A lightweight intelligent monitoring method and system for collaborative scheduling of edge-cloud resources

Abstract

This invention provides a lightweight intelligent monitoring method and system for collaborative scheduling of edge-cloud resources. The method includes: obtaining an edge autonomous region (AUR) based on the end-side device and the edge-side server; the AUR includes a resource monitoring agent and an initial LSTM load prediction model; updating the initial LSTM load prediction model to a final LSTM load prediction model based on real-time resource status data and a central coordination layer to obtain predicted resource status data; obtaining a comprehensive load index for the AUR; and when a real-time task is triggered in an AUR, obtaining the priority scores of historically executed tasks in that AUR, and generating a task scheduling strategy based on the predicted resource status data, the comprehensive load index, and the priority scores. By constructing a three-level collaborative mechanism of edge AUR – central coordination layer – global optimization, intelligent prediction is achieved in advance, and global state aggregation and dynamic cross-domain scheduling collaboration are realized while ensuring local response efficiency.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a lightweight intelligent monitoring method and system for collaborative scheduling of edge-cloud resources. Background Technology

[0002] In lightweight intelligent monitoring systems, a complex collaborative network is formed by a massive number of heterogeneous front-end devices and multi-level computing nodes. As intelligent video surveillance systems evolve towards a three-tier architecture of "device-edge-cloud," the deployment of AI algorithms and resource scheduling have become the core bottlenecks affecting system performance.

[0003] Currently, lightweight intelligent monitoring systems mainly rely on centralized decision-making in the central cloud or independent scheduling by edge nodes. However, this scheduling method has the following problems: 1. The scheduling mechanism is static and lacks dynamic adaptability. Most of the current mainstream scheduling strategies are based on preset rules (such as fixed task allocation according to device type) and cannot be adjusted according to dynamic factors such as real-time load, network status, and task priority. For example, when a sudden traffic incident in a certain area causes a surge in video streams, the computing power of edge nodes quickly becomes saturated, but the system cannot automatically migrate some tasks to idle edge nodes or the central cloud, resulting in task backlog and increased latency; 2. Resource status awareness is fragmented, and collaboration efficiency is low. There are delays and inconsistencies in the transmission of resource information between the end side, the edge side, and the central cloud. End devices have difficulty obtaining the global computing power distribution, and there is a lack of horizontal communication mechanisms between edge nodes, leading to "local optima and global imbalance"; 3. There is a lack of intelligent prediction and autonomous decision-making capabilities. Existing systems mostly adopt an "ex-post response" mode, that is, scheduling actions are triggered only after resource overload, which has obvious lag. At the same time, scheduling decisions rely on manual configuration or simple threshold judgments, lacking a comprehensive assessment of task complexity, data value, and QoS (quality of service) requirements, making it difficult to achieve refined management. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the present invention aims to provide a lightweight intelligent monitoring method and system for collaborative scheduling of edge-cloud resources. This system addresses the technical problems of existing technologies that rely on centralized decision-making in the central cloud or independent scheduling by edge nodes. These technologies suffer from static scheduling mechanisms, lack of dynamic adaptability, fragmented resource status awareness, low collaborative efficiency, and a lack of intelligent prediction and autonomous decision-making capabilities.

[0005] To achieve the above objectives, in a first aspect, embodiments of this application provide a lightweight intelligent monitoring method for collaborative scheduling of edge-cloud resources, comprising the following steps:

[0006] A combination of several geographically proximate end-side devices and at least one edge-side server forms an edge autonomous domain, which includes a resource monitoring agent and an initial LSTM load prediction model.

[0007] The resource monitoring agent obtains the real-time resource status data of the side server, updates the initial LSTM load prediction model to the final LSTM load prediction model based on the real-time resource status data and the central coordination layer, and obtains the predicted resource status data based on the final LSTM load prediction model and the real-time resource status data.

[0008] Obtain the basic state information of the edge autonomous region, and obtain the comprehensive load index of the edge autonomous region based on the basic state information;

[0009] When a real-time task is triggered in a certain edge autonomous region, the edge autonomous region is selected as the baseline autonomous region, and the remaining edge autonomous regions are selected as pending autonomous regions. The priority scores of historical tasks in the baseline autonomous region are obtained, and a task scheduling strategy is generated based on the predicted resource status data, the comprehensive load index, and the priority scores.

