AI edge server system with high-performance computing power

Abstract

The invention relates to the technical field of server computing, and discloses an AI edge server system with high-performance computing power. Comprising the following modules and components: an end-side cloud collaborative benchmark construction module, an edge computing power elastic capacity expansion module, an intelligent task shunting processing module, a result feedback and cloud collaborative module, a fault self-diagnosis and self-healing module and a full-life-cycle iterative optimization module which are linked in sequence. End-side cloud collaborative computing power scheduling and high-performance computing output of the AI edge server are realized; the invention aims to provide an AI edge server system with high-performance computing power so as to solve the problems that an existing edge server system generally has limitation of fixed computing power configuration, lacks a flexible collaboration and elastic capacity expansion mechanism with a cloud end, cannot dynamically adjust available computing power according to task loads, and cannot dynamically adjust the computing power according to the task loads. The problem that heavy-weight tasks with high computing power requirements are prone to falling into the bottleneck of insufficient computing power, and idle and waste of computing power resources are caused when light-weight tasks are processed is solved.

Description

Technical Field

[0001] This invention relates to the field of server computing technology, and in particular to a high-performance AI edge server system. Background Technology

[0002] AI edge server computing power refers to the computing power of heterogeneous hardware such as CPU, GPU, and NPU, deployed at edge nodes close to the data generation source, integrating the computing power and intelligent scheduling mechanism of these devices. It can receive and process various AI tasks with low latency, such as computer vision and natural language processing, and achieve efficient computing power output and rational resource utilization by coordinating with the cloud to optimize resource allocation. Its core value lies in balancing task processing efficiency, data transmission cost, and service response speed to meet the computing needs for real-time performance and reliability in edge scenarios. Existing edge server systems generally suffer from fixed computing power configurations, lack flexible collaboration and elastic scaling mechanisms with the cloud, and cannot dynamically adjust available computing power according to task load. When faced with heavyweight tasks with high computing power requirements, they are prone to bottlenecks due to insufficient computing power, while when handling lightweight tasks, they will cause idle and wasteful computing resources. Summary of the Invention

[0003] The purpose of this invention is to provide a high-performance AI edge server system to solve the problem that existing edge server systems generally have the limitation of fixed computing power configuration, lack flexible collaboration and elastic expansion mechanism with the cloud, cannot dynamically adjust available computing power according to task load, are prone to insufficient computing power bottleneck when facing heavyweight tasks with high computing power requirements, and will cause idle and wasteful computing power resources when handling lightweight tasks.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: A high-performance AI edge server system includes the following modules and components: The module consists of a baseline construction module for edge-cloud collaboration, an elastic scaling module for edge computing power, an intelligent task offloading and processing module, a result feedback and cloud collaboration module, a fault self-diagnosis and self-healing module, and a full lifecycle iterative optimization module. These modules work together in sequence to achieve edge-cloud collaborative computing power scheduling and high-performance computing output for AI edge servers.

[0005] As a further improvement to this technical solution: the edge-cloud collaborative benchmark construction module includes an encrypted communication unit, a model repository synchronization unit, a computing power benchmark pool initialization unit, and a model adaptation library construction unit; the encrypted communication unit includes an AES-256 encryption chip and a bidirectional communication interface; the model repository synchronization unit includes a cloud model indexer, a local model cache, and a version management tool; the computing power benchmark pool initialization unit includes a hardware computing power detector, a computing power binding engine, and a threshold configuration tool. The computing power benchmark pool initialization unit automatically detects the computing power limit (Cedge) of the local hardware on the edge server, synchronizes the peak elastic computing power (Ccloud) available in the cloud, establishes an integrated local and cloud computing power benchmark pool, sets computing power expansion trigger thresholds and contraction thresholds, and the computing power binding formula is: ,in This represents the maximum available computing power under edge-cloud collaboration, measured in TOPS. This represents the actual computing power limit of the local hardware on the edge server, measured in TOPS. This represents the peak elastic computing power that the cloud can allocate to this edge node, measured in TOPS. This is the computing power collaboration coefficient, ranging from 0.8 to 1.0, dynamically adjusted according to network latency. It is set to 1.0 when the latency is ≤100ms and 0.8 when the latency is >200ms.

[0006] As a further improvement to this technical solution: the model adaptation library construction unit includes a model computing power requirement assessment tool, a hardware compatibility detection module, and a tag generator. The model adaptation library construction unit evaluates the computing power requirement (Cmodel) and execution latency (Tmodel) of all models synchronized locally, detects the compatibility between the models and edge hardware, and labels each model with computing power level, compatible hardware, and latency range. The compatibility determination formula is: ; in This is a model fit indicator, where 1 = fit and 0 = misfit. This represents the single-task computing power requirement of the model, expressed in TOPS. This represents the upper limit of local edge computing power, measured in TOPS. The theoretical execution latency of the model on edge hardware is expressed in milliseconds (ms). This represents the maximum allowed delay for the current task, in milliseconds (ms).

