Computing power resource scheduling method and system based on large model intention perception

Abstract

The invention relates to the technical field of computing power scheduling, and discloses a computing power resource scheduling method and system based on large model intention awareness, a BERT architecture is adopted as a basic model to construct a task intention awareness large model, task features of different scenes are adapted through transfer learning, and feature vectors containing task types, priorities and computing power requirements are output; deploying a lightweight CNN model locally at an edge node, preprocessing local task data, and uploading the preprocessed local task data to a task intention perception large model for intention recognition; collecting real-time state parameters of an edge node cluster, and generating a computing power distribution scheme through a dynamic scheduling algorithm in combination with the feature vector output by the task intention sensing large model and the real-time state parameters of the edge nodes; computing power resource scheduling is executed based on the computing power distribution scheme, and a scheduling strategy is adjusted in real time through a feedback mechanism; according to the invention, calculation power demand prediction is realized.

Description

Technical Field

[0001] This invention relates to the field of computing power scheduling technology, specifically to a computing power resource scheduling method and system based on large model intent awareness. Background Technology

[0002] With the development of edge computing technology, collaborative task processing between edge nodes and the cloud has become the mainstream architecture. However, existing computing power scheduling methods rely on static rules and cannot be dynamically adjusted according to task intent, resulting in wasted computing resources. The allocation of computing power between edge nodes and the cloud lacks real-time performance, making it difficult to meet the high real-time requirements of scenarios such as low-altitude monitoring and industrial inspection. The matching degree between task security level and computing power channel is low, posing a risk to data transmission. For example, in low-altitude intelligent networks, the computing power requirements of real-time monitoring tasks and path planning tasks of aircraft differ significantly. Traditional scheduling methods cannot quickly allocate resources according to task intent, which may lead to monitoring delays or computing power overload. Summary of the Invention

[0003] The purpose of this invention is to solve the above problems by designing a computing resource scheduling method and system based on large model intent awareness.

[0004] The first aspect of this invention provides a computing resource scheduling method based on large model intent awareness, the method comprising the following steps: A large task intent perception model is built using the BERT architecture as the base model. Through transfer learning, it is adapted to the task features of different scenarios and outputs feature vectors containing task type, priority, and computing power requirements. A lightweight CNN model is deployed locally on the edge node. After preprocessing the local task data, it is uploaded to the task intent perception big model for intent recognition. Collect real-time status parameters of the edge node cluster, including current load rate, network bandwidth, storage capacity, and computing power redundancy. By combining the feature vectors output by the task intent perception model with the real-time state parameters of edge nodes, a computing power allocation scheme is generated through a dynamic scheduling algorithm. The computing power allocation scheme is used to schedule computing resources, and the scheduling strategy is adjusted in real time through a feedback mechanism.

[0005] Optionally, in the first implementation of the first aspect of the present invention, the step of using the BERT architecture as the basic model to construct a large task intent-aware model, adapting to the task features of different scenarios through transfer learning, and outputting a feature vector containing task type, priority, and computing power requirements includes: The BERT architecture was selected as the base model, and the parameters of the base model were initialized. The base model includes a word embedding layer, a multi-head attention mechanism layer, and a feedforward neural network layer. Collect multi-scenario task data as a training set, and divide the multi-scenario task data into source domain data and target domain data according to the scenario. The multi-scenario task data includes at least aircraft monitoring data of low-altitude intelligent network and equipment detection data of industrial Internet of Things. The BERT base model is pre-trained based on source domain data, and transfer learning is used to adapt it to the target domain scenario. After transfer learning, the model processes the input data and generates feature vectors containing task type, priority, and computing power requirements.

[0006] Optionally, in a second implementation of the first aspect of the present invention, the pre-training of the BERT base model based on source domain data and the adaptation of the target domain scenario using transfer learning include: First, the text-based task description data in the source domain data is segmented and masked. Then, the non-text-based data in the source domain data is serialized and converted into a format that the model can recognize. The preprocessed source domain data is input into the BERT base model in batches. In each round of training, the model predicts the masked samples and outputs the probability distribution of the masked words. The trained model is adjusted using target domain data. By calculating the loss value between the prediction results and the real task labels, the parameters are backpropagated until the model's intent recognition accuracy on the target domain task stabilizes, thus completing the transfer learning adaptation.

