Intelligent power distribution terminal based on native AI, construction method and power distribution network system

Abstract

The application provides a native AI-based intelligent power distribution terminal, a construction method and a power distribution network system, and relates to the technical field of power systems, comprising: a native AI computing and control unit, integrating at least one general CPU core and at least one AI computing core sharing memory with the general CPU core; a multi-source sensing module for writing collected monitoring data in the form of tensors or vectors into a shared multi-dimensional data pool; the general CPU core is used for deploying different microservices, calling determined target microservices in calling order, and the native AI computing and control unit is used for responding to inference requests, inferring target monitoring data through a target model, and writing inference results into a decision event queue; the general CPU core is also used for generating control instructions according to the decision event queue to control the deterministic control and execution module to execute corresponding control actions. The application effectively improves the AI task processing efficiency of the power distribution terminal.

Description

Technical Field

[0001] This application relates to the field of power system technology, specifically to a smart distribution terminal based on native AI, a method for constructing a smart distribution terminal based on native AI, a power distribution network system, a machine-readable storage medium, and a terminal device. Background Technology

[0002] With the increasing automation of power distribution networks, intelligent distribution terminals, as core monitoring and control devices for substations, lines, and transformer areas, directly affect power supply reliability due to their level of intelligence. Existing technical solutions are mainly divided into three categories: 1) Traditional embedded terminal solutions: The terminal is built on a general-purpose embedded processor (such as the ARM Cortex-A / M series) and runs embedded Linux or RTOS. Its core functions are data acquisition (telemetry, communication), logical judgment (based on fixed thresholds or simple logic), and switch control (remote control). AI analysis functions (such as defect identification and load prediction) are mainly implemented in the cloud or on the main station. The terminal itself does not have or only has very weak real-time AI computing capabilities.

[0003] 2) "External" AI Terminal Solution: Based on the traditional terminal architecture, specific AI models are executed through external expansion or by integrating independent AI acceleration cores (such as NPU or DSP) within the main control chip. AI models are typically trained in the cloud and then compiled and downloaded to the terminal. While this approach endows the terminal with certain AI capabilities, it has significant drawbacks: the AI ​​computing unit is relatively independent from the main control unit, leading to latency and bandwidth bottlenecks in data interaction; the software architecture is a patchwork of "main control application + AI coprocessor," resulting in complex AI task scheduling and resource management; and the AI ​​functions are fixed to specific models, making dynamic updates or expansion difficult.

[0004] 3) Cloud-Edge Collaborative AI Solution: The terminal is responsible for collecting data and uploading it to the edge gateway or cloud platform. The powerful computing power of the edge / cloud performs complex AI analysis and then sends the results or instructions back to the terminal. Although this solution can handle complex models, it is highly dependent on the network, has poor real-time performance (it is difficult to meet the millisecond-level requirements of relay protection and rapid fault isolation), and the terminal will lose advanced analysis capabilities when the network is interrupted. Data privacy and security are also at risk.

[0005] However, the core flaw of existing technologies lies in the failure to integrate AI as a core capability into the underlying logic of terminal design. This results in low AI computing efficiency, insufficient real-time performance, rigid functions, and high system complexity, which cannot meet the intelligent needs of future power distribution networks for massive terminals to "autonomously perceive, autonomously decide, and autonomously execute". Summary of the Invention

[0006] The purpose of this application is to provide a native AI-based intelligent power distribution terminal, a method for constructing a native AI-based intelligent power distribution terminal, a power distribution network system, a machine-readable storage medium, and a terminal device to solve the above-mentioned problems.

[0007] To achieve the above objectives, the first aspect of this application provides a smart power distribution terminal based on native AI, comprising: Native AI computing and control unit, multi-source sensing module, and deterministic control and execution module; The native AI computing and control unit integrates at least one general-purpose CPU core and at least one AI computing core that shares memory with the general-purpose CPU core; The multi-source sensing module is used to write the collected monitoring data into a shared multidimensional data pool in the form of tensors or vectors. The general-purpose CPU core is used to deploy different microservices, determine the target microservices and calling order corresponding to the target task, and call the determined target microservices according to the calling order. The microservices are configured to pull the target monitoring data required by the target task from the shared multidimensional data pool, or to initiate inference requests to the native AI computing and control unit. In response to the inference request, the native AI computing and control unit loads the target model from the specified storage location, infers the target monitoring data through the target model, and writes the inference result into the decision event queue. The general-purpose CPU core is also used to generate control instructions based on the decision event queue to control the deterministic control and execution module to perform corresponding control actions.

[0008] Optionally, the multi-source sensing module includes: One or more of the electrical quantity acquisition submodule, environmental perception submodule, visual perception submodule and audio perception submodule; The electrical quantity acquisition submodule is used to synchronously acquire voltage and current data, and write the synchronously sampled voltage and current data into the shared multidimensional data pool through the DMA channel; The environmental sensing submodule includes an integrated temperature sensor interface, a humidity sensor interface, and a partial discharge sensor interface, used to collect temperature data through the temperature sensor, collect humidity data through the humidity sensor, and collect partial discharge data through the partial discharge sensor, and write the collected temperature data, humidity data, and partial discharge data into the shared multidimensional data pool. The visual perception submodule is used to acquire image signals and write the acquired image signals into the shared multidimensional data pool. The audio perception submodule is used to collect audio signals, extract audio signal feature vectors, and write the extracted audio signal feature vectors into the shared multidimensional data pool.

[0009] Optionally, the target microservices corresponding to the target task and their invocation order are determined, including: Obtain the pipeline configuration file for the target task. The pipeline configuration file includes at least a sequential processing chain that describes the acquisition of target monitoring data for the target task, the inference performed on at least one task target, and the inference results of the inference performed on the at least one task target written into the decision event queue. The pipeline configuration file is parsed, and the target microservice and its invocation order corresponding to the target task are determined based on the sequential processing chain.