[0010] Furthermore, the step of updating the initial LSTM load prediction model to the final LSTM load prediction model based on the real-time resource status data and the central coordination layer includes:

[0011] The initial weight parameters are assigned to the initial LSTM load prediction model through the central coordination layer. The real-time resource status data is input into the initial LSTM load prediction model to obtain the initial parameter gradient tensor corresponding to the initial weight parameters. The initial parameter gradient tensor is then converted into a binary data stream.

[0012] The binary data stream is encrypted and transmitted to the central coordination layer. The central coordination layer decrypts and reverses the data stream to update the initial parameter gradient tensor to the updated parameter gradient tensor.

[0013] The updated weight parameters are obtained based on the initial weight parameters and the gradient tensor of the updated parameters, and the updated weight parameters are assigned to the initial LSTM load prediction model through the central coordination layer, so as to update the initial LSTM load prediction model into the final LSTM load prediction model.

[0014] Furthermore, the step of converting the initial parameter gradient tensor into a binary data stream includes:

[0015] The initial parameter gradient tensor is arranged continuously in memory to obtain a floating-point array, which includes several floating-point numbers. The floating-point numbers are then converted into binary data segments.

[0016] Several binary data segments are concatenated into an initial data stream, a magic number is assigned to the initial data stream, the total length of the sequence is obtained based on the length of the initial data stream and the length of the magic number, and the total length of the sequence, the magic number, and the initial data stream are combined into a binary data stream.

[0017] Furthermore, the formula for obtaining the gradient tensor of the update parameter is:

[0018] ,

[0019] in, Let the gradient tensor of the update parameters of the k-th edge autonomous region be denoted as . Let the gradient tensor of the initial parameters of the k-th marginal autonomous region be denoted as . This represents the total number of marginal autonomous regions. This represents the sample size of the real-time resource status data for the k-th edge autonomous region. This represents the sample size of real-time resource status data for all edge autonomous systems.

[0020] Furthermore, the basic status information includes current computing power utilization, maximum computing power capacity, current uplink bandwidth usage, total available bandwidth, and service health score. The formula for obtaining the comprehensive load index is:

[0021] ,

[0022] in, This represents the overall load index of the i-th edge autonomous region. This indicates the current computing power utilization rate. Indicates the maximum computing power capacity. Indicates the current uplink bandwidth usage. Indicates the total available bandwidth. This indicates the service health score. , , Both represent weighting coefficients, and + + =1.

[0023] Furthermore, the step of obtaining the priority scores of historically executed tasks in the baseline autonomous region includes:

[0024] Extract historical execution tasks from the baseline autonomous region and obtain the maximum allowable latency, data value coefficient, and task type weight of the historical execution tasks;

[0025] The priority score is obtained by using the maximum allowable latency, the data value coefficient, and the task type weight.

[0026] Furthermore, the formula for obtaining the priority score is:

[0027] ,

[0028] in, This represents the priority score of the j-th historically executed task. Indicates the maximum allowable delay. Indicates the data value coefficient. Indicates the task type weight. , , Both represent weight values, and + + =1.

[0029] Furthermore, the step of generating a task scheduling strategy based on the predicted resource status data, the comprehensive load index, and the priority score includes:

[0030] The predicted resource status data of the baseline autonomous region is compared with a preset threshold. If the predicted resource status data of the baseline autonomous region is greater than the preset threshold, it is determined that task scheduling is required.

[0031] Based on the priority scores of historically executed tasks in the benchmark autonomous region, several tasks to be scheduled are selected from several historically executed tasks in the benchmark autonomous region.

[0032] Based on the comprehensive load index, the load index difference between the baseline autonomous region and the undetermined autonomous region is obtained. The load index difference is compared with the difference threshold to select the final target autonomous region from a number of undetermined autonomous regions. The task to be scheduled and the final target autonomous region are combined into a task scheduling strategy.

[0033] Furthermore, the undetermined autonomous region whose difference from the load index is greater than the difference threshold is selected as the first target autonomous region, the transmission duration between the reference autonomous region and the first target autonomous region is obtained, and the processing delay of the first target autonomous region is obtained. The scheduling duration is obtained based on the transmission duration and the processing delay.

[0034] The scheduling duration is compared with the maximum allowable delay of the task to be scheduled in order to select the final target autonomous region from a plurality of first target autonomous regions.

