Multistage resource dynamic scheduling method based on heterogeneous edge framework

Abstract

The invention relates to the technical field of health management and intelligent operation and maintenance of large power generation equipment, in particular to a multistage resource dynamic scheduling method based on a heterogeneous edge framework, and the framework comprises a network switching unit, a signal processing unit, a storage unit and a plurality of core computing board units. Each core computing board unit comprises a plurality of heterogeneous computing modules; the method comprises the following steps: distributing tasks to corresponding core computing board units; decomposing and classifying the task to form a plurality of sub-tasks, and adaptively matching the plurality of sub-tasks with the characteristics of the heterogeneous computing module; distributing each subtask to a heterogeneous computing module which is matched with the subtask and is lower in load or higher in energy efficiency ratio; a calculation flow is constructed based on a sub-task dependency relationship, the sub-tasks are graded according to real-time performance, dynamic balanced distribution is achieved through an intelligent scheduling algorithm, and it is ensured that the load of each heterogeneous calculation module tends to be balanced. Through the scheduling and optimizing method, the efficiency of task scheduling and resource allocation can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of health management and intelligent operation and maintenance of large-scale power generation equipment, and in particular to a multi-level resource dynamic scheduling method based on a heterogeneous edge framework. Background Art

[0002] Traditional monitoring methods for the operation, maintenance, and overhaul of large generator sets rely primarily on manual inspections and periodic testing. These methods suffer from low monitoring frequency and incomplete local monitoring, making it difficult to promptly detect potential faults and abnormal conditions. Furthermore, the lack of information-based analysis methods makes it difficult to quantify equipment degradation and remaining lifespan, and to assess overall unit and component energy consumption.

[0003] In recent years, edge intelligent computing devices, featuring integrated hardware and software designs, have gradually become a new paradigm for online equipment status monitoring and fault diagnosis. They offer advantages such as low latency, decentralized resource consumption, high data security, and synchronized diagnosis and control, providing a new solution for intelligent management of industrial equipment. However, traditional edge intelligent computing devices still face numerous challenges in practical applications: their limited computing resources, storage capacity, and communication performance make it difficult to efficiently process complex evaluation and diagnostic algorithms and massive amounts of equipment time series data.

[0004] In the prior art, a Chinese invention patent document with publication number CN 114116157A and publication date of March 1, 2022 has been proposed. The technical solution disclosed in this patent document is as follows: a multi-edge cluster cloud structure and load balancing scheduling method in an edge environment, which allocates different types of tasks to reasonable clusters based on the spatiotemporal differences of clusters in the edge environment and the delay sensitivity of tasks. First, tasks are divided into delay-sensitive tasks and delay-insensitive tasks. Secondly, different scheduling strategies are adopted for the two different tasks. Delay-sensitive tasks are allocated to edge computing nodes in the shortest time using genetic algorithms; delay-insensitive tasks are selected and allocated based on the load conditions of the local cluster and the load conditions of the surrounding clusters.

[0005] The above technical solution is to realize task scheduling between general cloud-edge collaborative cluster servers. The scheduling level is extensive and simple, resulting in the need to improve the efficiency of task scheduling and resource allocation in actual use. Summary of the Invention

[0006] In order to solve the above technical problems, the present invention proposes a multi-level resource dynamic scheduling method based on a multi-heterogeneous edge framework. It is based on task scheduling between multiple core computing board units and between heterogeneous computing modules within the core computing board units, and can effectively improve the efficiency of task scheduling and resource allocation, thereby improving the computing power performance of edge intelligent computing devices, realizing edge deployment applications of high-dimensional complex algorithms in the industrial field, and improving the efficiency and accuracy of intelligent diagnosis.

[0007] The present invention is achieved by adopting the following technical solutions: A multi-level resource dynamic scheduling method based on a multi-heterogeneous edge framework, wherein the multi-heterogeneous edge framework includes a network switching unit, a signal processing unit, a storage unit, and multiple core computing board units; each core computing board unit includes multiple heterogeneous computing modules; the method includes the following steps: Step S1 real-time monitoring and calculation of the computing power resource load status of each core computing board unit, and assigning tasks to the corresponding core computing board unit through an intelligent scheduling algorithm; Step S2. Decompose and categorize the task according to its computational complexity into several subtasks, and adaptively match each of these subtasks to the characteristics of the heterogeneous computing modules. Step S3 real-time monitoring and calculation of the load status or energy consumption status of each heterogeneous computing module in the core computing board unit allocated in step S1; Step S4. Allocate each subtask to a matching heterogeneous computing module with lower load or higher energy efficiency. Step S5. Build a computing flow based on subtask dependencies and classify subtasks according to real-time performance. Dynamically balance the distribution of tasks through an intelligent scheduling algorithm to ensure that the load of each heterogeneous computing module is balanced.

