Intelligent fusion terminal acquisition success rate prediction system based on topology intelligent perception

Abstract

The invention discloses an intelligent fusion terminal acquisition success rate prediction system based on topology intelligent perception. The system is composed of a master station and an intelligent fusion terminal. A carrier APP container, a message queue module, a data center, an AI-APP container, a collection APP container, a display APP container and an upgrade management and data reporting module are arranged on the intelligent fusion terminal side; the carrier wave APP container obtains topology and operation characteristics related to an acquisition link, the topology and operation characteristics are sent to a data center through a message queue, the AI-APP container grabs characteristic data for forward reasoning, five-level acquisition success rate levels and confidence coefficients are output, and the acquisition APP container formulates an issuing strategy and displays the strategy according to the five-level acquisition success rate levels and the confidence coefficients. The master station side comprises a differential upgrading service, a training framework, an ONNX converter and a mirror image warehouse, the differential upgrading service, the training framework, the ONNX converter and the mirror image warehouse generate updating content through ONNX conversion and publish the updating content to the mirror image warehouse, and the intelligent fusion terminal is pulled by the upgrading management module and updated online. According to the system, real-time prediction and closed-loop optimization of the acquisition success rate are realized in a resource-constrained environment, the time delay is reduced, and the operation and maintenance efficiency is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of low-voltage power user power information collection, in particular to a kind of intelligent fusion terminal acquisition success rate prediction system based on topology intelligent perception, the system can be widely applied to smart grid, industrial internet of things, wisdom building etc. Scene, by real-time perception field topology information, millisecond level success rate evaluation is completed at the side of intelligent fusion terminal, and according to this instant adjustment deployment position or optimization collection strategy, significantly reduce manpower waste and operation cost. BACKGROUND

[0002] The statements in this section merely provide background information related to the present disclosure and can not constitute prior art.

[0003] With the continuous rise of low-voltage area user number, load type is increasingly diverse, and topology structure is increasingly complex. After the installation of power intelligent fusion terminal, it often has to go through the "meter waiting" stage. Only when the cumulative data reaches the main station statistical threshold, the real acquisition success rate can be obtained. Once the success rate is lower than the threshold set by the power grid company, the operation and maintenance personnel need to reselect the site or replace the hardware. The process involves multiple links such as vehicle round trip, climbing operation, user power failure, work order approval, etc. The cost of manpower, material resources and time is significantly magnified.

[0004] At the same time, the existing success rate calculation is completely completed by the main station remote end: the field intelligent fusion terminal needs to return the high frequency and large volume data such as original carrier frame, impedance scanning curve, event message through 4G, NB-IoT or optical fiber; due to network jitter, link congestion and package flow limit, the data integrity and real-time performance cannot be guaranteed, and the feedback cycle is often extended from several hours to several days. The above-mentioned "long link" mode causes that the field personnel cannot know the collection effect in time, and cannot complete the "prediction, adjustment and confirmation" closed loop in the installation moment, thereby causing a large amount of invalid on-site and repeated construction.

[0005] Therefore, there is an urgent need for a lightweight technology that can directly complete "millisecond-level prediction and second-level response" at the local intelligent fusion terminal: without returning sensitive data, it can continuously evolve with the field environment, thereby completely solving the "meter waiting" pain point, significantly reducing the operation and maintenance cost, and improving the intelligent level of power grid. SUMMARY

[0006] The purpose of the present application is to solve the above-mentioned pain points by providing an intelligent fusion terminal acquisition success rate prediction system based on topology intelligent perception.

[0007] The core innovation of the present application is embodied in three technical features of real-time topology perception-prediction closed loop, containerized one-key deployment and model weight self-adaptive evolution, which realizes millisecond-level local prediction, second-level field decision, cross-platform rapid migration and life cycle continuous evolution.

[0008] The application is based on AI computing capability, and a real-time topology perception-prediction closed loop compresses a long link of traditional "hanging table waiting, manual return, and cloud computing" into a short link of "on-site collection, local reasoning, and immediate decision", so that the deployment personnel can know a five-level collection success rate level at the moment of power-on, and immediately adjust the installation position or relay distribution point on this basis.