[0007] As a further improvement to this technical solution: the edge computing power elastic expansion module includes a computing load monitoring unit, an expansion judgment algorithm unit, a computing power channel establishment unit, and a computing power contraction control unit; the computing load monitoring unit includes CPU, GPU, and NPU load sensors, a task queue counter, and a latency timer. The computing load monitoring unit collects the local hardware's computing power utilization rate (Ucurrent), task queue length (Qlength), and average latency (Tavg) of executed tasks in real time at a frequency of 10Hz; the expansion judgment algorithm unit includes an algorithm processing chip, a threshold comparison engine, and a decision generation module. Based on the monitoring data, the expansion judgment algorithm unit calculates the expansion demand index using an expansion judgment algorithm. The expansion judgment formula is: ,in To expand the demand index, This represents the length of the current queue of tasks to be executed, in units of tasks. This represents the average computing power requirement per task, expressed in TOPS per task. It is obtained from the average computing power requirement of the corresponding model for this type of task, extracted from the model adaptation library. This represents the upper limit of local edge computing power, measured in TOPS. The current local computing power utilization rate is represented by a value of 0-1; the computing power channel establishment unit includes a high-speed data forwarding module, a cloud computing power request interface, and a link quality detection tool; the computing power contraction control unit includes a contraction condition detection module and a cloud computing power release interface.

[0008] As a further improvement to this technical solution: the intelligent task offloading and processing module includes a task parsing and classification unit, a local heterogeneous computing unit, a cloud collaborative computing unit, and a result fusion unit; the task parsing and classification unit includes a task type identifier, a computing power requirement estimation module, and a priority determination tool. The task parsing and classification unit receives external input tasks, identifies the task type, estimates the computing power required for the task (Ctask) and the allowable latency (Ttask), and determines the task priority (P1, P2, P3) based on business rules. Lightweight tasks (Ctask ≤ 0.5 TOPS and Ttask ≥ 50ms) are allocated to local processing, while heavyweight tasks (Ctask > 0.5 TOPS or Ttask < 50ms) are offloaded to cloud processing. P1-level tasks are preferentially allocated local computing power, and if local computing power is insufficient, cloud computing power is called synchronously. The local heterogeneous computing unit includes a CPU scheduling core, a GPU / NPU computing core, and a data preprocessing engine; the cloud collaborative computing unit includes a cloud computing power call interface, a data compression module, and a result simplification engine; the result fusion unit includes a result verification module and an information integration tool.

[0009] As a further improvement to this technical solution: the result feedback and cloud collaboration module includes a result receiving and integration unit, a status synchronization unit, and a cache management unit; the result receiving and integration unit includes a multi-source result receiving interface and a data integration engine; the status synchronization unit includes a status reporting module and a cloud command receiving interface; the cache management unit includes a Redis cache chip, a cache strategy engine, and an expiration determination module. The cache management unit identifies high-frequency repetitive tasks, stores their results in the cache, and sets the cache expiration time to be adjustable from 10-60 seconds according to the task type. A second call before expiration directly returns the cached result. The cache hit rate calculation formula is: ,in Cache hit rate, in percentage. This represents the number of tasks completed through caching, expressed in times. This represents the total number of times the task was executed, expressed in times.

[0010] As a further improvement to this technical solution: the fault self-diagnosis and self-healing module includes a multi-dimensional monitoring unit, a fault location algorithm unit, a self-healing strategy execution unit, and a cloud takeover trigger unit; the multi-dimensional monitoring unit includes a hardware status sensor, a computing power output detector, a network link monitor, and a model running error calculator; the fault location algorithm unit includes a fault tree algorithm engine and an abnormal feature matching module. Based on multi-dimensional monitoring data, the fault location algorithm unit locates the fault type using a fault tree algorithm, constructs a fault type and abnormal feature mapping library, and the fault probability calculation formula is: ,in The probability of a certain type of failure occurring, expressed as a percentage. This represents the number of currently detected abnormal features that match the fault feature, expressed in units of individual matches. This represents the total number of characteristics of the fault, expressed in units of individual items. The weighting coefficient for this fault ranges from 0.8 to 1.2; the self-healing strategy execution unit includes a strategy execution engine, a hardware control interface, and a software repair module; the cloud takeover triggering unit includes a severe fault determination module and a cloud takeover request interface.

[0011] As a further improvement to this technical solution: the full lifecycle iterative optimization module includes a hybrid storage unit, a federated learning iterative unit, and a scheduling strategy optimization unit; the hybrid storage unit includes a MySQL database, a MongoDB database, and a data retrieval engine; the federated learning iterative unit includes a federated learning client, a model parameter update module, and an accuracy verification tool. Each month, based on task data accumulated at the edge, the federated learning iterative unit participates in cloud model training through federated learning, receives lightweight model parameters optimized from the cloud, and updates the local model adaptation library. The model parameter iteration formula is: ,in These are the model parameters after iteration. These are the current model parameters before the iteration. The learning rate is 0.001. The number of edge nodes participating in federated learning, expressed in units. For the i-th edge node based on local data The gradient of the loss function is calculated. The local task data is for the i-th edge node; the scheduling strategy optimization unit includes a strategy analysis module and a parameter adjustment tool.

