process for initializing the lightweight recognition model (initial deployment) is as follows: Obtain a historical remote sensing sample set of power facilities containing sample images labeled with normal and several abnormal states; A ground-based high-precision verification model was trained using a historical remote sensing sample set of power facilities. The ground-based high-precision verification model is a deep neural network model deployed on a ground server. Its network structure and parameter scale are not limited by spaceborne resources. It is used to perform high-precision verification and identification of low-confidence image data packets to be verified. Using knowledge distillation technology, a simplified, lightweight on-board recognition model is derived by using a high-precision ground-based verification model as the teacher model, and then deployed on the satellite platform.

[0030] Before satellite launch, a high-performance teacher model and a student model adapted to the satellite environment are built on the ground. The student model is then pre-deployed to the satellite platform to provide an initial model foundation for subsequent on-orbit recognition tasks. This process mainly consists of two core steps: training a high-precision ground-based validation model and distilling and deploying a lightweight on-board recognition model. The goal is to train a high-precision ground-based validation model with the highest recognition accuracy, which is not limited by computing resources and will serve as the teacher model in the subsequent knowledge distillation process.

[0031] First, a large-scale, high-quality remote sensing sample set of historical power facilities is required. This sample set typically contains hundreds of thousands of high-resolution remote sensing images with a resolution between 0.5 meters and 2 meters, each meticulously annotated by a team of experts. The annotation information includes not only the precise location bounding boxes of power facilities such as transmission towers and substations, but also their status labels, such as "normal," "excessive vegetation around the tower base," "tower tilted," and "large construction machinery nearby," among several predefined normal and abnormal states.

[0032] After acquiring the sample set, a complex deep neural network model with a large number of parameters is selected as the basic architecture for the high-precision ground validation model, such as EfficientNet-B7 or a large-scale vision model based on Transformer. During training, a historical remote sensing sample set of power facilities is used as input, and a cross-entropy loss function with class weights is adopted to address the sample imbalance problem. The AdamW optimizer is used for iterative optimization. Training will be carried out on a high-performance GPU cluster for hundreds of training epochs until the model's mean average accuracy (mAP) on the validation set reaches over 98%, ensuring that it is qualified as a "teacher".

[0033] Based on the already trained high-precision ground-based validation model, a lightweight on-board recognition model with a simplified structure, low computational cost, and compliance with on-board processor resource constraints is generated using knowledge distillation technology, and then deployed on the satellite platform. This process uses the high-precision ground-based validation model as the teacher model and a lightweight network structure such as MobileNetV3 or a ShuffleNet V2 optimized with channel pruning and quantization as the student model, i.e., the lightweight on-board recognition model.

[0034] The training process for knowledge distillation utilizes not only hard labels (real labels) from a historical remote sensing sample set of power facilities, but also soft labels with class probability distributions output by a teacher model for inference on the same batch of image data. The training objective is to minimize a combined loss function. : ; in, The standard cross-entropy loss represents the difference between the student model output and the true label, and is used to ensure that the student model learns the correct classification knowledge. The Kullback-Leibler divergence loss, representing the soft label outputs of the student and teacher models, prompts the student model to mimic the class probability distribution of the teacher model and learn the teacher model's "insights" about the similarity between different classes. `t` is a weighting hyperparameter used to balance the importance of hard and soft label losses, and its value is typically between 0.1 and 0.5. `T` is the distillation temperature coefficient, a scalar greater than 1, used to smooth the output probability distribution of the teacher and student models, and is generally set to 2 to 10.

[0035] By backpropagating and optimizing the combined loss function, the resulting student model is the lightweight on-board recognition model. Before deployment, the model needs to be subjected to fixed-point quantization, such as INT8, to further compress the model size and accelerate inference. It is then embedded into the satellite platform's FPGA or a dedicated AI acceleration chip to complete model initialization.

[0036] One application example is in the model initialization process for power facility inspection tasks. First, a set of 300,000 historical remote sensing images with 0.5-meter resolution, labeled with abnormal conditions such as transmission tower tilt and cable icing, is acquired. A high-precision ground-based validation model is trained on a GPU cluster using EfficientNet-B7 as the infrastructure, achieving an mAP of 98.5% to serve as the teacher model. Subsequently, MobileNetV3 is selected as the student model, i.e., the satellite-based lightweight recognition model, and trained using knowledge distillation techniques.

[0037] In calculating the combination loss function At that time, preset weight hyperparameters The cross-entropy loss is 0.3, and the distillation temperature coefficient T is 2. If, in a certain training iteration, the student model produces a cross-entropy loss... The KL divergence loss is 0.15, representing the difference between the soft label outputs of the teacher and student models. The value is 0.08. Substituting this value into the formula, we can calculate the combined loss value at that moment. .

[0038] Finally, by performing INT8 fixed-point quantization on the optimized model, the parameter size was compressed from about 66MB in the teacher model to about 5MB in the student model, and then solidified into the dedicated AI acceleration chip of the satellite platform, completing the initial deployment. Figure 2 The accuracy distribution of the method in this embodiment for identifying four typical states of power facilities: normal, tilted tower, excessive vegetation, and construction machinery is shown. Figure 2 The diagonal values ​​in the table represent the accuracy of each category recognition.