[0009] Meanwhile, the application compresses a traditional deployment process into an agile link of "pulling through a container image" by means of containerized packaging, so that the operation and maintenance personnel can complete model upgrading or capacity expansion on any terminal without modifying the kernel to realize global synchronization.

[0010] The application optimizes the model weight upgrading mode, and compresses a traditional static deployment mode into an online closed loop of micro-weight return and hot replacement through weight evolution, so that the model can be continuously optimized with environmental changes in the whole life cycle to realize adaptive weight updating.

[0011] The technical scheme of the application is as follows: An intelligent fusion terminal collection success rate prediction system based on topology intelligent perception, comprising: a master station and at least one intelligent fusion terminal, wherein the intelligent fusion terminal is in communication connection with the master station; The intelligent fusion terminal comprises: A carrier APP container acquires topology information and operating characteristics related to the intelligent fusion terminal collection terminal communication link from the on-site low-voltage power distribution network according to the downlink strategy of the collection APP container, generates feature data, and encapsulates the feature data into a feature message and publishes it to a message queue; A message queue module is used for asynchronous buffering and decoupling transmission of the feature message; A data center is used for buffering and managing the feature data transmitted by the message queue module, and providing data services to the downstream; An AI-APP container is embedded with a general inference engine and loads a prediction model trained by the master station, lightened and converted by ONNX, performs forward inference on the feature data from the data center, outputs a five-level collection success rate level and its confidence, and synchronizes the results to the data center, the collection APP container, and the display APP container; The collection APP container is used for determining the downlink strategy of the collection instruction based on the five-level collection success rate level and its confidence output by the AI-APP container, and coordinating the carrier APP container to perform collection; The display APP container is used for presenting the five-level collection success rate level and its confidence information on the intelligent fusion terminal side; An upgrade management module is configured to pull the update content published by the master station from the mirror repository and complete the online update of the prediction model used by the AI-APP container without interrupting the service; A data reporting module is configured to return the incremental sample carrying the prediction result label and the corresponding topology feature to the master station after encryption; The master station comprises: A differential upgrade service is configured to receive the incremental data returned by the intelligent fusion terminal, process the incremental data, generate the differential update content relative to the existing deployment version based on the output of the training framework and the ONNX converter, and publish the update content to the mirror repository; A training framework is configured to train and fine-tune the prediction model based on the incremental data provided by the differential upgrade service and output the update weight; An ONNX converter is configured to convert the update weight or model output by the training framework into a general intermediate format for end-side inference; A mirror repository is configured to store the ONNX-converted model or container image containing the model and provide a pull service to the intelligent fusion terminal.

[0012] Further, the topology information and operating features include: One or more of node level, line impedance, signal-to-noise ratio, voltage fluctuation, and load type.

[0013] Further, the delivery strategy includes: Whether to deliver, delivery timing, and parameter settings.

[0014] Further, the prediction model is subjected to lightweight processing on the master station side, including: Performing structured pruning and dynamic INT8 quantization on the trained network to generate a candidate model with a volume less than a set threshold.

[0015] Further, the general inference engine embedded in the AI-APP container can execute the ONNX-converted model on the processor or neural network unit without modifying the firmware of the intelligent fusion terminal.

[0016] Further, the AI-APP container, general inference engine, ONNX-converted model, and its configuration file are packaged as a container image for deployment, and the container image is trimmed based on the base image to reduce the image size.

[0017] Further, the message queue module and the data center jointly realize decoupling of data collection and inference processing, adopt an asynchronous queue mechanism for data transmission, and the data center provides the data service to the AI-APP container to form an asynchronous pipeline.

[0018] Further, the processing flow of the incremental data and the update is as follows: The incremental data returned by the intelligent fusion terminal is received by the differential upgrade service and provided to the training framework, the training framework outputs the updated weight and is converted into a general intermediate format by the ONNX converter, the differential upgrade service generates the update content based on the conversion result and publishes it to the mirror repository, the intelligent fusion terminal pulls the update content from the mirror repository to complete the model update by the upgrade management module, and the AI-APP container enables the updated model.