[0008] The types of heterogeneous computing modules include general computing CPU, parallel computing GPU, neural network processing NPU and digital signal processing DSP.

[0009] Step S2 specifically includes: decomposing tasks according to computing types, splitting tasks into computing-intensive subtasks, memory-intensive subtasks, and low-latency subtasks, and matching them with parallel computing GPUs, general computing CPUs, neural network processing NPUs, or digital signal processing DSPs respectively.

[0010] The storage unit adopts a layered storage architecture, including cache, memory and persistent storage.

[0011] The step S5 also includes analyzing the trend of resource demand in advance based on load prediction and elastic scaling technology to achieve dynamic and balanced allocation of heterogeneous computing modules.

[0012] It also includes real-time monitoring of the load status and energy consumption of each heterogeneous computing module in the remaining core computing board units, and abstracting all heterogeneous computing modules into a unified resource pool through virtualization technology.

[0013] It also includes real-time monitoring of subtask execution and hardware status, dynamic adjustment of subtask allocation strategy, and adjustment of subtasks to corresponding heterogeneous computing modules in other core computing board units.

[0014] In step S5, grading the subtasks according to real-time performance specifically refers to setting differentiated time limits for the subtasks.

[0015] It also includes a priority scheduling mechanism to prioritize high-priority subtasks.

[0016] The intelligent scheduling algorithm in step S5 adopts a reinforcement learning-based approach to model the scheduling problem as a Markov decision process, and the reward function is: , in, are the corresponding weight factors respectively.

[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. Existing dynamic scheduling methods typically rely on static task allocation and a single type of edge computing resource, making it difficult to handle complex algorithmic tasks and massive amounts of time-series data. This invention, through hardware and software collaborative design, fully utilizes multiple core computing boards and the diverse heterogeneous computing modules within them (such as CPUs, GPUs, NPUs, DSPs, etc.), combined with intelligent scheduling algorithms. This achieves efficient task allocation and dynamic resource optimization, significantly improving overall scheduling performance. It also ensures that the heterogeneous edge framework maintains a stable operating state even under high loads or sudden failures, thereby enhancing system reliability.

[0018] Furthermore, compared with the Chinese invention patent document with publication number CN 114116157A, the present invention is aimed at task scheduling within heterogeneous computing modules. Based on load status, computing type, real-time performance, hardware adaptability, etc., it supports multi-level resource scheduling across core computing board units and within core computing board units, with higher scheduling granularity.

[0019] 2. In the present invention, the scheduling in step S3 is at a lower level, but has a faster response and is task-oriented; the scheduling in step S5 is at a higher level, has a relatively slower response, and is resource-oriented. The coordination of the two scheduling methods in steps S3 and S5 enables rapid response while meeting long-term optimization.

[0020] 3. Existing tasks, such as fault diagnosis tasks, are generally implemented through edge-cloud collaboration to achieve system diagnosis. However, problems such as long data transmission links and incomplete data affect the diagnostic efficiency. The method of the present invention, by constructing a multi-heterogeneous edge framework integrating hardware and software, combined with dynamic resource scheduling optimization methods such as multi-level scheduling of complex algorithm tasks and balanced allocation of heterogeneous computing power, can rapidly implement equipment status assessment, abnormal analysis and early warning, and fault diagnosis through edge deployment applications for high-dimensional complex algorithms in the industrial field. This can significantly improve the system diagnostic efficiency, thereby enabling equipment faults to be discovered earlier and handled more quickly, reducing downtime and maintenance costs.

[0021] 4. The hardware and software collaborative integrated architecture constructed by the multiple heterogeneous computing modules (CPU, GPU, NPU) and dedicated functional units (storage, switching, signal processing) in the present invention can achieve deep binding of hardware characteristics and algorithms. It is not just a simple resource sharing, but an integrated hardware architecture of computing resource collaboration, storage resource sharing, and data exchange and transmission is formed to achieve an overall improvement in the performance of the diagnostic system.

[0022] 5. This scheduling and optimization method enables the scheduling of complex algorithmic tasks and the balanced distribution of heterogeneous computing power across core computing board units and heterogeneous computing modules within core computing board units. It can also analyze task dependencies in real time and match the characteristics of heterogeneous computing modules, ensuring the stability and diagnostic performance of the multi-heterogeneous edge framework in high-concurrency and high-dynamic scenarios. Combined with the corresponding intelligent scheduling algorithm, dynamic task allocation and optimization can be achieved, effectively reducing task completion time, improving resource utilization, and adapting to dynamic changes in the edge environment.