[0010] Optionally, the general-purpose CPU core is further used for: The system receives a model package from a cloud management platform. The model package is generated by the cloud management platform through symmetric encryption of the original model file and its digital signature using a session key. The digital signature is obtained by the cloud management platform through encryption of the digest of the original model file using a cloud private key. The digest of the original model file is obtained by the cloud management platform through calculation of the original model file using a preset hash algorithm. The session key is asymmetrically encrypted by the cloud management platform using a terminal public key and then sent together with the model package. The designated storage location is determined to be a dedicated isolated storage area for the native AI computing and control unit. The session key is obtained by performing asymmetric decryption using the terminal private key that matches the terminal public key. The model package is symmetrically decrypted using the session key to obtain the original model file and its digital signature; The digital signature is decrypted using a cloud public key that matches the cloud private key to obtain a first digest. The original model file is then calculated using a preset hash algorithm to obtain a second digest. If the first digest and the second digest are consistent, the signature verification is confirmed, and the original model file is loaded into the designated storage location.

[0011] Optionally, the target model includes at least: A first inference network and a second learning network, wherein the first inference network is used to take the target monitoring data as input and output the inference result of the target monitoring data; The second learning network has a smaller network capacity than the first inference network, and the second learning network is in a dormant state when the first inference network infers the target monitoring data; The second learning network is used to activate the second learning network for local incremental training after detecting concept drift in the first inference network, so as to adjust the network parameters of the first inference network based on the network parameters of the trained second learning network.

[0012] Optionally, detecting concept drift in the first inference network includes: At the current moment, multiple consecutive historical inference results of the first inference network are obtained sequentially in chronological order through multiple preset time windows; For each time window, obtain the predicted probability of each historical inference result belonging to the corresponding true result within the current time window by the first inference network; Based on the prediction probability of each historical inference result of the first inference network within the current time window, calculate the uncertainty index that characterizes the average prediction uncertainty of the first inference network within the current time window. Based on the average value of the uncertainty index for each time window, the drift index of the first inference network at the current moment is determined. Based on all historical drift indicators prior to the current moment, the statistical distribution characteristics of the historical drift indicators are determined, and based on the statistical distribution characteristics and preset sensitivity control parameters, the dynamic threshold of the drift indicator at the current moment is generated. If the drift index at the current moment is greater than the dynamic threshold, it is determined that the first inference network has experienced concept drift.

[0013] Optionally, based on the prediction probabilities of the first inference network for each historical inference result within the current time window, an uncertainty index characterizing the average prediction uncertainty of the first inference network within the current time window is calculated, including: Calculate the logarithm of the prediction probability of the first inference network for each historical inference result within the current time window; Calculate the average of the logarithmic values ​​of all predicted probabilities within the current time window; The negative of the average of the logarithmic values ​​of all predicted probabilities within the current time window is used as the uncertainty index of the first inference network within the current time window.

[0014] Optionally, the statistical distribution characteristics include the mean and standard deviation of the historical drift index; based on the statistical distribution characteristics and preset sensitivity control parameters, a dynamic threshold for the drift index at the current moment is generated, including: Calculate the product of the mean of the historical drift index and the sensitivity control parameter, and use the sum of the product and the standard deviation of the historical drift index as the dynamic threshold.

[0015] Optionally, the second learning network is activated for local incremental training, including: Determine the target moment when the concept drift of the first inference network is detected, and use the target monitoring data and the corresponding real results of multiple adjacent moments before and after the target moment as target training samples; The second learning network is incrementally trained using the target training samples, and during the incremental training process, replay samples representing the historical data distribution of the target task are extracted from a preset historical sample database to train the second learning network.

[0016] Optionally, the network parameters of the first inference network are adjusted based on the network parameters of the trained second learning network, including: After incremental training is completed, obtain the updated network parameters of the second learning network; The parameters of the second learning network are fused into the first inference network through knowledge distillation or parameter smoothing to update the network parameters of the first inference network. In addition, the difference between the updated network parameters of the first inference network and the network parameters before the update of the first inference network is determined as the parameter increment; The parameter increments generated by this update will be encrypted and uploaded to the cloud management platform.

[0017] A second aspect of this application provides a method for constructing a smart power distribution terminal based on native AI, including: A native AI computing and control unit is constructed by integrating at least one general-purpose CPU core and at least one AI computing core that shares memory with the general-purpose CPU core; The multi-source sensing module is used to write the collected monitoring data into a shared multidimensional data pool in the form of tensors or vectors. Different microservices are deployed through the general-purpose CPU cores to determine the target microservices and their calling order corresponding to the target task. The determined target microservices are called according to the calling order. The microservices are configured to pull the target monitoring data required by the target task from the shared multidimensional data pool or to initiate inference requests to the native AI computing and control unit. The native AI computing and control unit responds to the inference request, loads the target model from the specified storage location, infers the target monitoring data using the target model, and writes the inference results into the decision event queue. The general-purpose CPU core generates control instructions based on the decision event queue, and the deterministic control and execution module responds to the control instructions and executes the corresponding control actions.

[0018] A third aspect of this application provides a power distribution network system, comprising: Cloud management platform; And at least one smart power distribution terminal based on native AI, as described above.

[0019] In a fourth aspect, this application provides a machine-readable storage medium storing instructions that, when executed by a processor, configure the processor to perform the aforementioned method for constructing a smart power distribution terminal based on native AI.

[0020] In a fifth aspect, this application provides a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method for constructing a smart power distribution terminal based on native AI.

[0021] This application constructs a smart terminal hardware architecture centered on a heterogeneous SoC integrating a dedicated AI computing core, enabling deep coupling between the AI ​​computing core, CPU, sensing module, and control module via a high-speed interconnect bus and shared memory. By constructing a heterogeneous computing software stack for native AI and a microservice encapsulation method for AI functions, it achieves unified scheduling, dynamic loading, and isolated execution of AI tasks. At the same time, based on the construction of a multi-AI task collaborative execution mechanism within the terminal using a data processing pipeline orchestration, it effectively improves the processing efficiency of AI tasks.