[0019] Further, the model training on the master station side includes: On the basis of the publicly released low-voltage topology data set, hierarchical sampling is performed according to the voltage level, cable material and user density of the target area, a small-scale sample matching the field working condition is selected to construct an initial training set, and the tensor size is constrained by the upper limit of the embedded device loading memory at one time and not triggering the exchange partition; Perform feature engineering on the initial training set, perform correlation test and mutual information screening to remove redundancy and retain features with information gain, combine L1 regularization to compress the dimension, and obtain a low-dimensional input space, which at least includes one or more of voltage level, cable material, user density, node level, impedance, signal-to-noise ratio and load type; Train the lightweight convolutional neural network and the graph neural network in parallel with shared input, so that each network independently outputs a five-level success rate probability, the training adopts a momentum adaptive learning rate strategy and sets an early stopping mechanism, and when there is no improvement for several consecutive rounds, the optimal weight is rolled back; After training, the network is structurally pruned and dynamically quantized, preferably INT8 quantization, a candidate model is generated, and a deployment version is determined by voting or weighted fusion to meet the millisecond-level inference requirements on the terminal side.

[0020] Further, the display APP container is used to present the five-level acquisition success rate level and its confidence information in the intelligent fusion terminal touch screen or mobile terminal in communication connection with the intelligent fusion terminal.

[0021] Compared with the existing technology, the beneficial effects of the present application are: 1. Extremely Fast Response Speed: This invention supports real-time acquisition of low-voltage topology information from the intelligent fusion terminal. Through rapid inference using a lightweight AI model, it instantly outputs five levels of acquisition success rate (A, B, C, D, E). After structured pruning, dynamic INT8 quantization, and operator-level fusion optimization, the AI ​​model performing the prediction function is compressed to less than 50 MB, with inference latency controlled in milliseconds. In traditional solutions, after deploying the intelligent fusion terminal to the site, installers still need to run the meters for several days and upload data to the main station for calculation before obtaining the acquisition success rate, a lengthy process. In contrast, this invention, based on the concept of real-time side-end computation, triggers a "prediction, adjustment, and confirmation" closed loop the instant the intelligent fusion terminal completes deployment and power-on, reducing on-site decision-making time from "days" to "seconds." Simultaneously, the lightweight design allows the model to run flexibly in a dedicated onboard NPU computing unit, with core tasks isolated in an independent sandbox, maintaining isolation from the original main business threads and ensuring no disruption to existing functions such as meter reading and event reporting.

[0022] 2. Simplified Deployment and High Cross-Platform Portability: This invention encapsulates the success rate inference app and all its runtime dependencies, such as model parameter files, forward computation graphs, and inference engine dependencies, into a single standard container image (Docker). It also provides multiple hardware executable layers, including ARMv7. In addition to smart converged terminals, the same image, supported by multiple architecture manifests, can run directly on heterogeneous hardware platforms such as converged terminals, smart circuit breakers, edge IoT agents, and smart converged terminals in distribution centers, achieving "one-stop development for multiple uses." When the image starts, the host machine's CPU architecture automatically selects the corresponding layer, eliminating the need for recompilation or manual adaptation. During on-site deployment and power-on, the operating system only needs to automatically execute a few container-based startup commands on the target hardware to complete the deployment within 30 seconds. In contrast, existing technologies struggle to achieve one-click migration and large-scale deployment of inference capabilities across heterogeneous terminals. Traditional non-containerized AI models are typically delivered as binary executables or dynamic libraries, and their dependency chains vary significantly depending on the hardware platform. Furthermore, non-container solutions lack a unified version management and differential upgrade mechanism; any updates to model weights, dependency libraries, or configuration files must be remotely upgraded or the entire package replaced.