[0012] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention effectively overcomes the limitations of fixed computing power on a single edge server by employing an edge-cloud collaborative architecture and an elastic computing power scheduling mechanism. It can dynamically call cloud resources according to task load, achieving elastic scaling of computing power. This satisfies the high computing power requirements of heavyweight tasks while avoiding the waste caused by lightweight tasks occupying cloud resources. The intelligent task diversion and cache reuse design allows different types of tasks to receive precisely matched processing resources, significantly improving the response speed of high-frequency repetitive tasks. At the same time, it reduces unnecessary computing power consumption and data transmission redundancy, significantly improving overall computing efficiency and resource utilization, and adapting to the diverse AI task processing needs in multiple scenarios.

[0013] 2. The fault self-diagnosis and self-healing functions integrated in this invention can quickly locate and resolve common operational faults. In the event of severe faults, cloud takeover can ensure service continuity, significantly reducing manual maintenance costs and business interruption risks. Relying on the iterative optimization mechanism of federated learning, while ensuring local data privacy and security, the model accuracy and scheduling strategy are continuously optimized, allowing the system to adapt to changes in business scenarios over the long term and continuously improve operational stability and processing performance. The overall solution forms a complete closed loop from computing power scheduling and task processing to fault response and long-term optimization, taking into account computing performance, operational reliability, and scenario adaptability, providing an efficient and sustainable solution for AI edge computing.

[0014] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it according to the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Specific embodiments of the present invention are given in detail below with reference to the accompanying drawings. Attached Figure Description

[0015] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings: Figure 1 This is a system diagram of a high-performance AI edge server system. Detailed Implementation

[0016] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are for illustrative purposes only and are not intended to limit the scope of the invention. The invention is described more specifically in the following paragraphs by way of example with reference to the accompanying drawings. It should be noted that the drawings are in a very simplified form and use non-precise proportions, and are only used to facilitate and clarify the illustration of the embodiments of the present invention.

[0017] Please see Figure 1 In this embodiment of the invention, a high-performance AI edge server system includes the following modules and components: The module consists of a baseline construction module for edge-cloud collaboration, an elastic expansion module for edge computing power, an intelligent task offloading and processing module, a result feedback and cloud collaboration module, a fault self-diagnosis and self-healing module, and a full lifecycle iterative optimization module. These modules work together in sequence to achieve edge-cloud collaborative computing power scheduling and high-performance computing output for AI edge servers. Specifically, the six modules cover the entire process of benchmark construction, computing power expansion, task processing, result feedback, fault self-healing, and iterative optimization, forming a closed-loop operation logic to ensure the integrity and high performance of the system from initialization to long-term operation.

[0018] The edge-cloud collaborative benchmark construction module includes an encrypted communication unit, a model repository synchronization unit, a computing power benchmark pool initialization unit, and a model adaptation library construction unit. The encrypted communication unit includes an AES-256 encryption chip and a bidirectional communication interface. The model repository synchronization unit includes a cloud-based model indexer, a local model cache, and a version management tool. The computing power benchmark pool initialization unit includes a hardware computing power detector, a computing power binding engine, and a threshold configuration tool. This unit automatically detects the computing power limit (Cedge) of the edge server's local hardware, synchronizes with the peak elastic computing power (Ccloud) available in the cloud, establishes an integrated local and cloud-based computing power benchmark pool, and sets the computing power expansion trigger threshold and contraction threshold. The computing power binding formula is: ,in This represents the maximum available computing power under edge-cloud collaboration, measured in TOPS. This represents the actual computing power limit of the local hardware on the edge server, measured in TOPS. This represents the peak elastic computing power that the cloud can allocate to this edge node, measured in TOPS. This is the computing power collaboration coefficient, ranging from 0.8 to 1.0, dynamically adjusted according to network latency. It is set to 1.0 when the latency is ≤100ms and 0.8 when the latency is >200ms. Specifically, the encrypted communication unit includes an AES-256 encryption chip and a two-way communication interface, which are used to establish a secure data channel between the edge and the cloud, ensuring that transmitted data is not stolen or tampered with, and supporting basic communication for end-edge-cloud collaboration. Model repository synchronization unit: includes cloud model indexer, local model cache, and version management tools, used to synchronize lightweight models and full models in the cloud, record model version and adapted hardware information, and provide diverse model selection for the edge side; The computing power benchmark pool initialization unit includes a hardware computing power detector, a computing power binding engine, and a threshold configuration tool; the hardware computing power detector automatically identifies the computing power limit of edge local hardware (CPU, GPU, NPU). Computing power binding engine synchronizes with cloud elastic computing power peak. The formula establishes the connection between local and cloud computing power; the threshold configuration tool sets expansion (local computing power utilization ≥ 85%) and contraction (local computing power utilization ≤ 60%) thresholds; Computing power binding formula annotation: ,in This represents the maximum available computing power under edge-cloud collaboration, measured in TOPS. This represents the actual computing power limit of the local hardware on the edge server, measured in TOPS. This represents the peak elastic computing power that the cloud can allocate to this edge node, measured in TOPS. This is the computing power collaboration coefficient, ranging from 0.8 to 1.0, dynamically adjusted according to network latency. It is set to 1.0 when the latency is ≤100ms and 0.8 when the latency is >200ms. The formula's purpose is to break the limitations of fixed computing power on a single edge server, quantify the upper limit of collaborative computing power between local and cloud environments, and provide a benchmark for subsequent elastic scaling.