[0007] Optionally, in a third implementation of the first aspect of the present invention, the step of deploying a lightweight CNN model locally at the edge node, preprocessing the local task data, and then uploading it to the task intent-aware large model for intent recognition includes: Deploy lightweight CNN models at edge nodes, where the CNN model structure includes an input layer, 2-3 convolutional layers, pooling layers, and an output layer; Edge nodes collect local task data, and a lightweight CNN model preprocesses the local task data, including data cleaning, feature extraction, and standardization. After preprocessing, the edge nodes use a symmetric encryption algorithm to encrypt the processed feature data and upload it to the task intent perception big model via the local network to send an intent recognition request.

[0008] Optionally, in a fourth implementation of the first aspect of the present invention, the edge node collects local task data, and a lightweight CNN model preprocesses the local task data, including: For numerical data, the lightweight CNN model identifies outliers using the 3σ rule, marking and removing values ​​that exceed the mean ± 3 times the standard deviation. For missing values, it uses forward imputation, replacing the missing position with the adjacent previous valid data point. For image data, it identifies damaged frames by pixel value range detection and deletes invalid images. The cleaned local task data is input into the convolution module of the lightweight CNN model. The local task data is processed by sliding window calculation using 3×3 convolution kernels to extract local features. The normalization layer standardizes the convolution results. The max pooling layer compresses the feature dimension by taking the maximum value of the local window. After multiple rounds of convolution and pooling operations, the features are mapped to a fixed-length one-dimensional feature vector through the global average pooling layer. The mean and standard deviation of each dimension in the feature vector obtained after feature extraction are calculated, and the feature values ​​of each dimension are standardized.

[0009] Optionally, in a fifth implementation of the first aspect of the present invention, the step of combining the feature vector output by the task intent perception large model with the real-time state parameters of the edge nodes to generate a computing power allocation scheme through a dynamic scheduling algorithm includes: Based on the priority in the feature vector output by the task intent perception model, the tasks waiting to be scheduled are sorted in descending order to determine the processing order. For each sorted task, select nodes from the edge node cluster that meet the conditions of computing power redundancy, network bandwidth, and storage capacity: For the selected candidate nodes, a comprehensive score is calculated using a dynamic scheduling algorithm, where the scoring dimensions include processing latency, resource utilization, and load balancing. Select the node with the highest overall score, determine the computing power allocation ratio, data transmission path and processing sequence, and generate a computing power allocation scheme.

[0010] Optionally, in a sixth implementation of the first aspect of the present invention, the step of performing computing resource scheduling based on a computing power allocation scheme and adjusting the scheduling strategy in real time through a feedback mechanism includes: The computing power allocation scheme is converted into executable instructions and sent to the selected edge nodes. The edge nodes adjust the resource allocation according to the instructions. Real-time monitoring of the task processing process and collection of feedback data, including at least the difference between the actual processing time and the predicted processing time, the load rate change of the node after processing the task, the packet loss rate during data transmission, and latency fluctuations. Based on feedback data, a state space and reward function are constructed, and the state and action value tables are updated through the Q-learning algorithm to optimize the node selection weight and computing power allocation ratio, and adjust the scheduling strategy in real time.

[0011] A second aspect of the present invention provides a computing resource scheduling system based on large model intent awareness, the system comprising: The building module is used to construct a large task intent perception model using the BERT architecture as the base model. It adapts to the task features of different scenarios through transfer learning and outputs feature vectors containing task type, priority, and computing power requirements. The upload module is used to deploy a lightweight CNN model locally on the edge node, preprocess the local task data, and then upload it to the task intent perception big model for intent recognition. The acquisition module is used to collect real-time status parameters of the edge node cluster, including current load rate, network bandwidth, storage capacity, and computing power redundancy. The generation module is used to combine the feature vectors output by the task intent perception large model with the real-time state parameters of the edge nodes, and generate a computing power allocation scheme through a dynamic scheduling algorithm. The adjustment module is used to perform computing resource scheduling based on the computing power allocation scheme and adjust the scheduling strategy in real time through a feedback mechanism.

[0012] A third aspect of the present invention provides a computing resource scheduling device based on large model intent awareness, the computing resource scheduling device based on large model intent awareness includes a memory and at least one processor, the memory storing instructions; the at least one processor invokes the instructions in the memory to cause the computing resource scheduling device based on large model intent awareness to perform the various steps of the computing resource scheduling method based on large model intent awareness as described in any of the preceding claims.