[0023] 3. Strong Model Adaptability: The AI ​​model described in this invention first undergoes initial training on the main station's x86 server based on a public dataset to form an initial weight file with general features, achieving an acceptable prediction baseline upon initial deployment. After the intelligent fusion terminal is put into operation, it continuously collects incremental features such as local topology, impedance, signal-to-noise ratio, and actual acquisition results, and transmits them back to the main station via an encrypted link. The main station uses the incremental data to perform transfer learning or incremental refitting on the original model, quickly generating a new weight file while maintaining the original knowledge. After digital signature and version verification, the newly generated file is redistributed to the designated path on the intelligent fusion terminal, completing a seamless hot replacement. This closed-loop process can be executed cyclically throughout the entire lifecycle of the intelligent fusion terminal. The model performance continuously optimizes with changes in the field environment, and the prediction accuracy shows an upward and converging trend, which is superior to the traditional static deployment method of "one-time training and permanent reset". Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments recorded in the embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0025] Figure 1 This is a swimlane diagram of the AI ​​workflow of this invention; Figure 2 This is a deployment diagram of the system of the present invention; Figure 3 This is a timing diagram of the system software execution of the present invention. Detailed Implementation

[0026] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0027] The features and performance of the present invention will be further described in detail below with reference to embodiments.

[0028] Example 1 Please seeFigure 1 First, it should be noted that the workflow of the AI ​​component in the intelligent fusion terminal acquisition success rate prediction system based on topological intelligent perception proposed in this embodiment consists of six steps: Initial dataset creation, feature engineering and selection, model training, model pruning, containerized deployment and porting, and transfer training, specifically as follows: Figure 1 As shown.

[0029] Please see Figure 2 and Figure 3 A system for predicting the success rate of data acquisition by an intelligent fusion terminal based on topological intelligent sensing includes: a main station and at least one intelligent fusion terminal, wherein the intelligent fusion terminal is communicatively connected to the main station; The intelligent fusion terminal includes: The carrier APP container, according to the distribution strategy of the acquisition APP container, acquires the topology information and operating characteristics related to the communication link between the intelligent fusion terminal acquisition terminal and the low-voltage power distribution network on site, generates feature data, and encapsulates the feature data into feature messages and publishes them to the message queue. The message queue module is used for asynchronous buffering and decoupled transmission of the feature messages; The data center is used to cache and manage the characteristic data transmitted by the message queue module and to provide data services to downstream users. The AI-APP container embeds a general-purpose inference engine and loads a prediction model trained by the main station, lightweighted, and transformed using ONNX. It captures feature data from the data center, performs millisecond-level forward inference, and outputs five levels of data acquisition success rate and their confidence scores. The results are synchronized to the data center, the acquisition APP container, and the display APP container. The display APP container is used to present the five levels of data acquisition success rate and their confidence scores as text prompts on the touch screen of the smart converged terminal or on a mobile terminal connected to the smart converged terminal. That is, the text prompts are presented instantly on the touch screen of the smart converged terminal or on the distribution network mobile terminal, allowing on-site personnel to intuitively understand the current data acquisition environment assessment conclusion at the moment of deployment. Furthermore, on-site personnel can manually choose whether to redeploy or optimize the strategy of the acquisition APP container based on the obtained prediction level. The acquisition APP container is used to determine the acquisition command issuance strategy based on the five-level acquisition success rate level and its confidence level output by the AI-APP container, and to coordinate the carrier APP container to execute acquisition. The display APP container is used to present five levels of data collection success rate and their confidence information on the intelligent converged terminal side; The upgrade management module is used to pull the updated content published by the main site from the image repository and complete the online update of the prediction model used by the AI-APP container without interrupting business operations; The data reporting module is used to encrypt and send incremental samples carrying prediction result labels and corresponding topological features back to the main station. The main station includes: The differential upgrade service is used to receive incremental data returned by the intelligent fusion terminal, process the incremental data, generate differential update content relative to the existing deployment version based on the output of the training framework and ONNX converter, and publish the update content to the mirror repository. The training framework is used to train and fine-tune the prediction model based on incremental data provided by the differential upgrade service, and output updated weights. The ONNX converter is used to convert the updated weights or model output by the training framework into a general intermediate format for edge inference. The image repository is used to store models converted by ONNX or container images containing the models, and to provide pull services to smart converged terminals.

[0030] In this embodiment, specifically, the topology information and operational characteristics include: One or more of the following: node level, line impedance, signal-to-noise ratio, voltage fluctuation, and load type; that is, after the intelligent fusion terminal is powered on, the carrier APP container starts up immediately, continuously scans the local low-voltage topology, gathers node level, path loss, signal-to-noise ratio, impedance, and load type in real time, and encapsulates the obtained features into JSON and other format messages, which are then pushed to the data center via message queue.