[0039] The lightweight onboard recognition model is a pruned and optimized convolutional neural network model, whose model size and computational complexity are adapted to the memory and computing power constraints of the onboard processor.

[0040] The lightweight onboard recognition model is a pruned and optimized convolutional neural network model that meets the high-precision recognition requirements of power facilities while strictly adapting to the limited memory and computing power (FLOPs, i.e., floating-point operations) constraints of the onboard processor. This optimized model can ensure that image processing and target recognition tasks are completed with low power consumption and real-time performance in the harsh environment of space, ensuring the continuous and efficient operation of the satellite platform.

[0041] Without significantly sacrificing recognition accuracy, it is necessary to drastically reduce the number of parameters and computational cost of convolutional neural network models, enabling their deployment on resource-constrained spaceborne processors. The lightweight spaceborne recognition model initially originates as a student model derived from a high-precision ground-based validation model using knowledge distillation techniques; this student model is typically already a lightweight network architecture. However, to further adapt to the demanding spaceborne resources, more in-depth pruning optimization is required. Pruning is a model compression technique, primarily employing unstructured and structured pruning methods. For hardware-friendly implementation, structured pruning, such as channel pruning or layer pruning, is typically used.

[0042] The engineering steps for channel pruning are as follows: For a convolutional layer C, its weight parameter W can be represented as a four-dimensional tensor, and the number of input channels is... The number of output channels is The kernel size is : ; For each output channel, the pruning algorithm evaluates its importance to the model output. A common method for importance evaluation is based on the L1 or L2 norm of the channel, or by adding a learnable scaling factor to the channel. This scaling factor is then used as a parameter in the Batch Normalization layer. During training, these scaling factors are encouraged to be close to zero.

[0043] After evaluation, channels with importance below a preset threshold Z are directly removed. The threshold Z is a key parameter obtained through iterative adjustment during pruning experiments, for example, set to 1% to 10% of the global maximum scaling factor. Simultaneously, the corresponding input channels in the next convolutional kernel connected to these channels are also removed, thus achieving a more compact network structure.

[0044] For example, if a convolutional layer has 128 output channels, pruning can reduce it to 64, which directly halves the number of parameters in that layer and the computational cost of the next layer. This process requires multiple iterations of training and pruning within a dedicated pruning framework until the model's accuracy loss on the validation set is within an acceptable range, while the model compression rate reaches the preset target.

[0045] The final model size is typically between several MB and tens of MB. The pruned and optimized convolutional neural network model is not only small in size, but its inference process can also be efficiently completed within the limited computing power constraints of the onboard processor, achieving real-time processing capabilities. After channel pruning, a series of computational complexity assessments and optimizations are required to adapt to the computing power of the onboard processor. The onboard processor has dedicated AI acceleration units, such as digital signal processors (DSPs), field-programmable gate arrays (FPGAs), or application-specific integrated circuits (ASICs), whose instruction sets and parallel computing capabilities are limited. Computational complexity is commonly measured in FLOPs (Floating Point Operations), and pruning can directly reduce the model's FLOPs. For a convolutional layer C, its computational cost... The following formula can be used to approximate the calculation: ; in, It is the size of the convolution kernel. This is the number of input channels. It is the number of output channels. and These are the height and width of the output feature map, respectively. Pruning reduces... and To reduce For example, when processing a frame When using pixel-level remote sensing images, the lightweight onboard recognition model needs to complete the recognition within 200 milliseconds after a frame of image input. This means that the overall computational load of the model needs to be controlled to hundreds of GOPs or less to adapt to onboard processors, such as processors based on the RISC-V architecture and equipped with specific instruction set extensions, which can achieve peak computing power of tens of TOPS. However, the actual continuous working computing power is limited by power consumption and heat dissipation, and is usually in the range of hundreds of GOPs to a few TOPS.

[0046] Furthermore, to further reduce computational and memory pressure during runtime, the model is typically quantized. For example, 32-bit floating-point weights and activation values ​​are quantized into 8-bit integers (INT8), or even binarized. This not only reduces the model size but also leverages the onboard accelerator's optimization of integer operations, further improving inference speed. Quantization optimization can increase model inference speed by 2-4 times and reduce memory usage by 4 times. Through these pruning and quantization optimization techniques, the lightweight onboard recognition model can be stably deployed and run efficiently on the onboard processor, meeting the requirements for high-timeliness recognition.

[0047] One application example is performing structured channel pruning on a lightweight student model distilled from a high-precision ground-based validation model, with respect to weight parameters. The convolutional layer, if the original output channel number The pruning algorithm evaluates the scaling factor in the Batch Normalization layer, setting the value to 128. The importance of this is that 64 channels below the preset threshold Z are directly removed, which halves the number of parameters in this layer.