[0019] The model adaptation library building unit includes a model computing power requirement assessment tool, a hardware compatibility detection module, and a label generator. For all models synchronized locally, this unit evaluates their single-frame, single-task computing power requirement (Cmodel) and execution latency (Tmodel), detects the model's compatibility with edge hardware, and labels each model with its computing power level, compatible hardware, and latency range. The compatibility determination formula is: ; in This is a model fit indicator, where 1 = fit and 0 = misfit. This represents the single-task computing power requirement of the model, expressed in TOPS. This represents the upper limit of local edge computing power, measured in TOPS. The theoretical execution latency of the model on edge hardware is expressed in milliseconds (ms). The maximum allowed delay for the current task, in milliseconds; Specifically, the model computational requirement assessment tool is used to evaluate the computational requirements of each model per frame / per task. Execution delay Quantify the resource consumption of the model; Hardware compatibility detection module: Detects the compatibility between the model and edge hardware (such as whether the GPU supports CUDA acceleration of the model) to avoid loading incompatible models; Tag generator: Labels each model with computing power level, compatible hardware, and latency range tags to facilitate quick matching during subsequent task splitting; Adaptability determination formula annotation: ; in This is a model fit indicator, where 1 = fit and 0 = misfit. This represents the single-task computing power requirement of the model, expressed in TOPS. This represents the upper limit of local edge computing power, measured in TOPS. The theoretical execution latency of the model on edge hardware is expressed in milliseconds (ms). This represents the maximum allowable latency for the current task, expressed in milliseconds. The purpose of this formula is to filter out models whose computing power requirements are within the edge's carrying capacity and whose latency meets the standards, thereby preventing task execution failures due to model incompatibility or excessive resource consumption and improving the efficiency of task processing on the edge side.

[0020] The edge computing power elastic expansion module includes a computing load monitoring unit, an expansion judgment algorithm unit, a computing power channel establishment unit, and a computing power contraction control unit. The computing load monitoring unit includes CPU, GPU, and NPU load sensors, a task queue counter, and a latency timer. It collects local hardware computing power utilization (Ucurrent), task queue length (Qlength), and average latency (Tavg) of executed tasks in real time at a frequency of 10Hz. The expansion judgment algorithm unit includes an algorithm processing chip, a threshold comparison engine, and a decision generation module. Based on the monitoring data, it calculates the expansion demand index using an expansion judgment algorithm. The expansion judgment formula is as follows: ,in To expand the demand index, This represents the length of the current queue of tasks to be executed, in units of tasks. This represents the average computing power requirement per task, expressed in TOPS per task. It is obtained from the average computing power requirement of the corresponding model for this type of task, extracted from the model adaptation library. This represents the upper limit of local edge computing power, measured in TOPS. The current local computing power utilization rate is represented by a value of 0-1; the computing power channel establishment unit includes a high-speed data forwarding module, a cloud computing power request interface, and a link quality detection tool; the computing power contraction control unit includes a contraction condition detection module and a cloud computing power release interface; Specifically, the computing load monitoring unit includes CPU, GPU, and NPU load sensors, a task queue counter, and a latency timer, which collect local computing power utilization in real time. Task queue length Average delay of executed tasks (Sampling frequency 10Hz) to provide real-time data support for capacity expansion decisions; Expansion determination algorithm unit: includes algorithm operation chip, threshold comparison engine, and decision generation module; the algorithm operation chip calculates the expansion demand index based on monitoring data; the threshold comparison engine compares the index with the preset threshold (1.2); the decision generation module outputs the decision of expansion, no expansion, and priority processing of high-priority tasks; Expansion determination formula annotation: ,in To expand the demand index, quantifying the current computing power gap, This represents the length of the current queue of tasks to be executed, in units of tasks. This represents the average computing power requirement per task, expressed in TOPS per task (taken from the average computing power requirement of the corresponding model for this type of task in the model adaptation library). This represents the upper limit of local edge computing power, measured in TOPS. The current local computing power utilization rate is represented by a value of 0-1. The computing power channel establishment unit includes a high-speed data forwarding module, a cloud computing power request interface, and a link quality detection tool. The computing power contraction control unit includes a contraction condition detection module and a cloud computing power release interface. The formula's function is to accurately determine the size of the computing power gap, avoid blind expansion or insufficient computing power without expansion, and achieve efficient utilization of computing power resources. Computing power channel establishment unit: includes a high-speed data forwarding module, a cloud computing power request interface, and a link quality detection tool. After triggering expansion, it establishes dedicated data channels between the edge and the cloud, and detects link bandwidth (≥100Mbps) and latency (≤200ms) to ensure data transmission efficiency. Computing power contraction control unit: Includes a contraction condition detection module and a cloud computing power release interface. When the local computing power utilization rate is ≤60% and the task queue length is ≤5, cloud computing power is released to reduce usage costs.