[0013] A fourth aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed by a processor, implement the steps of the computing resource scheduling method based on large model intent awareness as described in any of the preceding claims.

[0014] The technical solution provided by this invention employs the BERT architecture as the basic model to construct a large-scale task intent perception model. Through transfer learning, it adapts to the task characteristics of different scenarios, outputting feature vectors containing task type, priority, and computing power requirements. A lightweight CNN model is deployed locally on edge nodes to preprocess local task data before uploading it to the large-scale task intent perception model for intent recognition. Real-time status parameters of the edge node cluster are collected, including current load rate, network bandwidth, storage capacity, and computing power redundancy. Combining the feature vectors output by the large-scale task intent perception model with the real-time status parameters of the edge nodes, a computing power allocation scheme is generated through a dynamic scheduling algorithm. Computing power resource scheduling is executed based on the allocation scheme, and the scheduling strategy is adjusted in real time through a feedback mechanism. This invention achieves accurate prediction of computing power requirements and improves resource utilization through large-scale model intent perception. It supports unified scheduling of heterogeneous edge devices and dynamically allocates computing power channels based on security levels, improving data transmission security in scenarios such as low-altitude intelligent networks. The feedback mechanism ensures continuous optimization of the scheduling strategy to meet the dynamic needs of multiple scenarios. Attached Figure Description

[0015] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0016] Figure 1 A flowchart of a computing resource scheduling method based on large model intent awareness provided in an embodiment of the present invention; Figure 2 A schematic diagram of the structure of a computing resource scheduling system based on large model intent awareness provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a computing resource scheduling device based on large model intent awareness, provided in an embodiment of the present invention. Detailed Implementation

[0017] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” or “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, apparatus, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0018] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 The flowchart of the computing resource scheduling method based on large model intent awareness provided in this embodiment of the invention includes the following steps: Step 101: Use the BERT architecture as the base model to build a large task intent perception model. Adapt the task features of different scenarios through transfer learning and output feature vectors containing task type, priority, and computing power requirements. In this embodiment, the BERT architecture is selected as the basic model, and the basic model parameters are initialized. The basic model includes a word embedding layer, a multi-head attention mechanism layer, and a feedforward neural network layer. Multi-scenario task data is collected as a training set, and the multi-scenario task data is divided into source domain data and target domain data according to the scenario. The multi-scenario task data includes at least aircraft monitoring data of low-altitude intelligent network and equipment detection data of industrial Internet of Things. The BERT basic model is pre-trained based on the source domain data, and transfer learning is used to adapt to the target domain scenario. After transfer learning, the model processes the input data and generates feature vectors containing task type, priority, and computing power requirements.