[0022] Other features and advantages of the embodiments of this application will be described in detail in the following detailed description section. Attached Figure Description

[0023] The accompanying drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the following detailed description to explain the embodiments of this application, but do not constitute a limitation on the embodiments of this application. In the drawings: Figure 1 A schematic diagram of the hardware structure of a smart power distribution terminal based on native AI provided in a preferred embodiment of this application; Figure 2 A schematic diagram of a terminal device provided for a preferred embodiment of this application.

[0024] Explanation of reference numerals in the attached figures 10 - Terminal device, 100 - Processor, 101 - Memory, 102 - Computer program. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only for illustration and explanation of the embodiments of this application and are not intended to limit the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0026] It should be noted that the technical solutions of the various embodiments of this application can be combined with each other, but only if they are based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by this application.

[0027] In existing power distribution terminal architectures, AI computing is typically non-native and inefficient. As an add-on function, AI incurs data transfer and scheduling overhead with the main system, failing to meet the real-time requirements of the power distribution network (e.g., fault identification <10ms). Furthermore, the system architecture is rigid, making function updates difficult. AI models are embedded in the terminal software, requiring system-wide upgrades or even hardware replacements for updates, expansions, or additions of new AI functions, resulting in high maintenance costs. Resource utilization is low; traditional CPUs have poor energy efficiency in handling AI tasks and cannot cope with the increasingly complex multimodal AI analysis needs of the terminal side. The level of intelligence is limited, lacking local adaptation and learning capabilities, making it unable to respond to dynamic changes in the characteristics of power distribution network equipment and loads.

[0028] To address the aforementioned issues, the first aspect of this application provides a native AI-based intelligent power distribution terminal, comprising: a native AI computing and control unit, a multi-source sensing module, and a deterministic control and execution module; the native AI computing and control unit integrates at least one general-purpose CPU core and at least one AI computing core sharing memory with the general-purpose CPU core; the multi-source sensing module is used to write the collected monitoring data into a shared multi-dimensional data pool in tensor or vector form; the general-purpose CPU core is used to deploy different microservices, determine the target microservices corresponding to the target task and their calling order, and call the determined target microservices according to the calling order, wherein the microservices are configured to pull the target monitoring data required by the target task from the shared multi-dimensional data pool, or initiate inference requests to the native AI computing and control unit; the native AI computing and control unit responds to the inference request by loading the target model from a specified storage location, performing inference on the target monitoring data through the target model, and writing the inference results into a decision event queue; the general-purpose CPU core is also used to generate control instructions based on the decision event queue to control the deterministic control and execution module to perform corresponding control actions.

[0029] In this way, by constructing a native AI hardware architecture, this application transforms AI computing from an external component to an internal one, realizing a direct path from sensor data to AI computation. This significantly reduces latency, improves computational efficiency, and makes it possible for complex AI algorithms (such as transient waveform recognition) to run in real time on the terminal side. Simultaneously, through microservice-based AI functions and a dynamic loading mechanism, AI capabilities are modularized and service-oriented. Users can download, deploy, and combine different AI microservices on demand, achieving software-defined and dynamic reconfiguration of terminal functions, greatly enhancing the terminal's flexibility and scalability. By introducing edge incremental learning methods, the terminal possesses the ability to fine-tune models based on local data, improving adaptability to personalized scenarios, while enabling continuous evolution of the global model through cloud-edge collaboration.

[0030] Specifically, such as Figure 1 As shown in this application, the native AI computing and control unit integrates a main control CPU core group including at least one general-purpose CPU core and an AI computing core. The AI ​​computing core includes a VPU (Visual / Vector Processing Unit), a TPU (Tensor Processing Unit), and dedicated memory and cache. The main control CPU core group and the AI ​​computing core are connected through a heterogeneous computing interconnect bus.

[0031] The multi-source sensing module includes one or more of the following: electrical quantity acquisition submodule, environmental sensing submodule, visual sensing submodule, and audio sensing submodule. Among them, the electrical quantity acquisition submodule uses a high-precision ADC and a dedicated metering chip to synchronously acquire voltage and current data. The synchronously sampled voltage and current data is written into a shared multi-dimensional data pool through a DMA (Direct Memory Access) channel for real-time access by the AI ​​core.

[0032] The environmental sensing submodule includes an integrated temperature sensor interface, a humidity sensor interface, and a partial discharge (UHF / AE) sensor interface. It is used to collect temperature data through the temperature sensor, humidity data through the humidity sensor, and partial discharge data through the partial discharge sensor, and write the collected temperature data, humidity data, and partial discharge data into a shared multidimensional data pool.

[0033] The visual perception submodule is used to acquire image signals and write them into a shared multidimensional data pool. It integrates a MIPI CSI interface, which can be directly connected to a camera. The AI ​​computing core's ISP (Image Signal Processing) unit and VPU perform image preprocessing, while the TPU performs real-time target recognition.

[0034] The audio perception submodule is used to acquire audio signals, extract audio signal feature vectors, and write the extracted audio signal feature vectors into a shared multidimensional data pool. It integrates a high-performance digital microphone or analog microphone array interface, and acquires audio signals through an ADC or dedicated I2S interface. After the signal undergoes front-end preprocessing (including noise reduction, echo cancellation, etc.) and acoustic feature extraction (such as MFCC), the feature vectors are written into the shared data pool in real time for the AI ​​core to use for abnormal sound detection, acoustic fault diagnosis, and other analyses.

[0035] The deterministic control and execution module is built on highly reliable control logic (FPGA or hardware protection logic). For ultra-high real-time (microsecond level) control tasks such as switch opening and closing and protection tripping, it is implemented by independent FPGA or hardware logic circuits. It receives decision results from the AI ​​computing core or main CPU, executes the corresponding protection actions, and ensures the absolute determinism of the actions.

[0036] The intelligent power distribution terminal of this application also includes a secure and reliable communication module, which has an embedded security chip that provides hardware-level key storage and encryption operations to ensure the security of AI models, parameters, and collected data. Simultaneously, the secure and reliable communication module includes a multi-mode communication interface supporting fiber optic Ethernet, 5G / 4G, HPLC, etc., and has a built-in protocol stack supporting MQTT, IEC104, GOOSE, etc. The communication data stream can be directly analyzed by the AI ​​core (e.g., for traffic anomaly detection).