[0031] In this embodiment, specifically, the distribution strategy includes: Whether to issue the document, the timing of issuance, and parameter settings.

[0032] In this embodiment, specifically, the prediction model is lightweighted on the main station side, including: The trained network is subjected to structured pruning and dynamic INT8 quantization to generate candidate models with a volume smaller than a set threshold.

[0033] In this embodiment, specifically, the general inference engine embedded in the AI-APP container can execute the ONNX-converted model on the processor or neural network unit without modifying the intelligent fusion terminal firmware.

[0034] In this embodiment, specifically, the AI-APP container, the general inference engine, the model converted by ONNX and its configuration file are packaged into a container image for deployment, and the container image is trimmed from the base image to reduce the image size.

[0035] In this embodiment, specifically, the message queue module and the data center jointly decouple data acquisition and inference processing, and use an asynchronous queue mechanism for data transmission. The data center provides the data service to the AI-APP container to form an asynchronous pipeline.

[0036] In this embodiment, the specific processing flow for incremental data and updates is as follows: The incremental data returned by the intelligent fusion terminal is received by the differential upgrade service and provided to the training framework. After the training framework outputs updated weights, they are converted into a general intermediate format by the ONNX converter. The differential upgrade service generates updated content based on the conversion result and publishes it to the mirror repository. The intelligent fusion terminal pulls the updated content from the mirror repository by the upgrade management module to complete the model update, and the AI-APP container enables the updated model.

[0037] In this embodiment, specifically, the model training on the main station side includes: Based on publicly released datasets such as low-voltage topology, stratified sampling is performed according to the voltage level, cable material and user density of the target transformer area. Small-scale samples matching the on-site working conditions are selected to construct the initial training set. The tensor size is constrained by the upper limit of the embedded device being loaded into memory at one time without triggering the swap partition. Feature engineering is performed on the initial training set, and correlation testing and mutual information filtering are performed to remove redundancy and retain features with information gain. Combined with L1 regularization to compress the dimension, a low-dimensional input space is obtained. The space includes at least one or more of the following: voltage level, cable material, user density, node level, impedance, signal-to-noise ratio and load type. Lightweight convolutional neural networks and graph neural networks are trained in parallel with shared input, so that each network can independently output five levels of success probability. The training adopts an adaptive learning rate strategy with momentum and sets an early stopping mechanism to roll back the optimal weights when there is no improvement for several consecutive rounds. After training, the network is subjected to structured pruning and dynamic integer quantization, preferably INT8 quantization, to generate candidate models. The deployment version is then determined by voting or weighted fusion to meet the millisecond-level inference requirements of the edge.

[0038] Based on publicly released datasets such as low-voltage topology open datasets, stratified sampling is performed according to the characteristics of the target transformer area, such as voltage level, cable material, and user density. A small number of samples that highly match the on-site working conditions are selected to construct a small-scale initial training set. The tensor size is capped at the level that the embedded device can load into memory at one time without triggering a swap partition, ensuring efficient inference even in resource-constrained environments.

[0039] Example 2 Example 2 is a further explanation of the model training in the intelligent fusion terminal acquisition success rate prediction system based on topological intelligent perception proposed in Example 1.

[0040] In this embodiment, specifically, feature engineering screening is performed on the initial dataset. First, a correlation test is performed to remove redundant quantities with a linear correlation of less than a threshold with the success rate of data collection. Then, a mutual information metric is used to retain sensitive quantities that have a significant information gain with the target variable. Then, L1 regularization is used to embed and compress the dimension, mapping the high-dimensional original vector into a low-dimensional dense representation. Finally, core features of more than a dozen dimensions such as voltage level, cable material, user density, node level, impedance, signal-to-noise ratio, and load type are retained to form a concise, robust, and edge computing-adaptive low-dimensional input space, balancing prediction accuracy and real-time inference requirements.

[0041] Model training is completed on the main site, associating the aforementioned multi-dimensional feature vectors with the topological graph structure to generate a unified node-edge view. Several publicly validated lightweight convolutional neural networks and graph neural networks (TinyGNN-Lite, etc.) are selected for parallel training, with each network sharing input and independently outputting five levels of success probability. The optimizer employs a dynamic adaptive learning rate strategy and an early stopping mechanism to monitor relevant metrics on the validation set; if there is no improvement after three consecutive rounds, the optimal weights are rolled back.