[0019] In this embodiment, source domain data covering general task scenarios is collected, including task descriptions, instructions, and related context information in various text formats; the data is cleaned to remove duplicate, invalid, or incorrectly formatted content; a word segmentation tool adapted to the BERT model is used to segment the text data, splitting continuous text into lexical units that the model can recognize; according to the BERT pre-training requirements, a masked language model (MLM) training sample is constructed, and some words are randomly selected for masking, such as replacing them with special symbols. At the same time, a next sentence prediction (NSP) training sample is constructed, marking continuous sentence pairs as continuous and non-continuous sentence pairs as discontinuous. Load the BERT infrastructure, including word embedding layer, multiple Transformer encoder layers, pooling layer, and output layer; initialize word embedding layer parameters to map segmented words to fixed-dimensional vectors; set Transformer layer parameters, including the number of multi-head attention heads, hidden layer dimension, and feedforward neural network dimension; configure pre-training hyperparameters, including batch processing data volume, learning rate, number of training iterations, and gradient clipping threshold. The preprocessed source domain training data is input into the BERT base model in batches. In each training round, the model predicts MLM samples, outputs the probability distribution of the masked words, and calculates the cross-entropy loss between the predicted results and the real words as the MLM loss. At the same time, it predicts NSP samples, outputs the probability of whether the sentence pairs are continuous, and calculates the cross-entropy loss between the predicted results and the real labels as the NSP loss. The MLM loss and NSP loss are weighted and summed to obtain the total loss, and the parameters of all layers of the model are updated through the backpropagation algorithm. The training process is repeated until the preset number of iterations is reached or the loss of the model on the validation set tends to stabilize. Collect task data for specific scenarios in the target domain, such as aircraft mission instructions in the Low Altitude Internet of Things and equipment control commands in the Industrial Internet of Things; preprocess the target domain data, including cleaning and word segmentation using the same word segmentation tool as the source domain, and construct labeled training samples according to the target domain task requirements, such as intent classification and computing power requirement prediction, where the labels contain key information such as task type and priority; convert the processed target domain data into the BERT model input format, including adding special start symbols, delimiters, and sequence length padding; Load the pre-trained BERT model, retaining its word embedding layer and Transformer encoder layer as feature extractors; remove the original model's output layer and replace it with a new output layer adapted to the target domain task, such as a fully connected layer + softmax activation function for classification tasks; freeze some parameters of the BERT model's lower layers, such as the first 80% of the Transformer layer parameters, retaining only the top Transformer layer parameters and the new output layer parameters for updates, in order to retain the general feature extraction capability learned from the source domain; The target domain training samples are input into the adjusted model in batches. General features are extracted through the frozen bottom layer network, and then the target domain task is predicted through the trainable top layer network and the new output layer. The loss value between the prediction result and the true label is calculated, such as the cross-entropy loss of the classification task. The backpropagation algorithm is used to update only the unfrozen top layer parameters and the new output layer parameters. During the fine-tuning process, the model performance is evaluated using the target domain validation set at certain intervals. If the performance does not improve for several consecutive rounds, the training is stopped early. The fine-tuned model is evaluated using the target domain test set, and metrics such as task intent recognition accuracy and computational demand prediction deviation are calculated. If the evaluation results do not meet expectations, the number of frozen network layers is adjusted, such as reducing the number of frozen layers, or hyperparameters are fine-tuned, such as lowering the learning rate or increasing the number of training epochs, and the fine-tuning process is repeated until the model's performance on the target domain test set reaches a stable state, thus completing the transfer learning adaptation.

[0020] Step 102: Deploy a lightweight CNN model locally on the edge node, preprocess the local task data, and then upload it to the task intent perception big model for intent recognition. In this embodiment, a lightweight CNN model is deployed at the edge node. The CNN model structure includes an input layer, 2-3 convolutional layers, a pooling layer, and an output layer. The edge node collects local task data, and the lightweight CNN model preprocesses the local task data, including data cleaning, feature extraction, and standardization. After preprocessing, the edge node uses a symmetric encryption algorithm to encrypt the processed feature data and uploads it to the task intent perception model via the local network to send an intent recognition request.

[0021] In this embodiment, the edge node acquires raw task data in real time through built-in sensors or external devices, including but not limited to video frame sequences, sensor time-series data, text commands, etc.; the acquired information is classified and stored according to data type, for example, continuous video frames are saved as image sequences, sensor readings are sorted into a one-dimensional array by timestamp, and text commands are preserved in their original character format; a data acquisition frequency threshold is set to ensure that the amount of data matches the processing capacity of the edge node and to avoid storage overflow caused by too fast acquisition. A pre-trained lightweight CNN model is loaded locally at the edge node. This model includes an input layer, 2-3 convolutional modules (each module consisting of a convolutional layer, a batch normalization layer, a ReLU activation function layer, and a max pooling layer), and a feature output layer. The dimensions of the model's input layer are adjusted according to the local task data type. For example, for single-channel sensor data, the input dimension is set to time step, number of features, and 1; for color image data, it is set to height, width, and 3. The parameters of each layer of the model are initialized using pre-trained weights as initial values ​​to reduce the local training overhead at the edge node. Specific cleaning strategies are adopted for different types of data: For numerical data such as sensor readings, outliers are identified using the 3σ rule, that is, the mean and standard deviation of the data are calculated, and values ​​exceeding the mean ± 3 times the standard deviation are marked as outliers and removed; for missing values, the forward imputation method is used to replace the missing position with the adjacent previous valid data point; for image data, damaged frames are identified by pixel value range detection, such as 0-255, and invalid images are directly deleted; for text data, meaningless symbols, duplicate characters, and short sentences with a length below the threshold are removed, while retaining the core instruction content. The cleaned task data is input into the convolutional module of the lightweight CNN model: the convolutional layer performs sliding window calculations on the input data using convolutional kernels of preset sizes, such as 3×3 or 5×5, to extract local features such as abrupt changes in sensor data and edge features in images; the batch normalization layer standardizes the convolutional results to stabilize the data distribution; the ReLU activation function layer filters negative features, retains effective features, and introduces non-linear expressions; the max pooling layer compresses the feature dimension by taking the maximum value of a local window, reducing the amount of data; after multiple rounds of convolution and pooling operations, the features are mapped to a fixed-length one-dimensional feature vector through a global average pooling layer; Calculate the mean and standard deviation of each dimension in the feature vector obtained after feature extraction; standardize the feature values ​​of each dimension by subtracting the mean of that dimension and then dividing by the standard deviation, so that the feature value distribution of each dimension satisfies the condition that the mean is 0 and the standard deviation is 1; for extreme values ​​of the standardized values ​​that exceed the range of [-3,3], truncation is performed to limit them to this range; convert the standardized feature vector into the model output format and add data identifiers such as task ID and timestamp.