[0023] 6. Combining load prediction and elastic scaling technology, the present invention can predict changes in resource demand in advance and achieve dynamic and balanced distribution of computing power, thereby improving overall computing efficiency.

[0024] 7. The present invention adopts reinforcement learning (MDP modeling) to dynamically optimize the scheduling strategy and supports fault recovery and energy efficiency optimization.

[0025] 8. During dynamic scheduling, this invention constructs a computational flow based on subtask dependencies, ensuring efficient data transfer between subtasks and avoiding resource contention. Subtasks are also graded based on real-time performance, enabling the use of a priority scheduling mechanism. Hardware-aware scheduling (e.g., prioritizing AI inference on the NPU) can also be used to ensure critical path latency. DETAILED DESCRIPTION

[0026] Example 1 As a basic embodiment of the present invention, the present invention includes a multi-level resource dynamic scheduling method based on a multi-heterogeneous edge framework, wherein the multi-heterogeneous edge framework includes a network switching unit, a signal processing unit, a storage unit, and multiple core computing board units. Each core computing board unit contains multiple heterogeneous computing modules. The method includes the following steps: Step S1. Monitor and calculate the computing resource load status of each core computing board unit in real time, and assign tasks to the corresponding core computing board unit through an intelligent scheduling algorithm.

[0027] Step S2: Decompose and classify the task according to its computational complexity to form several subtasks, and adaptively match the subtasks with the characteristics of the heterogeneous computing modules.

[0028] Step S3: Monitor and calculate in real time the load status or energy consumption status of each heterogeneous computing module in the core computing board unit allocated in step S1.

[0029] Step S4: Allocate each subtask to a matching heterogeneous computing module with lower load or higher energy efficiency.

[0030] Step S5. Build a computing flow based on subtask dependencies and classify subtasks according to real-time performance. Dynamically balance the distribution of tasks through an intelligent scheduling algorithm to ensure that the load of each heterogeneous computing module is balanced.

[0031] Example 2 As a preferred embodiment of the present invention, the present invention includes a multi-level resource dynamic scheduling method based on a multi-heterogeneous edge framework, wherein the multi-heterogeneous edge framework specifically includes a network switching unit, a signal processing unit, a storage unit, and multiple core computing board units. The storage unit adopts a layered storage architecture, including cache, memory, and persistent storage. Each core computing board unit contains multiple heterogeneous computing modules. The method includes the following steps: Step S1. Monitor and calculate the computing resource load status of each core computing board unit in real time, and assign tasks to the corresponding core computing board unit through an intelligent scheduling algorithm.

[0032] Step S2: Decompose and classify the task according to its computational complexity to form several subtasks, and adaptively match the subtasks with the characteristics of the heterogeneous computing modules.

[0033] Step S3: Monitor and calculate in real time the load status or energy consumption status of each heterogeneous computing module in the core computing board unit allocated in step S1.

[0034] Step S4: Allocate each subtask to a matching heterogeneous computing module with lower load or higher energy efficiency.

[0035] Step S5. Build a computing flow based on subtask dependencies and classify subtasks according to real-time performance. Dynamically balance the distribution of tasks through an intelligent scheduling algorithm to ensure that the load of each heterogeneous computing module is balanced.

[0036] During this process, the subtask execution status and hardware status are monitored in real time, and the load status and energy consumption of each heterogeneous computing module in the remaining core computing board units are monitored in real time, and all heterogeneous computing modules are abstracted into a unified resource pool through virtualization technology. The remaining core computing board units specifically refer to core computing board units other than the core computing board units determined in step S1. When the core computing board unit determined in step S1 does not meet the preset conditions or requirements, the task allocation strategy can be further dynamically adjusted to adjust the subtasks to the corresponding heterogeneous computing modules in other core computing board units to achieve cross-board scheduling.

[0037] Example 3 As another preferred embodiment of the present invention, the present invention includes a multi-level resource dynamic scheduling method based on a multi-heterogeneous edge framework, wherein the multi-heterogeneous edge framework specifically includes a network switching unit, a signal processing unit, a storage unit, and a plurality of core computing board units. Each core computing board unit contains a plurality of heterogeneous computing modules. The types of heterogeneous computing modules include general computing CPU, parallel computing GPU, neural network processing NPU, and digital signal processing DSP. The method comprises the following steps: Step S1. Monitor and calculate the computing resource load status of each core computing board unit in real time, and assign tasks to the corresponding core computing board unit through an intelligent scheduling algorithm.