[0035] Secondly, embodiments of this application provide a lightweight intelligent monitoring edge-cloud resource collaborative scheduling system, applied to the lightweight intelligent monitoring edge-cloud resource collaborative scheduling method described in the first aspect above, the system comprising:

[0036] A building module is used to combine several geographically adjacent end-side devices with at least one edge-side server into an edge autonomous domain, wherein the edge autonomous domain includes a resource monitoring agent and an initial LSTM load prediction model;

[0037] The prediction module is used to obtain real-time resource status data of the edge server through the resource monitoring agent, update the initial LSTM load prediction model to the final LSTM load prediction model based on the real-time resource status data and the central coordination layer, and obtain predicted resource status data based on the final LSTM load prediction model and the real-time resource status data.

[0038] The acquisition module is used to acquire the basic state information of the edge autonomous region and acquire the comprehensive load index of the edge autonomous region based on the basic state information;

[0039] The execution module is used to select the edge autonomous region as the base autonomous region and select the remaining edge autonomous regions as pending autonomous regions when a real-time execution task is triggered in a certain edge autonomous region, obtain the priority score of the historical execution tasks in the base autonomous region, and generate a task scheduling strategy based on the predicted resource status data, the comprehensive load index and the priority score.

[0040] Thirdly, embodiments of this application provide a computer, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the lightweight intelligent monitoring edge-cloud resource collaborative scheduling method as described in the first aspect above.

[0041] Fourthly, embodiments of this application provide a storage medium storing a computer program thereon, which, when executed by a processor, implements the lightweight intelligent monitoring edge-cloud resource collaborative scheduling method as described in the first aspect above.

[0042] Compared with existing technologies, the beneficial effects of this invention are as follows: By combining the geographically proximate end-side devices and the edge-side servers into an edge autonomous region including the resource monitoring agent, it endows them with local resource awareness capabilities, significantly improving local response efficiency; by setting the LSTM load prediction model in the edge autonomous region and collaborating with the central coordination layer, a three-level collaborative mechanism of edge autonomous region-central coordination layer-global optimization is constructed, realizing a paradigm shift from centralized control to distributed autonomy, achieving intelligent prediction of resource requirements such as computing power and bandwidth of the edge autonomous region, changing "post-event response" to "pre-event judgment", significantly improving scheduling efficiency and system elasticity; at the same time, by combining real-time resource status data of different edge autonomous regions during the prediction process, global status aggregation and cross-domain scheduling collaboration are achieved while ensuring local response efficiency; global overall prediction based on the real-time resource status data transforms the static scheduling mechanism into a dynamic scheduling mechanism; when constructing the task scheduling strategy, the priority score is introduced, comprehensively considering the latency requirements of soft tasks, data value and type weight, realizing differentiated service guarantees, and further enhancing the dynamic adaptability of task scheduling. Attached Figure Description

[0043] Figure 1 This is a flowchart of the lightweight intelligent monitoring edge-cloud resource collaborative scheduling method in the first embodiment of the present invention;

[0044] Figure 2 This is a structural block diagram of the lightweight intelligent monitoring edge-cloud resource collaborative scheduling system in the second embodiment of the present invention;

[0045] The following detailed description, in conjunction with the accompanying drawings, will further illustrate the present invention. Detailed Implementation

[0046] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Several embodiments of the invention are illustrated in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0047] It should be noted that when a component is said to be "fixed to" another component, it can be directly on the other component or there may be an intervening component. When a component is said to be "connected to" another component, it can be directly connected to the other component or there may be an intervening component. The terms "vertical," "horizontal," "left," "right," and similar expressions used in this document are for illustrative purposes only.

[0048] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0049] Please see Figure 1 The first embodiment of the present invention provides a lightweight intelligent monitoring method for collaborative scheduling of edge-cloud resources, comprising the following steps:

[0050] S10: Combine several geographically adjacent end-side devices with at least one edge-side server to form an edge autonomous domain, wherein the edge autonomous domain includes a resource monitoring agent and an initial LSTM load prediction model;