[0037] The raw data collected by the sensing module is directly injected into a shared multidimensional data pool managed by the AI ​​computing core via a high-speed interconnect bus or DMA channel. The data format is standardized tensor or vector. The AI ​​computing core actively pulls data from the data pool for inference calculations based on pre-loaded or dynamically loaded AI models. The calculation process requires no deep intervention from the main CPU, achieving data-driven computation. AI inference results (such as three-phase imbalance exceeding limits, switchgear temperature anomalies, and suspected partial discharge signals) are output to a decision event queue. The main CPU or FPGA retrieves events from the queue and, combined with business logic (such as protection settings and control strategies), generates final control commands. The main CPU is responsible for the dynamic loading, version management, and lifecycle monitoring of the AI ​​model, and interacts with the cloud management platform through a secure communication module to achieve remote secure updates and collaborative training of the model.

[0038] The steps for constructing the intelligent power distribution terminal in this application include: Step 1: Heterogeneous SoC Selection and Hardware Abstraction Layer Design We selected a SoC integrating a high-performance CPU core and a dedicated AI acceleration core, and designed a unified hardware abstraction layer. The CPU is responsible for control logic and communication, while the AI ​​acceleration core is responsible for intensive model inference and incremental training. We encapsulated the AI ​​computing core, perception interface, and control interface, providing a unified API (such as ai_engine_infer(), sensor_fetch_tensor()) to shield the underlying hardware differences. Even if the chip is replaced later, there is no need to modify all the upper-layer AI microservices and applications.

[0039] Step 2: Building a native AI software stack: The microkernel operating system is built using a lightweight OS with strong real-time performance and optimized for AI task scheduling, supporting high-priority preemptive task scheduling for AI computing cores. For example, AI inference tasks are set to the highest preemptive priority. When the AI ​​acceleration core completes a round of prediction, it notifies the CPU via an interrupt. The operating system will immediately preempt the current low-priority task (such as log recording) and process the inference results as soon as possible to ensure real-time performance.

[0040] The construction of the heterogeneous computing framework includes deploying a lightweight inference framework (such as TensorFlow LiteMicro, Paddle Lite) on the terminal side, and performing deep operator optimization and model compilation for the AI ​​engine AI-Engine of the hardware in this application to generate an efficient executable file.

[0041] The microservice approach to AI capabilities involves encapsulating different AI capabilities (such as load forecasting, fault diagnosis, and image inspection) into independent microservices. Each microservice contains: a) preprocessing code; b) optimized model files; and c) post-processing and output formatting code. Each microservice is deployed and managed in isolation using containerization technologies (such as Docker) or lightweight sandboxes.

[0042] For example, the load forecasting microservice includes (a) preprocessing code: obtaining the A-phase current waveform from sensor_fetch_tensor("I_A"), normalizing it, and performing sliding window segmentation; (b) optimized model file: a short-term load forecasting model based on dual-network LSTM, compiled into load_forecast.nb; and (c) postprocessing code: inversely transforming the normalized load value output by the model into the actual power and formatting it into an IEC61850 MMS message.

[0043] The image inspection microservice includes (a) preprocessing: scaling the infrared thermal image and extracting the temperature matrix; (b) optimized model file: a lightweight target detection model thermal_inspect.nb based on MobileNet, which is a single network model here; and (c) postprocessing: binding the detected device hotspot temperature with the device number to generate an alarm structure.

[0044] Step 3: Dynamic Model Loading and Pipeline Orchestration The terminal has a built-in secure and trusted model repository, and the cloud management platform can distribute new, signed, and encrypted model / microservice packages. After verifying the signature, the terminal securely loads the new model / microservice package into the dedicated storage area of ​​the AI ​​computing core. For example, if the cloud management platform detects an improvement in the performance of the latest trained global load prediction model (integrating incremental uploads from multiple terminals), it packages the model into a file named load_forecast_v2.0.bin, signs it with the cloud private key, encrypts it with the session key, and then distributes it. Upon receiving the model, the terminal's model repository verifies the signature using a pre-configured public key. If successful, it decrypts the file and writes the new .nb file to the dedicated MMC isolated partition of the AI ​​acceleration core, while the old version is retained for rollback.

[0045] In this application, a hybrid security mechanism combining asymmetric and symmetric encryption is used between the cloud management platform and the power distribution terminal to ensure the confidentiality, integrity, and authenticity of the model package during transmission and loading. The general-purpose CPU core is also used for: receiving the model package from the cloud management platform, whereby the model package is generated by the cloud management platform through symmetric encryption of the original model file and its digital signature using a session key; the digital signature is obtained by the cloud management platform through encryption of the digest of the original model file using a cloud private key; the digest of the original model file is obtained by the cloud management platform through calculation of the original model file using a preset hash algorithm (such as SHA-256); and the session key is asymmetrically encrypted by the cloud management platform using the terminal public key, and then sent to the power distribution terminal along with the model package; and determining a designated storage location as a dedicated isolated storage area for the native AI computing and control unit, for example, specifying... The designated storage location can be a high-speed static random access memory integrated inside the AI ​​acceleration core chip. After the general-purpose CPU core receives the data, it first performs asymmetric decryption using the terminal private key matched with the terminal public key to obtain the session key; then it performs symmetric decryption on the model packet using the session key to obtain the original model file and its digital signature; finally, it decrypts the digital signature using the cloud public key matched with the cloud private key to obtain the first digest, and calculates the second digest on the original model file using a preset hash algorithm (such as SHA-256). If the first digest and the second digest are consistent, the signature verification is confirmed, and the original model file is loaded into the designated storage location.

[0046] The data processing pipeline orchestration process includes defining the data flow processing chain (Pipeline) through configuration files. For example, it might involve microservices such as acquiring electrical quantity data, data cleaning, short-term load forecasting, and finally reporting the results. The native AI computing and control unit automatically schedules different microservices to execute on the TPU / VPU based on the pipeline description, enabling parallel pipeline execution of multiple AI tasks.