[0042] After training, structured pruning and dynamic INT8 quantization are performed on each network to generate candidate models with a size smaller than a set threshold. The final deployment version is determined through voting or weighted fusion to ensure millisecond-level inference requirements at the edge.

[0043] During the containerized deployment and porting phase, the trained model is converted into a general intermediate format, and the open-source cross-platform inference engine ONNX Runtime is used as the execution backend. The model, engine library, feature extraction script and configuration parameters are packaged together into a single containerized image. The image is trimmed from a lightweight base image, and the overall size is less than a set threshold.

[0044] After the target device pulls and starts the container in its local operating environment, the inference engine automatically adapts to the intelligent converged terminal's CPU or onboard NPU to complete inference, without requiring modification to the original firmware or kernel. Subsequent upgrades only require replacing the model files within the image in a differential manner, with seamless switching of business processes.

[0045] The aforementioned AI model was initially trained on the main station and then deployed to the field as an initial weight file. After the intelligent fusion terminal was put into operation, it collected new field topology, impedance, signal-to-noise ratio, and corresponding success rate records locally at set intervals to build an incremental dataset, which was then encrypted and sent back to the main station. The main station performed incremental fine-tuning based on this batch of data, producing updated weights; the updated package, after being signed and verified, was distributed and overwrote the original file, achieving online upgrades with zero business interruption. As the number of iterations increased, the model performance gradually approached the optimal value for the scenario and eventually stabilized.

[0046] In summary, this invention focuses on real-time topology sensing, containerized deployment, and weight adaptation. It enables millisecond-level prediction and closed-loop optimization of data acquisition success rate on the local intelligent fusion terminal, significantly reducing manpower and resource waste, shortening deployment cycles, and improving the intelligence level of power grid operation and maintenance. It can be widely applied to diverse scenarios such as smart power distribution, industrial IoT, and smart buildings.

[0047] The embodiments described above merely illustrate specific implementation methods of this application, and while the descriptions are detailed and specific, they should not be construed as limiting the scope of protection of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the technical solution of this application, and these modifications and improvements all fall within the scope of protection of this application.

[0048] This background section is provided to generally present the context of the invention. The work of the currently named inventors, the work to the extent described in this background section, and aspects of this section that did not constitute prior art at the time of application are neither expressly nor impliedly acknowledged as prior art to the invention.

Claims

1. A system for predicting the success rate of data acquisition by an intelligent fusion terminal based on topological intelligent sensing, characterized in that, include: The main station is connected to at least one intelligent fusion terminal, and the intelligent fusion terminal is communicatively connected to the main station. The intelligent fusion terminal includes: The carrier APP container, according to the distribution strategy of the acquisition APP container, acquires the topology information and operating characteristics related to the communication link between the intelligent fusion terminal acquisition terminal and the low-voltage power distribution network on site, generates feature data, and encapsulates the feature data into feature messages and publishes them to the message queue. The message queue module is used for asynchronous buffering and decoupled transmission of the feature messages; The data center is used to cache and manage the characteristic data transmitted by the message queue module and to provide data services to downstream users. The AI-APP container embeds a general inference engine and loads a prediction model trained by the main station, which has been lightweighted and transformed by ONNX. It captures feature data from the data center to perform forward inference, outputs five levels of collection success rate and their confidence, and synchronizes the results to the data center, the collection APP container and the display APP container. The acquisition APP container is used to determine the acquisition command issuance strategy based on the five-level acquisition success rate level and its confidence level output by the AI-APP container, and to coordinate the carrier APP container to execute acquisition. The display APP container is used to present five levels of data collection success rate and their confidence information on the intelligent converged terminal side; The upgrade management module is used to pull the updated content published by the main site from the image repository and complete the online update of the prediction model used by the AI-APP container without interrupting business operations; The data reporting module is used to encrypt and send incremental samples carrying prediction result labels and corresponding topological features back to the main station. The main station includes: The differential upgrade service is used to receive incremental data returned by the intelligent fusion terminal, process the incremental data, generate differential update content relative to the existing deployment version based on the output of the training framework and ONNX converter, and publish the update content to the mirror repository. The training framework is used to train and fine-tune the prediction model based on incremental data provided by the differential upgrade service, and output updated weights. The ONNX converter is used to convert the updated weights or model output by the training framework into a general intermediate format for edge inference. The image repository is used to store models converted by ONNX or container images containing the models, and to provide pull services to smart converged terminals.