[0048] When evaluating computational complexity, utilize The calculation is performed using the formula, assuming the convolution kernel size... The number of input channels is 3. The number of output channels was reduced from 64 to 32 after pruning. The size of the output feature map is reduced from 128 to 64. Then the computational load of this layer is reduced from that before pruning. Sub-floating-point operations reduced to The sub-floating-point operations significantly reduced the computational burden. Finally, INT8 quantization converted the model weights from 32-bit floating-point numbers to 8-bit integers, reducing memory usage by 4 times and ensuring that the overall computational load of the model was kept within a few hundred GOPs. This made it suitable for spaceborne processors with limited computing power, enabling real-time image recognition within 200 milliseconds.

[0049] The on-board identification results are graded based on the initial confidence level, including: Obtain a first confidence threshold and a second confidence threshold to distinguish between high-confidence and low-confidence recognition results, wherein the first confidence threshold is higher than the second confidence threshold; For satellite identification results with an initial confidence level higher than the first confidence threshold, their location information and facility type identification results are extracted and packaged into high-confidence anomaly alarm information; The original image regions corresponding to the on-board recognition results with initial confidence levels between the first and second confidence thresholds are cropped to generate low-confidence image data packets to be verified.

[0050] Based on the confidence level of the identification results, a large number of power facility identification results processed on satellite are intelligently classified, thereby efficiently generating targeted high-confidence anomaly alarm information and low-confidence pending verification image data packets that require further confirmation by humans or ground models, thus optimizing communication bandwidth utilization and ground processing efficiency.

[0051] The tiered processing procedure primarily relies on two preset confidence thresholds. Two key confidence thresholds are set and acquired as decision boundaries to distinguish identification results at different confidence levels, guiding subsequent tiered processing. The first and second confidence thresholds are pre-set and stored. The settings of the first confidence threshold R1 and the second confidence threshold R2 are dynamically adjustable. Depending on different power inspection task requirements or the on-orbit performance evolution of the model, the thresholds can be remotely modified via ground commands or autonomously updated by the onboard mission scheduling unit.

[0052] For example, during periods of frequent meteorological disasters such as thunderstorms and severe convection, in order to maximize the capture of potential hazards to power facilities, the ground station can issue control commands to lower the values ​​of R1 and R2, thereby expanding the acquisition coverage of the image data packets to be inspected and improving the recall rate of detecting minor anomalies. On the other hand, as the number of on-orbit incremental learning iterations increases, when the overall discrimination confidence distribution of the on-board lightweight recognition model matures, the value of R1 can be increased, thereby making high-confidence anomaly alarm information more accurate, further reducing unnecessary low-confidence image downlink volume, and achieving optimal utilization of satellite communication bandwidth.

[0053] The flexible configuration of thresholds ensures that the method in this embodiment can achieve a dynamic balance between recognition accuracy and resource consumption based on real-time business priorities. For example, the first confidence threshold R1 can be set at 0.95, meaning that only results with very high confidence are considered reliable; the second confidence threshold R2 can be set at 0.7, defining a middle ground where results below R2 may be ignored or marked as unreliable, while results above R2 but below R1 are considered to require further verification. Crucially, the value of R1 must be strictly higher than R2 to ensure the logical order of confidence interval division.

[0054] From massive amounts of on-board identification results, highly confident anomalies are quickly filtered out and encapsulated into concise high-confidence anomaly alert messages for immediate transmission. After the lightweight on-board identification model infers the raw image data stream, it generates an initial confidence level C1 for each identified target, along with its corresponding location information P, such as bounding box coordinates or center point latitude and longitude, and facility type identification result T1, such as "transmission tower anomaly" or "substation oil leak." First, it is determined whether the initial confidence level C1 of the current target is greater than the first confidence threshold R1. If it meets the threshold... The conditions indicate that the identification result has a very high degree of credibility and is likely a real anomaly requiring an urgent response.

[0055] At this point, the precise location information P of the target and the facility type identification result T1 are immediately extracted and integrated into a structured, lightweight, high-confidence anomaly alert message. This message does not contain raw image data, but only transmits necessary decision-making information, such as: "Satellite Observation ID: [Serial Number], Geographic Location: [Latitude and Longitude], Facility Type: [Transmission Tower], Anomaly Type: [Tower Tilt], Detection Time: [UTC Time], Confidence Level: [C1]". This design ensures the immediacy and efficiency of information transmission, buying valuable time for ground dispatch.

[0056] The original image regions corresponding to intermediate recognition results that are not confident enough but not so much that they should be discarded are precisely cropped and packaged into low-confidence test image data packages for further ground processing. When a target is identified on-board and its confidence C1 is between the second confidence threshold R2 and the first confidence threshold R1, then the condition is satisfied. The conditions were met, and these results were considered to have potential value, but their confidence level was not yet high enough to issue an alarm directly. They require verification by a high-precision ground-based validation model or manual verification.

[0057] For this type of identification result, instead of immediately sending an alert, the system traces back to the original image data stream and, based on the previously identified target location information P, precisely crops a small image patch containing the target and its surrounding environment from the original remote sensing image. The cropped area is typically a fixed-size region, such as one based on the target's center. Pixel blocks or Pixel blocks, ensuring they contain sufficient contextual information.