[0021] The intelligent task offloading and processing module includes a task parsing and classification unit, a local heterogeneous computing unit, a cloud collaborative computing unit, and a result fusion unit. The task parsing and classification unit includes a task type identifier, a computing power requirement estimation module, and a priority determination tool. This unit receives external input tasks, identifies the task type, estimates the required computing power (Ctask) and allowable latency (Ttask), and determines the task priority (P1, P2, P3) based on business rules. Lightweight tasks (Ctask ≤ 0.5 TOPS and Ttask ≥ 50ms) are allocated to local processing, while heavyweight tasks (Ctask > 0.5 TOPS or Ttask < 50ms) are offloaded to cloud processing. P1-level tasks are prioritized for local computing power; if local computing power is insufficient, cloud computing power is simultaneously invoked. The local heterogeneous computing unit includes a CPU scheduling core, a GPU / NPU computing core, and a data preprocessing engine. The cloud collaborative computing unit includes a cloud computing power invocation interface, a data compression module, and a result simplification engine. The result fusion unit includes a result verification module and an information integration tool. Specifically, the task parsing and classification unit includes a task type identifier, a computing power requirement estimation module, and a priority determination tool. The task type identifier identifies the category to which the task belongs (computer vision, NLP, data statistics, etc.). The computing power requirement estimation module estimates the computing power required for the task (Ctask) and the allowable latency (Ttask). The priority determination tool determines the task priority (P1, P2, P3). Task allocation logic: Lightweight tasks with Ctask≤0.5TOPS and Ttask≥50ms are allocated to local processing, while heavyweight tasks with Ctask>0.5TOPS or Ttask<50ms are allocated to cloud processing; P1 level tasks prioritize local computing power, and cloud computing power is called up synchronously when local computing power is insufficient. Function: To enable fine-grained task allocation, avoid resource waste caused by lightweight tasks occupying cloud computing power or heavyweight tasks occupying local computing power, and improve overall task processing efficiency. Local heterogeneous computing unit: includes CPU scheduling core, GPU, NPU computing core, and data preprocessing engine. After loading the lightweight model, the CPU is responsible for task scheduling and data preprocessing, while the GPU and NPU are responsible for parallel computing, maximizing the advantages of local hardware heterogeneous computing. Cloud-based collaborative computing unit: includes cloud computing power call interface, data compression module, and result simplification engine. It compresses and transmits the simplified data after edge preprocessing to the cloud. After calling the full model for calculation, it receives and simplifies the results (removing redundant information) to reduce the amount of data transmission. The results fusion unit includes a results verification module and information integration tools. It verifies and compares local and cloud processing results, integrates key information to generate the final output (control commands, analysis reports), and ensures the accuracy of the results.

[0022] The result feedback and cloud collaboration module includes a result receiving and integration unit, a status synchronization unit, and a cache management unit. The result receiving and integration unit includes a multi-source result receiving interface and a data integration engine. The status synchronization unit includes a status reporting module and a cloud command receiving interface. The cache management unit includes a Redis cache chip, a cache strategy engine, and an expiration determination module. The cache management unit identifies high-frequency repetitive tasks, stores their results in the cache, and sets the cache expiration time (adjustable from 10-60 seconds) according to the task type. A second call before expiration directly returns the cached result. The cache hit rate is calculated using the following formula: ,in Cache hit rate, in percentage. This represents the number of tasks completed through caching, expressed in times. This represents the total number of times the task was executed, expressed in times. Specifically, the result receiving and integration unit includes a multi-source result receiving interface and a data integration engine. It receives local and cloud processing results, integrates them by task ID, and supplements metadata such as execution time and computing power consumption for easy subsequent traceability. The status synchronization unit includes a status reporting module and a cloud instruction receiving interface. It synchronizes the task execution status (pending execution, in execution, completed, failed) and computing power usage data to the cloud in real time, and receives cloud global scheduling strategy adjustment instructions to achieve end-edge-cloud status communication. The cache management unit includes a Redis cache chip, a cache strategy engine, and an expiration determination module. The Redis cache chip stores the results of frequently repeated tasks. The cache strategy engine sets an adjustable expiration time of 10-60 seconds according to the task type. If the cache has not expired when the task is called a second time, the expiration determination module will directly return the result. The formula for calculating cache hit rate is: ,in This represents the cache hit rate, expressed as a percentage (target ≥ 85%). This represents the number of tasks completed through caching, expressed in times. The total number of times the task is executed is expressed in times. The purpose of the formula is to quantify the effective utilization of the cache and dynamically adjust the cache expiration time (extend by 10 seconds for a hit rate <80%, shorten by 5 seconds for a hit rate >90%), thereby reducing the computational consumption and execution latency of high-frequency repetitive tasks.