[0022] Step 103: Collect real-time status parameters of the edge node cluster, including current load rate, network bandwidth, storage capacity, and computing power redundancy. Step 104: Combine the feature vector output by the task intent perception large model with the real-time state parameters of the edge nodes, and generate a computing power allocation scheme through a dynamic scheduling algorithm. In this embodiment, the priority parameter of each waiting task is extracted from the feature vector output by the task intent perception model. This parameter is a quantized value in the range of 0-1, with a larger value indicating a higher urgency of the task. All tasks in the waiting state are collected to form a task queue. By comparing the size of the priority parameters of each task, the task queue is sorted in descending order, with tasks with higher priority values ​​placed at the front. If there are tasks with the same priority value, the real-time requirement parameters in the feature vector, such as time thresholds, are further referenced, and tasks with stricter real-time requirements are prioritized. After sorting, the processing order of the tasks is determined, and subsequent scheduling operations will be executed in this order. Following the processing order determined in the first step, the filtering operation is performed on each task sequentially. For the current task, the computing power redundancy, network bandwidth, and storage margin parameters of each node are extracted from the real-time status matrix of the edge node cluster. Filtering thresholds are set: computing power redundancy must be no less than 60% of the task's computing power requirement, reserving resource fluctuation space; network bandwidth must be no less than the minimum bandwidth required for task data transmission, calculated based on the task data volume and processing time limit; and storage margin must be no less than the task data storage capacity including temporary file space. Each node is checked to see if it meets all three conditions simultaneously, and all qualified nodes are included in the candidate node set. If the candidate node set is empty, the task is marked as "requiring edge-cloud collaborative processing" and proceeds to the subsequent cloud collaborative process. For the three dimensions of latency, resource utilization, and load balancing, scoring weights are assigned respectively, such as latency accounting for 40%, resource utilization for 30%, and load balancing for 30%. Scores for each dimension are calculated as follows: For latency, the score is calculated based on the node's current load rate (lower load rates result in higher scores) and the network latency with the task initiating node (lower latency results in higher scores). The actual values ​​are converted to a score of 0-10 using a linear mapping. For resource utilization, the matching degree between node computing power redundancy and task computing power requirements is calculated: matching degree = task requirements / node redundancy; the closer to 1, the higher the score. This is converted to a score of 0-10. For load balancing, the score is calculated based on the node's current load rate compared to the cluster average load rate; the greater the difference between the current load rate and the average load rate, the higher the score. This is also converted to a score of 0-10. The scores for the three dimensions are then weighted and summed to obtain a comprehensive score of 0-10 for each candidate node. The node with the highest comprehensive score is selected from the candidate nodes as the target processing node. Based on the task's computing power requirements and the target node's computing power redundancy, the computing power allocation ratio is determined: if the node has sufficient computing power redundancy (≥120% of the task requirements), computing power equal to the task requirements is allocated; if the redundancy is limited (60%-120% of the task requirements), computing power is allocated at 90% of the redundancy, reserving 10% redundancy to handle fluctuations. Data transmission paths are planned: priority is given to communication links with the fewest hops and the lowest historical packet loss rate between nodes, clearly defining the forwarding node sequence from the initiating node to the target node. Processing sequences are defined: if the task can be fragmented, it is divided into multiple subtasks based on data correlation, and the processing order of the subtasks is set, such as according to the data collection time sequence and the time interval between subtasks. The target node identifier, computing power allocation ratio, transmission path, and processing sequence are integrated into a standardized computing power allocation scheme.