[0058] Subsequently, the cropped image data, along with corresponding preliminary identification information such as target location P, facility type identification result T1, and confidence level C1, are packaged into a single low-confidence image data packet. These packets are then transmitted to the ground station via a broadband channel, providing foundational data for subsequent detailed analysis and model learning. This process optimizes bandwidth utilization, avoids transmitting large amounts of high-confidence images, and preserves image evidence of potentially important information.

[0059] One application example is to pre-set and obtain the first confidence threshold. Second confidence threshold In an on-orbit inference mission involving a power transmission line, the onboard lightweight identification model identified two suspected anomalous targets. The initial confidence level of target A was... Initial confidence level of target B .

[0060] For objective A, since the condition is satisfied... It was identified as a highly reliable emergency anomaly, and its "tower tilt" identification result and geographical location latitude and longitude were extracted and packaged into a lightweight, high-confidence anomaly alarm message without image data for immediate transmission via the telemetry channel; for target B, since the conditions are met... It was determined that it had potential value but needed verification, so it was precisely cropped from the original image based on its location information. Local slices of pixels, along with their preliminary identification information, are packaged to generate low-confidence image data packets to be verified. Through differentiated processing, the utilization rate of satellite communication bandwidth is optimized while ensuring that no key anomalies are missed.

[0061] The differential identification results data are transmitted to the ground station via satellite communication link, including: High-confidence anomaly alarm information is transmitted in real time via the satellite's real-time telemetry channel; Low-confidence image data packets to be verified are downloaded non-real-time via the satellite's broadband data transmission channel.

[0062] The differential identification results are transmitted to the ground station via satellite communication link. Based on the data's real-time nature, importance, and volume, the optimal satellite communication channel and transmission strategy are selected to ensure the immediate delivery of high-priority anomaly alerts and the reliable transmission of low-priority images awaiting verification, while maximizing the utilization efficiency of satellite communication resources. This process involves the separate transmission of two different types of data.

[0063] It is essential to ensure that high-confidence anomaly alerts are transmitted from the satellite platform to the ground station with minimal latency, enabling rapid response to anomalies. Once the onboard hierarchical processing module generates a high-confidence anomaly alert, it is treated as the highest priority data. This information is typically highly compressed structured text or a small amount of binary data, with a very small data volume, for example, each alert is between a few hundred bytes and a few thousand bytes. The satellite communication link's real-time telemetry channel is dedicated to this purpose. Telemetry channels typically have low bandwidth, for example, hundreds of bits per second to tens of kilobits per second, but they are characterized by rapid link establishment, extremely low transmission latency, and high reliability, and are commonly used for satellite status monitoring and emergency command transmission.

[0064] The system triggers a pre-defined downlink task, encapsulating each generated high-confidence anomaly alarm message into a telemetry data packet, which is then modulated by the onboard telemetry unit and sent to the downlink real-time telemetry channel. These data packets undergo error-correcting coding, such as Reed-Solomon coding, to cope with the complex space communication environment.

[0065] The telemetry receiver at the ground station continuously listens to and decodes these data streams. Once it receives an abnormal alarm message, it will immediately analyze and process it. Its end-to-end latency is usually controlled within a few seconds to tens of seconds.

[0066] During communication, it is necessary to efficiently and reliably transmit large amounts of low-confidence image data packets from the satellite platform to the ground station for subsequent detailed analysis and model updates, while avoiding the consumption of scarce real-time communication resources. Unlike high-confidence anomaly alarms, low-confidence image data packets, although important, have relatively lower real-time requirements, and each data packet is large in size; for example, each cropped image can reach hundreds of kilobytes to several megabytes.

[0067] Therefore, the satellite's broadband data transmission channel will be selected for transmission. Broadband data transmission channels are designed specifically for large data volume transmission and have high bandwidth, for example, tens to hundreds of megabits per second. However, there may be some delay in link establishment and data queuing processing, so transmission is typically performed in batches when the satellite passes over the ground station. The onboard data transmission unit will store, compress, and encrypt the collected low-confidence image data packets, such as using lossless or near-lossless JPEG2000 compression, with compression ratios reaching 5:1 to 20:1, and place them into the transmission queue according to a predetermined transmission plan.

[0068] When the satellite passes through the communication window above the ground station, the onboard data transmission unit activates the broadband data transmission link and transmits the data in batches to the ground station via a high-gain antenna at a preset data transmission rate. This non-real-time transmission strategy effectively utilizes the advantages of broadband data transmission, reduces the unit data transmission cost, and ensures the reliable recovery of large amounts of image data.

[0069] One application example is generating a high-confidence anomaly alarm message about an oil leak in a substation. The data size is only 500 bytes. It is then encapsulated into a telemetry data packet and error-corrected using Reed-Solomon codes. The data is then transmitted down through a fast and low-latency real-time telemetry channel established via a link, ensuring that the ground station receives and responds to the disaster within 30 seconds. Meanwhile, for suspected inspection targets with low confidence, a 20MB image data packet is generated for inspection. This packet is then compressed using the JPEG2000 algorithm at a compression ratio of 10:1, resulting in near-lossless compression and reducing the actual transmitted data size to [a smaller value]. The data is stored in a queue to be transmitted until the satellite flies over the ground station and is within a communication window. Then, it is used to perform batch non-real-time downlinking using a broadband data transmission channel with a bandwidth of up to 100Mbps. Through this differentiated transmission strategy based on data priority, the efficiency of satellite communication resources is maximized while ensuring that emergency alarms are delivered in a timely manner.