[0023] The fault self-diagnosis and self-healing module includes a multi-dimensional monitoring unit, a fault location algorithm unit, a self-healing strategy execution unit, and a cloud takeover trigger unit. The multi-dimensional monitoring unit includes hardware status sensors, a computing power output detector, a network link monitor, and a model execution error calculator. The fault location algorithm unit includes a fault tree algorithm engine and an anomaly feature matching module. Based on multi-dimensional monitoring data, the fault location algorithm unit uses a fault tree algorithm to locate the fault type and constructs a fault type-anomaly feature mapping library. The fault probability calculation formula is as follows: ,in The probability of a certain type of failure occurring, expressed as a percentage. This represents the number of currently detected abnormal features that match the fault feature, expressed in units of individual matches. This represents the total number of characteristics of the fault, expressed in units of individual items. The weighting coefficient for this fault ranges from 0.8 to 1.2; the self-healing strategy execution unit includes a strategy execution engine, a hardware control interface, and a software repair module; the cloud takeover triggering unit includes a severe fault determination module and a cloud takeover request interface. Specifically, the multi-dimensional monitoring unit includes hardware status sensors, computing power output detectors, network link monitors, and model running error calculators, which collect hardware operating status, computing power output stability, network connection status, and model running errors in real time to comprehensively capture system anomalies. Fault location algorithm unit: includes fault tree algorithm engine and abnormal feature matching module; the fault tree algorithm engine locates fault types (hardware, software, network faults) based on monitoring data; the abnormal feature matching module builds a fault type and abnormal feature mapping library and calculates the probability of fault occurrence through formulas; Failure probability formula annotation: ,in The probability of a certain type of failure occurring, expressed as a percentage. This represents the number of currently detected abnormal features that match the fault feature, expressed in units of individual matches. This represents the total number of characteristics of the fault, expressed in units of individual items. The weighting coefficient for this fault ranges from 0.8 to 1.2, set according to the frequency of occurrence; the self-healing strategy execution unit includes a strategy execution engine, a hardware control interface, and a software repair module; the cloud takeover triggering unit includes a severe fault determination module and a cloud takeover request interface; the formula's function is to quickly and accurately locate the fault type (location time ≤ 3s), solving the problems of difficult and time-consuming fault location in traditional edge servers; Self-healing strategy execution unit: includes strategy execution engine, hardware control interface and software repair module. It executes self-healing strategy according to fault type (hardware overheating frequency reduction, network interruption link switching, model error exceeding standard loading backup model, etc.) to solve common faults without manual intervention. Cloud takeover trigger unit: Includes a severe fault determination module and a cloud takeover request interface. When a severe fault (hardware damage, system crash) is detected, all tasks are diverted to the cloud to ensure service continuity.

[0024] The full lifecycle iterative optimization module includes a hybrid storage unit, a federated learning iterative unit, and a scheduling strategy optimization unit. The hybrid storage unit includes a MySQL database, a MongoDB database, and a data retrieval engine. The federated learning iterative unit includes a federated learning client, a model parameter update module, and an accuracy verification tool. Each month, based on task data accumulated at the edge, the federated learning iterative unit participates in cloud model training via federated learning, receives the optimized lightweight model parameters from the cloud, and updates the local model adaptation library. The model parameter iteration formula is: ,in These are the model parameters after iteration. These are the current model parameters before the iteration. The learning rate is 0.001. The number of edge nodes participating in federated learning, expressed in units. For the i-th edge node based on local data The gradient of the loss function is calculated. This represents the local task data for the i-th edge node; the scheduling strategy optimization unit includes a strategy analysis module and a parameter adjustment tool. Specifically, the hybrid storage unit includes a MySQL database, a MongoDB database, and a data retrieval engine. It categorizes and stores task data, computing power scheduling records, and fault logs (retained for 1 year). It supports multi-dimensional retrieval by device, time, and task type, providing data support for iterative optimization. Federated Learning Iteration Unit: Includes a federated learning client, a model parameter update module, and an accuracy verification tool; the federated learning client participates in cloud model training based on edge local task data (only transmitting parameter gradients, without leaking the original data); the model parameter update module receives the optimized model parameters from the cloud and updates the local model adaptation library; the accuracy verification tool verifies the computational accuracy of the model after iteration; Model parameter iteration formula annotation: ,in These are the model parameters after iteration (such as convolution kernel weights and threshold parameters). These are the current model parameters before the iteration. The learning rate is set to 0.001 (to control the step size of the parameter update). The number of edge nodes participating in federated learning, expressed in units. For the i-th edge node based on local data The gradient of the loss function is calculated. This is the local task data of the i-th edge node (not transmitted externally); the scheduling strategy optimization unit includes a strategy analysis module and a parameter adjustment tool; the formula's function is to continuously optimize the computational accuracy of the local lightweight model (target recognition accuracy ≥ 98%) while ensuring data privacy, so that the model can adapt to the long-term changes in edge business scenarios; Scheduling strategy optimization unit: Includes strategy analysis module and parameter adjustment tool. Based on global data in the cloud and local storage data, it analyzes computing power scheduling efficiency and task diversion rationality, and adjusts expansion threshold, task diversion judgment parameters, cache expiration time, etc., to maintain the optimal operating state of the system.