[0023] Step 105: Execute computing resource scheduling based on the computing power allocation scheme, and adjust the scheduling strategy in real time through the feedback mechanism.

[0024] In this embodiment, the computing power allocation scheme is converted into executable instructions and resource allocation is adjusted. The scheduling center first parses the computing power allocation scheme and extracts key information, including the target edge node identifier, the allocated computing power ratio, data transmission path, and processing timing requirements. This information is then converted into a standardized instruction format, such as a binary instruction containing an opcode, parameter value, and checksum. The opcode corresponds to a specific resource adjustment type, such as CPU core allocation or memory quota setting. The parameter value specifies the adjustment range, such as increasing the CPU utilization limit of node A from 60% to 80%. The instructions are encrypted and sent to the selected edge nodes through edge cluster communication protocols such as CoAP or HTTP / 2 to ensure the integrity of instruction transmission. After receiving the instructions, the edge nodes parse the instruction content through their local resource managers and execute resource adjustment operations in sequence: adjusting the process scheduling priority by setting the target task process priority to real-time level, allocating dedicated memory blocks to reserve continuous memory space matching the task data volume, activating idle computing power cores, such as waking up dormant GPU cores to participate in computation, and generating an instruction execution receipt to be sent back to the scheduling center. Deploy distributed monitoring points in the task processing chain: Set start / end timers at the task initiation node to record the actual time from receiving the instruction to completion, and compare it with the predicted time in the scheduling phase to calculate the difference: actual time - predicted time; Deploy resource monitoring agents at the target edge nodes to collect CPU and GPU load rates every 50ms, and calculate the load change before and after task processing: load rate after processing - load rate before processing; Enable traffic analysis tools at key nodes in the data transmission path, such as routers and switches, to count the total number of data packets transmitted per unit time and the number of lost data packets, and calculate the packet loss rate: number of lost data packets / total number of data packets; Record the transmission delay from the sender to the receiver through a timestamp synchronization mechanism, continuously collect 100 delay values ​​to calculate the standard deviation, and use it as a delay fluctuation indicator; All collected feedback data are associated by timestamp, packaged, and uploaded to the feedback database of the scheduling center in real time; A state space is constructed, including the current task feature vector type, priority, computing power requirement, real-time status of each node in the edge node cluster, load rate, bandwidth, computing power redundancy, feedback data processing time difference, load change, etc., and this information is quantified into a multi-dimensional state vector. A reward function is designed: when the processing time difference is negative, the actual time is less than the prediction, the load change is within ±10%, the packet loss rate is less than 1%, and the latency fluctuation standard deviation is less than 5ms, a positive reward is given, and the reward value is positively correlated with the degree of achievement of the indicator; conversely, if any indicator exceeds the threshold, a negative reward is given, and the penalty value is positively correlated with the degree of deviation. The Q table is initialized, where actions include node selection combination and computing power allocation ratio. After each round of task processing, the Q table is updated using the Q-learning algorithm based on the current state, the executed actions, and the obtained reward: new Q value = old Q value + learning rate × reward + discount factor × maximum Q value of the next state - old Q value. Based on the updated Q table, the node selection weight and computing power allocation ratio are adjusted to generate a new scheduling strategy for computing power allocation in the next round of tasks.