[0070] The specific verification results reported by the ground station are as follows: The ground station receives low-confidence image data packets and inputs them into a high-precision ground verification model for fine identification to obtain corrected labels. The low-confidence image data to be verified and its corresponding correction labels are packaged together to generate a mini incremental training sample package as the verification result.

[0071] The low-confidence results of the on-board recognition model are meticulously reviewed to obtain high-quality, accurate corrected labels. These corrected labels are combined with the corresponding image data packets to form a micro-incremental training sample package, which is specifically used for subsequent on-orbit incremental learning of the on-board model, thereby continuously improving the performance and reliability of the lightweight on-board recognition model. This process involves the ground station receiving low-confidence data packets, high-precision recognition, and the final sample package generation.

[0072] The ground station receives low-confidence imagery data packets from the satellite platform and uses a high-precision verification model deployed on the ground to perform fine-grained identification of these images, generating accurate correction labels. The ground station receives all low-confidence imagery data packets from the satellite platform via a broadband data transmission channel. Each data packet contains one or more cropped remote sensing images of power facilities, along with the initial confidence level, location information, and preliminary identification type generated during onboard identification. Once reception is complete and passes data integrity verification, this imagery data is immediately sent to the ground-based high-precision verification model pre-deployed on the ground server.

[0073] The ground-based high-precision verification model is a large-scale deep neural network model deployed on a GPU cluster. It has a huge number of parameters, such as hundreds of millions or even billions of parameters. Its network structure and computational complexity are not limited by spaceborne resources, thus it has extremely high recognition accuracy and robustness.

[0074] The ground-based high-precision validation model performs comprehensive reasoning analysis on each low-confidence image to be verified, outputting more accurate facility type identification results and fault status judgments than the satellite-based model. For example, if the satellite-based model only gives a preliminary judgment of "tower anomaly," the ground-based model may accurately identify it as "tower beam deformation" or "tower top corrosion," and generate corresponding precise confidence levels. These accurate and highly confident identification results output by the ground-based high-precision validation model are considered as "correction labels" for the image.

[0075] Correction labels may include new target categories, more refined anomaly types, corrected bounding box coordinates, and negative corrections to the original on-board identification results. For example, if the ground model determines that there are no anomalies in the area, it corrects the false alarms of the on-board model.

[0076] To further improve the reliability of the corrected labels, human experts will be introduced for review in some scenarios, especially when the confidence level of the ground high-precision verification model itself is not high or when business rules require it. The corrected labels obtained from the ground high-precision verification model are effectively bundled with their corresponding original low-confidence image data packets to generate structured micro-incremental training sample packets, which will facilitate subsequent on-orbit learning by the satellite.

[0077] After obtaining the corrected label for each low-confidence image to be examined, the ground data processing system encapsulates the corrected label with the original received low-confidence image data. The corrected label includes, but is not limited to, detailed information such as target category, state, and bounding box. The micro-incremental training sample bag structure is defined as tuple M: ; Here, I1 represents one or more original image datasets used for incremental training. These datasets are images extracted from low-confidence test image datasets. L represents the corrected labels generated by a ground-based high-precision validation model or human experts, precisely corresponding to I1. L is a structure containing multiple fields, such as classification labels. Its values ​​are derived from the precise categories identified by the high-precision ground-based validation model, which may cover from 0 to hundreds of different categories, as well as bounding boxes. , which represents the precise pixel coordinates of the target in the image.

[0078] The generation of micro-incremental training sample packs follows rigorous data association and consistency checks to ensure that each image has its correct and unique correction label. For example, the size of a single sample pack may range from tens of KB to several MB, depending on the image size and the complexity of the correction label. After these sample packs are accumulated, they will be sent back to the satellite platform according to a predetermined transmission plan, serving as the core data input for the onboard lightweight recognition model to perform incremental learning.

[0079] One application example is that a ground station receives a low-confidence image data packet containing two cropped remote sensing images of power facilities and inputs it into a high-precision ground-based validation model with billions of parameters for deep analysis. For one image initially labeled as "tower anomaly" by the onboard model, the ground model precisely identifies it as "tower beam deformation" and generates a correction label L containing the accurate category and the corrected bounding box coordinates. Subsequently, the original image data I1 and the corresponding correction label L are encapsulated into a structured micro-incremental training sample package M. If this sample package contains two images... Given an image of pixels, approximately 192KB per image, and a total of 2KB of tag data, the total data size of tuple M is calculated as follows: The generated sample packages undergo rigorous data association and consistency checks to ensure that each image corresponds to a unique correction result. After accumulation, they are sent back to the satellite platform according to the transmission plan.

[0080] In-orbit incremental learning of the lightweight on-board recognition model using the validation results includes: Satellite receives mini-incremental training sample packets; On the onboard processor, the current lightweight recognition model is trained using micro-incremental training sample packages, thereby improving the lightweight recognition model's ability to distinguish scenes related to low-confidence image data packets. Update the on-board lightweight recognition model with the new parameters after training.