2. The intelligent fusion terminal acquisition success rate prediction system based on topological intelligent sensing according to claim 1, characterized in that, The topology information and operational characteristics include: One or more of the following: node level, line impedance, signal-to-noise ratio, voltage fluctuation, and load type.

3. The intelligent fusion terminal acquisition success rate prediction system based on topological intelligent sensing according to claim 1, characterized in that, The distribution strategy includes: Whether to issue the document, the timing of issuance, and parameter settings.

4. The intelligent fusion terminal acquisition success rate prediction system based on topological intelligent sensing according to claim 1, characterized in that, The prediction model is lightweighted on the main station side, including: The trained network is subjected to structured pruning and dynamic INT8 quantization to generate candidate models with a volume smaller than a set threshold.

5. The intelligent fusion terminal acquisition success rate prediction system based on topological intelligent sensing according to claim 1, characterized in that, The general inference engine embedded in the AI-APP container can execute ONNX-transformed models on processors or neural network units without modifying the firmware of the intelligent fusion terminal.

6. The intelligent fusion terminal acquisition success rate prediction system based on topological intelligent sensing according to claim 1, characterized in that, The AI-APP container, general inference engine, ONNX-converted model and its configuration file are packaged into a container image for deployment, and the container image is trimmed from the base image to reduce the image size.

7. The intelligent fusion terminal acquisition success rate prediction system based on topological intelligent sensing according to claim 1, characterized in that, The message queue module and the data center work together to decouple data acquisition and inference processing, and use an asynchronous queue mechanism for data transmission. The data center provides the data service to the AI-APP container to form an asynchronous pipeline.

8. The intelligent fusion terminal acquisition success rate prediction system based on topological intelligent sensing according to claim 1, characterized in that, The processing flow for incremental data and updates is as follows: The incremental data returned by the intelligent fusion terminal is received by the differential upgrade service and provided to the training framework. After the training framework outputs updated weights, they are converted into a general intermediate format by the ONNX converter. The differential upgrade service generates updated content based on the conversion result and publishes it to the mirror repository. The intelligent fusion terminal pulls the updated content from the mirror repository by the upgrade management module to complete the model update, and the AI-APP container enables the updated model.

9. The intelligent fusion terminal acquisition success rate prediction system based on topological intelligent sensing according to claim 1, characterized in that, The model training on the main site side includes: Based on the publicly released low-voltage topology dataset, stratified sampling is performed according to the voltage level, cable material and user density of the target transformer area. Small-scale samples matching the on-site working conditions are selected to construct the initial training set. The tensor size is constrained by the upper limit of the embedded device being loaded into memory at one time without triggering the swap partition. Feature engineering is performed on the initial training set, and correlation testing and mutual information filtering are performed to remove redundancy and retain features with information gain. Combined with L1 regularization to compress the dimension, a low-dimensional input space is obtained. The space includes at least one or more of the following: voltage level, cable material, user density, node level, impedance, signal-to-noise ratio and load type. Lightweight convolutional neural networks and graph neural networks are trained in parallel with shared input, so that each network can independently output five levels of success probability. The training adopts an adaptive learning rate strategy with momentum and sets an early stopping mechanism to roll back the optimal weights when there is no improvement for several consecutive rounds. After training, the network is subjected to structured pruning and dynamic integer quantization, INT8 quantization is performed, candidate models are generated, and the deployment version is determined by voting or weighted fusion to meet the millisecond-level inference requirements of the edge.

10. The intelligent fusion terminal acquisition success rate prediction system based on topological intelligent sensing according to claim 1, characterized in that, The display APP container is used to present five levels of data acquisition success rate and their confidence information in the form of text prompts on the touch screen of the smart converged terminal or on a mobile terminal that is connected to the smart converged terminal.

Images