[0051] Understandably, the end-side device is a smart camera, and the edge-side server is an edge computing node. For example, in an urban traffic intersection scenario, it consists of several smart cameras and one edge computing node with GPU computing power, covering the omnidirectional video acquisition and preliminary analysis tasks of a single intersection. In this embodiment, the edge autonomous region includes four end-side devices and one edge-side server. Preferably, the edge autonomous region also includes a local scheduling decision module, a control service mesh, and a service mesh data plane. The local scheduling decision module is used to receive monitoring data and task requests, and, combined with a load prediction model, completes the generation of scheduling strategies and task allocation decisions. Its output can be obtained through the traffic routing rules of the control service mesh without relying on cloud instructions. The service mesh data plane is deployed in the edge autonomous region using a lightweight proxy as a communication intermediary for service instances. It is used to intercept all traffic entering and leaving the service, supports fine-grained routing control based on conditions such as weight, version tags, and request headers, and provides execution channels for dynamic scheduling. Specifically, the service mesh data plane is used to implement: 1. Automatic registration of AI service instances: When the edge server starts an AI inference service (such as "face detection v1"), its metadata (IP, port, version, QoS level) is automatically registered to the service registry center, and the registration information is synchronized to the control plane; 2. Real-time health check: The control plane sends a heartbeat probe request to each service instance every second. If no response is received for 3 consecutive times, the instance is determined to be abnormal, and the control service mesh is triggered to automatically remove it from the list of available instances to prevent traffic from being routed to the faulty node; 3. Dynamic traffic routing: Supports multiple routing strategies. For example, "weighted routing" can be configured: 70% of face detection requests are sent to the local instance, and 30% are sent to similar services in neighboring edge autonomous systems to achieve load balancing; or "priority routing" can be configured, with high-priority tasks (such as police vehicle identification) forcibly routed to highly available service instances; 4. Automatic circuit breaking and retrying: When the response latency of a service instance exceeds 500ms or the error rate is higher than 5%, the control service mesh automatically "circuits" traffic to that instance and retryes the request to a backup instance. The circuit breaker state lasts for 10 seconds before attempting to recover to avoid a cascading failure effect. The edge autonomous systems are interconnected via gigabit or 10-gigabit LANs, with communication latency between devices less than 1 millisecond, ensuring that resource status is synchronized with the real-time issuance of scheduling instructions. This architecture significantly reduces dependence on the cloud and improves local response speed and overall system robustness.

[0052] S20: Obtain real-time resource status data of the side server through the resource monitoring agent, update the initial LSTM load prediction model to the final LSTM load prediction model based on the real-time resource status data and the central coordination layer, and obtain predicted resource status data based on the final LSTM load prediction model and the real-time resource status data.

[0053] The resource monitoring agent continuously collects key resource indicators of the edge server, such as CPU utilization, GPU memory usage, memory utilization, disk I / O throughput, and uplink network bandwidth, with a sampling period of 1 second. This acquires the real-time resource status data. Simultaneously, it collects the video stream resolution, frame rate, encoding format, and currently running AI model type (such as face detection or license plate recognition) from each end-device, forming a complete local resource profile. Specifically, step S20 includes:

[0054] S210: Assign initial weight parameters to the initial LSTM load prediction model through the central coordination layer, input the real-time resource status data into the initial LSTM load prediction model to obtain the initial parameter gradient tensor corresponding to the initial weight parameters, and convert the initial parameter gradient tensor into a binary data stream.

[0055] Understandably, the central coordination layer is located in the cloud and connects all the edge autonomous systems. After obtaining the initial parameter gradient tensor, the initial parameter gradient tensor is continuously arranged in memory to obtain a floating-point array, which includes several floating-point numbers. The floating-point numbers are converted into binary data segments. The several binary data segments are concatenated into an initial data stream, and a magic number is assigned to the initial data stream. The total sequence length is obtained based on the length of the initial data stream and the length of the magic number. The total sequence length, the magic number, and the initial data stream are combined into a binary data stream.

[0056] The magic number refers to a specific byte sequence located at the beginning of a file or data stream that uniquely identifies the data format. It is used by the central coordination layer to identify the binary data stream when it is transmitted to the central coordination layer and to call the correct decoding program.

[0057] S220: The binary data stream is encrypted and transmitted to the central coordination layer. The central coordination layer decrypts and reverses the binary data stream to update the initial parameter gradient tensor to the updated parameter gradient tensor.