[0038] Step S2. Decompose and categorize the task based on its computational complexity into several subtasks. Each of these subtasks is adaptively matched to the characteristics of the heterogeneous computing modules. Specifically, tasks are decomposed based on computational type, splitting them into compute-intensive, memory-intensive, and low-latency subtasks. These are then matched to parallel computing GPUs, general-purpose computing CPUs, neural network processing NPUs, or digital signal processing DSPs, respectively.

[0039] Step S3: Monitor and calculate in real time the load status or energy consumption status of each heterogeneous computing module in the core computing board unit allocated in step S1.

[0040] Step S4: Allocate each subtask to a matching heterogeneous computing module with lower load or higher energy efficiency.

[0041] Step S5. A computational flow is constructed based on subtask dependencies, and subtasks are classified by real-time requirements. Dynamically balanced allocation is achieved through an intelligent scheduling algorithm, ensuring load balancing across heterogeneous computing modules. This real-time classification of subtasks specifically involves setting differentiated deadlines for subtasks and prioritizing high-priority subtasks. Based on this, a priority scheduling mechanism can be employed during dynamic balanced allocation to prioritize high-priority subtasks.

[0042] Example 4 As another preferred embodiment of the present invention, the present invention includes a multi-level dynamic resource scheduling method based on a multi-heterogeneous edge framework. The multi-heterogeneous edge framework utilizes an integrated hardware and software design, including a network switching unit, a signal processing unit, a storage unit, and multiple core computing board units. This forms an integrated hardware architecture for collaborative computing resources, shared storage resources, and data exchange and transmission, thereby improving the overall performance of the diagnostic system.

[0043] Each core computing board unit contains multiple heterogeneous computing modules. The types of heterogeneous computing modules include general-purpose computing CPUs, parallel computing GPUs, neural network processing NPUs, and digital signal processing DSPs. Each heterogeneous computing module specializes in processing specific types of tasks based on its characteristics, such as general-purpose computing GPUs for image processing, neural network processing NPUs for AI reasoning, and digital signal processing DSPs for signal processing.

[0044] The network switching unit incorporates a high-performance network switching module, supporting low-latency, high-bandwidth data transmission. This ensures efficient communication between heterogeneous computing modules and data exchange between the device and external devices. The storage unit utilizes a tiered storage architecture, including cache (e.g., SRAM), memory (e.g., DRAM), and persistent storage (e.g., SSD) to meet the storage speed and capacity requirements of different tasks. The storage unit supports data sharing and fast access, reducing data transmission overhead. The signal processing unit primarily preprocesses sensor data and information, providing processed data for computation.

[0045] Based on the above-mentioned multi-heterogeneous edge framework, integrated software and hardware collaborative optimization can be achieved. For example, the load, energy consumption and performance status of each heterogeneous computing module can be monitored in real time through the hardware status monitoring and perception mechanism, and tasks can be dynamically assigned based on the characteristics of the task. For example, computing-intensive subtasks are assigned to the GPU, and low-latency subtasks are assigned to the NPU or DSP. Through virtualization technology, hardware resources are abstracted into a unified resource pool, supporting dynamic allocation and elastic expansion. Tasks can dynamically occupy or release resources according to demand, improving resource utilization. Local data processing is achieved on the edge side to reduce data transmission delay and bandwidth consumption. The data caching and preprocessing functions of the storage unit support rapid response of real-time tasks.

[0046] Specifically, the method for implementing multi-level resource dynamic scheduling based on the multi-heterogeneous edge framework may include the following steps: Step S1. Monitor and calculate the computing resource load status of each core computing board unit in real time, and collect performance indicators (such as utilization, temperature, power consumption, etc.). Through status evaluation, we can fully understand the current working status of each core computing board unit and provide data support for dynamic scheduling.

[0047] Tasks are assigned to corresponding core computing board units through an intelligent scheduling algorithm. Specifically, hardware resources can be abstracted into measurable minimum units, with resource utilization as the scoring criterion and scheduling taking into account the minimum resource guarantees required by the algorithm.

[0048] Step S2. Decompose and categorize the task based on its computational complexity into several subtasks. These subtasks are then adaptively matched to the characteristics of heterogeneous computing modules. For example, an AI inference task can be broken down into three phases: data preprocessing, model inference, and post-processing. Tasks are also categorized to identify their appropriate heterogeneous computing module types (e.g., CPU, GPU, NPU, etc.).