[0025] Please see Figure 2 A schematic diagram of the structure of a computing resource scheduling system based on large model intent awareness provided in this embodiment of the invention. The system includes: The building module is used to construct a large task intent perception model using the BERT architecture as the base model. It adapts to the task features of different scenarios through transfer learning and outputs feature vectors containing task type, priority, and computing power requirements. The upload module is used to deploy a lightweight CNN model locally on the edge node, preprocess the local task data, and then upload it to the task intent perception big model for intent recognition. The acquisition module is used to collect real-time status parameters of the edge node cluster, including current load rate, network bandwidth, storage capacity, and computing power redundancy. The generation module is used to combine the feature vectors output by the task intent perception large model with the real-time state parameters of the edge nodes, and generate a computing power allocation scheme through a dynamic scheduling algorithm. The adjustment module is used to perform computing resource scheduling based on the computing power allocation scheme and adjust the scheduling strategy in real time through a feedback mechanism. Figure 3This is a schematic diagram of a computing resource scheduling device based on large model intent awareness provided in an embodiment of the present invention. The computing resource scheduling device 300 based on large model intent awareness can vary significantly due to different configurations or performance. It may include one or more central processing units (CPUs) 310 (e.g., one or more processors) and a memory 320, and one or more storage media 330 (e.g., one or more mass storage devices) for storing application programs 333 or data 332. The memory 320 and storage media 330 can be temporary or persistent storage. The program stored in the storage media 330 may include one or more modules (not shown in the diagram), each module including a series of instruction operations on the computing resource scheduling device 300 based on large model intent awareness. Furthermore, the processor 310 may be configured to communicate with the storage media 330 and execute the series of instruction operations in the storage media 330 on the computing resource scheduling device 300 based on large model intent awareness to implement the method provided in the above embodiment.

[0026] The large-model intent-aware computing resource scheduling device 300 may also include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input / output interfaces 360, and / or one or more operating systems 331, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 3 The illustrated computing resource scheduling device structure based on large model intent awareness does not constitute a limitation on the computer device provided by this invention. It may include more or fewer components than illustrated, or combine certain components, or have different component arrangements.

[0027] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when executed on a computer, cause the computer to perform the various steps of the computing resource scheduling method based on large model intent awareness provided in the above embodiments.

[0028] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the above-described equipment or apparatus / unit can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0029] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0030] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A computing resource scheduling method based on large-scale model intent awareness, characterized in that, The method includes the following steps: A large task intent perception model is built using the BERT architecture as the base model. Through transfer learning, it is adapted to the task features of different scenarios and outputs feature vectors containing task type, priority, and computing power requirements. A lightweight CNN model is deployed locally on the edge node. After preprocessing the local task data, it is uploaded to the task intent perception big model for intent recognition. Collect real-time status parameters of the edge node cluster, including current load rate, network bandwidth, storage capacity, and computing power redundancy. By combining the feature vectors output by the task intent perception model with the real-time state parameters of edge nodes, a computing power allocation scheme is generated through a dynamic scheduling algorithm. The computing power allocation scheme is used to schedule computing resources, and the scheduling strategy is adjusted in real time through a feedback mechanism.

2. The computing resource scheduling method based on large model intent awareness as described in claim 1, characterized in that, The system employs the BERT architecture as the base model to construct a large-scale task intent-aware model. Through transfer learning, it adapts to the task features of different scenarios, outputting feature vectors containing task type, priority, and computational requirements, including: The BERT architecture was selected as the base model, and the parameters of the base model were initialized. The base model includes a word embedding layer, a multi-head attention mechanism layer, and a feedforward neural network layer. Collect multi-scenario task data as a training set, and divide the multi-scenario task data into source domain data and target domain data according to the scenario. The multi-scenario task data includes at least aircraft monitoring data of low-altitude intelligent network and equipment detection data of industrial Internet of Things. The BERT base model is pre-trained based on source domain data, and transfer learning is used to adapt it to the target domain scenario. After transfer learning, the model processes the input data and generates feature vectors containing task type, priority, and computing power requirements.

3. The computing resource scheduling method based on large model intent awareness as described in claim 2, characterized in that, The process of pre-training the BERT base model based on source domain data and adapting it to the target domain scenario using transfer learning includes: First, the text-based task description data in the source domain data is segmented and masked. Then, the non-text-based data in the source domain data is serialized and converted into a format that the model can recognize. The preprocessed source domain data is input into the BERT base model in batches. In each round of training, the model predicts the masked samples and outputs the probability distribution of the masked words. The trained model is adjusted using target domain data. By calculating the loss value between the prediction results and the real task labels, the parameters are backpropagated until the model's intent recognition accuracy on the target domain task stabilizes, thus completing the transfer learning adaptation.