[0058] After the encryption of the binary data stream is completed, an initialization vector corresponding to the binary database is generated synchronously. This initialization vector is essentially a 16-byte data block. The encryption process is as follows: the binary data stream is divided into several binary data blocks; one of these blocks is XORed with the initialization vector, and then encrypted using an AES key to obtain a first ciphertext data block; another binary block is XORed with the first ciphertext data block, and then encrypted using an AES key to obtain a second ciphertext data block. This process is repeated for each resulting binary data block to complete the encryption. When the encrypted binary data stream is transmitted to the central coordination layer, the central coordination layer decrypts it using the initialization vector and converts the decrypted binary data stream back into the initial gradient tensor. This encryption process improves data security during data transmission.

[0059] The formula for obtaining the gradient tensor of the update parameter is:

[0060] ,

[0061] in, Let the gradient tensor of the update parameters of the k-th edge autonomous region be denoted as . Let the gradient tensor of the initial parameters of the k-th marginal autonomous region be denoted as . This represents the total number of marginal autonomous regions. This represents the sample size of the real-time resource status data for the k-th edge autonomous region. This represents the sample size of real-time resource status data for all edge autonomous systems.

[0062] S230: Based on the initial weight parameters and the gradient tensor of the update parameters, obtain the updated weight parameters, and assign the updated weight parameters to the initial LSTM load prediction model through the central coordination layer, so as to update the initial LSTM load prediction model into the final LSTM load prediction model.

[0063] Understandably, based on the preset learning rate and the gradient tensor of the update parameter, the change value is obtained, and the difference between the initial weight parameter and the change value is selected as the update weight parameter.

[0064] S30: Obtain the basic state information of the edge autonomous region, and obtain the comprehensive load index of the edge autonomous region based on the basic state information;

[0065] The basic status information includes current computing power utilization, maximum computing power capacity, current uplink bandwidth usage, total available bandwidth, and service health score. The formula for obtaining the comprehensive load index is:

[0066] ,

[0067] in, This represents the overall load index of the i-th edge autonomous region. This indicates the current computing power utilization rate. Indicates the maximum computing power capacity. Indicates the current uplink bandwidth usage. Indicates the total available bandwidth. This indicates the service health score. , , Both represent weighting coefficients, and + + =1. In this embodiment, The value is 0.5. The value is 0.3. The value is 0.2. Understandably, the service health score ranges from 0 to 1. When the service health score is 1, it means that all instances are healthy.

[0068] S40: When a real-time execution task is triggered in a certain edge autonomous region, the edge autonomous region is selected as the base autonomous region, and the remaining edge autonomous regions are selected as pending autonomous regions. The priority scores of historical execution tasks in the base autonomous region are obtained, and a task scheduling strategy is generated based on the predicted resource status data, the comprehensive load index and the priority scores.

[0069] Specifically, step S40 includes:

[0070] S410: Extract historical execution tasks from the baseline autonomous region and obtain the maximum allowable latency, data value coefficient, and task type weight of the historical execution tasks;

[0071] The maximum allowable latency, data value coefficient, and task type weight are different for different historical execution tasks. For example, the maximum allowable latency is 2 seconds for emergency alarm tasks (such as fire and intrusion), and 8 seconds for routine video analysis (such as people counting).

[0072] S420: Obtain the priority score by using the maximum allowable latency, the data value coefficient, and the task type weight;

[0073] The formula for obtaining the priority score is:

[0074] ,

[0075] in, This represents the priority score of the j-th historically executed task. Indicates the maximum allowable delay. Indicates the data value coefficient. Indicates the task type weight. , , Both represent weight values, and + + =1.

[0076] S430: Compare the predicted resource status data of the baseline autonomous region with a preset threshold. If the predicted resource status data of the baseline autonomous region is greater than the preset threshold, it is determined that task scheduling is required.

[0077] In this embodiment, the preset threshold is 80%. If it is found that the GPU utilization of the baseline autonomous region will reach 85% in the future, exceeding 80%, task scheduling is required to avoid resource overload.

[0078] S440: Based on the priority scores of historically executed tasks in the baseline autonomous region, select several tasks to be scheduled from several historically executed tasks in the baseline autonomous region;

[0079] The historical execution tasks are sorted based on the priority scores, and the historical execution tasks with lower priority scores are selected as tasks to be scheduled. The historical execution tasks with higher priority scores are more important and are retained in the baseline autonomous region to ensure the timeliness of task execution.

[0080] S450: Based on the comprehensive load index, obtain the load index difference between the benchmark autonomous region and the undetermined autonomous region, compare the load index difference with the difference threshold, select the final target autonomous region from a number of undetermined autonomous regions, and combine the task to be scheduled with the final target autonomous region into a task scheduling strategy.