[0049] Specifically, tasks are decomposed based on computational type, splitting them into compute-intensive subtasks (such as FFT and matrix operations), memory-intensive subtasks (such as thermal field simulation), and low-latency subtasks (such as real-time filtering). These subtasks are then matched to the hardware characteristics of heterogeneous computing modules, such as parallel computing GPUs, general-purpose computing CPUs, neural network processing NPUs, or digital signal processing DSPs. For example, compute-intensive subtasks are assigned to the GPU, low-latency subtasks are assigned to the NPU or DSP, and general-purpose tasks are handled by the CPU. Through this allocation strategy, the system can maximize the utilization of hardware resources, avoid overloading or idleness of some units, and achieve a balanced distribution of computing power.

[0050] Step S3: Monitor and calculate the load and energy consumption of each heterogeneous computing module in the core computing board unit assigned in step S1 in real time, collecting performance indicators (such as utilization, temperature, and power consumption). This status evaluation provides a comprehensive understanding of the current operating status of each heterogeneous computing module, providing data support for dynamic scheduling. This scheduling approach is low-level, but more responsive and task-oriented.

[0051] Step S4: Allocate each subtask to a matching heterogeneous computing module with lower load or higher energy efficiency.

[0052] Step S5. A computational flow is constructed based on subtask dependencies, and subtasks are classified by real-time requirements. Intelligent scheduling algorithms are used to achieve dynamic and balanced distribution, ensuring a balanced load across heterogeneous computing modules. For example, in high-concurrency scenarios, the system evenly distributes subtasks across multiple corresponding heterogeneous computing modules to avoid overloading a single point. Furthermore, taking into account both energy consumption and performance, heterogeneous computing modules with high energy efficiency are prioritized for task execution, reducing overall energy consumption and improving system efficiency.

[0053] Specifically, building a computing flow based on subtask dependencies refers to: building a computing flow based on task dependencies (such as sensor data → preprocessing → feature extraction → decision-making) to ensure efficient data transmission between subtasks and avoid resource competition. Grading subtasks by real-time performance specifically refers to: setting differentiated time limits such as milliseconds (DSP processing) and seconds (GPU computing) for subtasks, and ensuring critical path delays through hardware-aware scheduling (such as NPU prioritizing AI inference). At the same time, a priority scheduling mechanism can be used to prioritize high-priority subtasks, ensuring that high-priority subtasks (such as real-time diagnosis) can obtain resources first to meet low-latency requirements.

[0054] The intelligent scheduling algorithm uses a reinforcement learning-based approach to model the scheduling problem as a Markov decision process, and the reward function is: , in, are the corresponding weight factors respectively.

[0055] This scheduling level is higher and the response is relatively slower, and it is resource-oriented. By cooperating with step S3, it can achieve a fast response in the short term while meeting the long-term optimality.

[0056] Furthermore, by combining load prediction algorithms and elastic scaling technology, the trend of resource demand can be analyzed in advance based on historical processing tasks, making it possible to predict changes in resource demand in advance and avoid overloading or idleness of heterogeneous computing modules.

[0057] When a heterogeneous computing module is detected to be overloaded, some subtasks can be migrated to other idle heterogeneous computing modules to achieve load balancing and optimize energy consumption. For example, when the GPU is detected to be overloaded, the system will migrate some subtasks to the NPU or CPU for execution, which is also a way to improve computing performance under the current state. If you queue for the GPU to match hardware, it may reduce current computing performance.

[0058] This dynamic balancing process also includes real-time monitoring of the load and energy consumption of the heterogeneous computing modules in the remaining core computing board units, as well as the execution of subtasks. Using virtualization technology, all heterogeneous computing modules are abstracted into a unified resource pool. When necessary, subtasks are reassigned to corresponding heterogeneous computing modules in other core computing board units. In other words, resources within the core computing board unit are prioritized for currently executing tasks, followed by allocation to other core computing board units.

[0059] Example 5 As a specific embodiment of the present invention, a multi-level dynamic resource scheduling method based on a multi-heterogeneous edge framework is used in a real-time wind turbine bearing fault diagnosis system. Monitoring and diagnosing wind turbine bearing anomalies requires real-time processing of time series data from multiple vibration, temperature, and current sensors, performing multiple tasks in parallel: high-frequency vibration signal analysis (FFT + wavelet transform, computationally intensive), three-dimensional temperature field reconstruction (based on finite element simulation, memory intensive), current harmonic detection (DSP real-time filtering, requiring low latency), and fault pattern matching (based on deep learning models, accelerated by NPUs / GPUs).