[0081] By receiving new data samples that have been precisely corrected by the ground station, the lightweight recognition model already deployed on the satellite is dynamically optimized, enabling it to better adapt to new recognition scenarios and correct past recognition errors. This continuously improves the model's ability to discriminate complex and abnormal scenarios, rather than redeploying the onboard model or performing global training. This process involves receiving micro-incremental training sample packets, incremental training on the onboard processor, and updating model parameters. The satellite platform can accurately and securely receive the micro-incremental training sample packets transmitted from the ground station, providing high-quality data input for subsequent on-orbit learning.

[0082] Once the ground station has completed the detailed identification of low-confidence images to be examined and generated micro-incremental training sample packets, these sample packets are uploaded to the satellite platform via the satellite's broadband data transmission channel or dedicated uplink during the communication window when the satellite flies over the ground station. The data receiving and processing unit in the satellite payload is responsible for receiving this uplink data and performing integrity verification, error detection, and decryption. To prevent data corruption or tampering during transmission, Cyclic Redundancy Check (CRC) and encryption algorithms are typically used to ensure that the data is reliable and secure before unpacking.

[0083] The received sample packets are stored in the persistent storage units of the onboard processor, such as high-reliability solid-state drives (NAND Flash), awaiting invocation by the model update module. A single micro-incremental training sample packet may contain 1 to several thousand images with correction labels, with the total data volume typically controlled within tens to hundreds of MB to accommodate uplink transmission bandwidth and onboard storage capabilities.

[0084] By utilizing the received micro-incremental training sample packages, and with limited onboard computing resources, the current lightweight onboard recognition model is efficiently optimized locally to improve its recognition accuracy for previously low-confidence scenarios. The lightweight incremental training task is initiated on the satellite's onboard processor, typically a System-on-Chip (SoC) chip equipped with an FPGA or ASIC accelerator.

[0085] Unlike the large-scale GPU clusters required for training terrestrial models, spaceborne processors have severely limited memory and computing power. Therefore, incremental training does not involve retraining the entire model, but rather employing specific optimization strategies. A common approach is fine-tuning in "transfer learning," which involves updating only a small number of parameters in the top or specific layers of the model, while keeping most of the parameters in the lower-level feature extraction layers unchanged.

[0086] During training, the images in each received mini-incremental training sample packet are used as input X, and the corresponding corrected labels are used as target Y. The training algorithm employs mini-batch gradient descent, and the loss function is typically the corrected cross-entropy loss. This is used to measure the difference between the model output V and the corrected label Y: ; in, This represents the true probability of the j-th category in the corrected label, usually 0 or 1. This represents the probability that the model predicts for the j-th class. Cross-entropy loss. Real-time computation is performed on the onboard processor, utilizing either the Adam optimizer or the SGD optimizer with a very small learning rate. Update the model weights. Y represents the corrected label corresponding to X, which is usually a one-hot encoded vector.

[0087] Incremental training cycles typically last from several to tens of epochs to avoid overfitting to a small number of new samples. This process can be triggered by receiving a certain number of sample packets or by periodically scheduled tasks. The optimized model parameters obtained from incremental training are then deployed to the runtime environment of the lightweight on-board recognition model, making them effective immediately and thus giving the model stronger recognition capabilities.

[0088] Once the onboard processor completes incremental training of the current lightweight onboard recognition model, it generates a new set of optimized model parameters. Perform a parameter update operation to update the current onboard lightweight recognition model. old parameters Replace with Updates typically involve loading the model weights file stored in persistent storage into memory, overwriting it with the new parameters, restarting the model inference service, or hot-loading the new parameters.

[0089] To ensure the stability and reliability of the update, a brief internal validation of the new parameters may be performed before the update, or a dual-model backup strategy may be adopted, with the old model continuing to run until the new model passes validation. Once the update is successful, all onboard real-time inference will use this updated model. This process improves the model's accuracy in distinguishing low-confidence scenarios, long-tailed anomalies, and changes in the on-orbit environment, and is a key step in achieving adaptive and continuous evolution of satellite artificial intelligence models.

[0090] One application example is that during the communication window when the satellite is flying over the ground station, it receives a set of micro-incremental training sample packets containing 1,000 images to be examined via a broadband data transmission channel. After CRC verification and decryption, the packets are stored in the onboard NAND Flash. Subsequently, the onboard SoC chip initiates a lightweight incremental training task, using a fine-tuning strategy to optimize the top-level parameters of the model.

[0091] During training, the modified cross-entropy loss function is used. Real-time computation of the difference between the predicted and corrected labels by the model. For a certain image of power facility anomalies, the corrected label Y is a one-hot encoded vector. Wherein represents the true probability of the "abnormal" category. The current inference output probability V of the model is The single-sample loss is then calculated as follows: The Adam optimizer was used with an extremely small learning rate of 0.0001. The model weights are updated iteratively, and new model parameters are generated after training for dozens of epochs. By using a "dual-model backup" strategy, old parameters are... Silent replacement allows the updated model to more accurately identify power facility disaster scenarios that previously had low confidence levels.