[0081] The difference in the comprehensive load index between the baseline autonomous region (AUR) and the AUR is the load index difference. The AURs whose load index difference is greater than a threshold are selected as the first target AURs. The transmission duration between the baseline AUR and the first target AUR is obtained, as is the processing delay of the first target AUR. A scheduling duration is obtained based on the transmission duration and the processing delay. The scheduling duration is compared with the maximum allowable delay of the task to be scheduled to select the final target AUR from several first target AURs. Specifically, the first target AUR corresponding to a scheduling duration less than the maximum allowable delay of the task to be scheduled is selected as the second target AUR. From several second target AURs, the second target AUR with the smallest scheduling duration is selected as the final target AUR.

[0082] By combining geographically proximate end-side devices and edge servers into an edge autonomous region (AUR), including the resource monitoring agent, local resource awareness is granted, significantly improving local response efficiency. By setting up the LSTM load prediction model within the AUR and collaborating with the central coordination layer, a three-tiered collaborative mechanism of edge AUR, central coordination layer, and global optimization is constructed. This achieves a paradigm shift from centralized control to distributed autonomy, enabling intelligent prediction of resource demands such as computing power and bandwidth within the AUR, transforming "post-event response" into "pre-event judgment," significantly improving scheduling efficiency and system resilience. Simultaneously, by combining real-time resource status data from different AURs during the prediction process, global status aggregation and cross-domain scheduling collaboration are achieved while ensuring local response efficiency. Global overall prediction based on the real-time resource status data transforms the static scheduling mechanism into a dynamic one. Incorporating priority scores when constructing the task scheduling strategy, and comprehensively considering soft task latency requirements, data value, and type weights, differentiated service guarantees are achieved, further enhancing the dynamic adaptability of task scheduling.

[0083] Please see Figure 2 The second embodiment of the present invention provides a lightweight intelligent monitoring edge-cloud resource collaborative scheduling system. This system is applied to the lightweight intelligent monitoring edge-cloud resource collaborative scheduling method described in the above embodiments, and will not be repeated hereafter. As used below, the terms "module," "unit," "subunit," etc., can refer to a combination of software and / or hardware that performs a predetermined function. Although the apparatus described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0084] The system includes:

[0085] Module 10 is used to combine several geographically adjacent end-side devices with at least one edge-side server into an edge autonomous domain, wherein the edge autonomous domain includes a resource monitoring agent and an initial LSTM load prediction model.

[0086] Prediction module 20 is used to obtain real-time resource status data of the edge server through the resource monitoring agent, update the initial LSTM load prediction model to the final LSTM load prediction model based on the real-time resource status data and the central coordination layer, and obtain predicted resource status data based on the final LSTM load prediction model and the real-time resource status data.

[0087] The prediction module 20 includes:

[0088] The first unit is used to assign initial weight parameters to the initial LSTM load prediction model through the central coordination layer, input the real-time resource status data into the initial LSTM load prediction model to obtain the initial parameter gradient tensor corresponding to the initial weight parameters, and convert the initial parameter gradient tensor into a binary data stream.

[0089] The first unit is further configured to continuously arrange the initial parameter gradient tensor in memory to obtain a floating-point array, the floating-point array including several floating-point numbers, convert the floating-point numbers into binary data segments; concatenate several binary data segments into an initial data stream, assign a magic number to the initial data stream, obtain the total sequence length based on the length of the initial data stream and the length of the magic number, and combine the total sequence length, the magic number, and the initial data stream into a binary data stream;

[0090] The second unit is used to encrypt the binary data stream and transmit the encrypted binary data stream to the central coordination layer, where the central coordination layer decrypts and reverses the data stream to update the initial parameter gradient tensor to the updated parameter gradient tensor.

[0091] The third unit is used to obtain updated weight parameters based on the initial weight parameters and the gradient tensor of the updated parameters, and to assign the updated weight parameters to the initial LSTM load prediction model through the central coordination layer, so as to update the initial LSTM load prediction model into the final LSTM load prediction model.

[0092] The acquisition module 30 is used to acquire the basic state information of the edge autonomous region and acquire the comprehensive load index of the edge autonomous region based on the basic state information;

[0093] The execution module 40 is used to select the edge autonomous region as the base autonomous region and select the remaining edge autonomous regions as pending autonomous regions when a real-time execution task is triggered in a certain edge autonomous region, obtain the priority score of the historical execution task in the base autonomous region, and generate a task scheduling strategy based on the predicted resource status data, the comprehensive load index and the priority score.