[0060] The multi-heterogeneous edge framework includes multiple core computing board units, the specific number of which can be dynamically configured according to demand. A single core computing board unit adopts a CPU+GPU+NPU+DSP architecture, supporting general computing and task scheduling coordination, high-throughput parallel computing, and neural network algorithm inference acceleration. The signal processing unit includes an ADC acquisition module, which supports synchronous acquisition of multiple sensors (temperature / vibration / current) and can be configured with FIR / IIR filters to achieve noise suppression. The network switching unit includes a switching chip that supports time-sensitive networks, and the protocol stack can use IEEE802.1Qbv (time-aware shaping) to ensure low-latency transmission of mission-critical data (such as alarm signals). The storage unit includes SRAM cache, DRAM memory, and SSD persistent storage to implement a data lifecycle classification and tiered storage strategy.

[0061] The hardware configuration of a single core computing board unit is: CPU: 4-core A76, video decoding; NPU: 4TOPS computing power, performing fault classification; GPU: 2TFLOPS, processing FFT; DSP: 500GOPS, signal noise reduction.

[0062] Task decomposition and classification: The task is broken down into five subtasks: vibration signal denoising, vibration spectrum analysis (FFT), fault classification (CNN model), thermal imaging video decoding, and thermal region segmentation (U-Net). These subtasks correspond to the computational types of digital filtering, parallel computing, matrix computing, lightweight computing, and image processing, respectively.

[0063] Dynamic scheduling of heterogeneous resources within core computing board units. This approach uses reinforcement learning-based hardware dynamic scheduling, modeling the scheduling problem as a Markov decision process (MDP). The state space includes the real-time load, energy consumption, and task queue length of resources (CPU / GPU / NPU) between and within core computing board units. The action space is the task allocation decision, including resource allocation between and within core computing board units.

[0064] State space: [NPU_load, GPU_load, DSP_load, CPU_load, vibration signal queue length, thermal imaging queue length]; Action space: {Assign vibration FFT to GPU, assign fault classification to NPU, migrate hot segmentation to GPU, downclock NPU by 10%, downclock GPU by 10%}.

[0065] Design reward function for multi-objective optimization: in, are the corresponding weight factors, which are determined by domain knowledge or offline parameter adjustment.

[0066] Load balancing across core computing boards. Multiple core computing boards and their heterogeneous hardware resources (CPU / GPU / NPU) are abstracted into a unified resource pool, supporting dynamic allocation and elastic scaling. Based on load status, fault pattern matching tasks are dynamically assigned to boards with idle NPU resources (e.g., board 2 - NPU1, board 4 - NPU2) via the PCIe bus.

[0067] Example 6 Another specific embodiment of the present invention includes a multi-level dynamic resource scheduling method based on a multi-heterogeneous edge framework. This method is a heterogeneous computing power scheduling system for multimodal fault diagnosis in gas turbines. This method integrates acoustic, thermal infrared, and vibration data to enable real-time diagnosis of complex faults such as blade cracks, bearing wear, and combustion instability. The method features high dimensionality, multi-algorithm coupling, and real-time grading.

[0068] The multi-heterogeneous edge framework includes multiple core computing board units, the specific number of which can be dynamically configured based on demand. A single core computing board unit adopts a CPU+GPU+NPU+DSP architecture, supporting general computing and task scheduling coordination, high-throughput parallel computing, and neural network algorithm inference acceleration. The signal processing unit includes an ADC acquisition module, supports simultaneous acquisition of multiple sensors (voiceprint / infrared / vibration), and can be configured with FIR / IIR filters for noise suppression. The network switching unit uses PCIe5.0 interconnection. The storage unit includes SRAM cache, DRAM memory, and SSD persistent storage, implementing a data lifecycle classification and tiered storage strategy.

[0069] The hardware configuration of a single core computing board unit is: CPU: 8-core ARM Cortex-A78 (2.8GHz), task scheduling / lightweight logic control; GPU: Mali-G78 (4TFLOPS), vibration spectrum analysis / image reconstruction; NPU: 4TOPS (INT8), fault classification model inference; DSP: TI C66x (1GHz, 8MAC / cycle), real-time acoustic signal filtering.

[0070] Task decomposition and classification: Decompose the gas turbine fault diagnosis task into subtasks that can be executed in parallel, and match the optimal computing unit based on computing characteristics.