[0092] The method in this embodiment is executed iteratively. As the number of iterations increases, the accuracy of the on-board lightweight recognition model in recognizing complex and abnormal scenes continues to improve, resulting in a decrease in the proportion of low-confidence image data packets to be verified.

[0093] As the number of iterations increases, the accuracy of the on-board lightweight recognition model in identifying complex and abnormal scenes continues to improve, leading to a decrease in the proportion of low-confidence image data packets. By designing and implementing a closed-loop, adaptive learning mechanism, the on-board lightweight recognition model can continuously improve itself after deployment without frequent manual intervention or large-scale retraining and redeployment of ground models, ultimately enhancing the intelligence level and autonomous operation capability of the on-board AI system. The core of this process is the synergistic effect of "on-orbit incremental learning" and "feedback loop".

[0094] By establishing a continuous adaptive learning loop, the lightweight on-board recognition model is ensured to continuously learn from new data and complex scenarios encountered in actual operation, correcting its own recognition biases. The entire on-orbit recognition method for power facilities, from the acquisition of remote sensing images, on-board inference, result classification, data downlink, to ground verification and generation of corrected labels, is ultimately fed back to the on-board model in the form of micro-incremental training sample packages for incremental learning, forming a complete closed loop.

[0095] This method is not a one-time deployment and remains unchanged, but rather a series of steps executed periodically or automatically or semi-automatically when conditions are met. Each incremental learning and parameter update of the model in orbit constitutes an "iteration." In each iteration, the updated lightweight on-board recognition model processes subsequent real-time remote sensing imagery, and based on its improved recognition capabilities, the identification results for anomalies will be more accurate. Through the on-board incremental learning mechanism, the ability of the lightweight on-board recognition model to handle diverse and complex anomaly scenarios is gradually improved, thereby reducing the uncertainty of the model itself.

[0096] When the onboard lightweight recognition model receives micro-incremental training sample packages through incremental learning, which include "correction labels" given by ground-based high-precision validation models or human experts for the original "low-confidence" images, the onboard processor uses these samples to adjust local parameters. For example, if the model previously frequently misidentified "small bird nests" as "tower top corrosion," and these false alarms were labeled as low-confidence, after ground correction, the correction label will clearly indicate that it is a "normal small bird nest." When these samples with corrected labels are fed back to the satellite for incremental learning, the model uses a new loss function to adjust the relevant weights, enabling the model to more accurately identify the true category of features similar to "small bird nests" and no longer incorrectly classify them as an anomaly.

[0097] With the cumulative learning of a large number of such complex or ambiguous samples, the weight parameters of the on-board lightweight recognition model... From the initial state Gradually evolved into , where n is the number of iterations, and each iteration makes the model converge in a more accurate direction.

[0098] The effect of quantifying on-orbit incremental learning, that is, observing the improvement in the recognition capability of the lightweight recognition model on the satellite, is directly reflected in the reduction of the amount of low-confidence image data to be verified by ground station intervention over time.

[0099] In the initial stage, due to the potential limitations of the on-board lightweight recognition model in identifying certain uncommon or boundary cases, it may produce a large number of results with moderate confidence. The identification results within the specified range are cropped and packaged into low-confidence pending image data packets for transmission to the ground station. However, as iterative learning progresses, the model learns a large number of previously blurred scenes, significantly improving its ability to distinguish these scenes. This means that the model's output confidence level C1 will gradually increase for the identification of equally complex scenes. This increase in confidence leads to two changes: Some identification results that originally had confidence levels between R2 and R1 will now have a confidence level C1 that exceeds R1, thus being directly identified as high-confidence anomaly alarm messages and no longer needing to be transmitted as images to be examined. Other identification results that originally had confidence levels between R2 and R1, after model learning, may be more accurately identified as normal, and their confidence level will also be improved, no longer meeting the criteria for low-confidence images to be examined.

[0100] Therefore, through continuous iteration, the on-board lightweight recognition model can autonomously handle more anomalies and complex scenarios, resulting in a gradual decrease in the number of low-confidence image data packets requiring ground verification. This downward trend can be quantified as the percentage or absolute number of low-confidence image data packets generated in each iteration relative to the total number of images processed. For example, initially this percentage might reach 20%-30%, but after dozens of iterations, it may stabilize at 5%-10%, or even lower, indicating that the model has become highly autonomous and accurate, significantly reducing ground processing load and communication bandwidth requirements.

[0101] One application example is the use of an adaptive learning loop to optimize the weight parameters of the lightweight on-board recognition model. From the initial state As the number of iterations n gradually evolves into In the initial iteration phase (n=1), due to the model's insufficient ability to distinguish complex scenes such as "small bird's nests," its recognition confidence C1 often falls within the threshold range. Within this context, assuming that processing 1000 power facilities generates 250 low-confidence image data packets, the data volume percentage is as follows: After dozens of iterations of incremental learning, the model accurately identified "small bird nests" as normal situations, and the confidence level C1 significantly improved, exceeding R1. At this point, processing the same 1000 power facilities only generated 60 low-confidence image data packets, reducing the data volume ratio to a much smaller percentage. The downward trend indicates that the model's accuracy in handling complex and abnormal scenarios continues to improve, reducing the ground processing load and communication bandwidth requirements. Figure 3 The study demonstrates the increasing accuracy of the on-board lightweight recognition model in complex scenes as the number of on-orbit incremental learning iterations increases, as well as the decreasing proportion of corresponding low-confidence test image data packets in the total data volume.