[0094] The execution module 40 includes:

[0095] The fourth unit is used to extract historical execution tasks from the baseline autonomous region and obtain the maximum allowable latency, data value coefficient, and task type weight of the historical execution tasks.

[0096] The fifth unit is used to obtain a priority score by means of the maximum allowable latency, the data value coefficient, and the task type weight;

[0097] The sixth unit is used to compare the predicted resource status data of the benchmark autonomous region with a preset threshold. If the predicted resource status data of the benchmark autonomous region is greater than the preset threshold, it is determined that task scheduling is required.

[0098] The seventh unit is used to select several tasks to be scheduled from several historical tasks in the benchmark autonomous region based on the priority scores of historically executed tasks in the benchmark autonomous region.

[0099] The eighth unit is used to obtain the load index difference between the benchmark autonomous region and the undetermined autonomous region based on the comprehensive load index, compare the load index difference with the difference threshold, select the final target autonomous region from a number of undetermined autonomous regions, and combine the task to be scheduled with the final target autonomous region into a task scheduling strategy.

[0100] The eighth unit is further configured to select the undetermined autonomous region whose difference with the load index is greater than the difference threshold as the first target autonomous region, obtain the transmission duration between the reference autonomous region and the first target autonomous region, and obtain the processing delay of the first target autonomous region, obtain the scheduling duration based on the transmission duration and the processing delay; compare the scheduling duration with the maximum allowable delay of the task to be scheduled, so as to select the final target autonomous region from a plurality of the first target autonomous regions.

[0101] The present invention also provides a computer, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the lightweight intelligent monitoring edge-cloud resource collaborative scheduling method as described in the above technical solutions.

[0102] The present invention also provides a storage medium storing a computer program thereon, which, when executed by a processor, implements the lightweight intelligent monitoring edge-cloud resource collaborative scheduling method as described in the above technical solution.

[0103] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0104] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for edge-cloud resource collaborative scheduling of lightweight intelligent monitoring, characterized in that, The method comprises the following steps: combining several end-side devices adjacent in geographical position with at least one side server into an edge autonomous domain, the edge autonomous domain comprising a resource monitoring agent and an initial LSTM load prediction model; obtaining real-time resource state data of the side server through the resource monitoring agent, updating the initial LSTM load prediction model into a final LSTM load prediction model based on the real-time resource state data and a central coordination layer, and obtaining predicted resource state data based on the final LSTM load prediction model and the real-time resource state data; the step of updating the initial LSTM load prediction model into a final LSTM load prediction model based on the real-time resource state data and the central coordination layer comprises: assigning an initial weight parameter to the initial LSTM load prediction model through the central coordination layer, inputting the real-time resource state data into the initial LSTM load prediction model to obtain an initial parameter gradient tensor corresponding to the initial weight parameter, and converting the initial parameter gradient tensor into a binary data stream; encrypting the binary data stream, transmitting the encrypted binary data stream to the central coordination layer, decrypting and inversely converting the binary data stream through the central coordination layer to update the initial parameter gradient tensor into an updated parameter gradient tensor; obtaining an updated weight parameter based on the initial weight parameter and the updated parameter gradient tensor, and assigning the updated weight parameter to the initial LSTM load prediction model through the central coordination layer to update the initial LSTM load prediction model into a final LSTM load prediction model; obtaining basic state information of the edge autonomous domain, and obtaining a comprehensive load index of the edge autonomous domain based on the basic state information; when a real-time execution task is triggered in a certain edge autonomous domain, selecting the edge autonomous domain as a reference autonomous domain and selecting the remaining edge autonomous domains as pending autonomous domains, obtaining a priority score of a historical execution task in the reference autonomous domain, and generating a task scheduling strategy based on the predicted resource state data, the comprehensive load index and the priority score.

2. The method of claim 1, wherein, the step of converting the initial parameter gradient tensor into a binary data stream comprises: continuously arranging the initial parameter gradient tensor in the memory to obtain a floating-point number array, the floating-point number array comprising several floating-point numbers, and converting the floating-point numbers into binary data segments; splicing the several binary data segments into an initial data stream, assigning a magic number to the initial data stream, obtaining a total sequence length based on the length of the initial data stream and the length of the magic number, and combining the total sequence length, the magic number and the initial data stream into a binary data stream.