[0071] The task is decomposed into five subtasks: acoustic signal filtering, vibration spectrum analysis, infrared image super-resolution reconstruction, exhaust composition time series prediction, and fault decision fusion. The corresponding computing features are streaming data (FIR / IIR filtering), large-scale FFT calculation, convolutional neural network reasoning, LSTM model reasoning, rule engine and lightweight reasoning.

[0072] Dynamic scheduling of heterogeneous resources within core computing board units. This approach uses reinforcement learning-based hardware dynamic scheduling, modeling the scheduling problem as a Markov decision process (MDP). The state space includes the real-time load, energy consumption, and task queue length of resources (CPU / GPU / NPU) between and within core computing board units. The action space is the task allocation decision, including resource allocation between and within core computing board units.

[0073] State space: [NPU_load, GPU_load, DSP_load, CPU_load, vibration signal queue length, infrared image queue length, exhaust component queue length]; Action space: {Assign vibration FFT to GPU, assign fault classification to NPU, move image reconstruction to GPU, downclock NPU_10%, downclock GPU_10%}.

[0074] Design reward functions for multi-objective optimization; , in, are the corresponding weight factors, which are determined by domain knowledge or offline parameter adjustment.

[0075] Load balancing across core computing board units. Multiple core computing board units and the heterogeneous hardware resources (CPU / GPU / NPU) within these units are abstracted into a unified resource pool, supporting dynamic allocation and elastic scaling. For example, in the case of dynamic task migration, when the GPU load on core computing board unit 1 exceeds 90%, 50% of FFT tasks are migrated to the GPU on core computing board unit 3. The main control board dynamically allocates NPU tasks: the NPU on core computing board unit 2 processes 40% of the samples, and the NPU on core computing board unit 4 processes 60%.

[0076] Example 7 Another specific embodiment of the present invention includes a multi-level dynamic resource scheduling method based on a multi-heterogeneous edge framework for use in a large-scale generator winding operating status assessment and diagnosis system. This system integrates multiple data sources, including vibration, partial discharge, temperature field, and electromagnetic signals, to diagnose complex faults such as winding insulation aging, partial discharge, and thermal imbalance.

[0077] The multi-heterogeneous edge framework includes multiple core computing board units, the specific number of which can be dynamically configured based on demand. A single core computing board unit adopts a CPU+GPU+NPU+DSP architecture, supporting general computing and task scheduling coordination, high-throughput parallel computing, and neural network algorithm inference acceleration. The signal processing unit includes an ADC acquisition module, supports simultaneous acquisition of multiple sensors (voiceprint / infrared / vibration), and can be configured with FIR / IIR filters for noise suppression. The network switching unit uses PCIe5.0 interconnection. The storage unit includes SRAM cache, DRAM memory, and SSD persistent storage, implementing a data lifecycle classification and tiered storage strategy.

[0078] The hardware configuration of a single core computing board unit is: CPU, 8-core ARM Cortex-A78AE (lockstep mode, 2.5GHz), task scheduling, lightweight logic control; GPU, NVIDIA Jetson AGX Orin (32TOPSINT8), finite element thermal field simulation, image reconstruction; NPU, MLU220 (16TOPSINT8), insulation aging model inference; DSP, TI C6678 (1.25GHz, 8MAC / cycle), real-time partial discharge (PD) pulse processing.

[0079] Task decomposition and classification: The task is decomposed into five subtasks: PD pulse detection, temperature field finite element calculation, insulation aging prediction, temperature data compression, and fault decision fusion. These subtasks correspond to the computational features of streaming filtering / peak detection, sparse matrix solving, model inference, wavelet compression, and rule engine.

[0080] Dynamic scheduling of heterogeneous resources within core computing board units. This approach uses reinforcement learning-based hardware dynamic scheduling, modeling the scheduling problem as a Markov decision process (MDP). The state space includes the real-time load, energy consumption, and task queue length of resources (CPU / GPU / NPU) between and within the core computing board units. The action space is the task allocation decision, including resource allocation between and within the core computing board units.

[0081] State space: [NPU computing core utilization NPU_load, GPU core occupancy GPU_load, DSP instruction cycle utilization DSP_load, CPU all core average load CPU_load, partial discharge (PD) pulse queue length, temperature field simulation queue length, temperature data queue length]; Action space: {Assign PD classification to the NPU, migrate thermal field calculation to the GPU, enable DSP-assisted compression, switch insulation model accuracy, reduce NPU frequency by 10%, reduce GPU frequency by 10%}.