[0047] In this application, determining the target microservices and their invocation order corresponding to the target task includes: obtaining the pipeline configuration file of the target task, which at least includes a sequential processing chain describing the target task's acquisition of target monitoring data, inference on at least one task target, and writing the inference results of the inference on at least one task target into a decision event queue; parsing the pipeline configuration file and determining the target microservices and their invocation order corresponding to the target task based on the sequential processing chain. For example, the configuration file defines an end-to-end real-time analysis pipeline, which includes data acquisition, data cleaning, model inference, and result reporting, with each stage corresponding to a microservice. After parsing the configuration, the pipeline scheduler of the general-purpose CPU core sequentially calls each microservice to execute the corresponding task. For example, the data processing in the data acquisition and data cleaning stages runs on the general-purpose CPU core, the computation of model inference is allocated to the AI ​​computing core, and the communication processing of result reporting runs on the general-purpose CPU core. In this way, multiple microservices of a complete task can run asynchronously on the CPU and AI computing core respectively, enabling multiple AI tasks to overlap and run in parallel in time and space, effectively improving the data processing efficiency of the power distribution terminal.

[0048] Step 4: Self-learning and Adaptive Mechanisms In this application, the power distribution terminal maintains a lightweight student model locally. When it is detected that the model is continuously inaccurate in a specific local scenario (such as the access of a new type of load), it can use new local data to fine-tune the model under the premise of protecting user privacy, thereby improving the model's adaptability in the scenario. The incremental parameters are then encrypted and uploaded to the cloud for global model optimization.

[0049] In this application, the target model includes at least: a first inference network and a second learning network. The first inference network is used to take target monitoring data as input and output the inference result of the target monitoring data. The network capacity of the second learning network is smaller than that of the first inference network. The second learning network is in a dormant state when the first inference network infers the target monitoring data. The second learning network is used to activate the second learning network for local incremental training after detecting concept drift in the first inference network, so as to adjust the network parameters of the first inference network based on the network parameters of the trained second learning network.

[0050] The model consists of two main networks: a first inference network (stable knowledge network) and a second learning network. The first inference network is the backbone network for normal inference. When a microservice receives an inference request, it performs forward computation on the input data and outputs the prediction result. The second learning network (flexible learning network) is a lightweight auxiliary network designed for rapid learning. It does not participate in computation during normal inference and remains dormant. It is only activated when the first inference network triggers a concept drift alarm. The second learning network then performs rapid incremental training using new local data. After learning, the new knowledge it acquires is unidirectionally fused back into the first inference network through knowledge distillation or parameter smoothing. For example, for a load prediction microservice A, its model file includes a first inference network with a 2-layer LSTM structure and a parameter set size of 500KB, and a second learning network with a 1-layer LSTM structure and a parameter set size of 150KB. During normal inference, only the first inference network is used. After a probability drift is triggered, the second learning network is activated to fine-tune the network parameters, and the knowledge is then fused back into the first inference network. Understandably, for AI tasks that do not require online adaptive learning, the corresponding microservice model does not need to adopt a dual-network architecture; a conventional single-network architecture is sufficient.

[0051] In this application, concept drift refers to the fact that the AI ​​model built into the power distribution terminal was initially trained on historical data and learned to recognize various operating states (such as normal, overload, and short-circuit waveforms). However, in actual operation, due to the dynamic changes in the environment in which the terminal is located, the patterns previously learned by the AI ​​gradually become invalid. For example, changes in load patterns, such as the construction of a new data center in the area, which changes the power consumption pattern from residential electricity consumption to 24-hour stable industrial electricity consumption, altering the normal waveform characteristics of current and power; equipment access and disconnection, with new distributed photovoltaic and charging piles connecting to the grid, generating new electrical characteristics that the model has never seen before; slow aging of equipment, such as transformers and lines aging over time, causing the signal baseline in their healthy state to gradually shift; and seasonal effects, with significant differences between summer air conditioning load and winter heating load.

[0052] Based on the acquired monitoring data, the terminal believes that waveform A represents normal operation based on past learning. However, due to changes in the terminal's environment, the same waveform A may now indicate an early fault, meaning that the correspondence between data (waveform) and results (status label) has drifted.

[0053] Therefore, this application constructs a model based on a drift-aware dual-network online incremental learning and adaptive playback architecture, which can achieve high-performance and highly stable local adaptive learning under resource-constrained conditions to address the concept drift problem caused by dynamic changes in the operating environment of power distribution terminals (such as load mode changes, new equipment access, and equipment aging).

[0054] Understandably, when deploying microservices, the CPU first reads the uninstantiated model data source file from the microservice, calls the model compiler to perform deep optimization for the AI ​​acceleration core, generates an executable file, and then loads it into the AI ​​acceleration core's dedicated isolated storage area. Simultaneously, a mapping from microservice to model is established. After deployment, when the AI ​​acceleration core executes inference tasks, it simply loads the instance from the dedicated isolated storage area based on the pre-established microservice-to-model mapping.

[0055] In this application, concept drift in the first inference network is detected, including: Step 1: At the current moment, obtain multiple consecutive historical inference results of the first inference network in chronological order through multiple preset time windows. First, define a sliding window W containing the most recent M batches of data, with each batch including B data samples.

[0056] Step 2: For each time window, obtain the predicted probability of each historical inference result belonging to the corresponding true result within the current time window by the first inference network. For example, a batch includes 3 samples, namely normal waveform, overload waveform, and short-circuit waveform data. Using the 3 samples of this batch as input, the AI ​​computing core loads the first inference network to perform forward inference. For each sample, the first inference network outputs a set of raw scores, which are normalized to obtain a predicted probability vector. For example, for sample 1, its output is [normal: 0.87, overload: 0.09, short-circuit: 0.04], for sample 2 it is [normal: 0.15, overload: 0.72, short-circuit: 0.13], and for sample 3 it is [normal: 0.02, overload: 0.05, short-circuit: 0.93]. Then, the predicted probabilities of the inference results of the 3 samples belonging to the true result label are 0.87, 0.72, and 0.93, respectively.