3. The method of claim 1, wherein, the formula for obtaining the updated parameter gradient tensor is: ， wherein, denotes the updated parameter gradient tensor of the kth edge autonomous domain, denotes the initial parameter gradient tensor of the kth edge autonomous domain, denotes the total number of edge autonomous domains, denotes the sample size of real-time resource status data of the kth edge autonomous domain, denotes the sample size of real-time resource status data of all edge autonomous domains.

4. The method of claim 1, wherein, the basic state information comprises a current computing power usage rate, a maximum computing power capacity, a current uplink bandwidth occupation, a total available bandwidth and a service health score, and the formula for obtaining the comprehensive load index is: ， wherein, represents the comprehensive load index of the ith edge autonomous domain, represents the current computing power usage rate, represents the maximum computing power capacity, represents the current uplink bandwidth occupation, represents the total available bandwidth, represents the service health score, , , all represent weighting coefficients, and + + = 1.

5. The method of claim 1, wherein, the step of obtaining the priority score of the historical execution task in the reference autonomous domain comprises: extracting a historical execution task in the reference autonomous domain, and obtaining a maximum allowed time delay, a data value coefficient, and a task type weight of the historical execution task; obtaining a priority score through the maximum allowed time delay, the data value coefficient, and the task type weight.

6. The method of claim 5, wherein, The priority score is obtained according to the following formula: ， wherein, represents a priority score of the jth historical executed task, represents a maximum allowed latency, represents a data value coefficient, represents a task type weight, , , all represent weight values, and + + = 1.

7. The method of claim 1, wherein, The step of generating a task scheduling strategy based on the predicted resource state data, the comprehensive load index, and the priority score includes: comparing the predicted resource state data of the reference autonomous domain with a preset threshold value, and determining that task scheduling needs to be performed if the predicted resource state data of the reference autonomous domain is greater than the preset threshold value; selecting a plurality of to-be-scheduled tasks from a plurality of historical execution tasks in the reference autonomous domain based on the priority scores of the historical execution tasks in the reference autonomous domain; obtaining a load index difference value between the reference autonomous domain and the to-be-determined autonomous domain based on the comprehensive load index, comparing the load index difference value with a difference value threshold, and selecting a final target autonomous domain from a plurality of to-be-determined autonomous domains, and combining the to-be-scheduled tasks and the final target autonomous domain into a task scheduling strategy.

8. The method of claim 7, wherein, The step of comparing the load index difference value with the difference value threshold to select the final target autonomous domain from the plurality of to-be-determined autonomous domains includes: selecting a first target autonomous domain from the to-be-determined autonomous domains whose load index difference value is greater than the difference value threshold, obtaining a transmission time length between the reference autonomous domain and the first target autonomous domain, and obtaining a processing time delay of the first target autonomous domain, and obtaining a scheduling time length based on the transmission time length and the processing time delay; comparing the scheduling time length with a maximum allowed time delay of the to-be-scheduled tasks to select a final target autonomous domain from a plurality of first target autonomous domains.

9. A system for lightweight intelligent monitoring of end-edge-cloud resource collaborative scheduling, applied to the lightweight intelligent monitoring of end-edge-cloud resource collaborative scheduling method of any one of claims 1-8, characterized in that, The system includes: a construction module configured to combine a plurality of edge-side devices adjacent in geographical position and at least one side server into an edge autonomous domain, the edge autonomous domain including a resource monitoring agent and an initial LSTM load prediction model; a prediction module configured to obtain real-time resource state data of the side server through the resource monitoring agent, update the initial LSTM load prediction model into a final LSTM load prediction model based on the real-time resource state data and a central coordination layer, and obtain predicted resource state data based on the final LSTM load prediction model and the real-time resource state data; an acquisition module configured to acquire basic state information of the edge autonomous domain, and obtain a comprehensive load index of the edge autonomous domain based on the basic state information; an execution module configured to select a reference autonomous domain from a plurality of edge autonomous domains when a real-time execution task is triggered in a certain edge autonomous domain, select remaining edge autonomous domains as to-be-determined autonomous domains, obtain a priority score of a historical execution task in the reference autonomous domain, and generate a task scheduling strategy based on predicted resource state data, a comprehensive load index, and the priority score.

Images