[0025] The method of use and working principle of this invention are as follows: Usage: Before use, system deployment and initialization must be completed, a secure connection between the edge server and the cloud must be established, the cloud model repository and computing power scheduling strategy must be synchronized, and local computing power benchmark thresholds and model adaptation rules must be configured. After the system starts, it automatically collects various input tasks and hardware operation data. Task parsing, computing power status assessment, and intelligent task allocation can be completed without manual intervention. Lightweight tasks are processed in local heterogeneous computing units, while heavyweight tasks are allocated to cloud collaborative computing. The processing results are directly output after integration, and the results of high-frequency repetitive tasks are automatically cached to improve response speed. During operation, the system monitors hardware status, computing power load, and task execution in real time. Common faults can be automatically self-healed, and in the event of a severe fault, cloud takeover will be triggered to ensure uninterrupted service. In long-term use, the system will automatically archive various operation data and iteratively optimize the local model and scheduling strategy monthly through federated learning to maintain long-term stable operation of the system.

[0026] Working Principle: With edge-cloud collaboration as its core logic, the baseline construction module establishes a computing power correlation and model adaptation system between the local and cloud environments, laying the foundation for subsequent operation. When a task is input, the load monitoring module senses the computing power status and task queue status in real time. The expansion judgment unit accurately identifies computing power gaps and calls upon cloud elastic computing power as needed. The distribution module allocates local or cloud processing resources according to task characteristics and priorities, achieving elastic scaling of computing power. The processing results from the local and cloud environments are integrated and verified by the feedback module before being output. Simultaneously, the caching module stores and reuses results from high-frequency tasks. The fault self-diagnosis module monitors the system's operating status throughout the process, quickly resolving common problems through fault location and self-healing strategies, triggering a cloud takeover mechanism in case of severe faults. Finally, the iterative optimization module continuously optimizes model accuracy and scheduling strategies based on historical operating data through federated learning, forming a closed-loop operating mechanism of baseline construction, computing power scheduling, task processing, fault self-healing, and iterative optimization, ultimately achieving high-performance and highly reliable edge computing output.

[0027] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Those skilled in the art can readily implement the present invention based on the accompanying drawings and the description above. However, any modifications, alterations, or variations made by those skilled in the art without departing from the scope of the present invention, using the disclosed technical content, are equivalent embodiments of the present invention. Furthermore, any modifications, alterations, or variations made to the above embodiments based on the essential technology of the present invention are still within the protection scope of the present invention.

Claims

1. A high-performance AI edge server system, characterized in that, Includes the following modules and components: The module consists of a baseline construction module for edge-cloud collaboration, an elastic scaling module for edge computing power, an intelligent task offloading and processing module, a result feedback and cloud collaboration module, a fault self-diagnosis and self-healing module, and a full lifecycle iterative optimization module. These modules work together in sequence to achieve edge-cloud collaborative computing power scheduling and high-performance computing output for AI edge servers.

2. The high-performance computing AI edge server system according to claim 1, characterized in that, The edge-cloud collaborative benchmark construction module includes an encrypted communication unit, a model repository synchronization unit, a computing power benchmark pool initialization unit, and a model adaptation library construction unit. The encrypted communication unit includes an AES-256 encryption chip and a bidirectional communication interface. The model repository synchronization unit includes a cloud-based model indexer, a local model cache, and a version management tool. The computing power benchmark pool initialization unit includes a hardware computing power detector, a computing power binding engine, and a threshold configuration tool. This unit automatically detects the computing power limit (Cedge) of the edge server's local hardware, synchronizes with the peak elastic computing power (Ccloud) available in the cloud, establishes an integrated local and cloud-based computing power benchmark pool, and sets computing power expansion and contraction thresholds. The computing power binding formula is: ,in This represents the maximum available computing power under edge-cloud collaboration, measured in TOPS. This represents the actual computing power limit of the local hardware on the edge server, measured in TOPS. This represents the peak elastic computing power that the cloud can allocate to this edge node, measured in TOPS. This is the computing power collaboration coefficient, ranging from 0.8 to 1.0, dynamically adjusted according to network latency. It is set to 1.0 when the latency is ≤100ms and 0.8 when the latency is >200ms.

3. The high-performance computing AI edge server system according to claim 2, characterized in that, The model adaptation library construction unit includes a model computing power requirement assessment tool, a hardware compatibility detection module, and a tag generator. This unit evaluates the computing power requirement (Cmodel) and execution latency (Tmodel) of all models synchronized locally, assesses the compatibility between the models and edge hardware, and labels each model with its computing power level, compatible hardware, and latency range. The compatibility determination formula is as follows: ； in This is a model fit indicator, where 1 = fit and 0 = misfit. This represents the single-task computing power requirement of the model, expressed in TOPS. This represents the upper limit of local edge computing power, measured in TOPS. The theoretical execution latency of the model on edge hardware is expressed in milliseconds (ms). This represents the maximum allowed delay for the current task, in milliseconds (ms).