[0102] Example 2: An on-orbit identification system for power facilities, used to implement the method in Example 1, comprising: The image acquisition module is used to acquire remote sensing images of power facilities in the target area in real time collected by the satellite platform and generate raw image data streams; The intelligent analysis module is used to input the raw image data stream into the on-board lightweight recognition model for real-time inference, and generate on-board recognition results that include the location information of the target power facility, the facility type recognition result, and the initial confidence level. The hierarchical processing module is used to perform hierarchical processing on the on-board recognition results according to the initial confidence level, and generate differentiated recognition result data containing high-confidence anomaly alarm information and low-confidence pending image data packets. The communication and transmission module is used to transmit the differential identification result data to the ground station via a satellite communication link; The feedback receiving module is used to receive the verification results fed back by the ground station based on the differential identification result data. The verification results are generated by the ground station after verifying the low-confidence image data packets to be verified, and include correction labels for the low-confidence image data packets to be verified. The model update module is used to perform on-orbit incremental learning of the on-board lightweight recognition model using the validation results, and update the on-board lightweight recognition model.

[0103] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can make equivalent substitutions or modifications based on the technical solution and concept of the present invention within the scope of the technology disclosed in the present invention, and such modifications should also be considered to fall within the scope of protection of the present invention.

Claims

1. A method for on-orbit identification of power facilities, characterized in that the steps include... include: Acquire remote sensing images of power facilities in the target area in real time from the satellite platform and generate raw image data stream; The raw image data stream is input into the on-board lightweight recognition model for real-time inference, generating on-board recognition results that include the location information of the target power facility, the facility type recognition result, and the initial confidence level. The on-board identification results are graded based on the initial confidence level to generate differentiated identification result data containing high-confidence anomaly alarm information and low-confidence image data packets to be verified. The grading process based on the initial confidence level includes: obtaining a first confidence threshold and a second confidence threshold to distinguish between high-confidence and low-confidence identification results, wherein the first confidence threshold is higher than the second confidence threshold; for on-board identification results with an initial confidence level higher than the first confidence threshold, extracting their location information and facility type identification results, and encapsulating them into high-confidence anomaly alarm information; and cropping the original image regions corresponding to on-board identification results with an initial confidence level between the first and second confidence thresholds to generate low-confidence image data packets to be verified. The differential identification result data is transmitted to the ground station via a satellite communication link. This transmission includes: transmitting high-confidence anomaly alarm information in real time via the satellite's real-time telemetry channel; and transmitting low-confidence image data packets to be verified non-real-time via the satellite's broadband data transmission channel. The system receives verification results from ground stations based on differentiated identification data. These verification results are generated by the ground stations after verifying low-confidence image data packets and include corrected labels for the low-confidence image data packets. Specifically, the verification results fed back by the ground stations are as follows: the ground stations receive low-confidence image data packets and input them into a high-precision ground verification model for fine identification to obtain corrected labels; the low-confidence image data packets and their corresponding corrected labels are packaged together to generate a micro-incremental training sample package as the verification results. The on-orbit lightweight recognition model is updated by using the validation results for in-orbit incremental learning. This process includes: the satellite receiving micro-incremental training sample packets; using these samples on the onboard processor to perform lightweight incremental training on the current on-orbit lightweight recognition model, thereby improving its ability to distinguish scenes related to low-confidence test image data packets; and updating the model with the new trained parameters.

2. The on-orbit identification method for power facilities according to claim 1, characterized in that, The acquisition of real-time remote sensing images of power facilities in the target area by the satellite platform is triggered by the onboard mission scheduling unit. When the satellite flies over the power facility monitoring area or receives an emergency observation command sent from the ground, it automatically controls the satellite payload to image the target area.

3. The on-orbit identification method for power facilities according to claim 1, characterized in that, The deployment process for initializing the lightweight recognition model is as follows: Obtain a historical remote sensing sample set of power facilities containing sample images labeled with normal and several abnormal states; A ground-based high-precision verification model was trained using a historical remote sensing sample set of power facilities. The ground-based high-precision verification model is a deep neural network model deployed on a ground server. Its network structure and parameter scale are not limited by spaceborne resources. It is used to perform high-precision verification and identification of low-confidence image data packets to be verified. Using knowledge distillation technology, a simplified, lightweight on-board recognition model is derived by using a high-precision ground-based verification model as the teacher model, and then deployed on the satellite platform.

4. The on-orbit identification method for power facilities according to claim 3, characterized in that, The lightweight onboard recognition model is a pruned and optimized convolutional neural network model, whose model size and computational complexity are adapted to the memory and computing power constraints of the onboard processor.

Images