[0057] Step 3: Based on the prediction probabilities of the first inference network for each historical inference result within the current time window, calculate the uncertainty index characterizing the average prediction uncertainty of the first inference network within the current time window. This includes: calculating the logarithm of the prediction probabilities of the first inference network for each historical inference result within the current time window; calculating the average of the logarithms of all prediction probabilities within the current time window; and using the negative value of the average of the logarithms of all prediction probabilities within the current time window as the uncertainty index of the first inference network within the current time window. This application uses negative log-likelihood to calculate the average prediction uncertainty of all data batches within this window as the drift index S. t .

[0058] Step 4: Based on the average value of the uncertainty index for each time window, determine the drift index of the first inference network at the current moment. Specifically, the drift index S of this application... t It can be represented as:

[0059] Where t represents the current time; M represents the sliding window size, used to smooth fluctuations (e.g., M=10); and B represents the number of data samples in each batch. This represents the i-th data sample and its label (the actual result); Represents a stable knowledge network (with parameters). ) for samples The predicted probability, that is, given an input sample Parameters of stable knowledge networks Under the given conditions, the model correctly classifies the sample as its true label. The predicted probability. Wherein, the index S... t The larger the value, the lower the average prediction confidence of the model on recent data, and the higher the uncertainty.

[0060] Step 5: Based on all historical drift indicators prior to the current moment, determine the statistical distribution characteristics of the historical drift indicators, and generate the dynamic threshold of the drift indicator at the current moment according to the statistical distribution characteristics and preset sensitivity control parameters. The statistical distribution characteristics include the mean and standard deviation of the historical drift indicators. Generating the dynamic threshold of the drift indicator at the current moment according to the statistical distribution characteristics and preset sensitivity control parameters includes: calculating the product of the mean of the historical drift indicator and the sensitivity control parameter, and using the sum of the product and the standard deviation of the historical drift indicator as the dynamic threshold.

[0061] Specifically, the dynamic threshold of this application The calculation formula can be expressed as:

[0062] in, This represents the mean of all historical indicators S from the start time to time t-1; This represents the standard deviation of the corresponding historical indicator S; Sensitivity coefficient (e.g.) =2), used to control the sensitivity of the alarm. The larger the value, the lower the tolerance to fluctuations, and the more sensitive the detection.

[0063] Step 6: If the drift index at the current moment is greater than the dynamic threshold, determine that concept drift has occurred in the first inference network. Specifically, the currently calculated drift index S... t With dynamic threshold Compare, if > If this occurs, a concept drift alert is triggered, indicating a statistically significant decline in the model's predictive performance on recent data, potentially signifying a shift in data distribution. Once the alert is triggered, the system will initiate subsequent adaptive memory playback and knowledge consolidation processes to adjust the model to adapt to the new data distribution.

[0064] This application detects concept drift by monitoring changes in uncertainty predicted by the model itself (a stable knowledge network), and can run online without requiring real labeled data streams. Based on a dynamically calculated threshold, it adapts to inherent data fluctuations, reducing false alarms, thus achieving efficient and adaptive concept drift detection on resource-constrained terminals.

[0065] After triggering the concept drift alarm, the system enters the adaptive memory playback and knowledge consolidation process, activating the second learning network for local incremental training. Specifically, this includes: determining the target moment when concept drift of the first inference network is detected, using target monitoring data and their corresponding real results from multiple adjacent moments before and after the target moment as target training samples; incrementally training the second learning network using the target training samples; and during the incremental training process, extracting playback samples representing the historical data distribution of the target task from a preset historical sample database to train the second learning network.

[0066] The replay samples in the historical sample database are a small batch of representative samples selected from historical data representing the old data distribution before the concept drift occurred and stored locally on the terminal. These reflect the data distribution patterns learned by the model during its initial training, i.e., the electrical environment and load patterns before the drift occurred. For example, in a residential distribution area over the past two years, electricity consumption patterns were primarily residential: low load during the day, high load at night, and extremely low load late at night. The model was trained on a large amount of data collected during these two years, learning to identify normal, overload, and short-circuit waveform characteristics under residential load patterns. From this historical data distribution, a small batch of the most representative samples (e.g., 500 waveform samples) covering various typical operating conditions was selected and stored in the terminal's local memory; these are the replay samples representing the historical data distribution.

[0067] During incremental training, monitoring data and corresponding real results from multiple adjacent time points before and after the detected probability drift are first collected to construct training samples for training the second learning network. Simultaneously, replay samples are extracted to train the second learning network. In this hybrid training approach, the flexible learning network learns to recognize new load patterns without forgetting the characteristics of old load patterns, thus achieving a balance between stability and flexibility.

[0068] In this application, the network parameters of the first inference network are adjusted based on the network parameters of the trained second learning network, including: after incremental training is completed, obtaining the updated network parameters of the second learning network; fusing the parameters of the second learning network into the first inference network through knowledge distillation or parameter smoothing to update the network parameters of the first inference network; determining the difference between the updated network parameters of the first inference network and the network parameters before the update as the parameter increment; and encrypting the parameter increment generated by this update and uploading it to the cloud management platform.

[0069] For example, after obtaining the updated network parameters of the second learning network, it is fused into the first inference network using the exponential moving average method. The fusion process can be represented as follows:

[0070] in, This represents the network parameters after the first inference network is updated. This represents the network parameters before the first inference network update. This represents the updated network parameters after the second learning network is established. Indicates the weight.

[0071] After local adaptive learning is completed, the difference between the new parameter values ​​and the old parameter values ​​of the first inference network is calculated to obtain the parameter increment of each network parameter. After receiving the encrypted parameter increments uploaded by each power distribution terminal, the cloud management platform can aggregate these parameter increments to continuously improve a globally shared AI model. For example, aggregation can be performed through a federated averaging method, which is not limited here.