4. The computing resource scheduling method based on large model intent awareness as described in claim 1, characterized in that, The step of deploying a lightweight CNN model locally on the edge node, preprocessing the local task data, and then uploading it to the task intent perception model for intent recognition includes: Deploy lightweight CNN models at edge nodes, where the CNN model structure includes an input layer, 2-3 convolutional layers, pooling layers, and an output layer; Edge nodes collect local task data, and a lightweight CNN model preprocesses the local task data, including data cleaning, feature extraction, and standardization. After preprocessing, the edge nodes use a symmetric encryption algorithm to encrypt the processed feature data and upload it to the task intent perception big model via the local network to send an intent recognition request.

5. The computing resource scheduling method based on large model intent awareness as described in claim 4, characterized in that, The edge nodes collect local task data, and a lightweight CNN model preprocesses the local task data, including: For numerical data, the lightweight CNN model identifies outliers using the 3σ rule, marking and removing values ​​that exceed the mean ± 3 times the standard deviation. For missing values, it uses forward imputation, replacing the missing position with the adjacent previous valid data point. For image data, it identifies damaged frames by pixel value range detection and deletes invalid images. The cleaned local task data is input into the convolution module of the lightweight CNN model. The local task data is processed by sliding window calculation using 3×3 convolution kernels to extract local features. The normalization layer standardizes the convolution results. The max pooling layer compresses the feature dimension by taking the maximum value of the local window. After multiple rounds of convolution and pooling operations, the features are mapped to a fixed-length one-dimensional feature vector through the global average pooling layer. The mean and standard deviation of each dimension in the feature vector obtained after feature extraction are calculated, and the feature values ​​of each dimension are standardized.

6. The computing resource scheduling method based on large model intent awareness as described in claim 1, characterized in that, The process of combining the feature vector output by the task intent perception model with the real-time state parameters of edge nodes, and generating a computing power allocation scheme through a dynamic scheduling algorithm, includes: Based on the priority in the feature vector output by the task intent perception model, the tasks waiting to be scheduled are sorted in descending order to determine the processing order. For each sorted task, select nodes from the edge node cluster that meet the conditions of computing power redundancy, network bandwidth, and storage capacity: For the selected candidate nodes, a comprehensive score is calculated using a dynamic scheduling algorithm, where the scoring dimensions include processing latency, resource utilization, and load balancing. Select the node with the highest overall score, determine the computing power allocation ratio, data transmission path and processing sequence, and generate a computing power allocation scheme.

7. The computing resource scheduling method based on large model intent awareness as described in claim 1, characterized in that, The process of scheduling computing resources based on a computing power allocation scheme and adjusting the scheduling strategy in real time through a feedback mechanism includes: The computing power allocation scheme is converted into executable instructions and sent to the selected edge nodes. The edge nodes adjust the resource allocation according to the instructions. Real-time monitoring of the task processing process and collection of feedback data, including at least the difference between the actual processing time and the predicted processing time, the load rate change of the node after processing the task, the packet loss rate during data transmission, and latency fluctuations. Based on feedback data, a state space and reward function are constructed, and the state and action value tables are updated through the Q-learning algorithm to optimize the node selection weight and computing power allocation ratio, and adjust the scheduling strategy in real time.

8. A computing resource scheduling system based on large-scale model intent awareness, characterized in that, The system includes: The building module is used to construct a large task intent perception model using the BERT architecture as the base model. It adapts to the task features of different scenarios through transfer learning and outputs feature vectors containing task type, priority, and computing power requirements. The upload module is used to deploy a lightweight CNN model locally on the edge node, preprocess the local task data, and then upload it to the task intent perception big model for intent recognition. The acquisition module is used to collect real-time status parameters of the edge node cluster, including current load rate, network bandwidth, storage capacity, and computing power redundancy. The generation module is used to combine the feature vectors output by the task intent perception large model with the real-time state parameters of the edge nodes, and generate a computing power allocation scheme through a dynamic scheduling algorithm. The adjustment module is used to perform computing resource scheduling based on the computing power allocation scheme and adjust the scheduling strategy in real time through a feedback mechanism.

9. A computing resource scheduling device based on large-scale model intent awareness, characterized in that, The large model intent-aware computing resource scheduling device includes a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the large model intent-aware computing resource scheduling device to execute each step of the large model intent-aware computing resource scheduling method as described in any one of claims 1-7.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When the instructions are executed by the processor, they implement the various steps of the computing resource scheduling method based on large model intent awareness as described in any one of claims 1-7.

Images