[0082] Design reward function for multi-objective optimization: , in, are the corresponding weight factors, which are determined by domain knowledge or offline parameter adjustment.

[0083] Cross-core computing board load balancing. Multiple core computing boards and the heterogeneous hardware resources (CPU / GPU / NPU) within them are abstracted into a unified resource pool, supporting dynamic allocation and elastic scaling. For example, thermal field computing migration: When the GPU load of core computing board unit 1 exceeds 85%, 50% of the grid partitions are migrated to the GPU of core computing board unit 4. PD processing disaster recovery: If the NPU of core computing board unit 3 fails, the NPU of core computing board unit 2 takes over its PD channel. Energy efficiency balancing: During low-load periods, the GPUs of core computing board units 3 and 4 are shut down, and only the NPUs are used for lightweight inference.

[0084] In summary, after reading the present invention document, ordinary technicians in this field can make various other corresponding transformation schemes based on the technical solutions and technical concepts of the present invention without creative mental work, which all fall within the scope of protection of the present invention.

Claims

1. A multi-level resource dynamic scheduling method based on a multi-heterogeneous edge framework, characterized by: The multi-heterogeneous edge framework includes a network switching unit, a signal processing unit, a storage unit, and multiple core computing board units; each core computing board unit includes multiple heterogeneous computing modules; the method includes the following steps: Step S1 real-time monitoring and calculation of the computing power resource load status of each core computing board unit, and assigning tasks to the corresponding core computing board unit through an intelligent scheduling algorithm; Step S2. Decompose and categorize the task according to its computational complexity into several subtasks, and adaptively match each of these subtasks to the characteristics of the heterogeneous computing modules. Step S3 real-time monitoring and calculation of the load status or energy consumption status of each heterogeneous computing module in the core computing board unit allocated in step S1; Step S4. Allocate each subtask to a matching heterogeneous computing module with lower load or higher energy efficiency. Step S5. Build a computing flow based on subtask dependencies and classify subtasks according to real-time performance. Dynamically balance the distribution of tasks through an intelligent scheduling algorithm to ensure that the load of each heterogeneous computing module is balanced.

2. The multi-level resource dynamic scheduling method based on a multi-heterogeneous edge framework according to claim 1 is characterized by: The types of heterogeneous computing modules include general computing CPU, parallel computing GPU, neural network processing NPU and digital signal processing DSP.

3. The multi-level resource dynamic scheduling method based on a multi-heterogeneous edge framework according to claim 2 is characterized by: Step S2 specifically includes: decomposing tasks according to computing types, splitting tasks into computing-intensive subtasks, memory-intensive subtasks, and low-latency subtasks, and matching them with parallel computing GPUs, general computing CPUs, neural network processing NPUs, or digital signal processing DSPs respectively.

4. The multi-level resource dynamic scheduling method based on a multi-heterogeneous edge framework according to claim 1 is characterized by: The storage unit adopts a layered storage architecture, including cache, memory and persistent storage.

5. The multi-level resource dynamic scheduling method based on a multi-heterogeneous edge framework according to claim 1 is characterized by: The step S5 also includes analyzing the trend of resource demand in advance based on load prediction and elastic scaling technology to achieve dynamic and balanced allocation of heterogeneous computing modules.

6. The multi-level resource dynamic scheduling method based on a multi-heterogeneous edge framework according to claim 5, characterized in that: It also includes real-time monitoring of the load status and energy consumption of each heterogeneous computing module in the remaining core computing board units, and abstracting all heterogeneous computing modules into a unified resource pool through virtualization technology.

7. The multi-level resource dynamic scheduling method based on a multi-heterogeneous edge framework according to claim 6, characterized in that: It also includes real-time monitoring of subtask execution and hardware status, dynamic adjustment of subtask allocation strategy, and adjustment of subtasks to corresponding heterogeneous computing modules in other core computing board units.

8. The multi-level resource dynamic scheduling method based on a multi-heterogeneous edge framework according to claim 1 is characterized by: In step S5, grading the subtasks according to real-time performance specifically refers to setting differentiated time limits for the subtasks.

9. The multi-level resource dynamic scheduling method based on a multi-heterogeneous edge framework according to claim 8, characterized in that: It also includes a priority scheduling mechanism to prioritize high-priority subtasks.

10. The multi-level resource dynamic scheduling method based on a multi-heterogeneous edge framework according to claim 1, characterized in that: The intelligent scheduling algorithm in step S5 adopts a reinforcement learning-based approach to model the scheduling problem as a Markov decision process, and the reward function is: ， in, 、 and are the corresponding weight factors respectively.

Images