[0072] In summary, this application constructs a native AI hardware architecture, transforming AI computing from an external component to an internal one. This establishes a direct path from sensor data to AI computation, significantly reducing latency and improving computational efficiency. This enables complex AI algorithms (such as transient waveform recognition) to run in real-time on the terminal side, solving the real-time performance and energy efficiency issues of AI computing. Through microservice-based AI functions and a dynamic loading mechanism, AI capabilities are modularized and service-oriented. Users can download, deploy, and combine different AI microservices as needed, enabling software definition and dynamic reconfiguration of terminal functions. This greatly enhances the flexibility and scalability of the terminal, solving the problem of rigid and unupdable terminal AI functions. By introducing edge incremental learning methods, the terminal gains the ability to fine-tune models based on local data, improving adaptability to personalized scenarios. Simultaneously, cloud-edge collaboration enables continuous evolution of the global model, addressing the issue of insufficient terminal adaptability.

[0073] A second aspect of this application provides a method for constructing a smart power distribution terminal based on native AI, including: A native AI computing and control unit is built by integrating at least one general-purpose CPU core and at least one AI computing core that shares memory with the general-purpose CPU core; The multi-source sensing module is used to write the collected monitoring data into a shared multidimensional data pool in the form of tensors or vectors. Different microservices are deployed through general-purpose CPU cores. The target microservices and their calling order corresponding to the target tasks are determined. The target microservices are called in the order they are called. The microservices are configured to pull the target monitoring data required by the target tasks from the shared multidimensional data pool or to initiate inference requests to the native AI computing and control unit. The native AI computing and control unit responds to inference requests, loads the target model from the specified storage location, infers the target monitoring data through the target model, and writes the inference results into the decision event queue. The general-purpose CPU core generates control instructions based on the decision event queue, and the deterministic control and execution module responds to the control instructions and executes the corresponding control actions.

[0074] A third aspect of this application provides a power distribution network system, comprising: Cloud management platform; And at least one smart power distribution terminal based on native AI, as described above.

[0075] In a fourth aspect, this application provides a machine-readable storage medium storing instructions that, when executed by a processor, cause the processor to be configured to perform the aforementioned method for constructing a smart power distribution terminal based on native AI.

[0076] In a fifth aspect, this application provides a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the above-described method for constructing a smart power distribution terminal based on native AI.

[0077] like Figure 2 The diagram shown is a schematic representation of a terminal device provided in an embodiment of this application. Figure 2 As shown, the terminal device 10 of this embodiment includes a processor 100, a memory 101, and a computer program 102 stored in the memory 101 and executable on the processor 100. When the processor 100 executes the computer program 102, it implements the steps in the above method embodiments. Alternatively, when the processor 100 executes the computer program 102, it implements the functions of each module / unit in the above device embodiments.

[0078] For example, computer program 102 may be divided into one or more modules / units, one or more of which are stored in memory 101 and executed by processor 100 to complete this application. One or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of computer program 102 in terminal device 10.

[0079] Terminal device 10 may be a computing device such as a desktop computer, laptop, handheld computer, or cloud server. Terminal device 10 may include, but is not limited to, a processor 100 and a memory 101. Those skilled in the art will understand that... Figure 2 This is merely an example of terminal device 10 and does not constitute a limitation on terminal device 10. It may include more or fewer components than shown, or combine certain components, or different components. For example, terminal device may also include input / output devices, network access devices, buses, etc.

[0080] The processor 100 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.

[0081] The memory 101 can be an internal storage unit of the terminal device 10, such as a hard disk or RAM of the terminal device 10. The memory 101 can also be an external storage device of the terminal device 10, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, or FlashCard equipped on the terminal device 10. Furthermore, the memory 101 can include both internal and external storage units of the terminal device 10. The memory 101 is used to store computer programs and other programs and data required by the terminal device 10. The memory 101 can also be used to temporarily store data that has been output or will be output.

[0082] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0083] It should also be noted that 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 process, method, article, or apparatus. Unless otherwise specified, 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 that element.

[0084] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A smart power distribution terminal based on native AI, characterized in that, include: Native AI computing and control unit, multi-source sensing module, and deterministic control and execution module; The native AI computing and control unit integrates at least one general-purpose CPU core and at least one AI computing core that shares memory with the general-purpose CPU core; The multi-source sensing module is used to write the collected monitoring data into a shared multidimensional data pool in the form of tensors or vectors. The general-purpose CPU core is used to deploy different microservices, determine the target microservices and calling order corresponding to the target task, and call the determined target microservices according to the calling order. The microservices are configured to pull the target monitoring data required by the target task from the shared multidimensional data pool, or to initiate inference requests to the native AI computing and control unit. In response to the inference request, the native AI computing and control unit loads the target model from the specified storage location, infers the target monitoring data through the target model, and writes the inference result into the decision event queue. The general-purpose CPU core is also used to generate control instructions based on the decision event queue to control the deterministic control and execution module to perform corresponding control actions.

2. The intelligent power distribution terminal based on native AI according to claim 1, characterized in that, The multi-source sensing module includes: One or more of the electrical quantity acquisition submodule, environmental perception submodule, visual perception submodule and audio perception submodule; The electrical quantity acquisition submodule is used to synchronously acquire voltage and current data, and write the synchronously sampled voltage and current data into the shared multidimensional data pool through the DMA channel; The environmental sensing submodule includes an integrated temperature sensor interface, a humidity sensor interface, and a partial discharge sensor interface, used to collect temperature data through the temperature sensor, collect humidity data through the humidity sensor, and collect partial discharge data through the partial discharge sensor, and write the collected temperature data, humidity data, and partial discharge data into the shared multidimensional data pool. The visual perception submodule is used to acquire image signals and write the acquired image signals into the shared multidimensional data pool. The audio perception submodule is used to collect audio signals, extract audio signal feature vectors, and write the extracted audio signal feature vectors into the shared multidimensional data pool.

3. The intelligent power distribution terminal based on native AI according to claim 1, characterized in that, Determine the target microservices and their invocation order corresponding to the target task, including: Obtain the pipeline configuration file for the target task. The pipeline configuration file includes at least a sequential processing chain that describes the acquisition of target monitoring data for the target task, the inference performed on at least one task target, and the inference results of the inference performed on the at least one task target written into the decision event queue. The pipeline configuration file is parsed, and the target microservice and its invocation order corresponding to the target task are determined based on the sequential processing chain.