4. The high-performance computing AI edge server system according to claim 1, characterized in that, The edge computing power elastic expansion module includes a computing load monitoring unit, an expansion judgment algorithm unit, a computing power channel establishment unit, and a computing power contraction control unit. The computing load monitoring unit includes CPU, GPU, and NPU load sensors, a task queue counter, and a latency timer. This unit collects local hardware computing power utilization (Ucurrent), task queue length (Qlength), and average latency (Tavg) of executed tasks in real time at a frequency of 10Hz. The expansion judgment algorithm unit includes an algorithm processing chip, a threshold comparison engine, and a decision generation module. Based on the monitoring data, the expansion judgment algorithm unit calculates the expansion demand index using an expansion judgment algorithm. The expansion judgment formula is: ,in To expand the demand index, This represents the length of the current queue of tasks to be executed, in units of tasks. This represents the average computing power requirement per task, expressed in TOPS per task. It is obtained from the average computing power requirement of the corresponding model for this type of task, extracted from the model adaptation library. This represents the upper limit of local edge computing power, measured in TOPS. The current local computing power utilization rate is represented by a value of 0-1; the computing power channel establishment unit includes a high-speed data forwarding module, a cloud computing power request interface, and a link quality detection tool; the computing power contraction control unit includes a contraction condition detection module and a cloud computing power release interface.

5. The high-performance computing AI edge server system according to claim 1, characterized in that, The intelligent task offloading and processing module includes a task parsing and classification unit, a local heterogeneous computing unit, a cloud collaborative computing unit, and a result fusion unit. The task parsing and classification unit includes a task type identifier, a computing power requirement estimation module, and a priority determination tool. The task parsing and classification unit receives external input tasks, identifies the task type, estimates the computing power required for the task (Ctask) and the allowable latency (Ttask), and determines the task priority (P1, P2, P3) based on business rules. Lightweight tasks (Ctask ≤ 0.5 TOPS and Ttask ≥ 50ms) are allocated to local processing, while heavyweight tasks (Ctask > 0.5 TOPS or Ttask < 50ms) are offloaded to cloud processing. P1-level tasks are prioritized for local computing power; if local computing power is insufficient, cloud computing power is synchronously called. The local heterogeneous computing unit includes a CPU scheduling core, a GPU / NPU computing core, and a data preprocessing engine. The cloud-based collaborative computing unit includes a cloud computing power call interface, a data compression module, and a result simplification engine; the result fusion unit includes a result verification module and an information integration tool.

6. The high-performance computing AI edge server system according to claim 1, characterized in that, The result feedback and cloud collaboration module includes a result receiving and integration unit, a status synchronization unit, and a cache management unit. The result receiving and integration unit includes a multi-source result receiving interface and a data integration engine. The status synchronization unit includes a status reporting module and a cloud command receiving interface. The cache management unit includes a Redis cache chip, a cache strategy engine, and an expiration determination module. The cache management unit identifies high-frequency repetitive tasks, stores their results in the cache, and sets the cache expiration time (adjustable from 10-60 seconds) according to the task type. A second call before expiration directly returns the cached result. The cache hit rate is calculated using the following formula: ,in Cache hit rate, in percentage. This represents the number of tasks completed through caching, expressed in times. This represents the total number of times the task was executed, expressed in times.

7. The high-performance computing AI edge server system according to claim 1, characterized in that, The fault self-diagnosis and self-healing module includes a multi-dimensional monitoring unit, a fault location algorithm unit, a self-healing strategy execution unit, and a cloud takeover trigger unit; the multi-dimensional monitoring unit includes a hardware status sensor, a computing power output detector, a network link monitor, and a model running error calculator. The fault location algorithm unit includes a fault tree algorithm engine and an anomaly feature matching module. Based on multi-dimensional monitoring data, the fault location algorithm unit locates the fault type through a fault tree algorithm, constructs a fault type-anomaly feature mapping library, and the fault probability calculation formula is: ,in The probability of a certain type of failure occurring, expressed as a percentage. This represents the number of currently detected abnormal features that match the fault feature, expressed in units of individual matches. This represents the total number of features of the fault, expressed in units of individual items. The weighting coefficient for this fault ranges from 0.8 to 1.2; the self-healing strategy execution unit includes a strategy execution engine, a hardware control interface, and a software repair module. The cloud takeover triggering unit includes a severe fault determination module and a cloud takeover request interface.

8. The high-performance computing AI edge server system according to claim 1, characterized in that, The full lifecycle iterative optimization module includes a hybrid storage unit, a federated learning iterative unit, and a scheduling strategy optimization unit. The hybrid storage unit includes a MySQL database, a MongoDB database, and a data retrieval engine. The federated learning iterative unit includes a federated learning client, a model parameter update module, and an accuracy verification tool. Each month, based on task data accumulated at the edge, the federated learning iterative unit participates in cloud model training via federated learning, receives the optimized lightweight model parameters from the cloud, and updates the local model adaptation library. The model parameter iteration formula is: ,in These are the model parameters after iteration. These are the current model parameters before the iteration. The learning rate is 0.

001. The number of edge nodes participating in federated learning, expressed in units. For the i-th edge node based on local data The gradient of the loss function is calculated. This represents the local task data for the i-th edge node; The scheduling strategy optimization unit includes a strategy analysis module and a parameter adjustment tool.

Images