4. The intelligent power distribution terminal based on native AI according to claim 1, characterized in that, The general-purpose CPU core is also used for: The system receives a model package from a cloud management platform. The model package is generated by the cloud management platform through symmetric encryption of the original model file and its digital signature using a session key. The digital signature is obtained by the cloud management platform through encryption of the digest of the original model file using a cloud private key. The digest of the original model file is obtained by the cloud management platform through calculation of the original model file using a preset hash algorithm. The session key is asymmetrically encrypted by the cloud management platform using a terminal public key and then sent together with the model package. The designated storage location is determined to be a dedicated isolated storage area for the native AI computing and control unit. The session key is obtained by performing asymmetric decryption using the terminal private key that matches the terminal public key. The model package is symmetrically decrypted using the session key to obtain the original model file and its digital signature; The digital signature is decrypted using a cloud public key that matches the cloud private key to obtain a first digest. The original model file is then calculated using a preset hash algorithm to obtain a second digest. If the first digest and the second digest are consistent, the signature verification is confirmed, and the original model file is loaded into the designated storage location.

5. The intelligent power distribution terminal based on native AI according to claim 1, characterized in that, The target model includes at least: A first inference network and a second learning network, wherein the first inference network is used to take the target monitoring data as input and output the inference result of the target monitoring data; The second learning network has a smaller network capacity than the first inference network, and the second learning network is in a dormant state when the first inference network infers the target monitoring data; The second learning network is used to activate the second learning network for local incremental training after detecting concept drift in the first inference network, so as to adjust the network parameters of the first inference network based on the network parameters of the trained second learning network.

6. The intelligent power distribution terminal based on native AI according to claim 5, characterized in that, Detecting concept drift in the first inference network includes: At the current moment, multiple consecutive historical inference results of the first inference network are obtained sequentially in chronological order through multiple preset time windows; For each time window, obtain the predicted probability of each historical inference result belonging to the corresponding true result within the current time window by the first inference network; Based on the prediction probability of each historical inference result of the first inference network within the current time window, calculate the uncertainty index that characterizes the average prediction uncertainty of the first inference network within the current time window. Based on the average value of the uncertainty index for each time window, the drift index of the first inference network at the current moment is determined. Based on all historical drift indicators prior to the current moment, the statistical distribution characteristics of the historical drift indicators are determined, and based on the statistical distribution characteristics and preset sensitivity control parameters, the dynamic threshold of the drift indicator at the current moment is generated. If the drift index at the current moment is greater than the dynamic threshold, it is determined that the first inference network has experienced concept drift.

7. The intelligent power distribution terminal based on native AI according to claim 6, characterized in that, Based on the prediction probabilities of the first inference network for each historical inference result within the current time window, an uncertainty index characterizing the average prediction uncertainty of the first inference network within the current time window is calculated, including: Calculate the logarithm of the prediction probability of the first inference network for each historical inference result within the current time window; Calculate the average of the logarithmic values ​​of all predicted probabilities within the current time window; The negative of the average of the logarithmic values ​​of all predicted probabilities within the current time window is used as the uncertainty index of the first inference network within the current time window.

8. The intelligent power distribution terminal based on native AI according to claim 6, characterized in that, The statistical distribution characteristics include the mean and standard deviation of historical drift indicators; based on the statistical distribution characteristics and preset sensitivity control parameters, a dynamic threshold for the drift indicator at the current moment is generated, including: Calculate the product of the mean of the historical drift index and the sensitivity control parameter, and use the sum of the product and the standard deviation of the historical drift index as the dynamic threshold.

9. The intelligent power distribution terminal based on native AI according to claim 5, characterized in that, Activate the second learning network for local incremental training, including: Determine the target moment when the concept drift of the first inference network is detected, and use the target monitoring data and the corresponding real results of multiple adjacent moments before and after the target moment as target training samples; The second learning network is incrementally trained using the target training samples, and during the incremental training process, replay samples representing the historical data distribution of the target task are extracted from a preset historical sample database to train the second learning network.

10. The intelligent power distribution terminal based on native AI according to claim 5, characterized in that, The network parameters of the first inference network are adjusted based on the network parameters of the trained second learning network, including: After incremental training is completed, obtain the updated network parameters of the second learning network; The parameters of the second learning network are fused into the first inference network through knowledge distillation or parameter smoothing to update the network parameters of the first inference network. In addition, the difference between the updated network parameters of the first inference network and the network parameters before the update of the first inference network is determined as the parameter increment; The parameter increments generated by this update will be encrypted and uploaded to the cloud management platform.

11. A method for constructing an intelligent power distribution terminal based on native AI, characterized in that, include: A native AI computing and control unit is constructed by integrating at least one general-purpose CPU core and at least one AI computing core that shares memory with the general-purpose CPU core; The multi-source sensing module is used to write the collected monitoring data into a shared multidimensional data pool in the form of tensors or vectors. Different microservices are deployed through the general-purpose CPU cores to determine the target microservices and their calling order corresponding to the target task. The determined target microservices are called according to the calling order. The microservices are configured to pull the target monitoring data required by the target task from the shared multidimensional data pool or to initiate inference requests to the native AI computing and control unit. The native AI computing and control unit responds to the inference request, loads the target model from the specified storage location, infers the target monitoring data using the target model, and writes the inference results into the decision event queue. The general-purpose CPU core generates control instructions based on the decision event queue, and the deterministic control and execution module responds to the control instructions and executes the corresponding control actions.

12. A power distribution network system, characterized in that, include: Cloud management platform; And at least one AI-based smart power distribution terminal as described in any one of claims 1-10.

13. A machine-readable storage medium storing instructions thereon, characterized in that, When executed by the processor, the instruction causes the processor to be configured to perform the intelligent power distribution terminal construction method based on native AI as described in claim 11.

14. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the intelligent power distribution terminal construction method based on native AI as described in claim 11.

Images