Edge computing telemedicine task unloading method based on improved PPO algorithm

Abstract

The invention provides an edge computing remote medical task offloading method based on an improved PPO algorithm, and the method comprises the steps: dynamically selecting an optimal offloading mode through a PPO-MED intelligent decision module according to the characteristics of a medical task and a system resource state, and enabling the offloading mode to comprise a local computing offloading mode (LCO), an edge server offloading mode (ESO) and an edge intelligent preprocessing offloading mode (EIPO); a PPO-MED algorithm models a medical task unloading problem into a Markov decision process, and an unloading strategy is optimized through deep reinforcement learning. The improved PPO-MED algorithm realizes intelligent decision, can dynamically select the optimal unloading strategy according to the real-time system state, task urgency and resource availability, significantly improves the adaptive capacity and service quality of the system in a complex environment, compresses the original medical data into a plurality of orders of magnitude, and improves the reliability of the system. On the premise that the diagnosis accuracy is guaranteed, the transmission delay is remarkably reduced, and the method is particularly suitable for remote diagnosis application of basic medical institutions under the limited bandwidth condition.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of edge computing, in particular to an edge computing remote medical task offloading method based on an improved PPO algorithm, and more particularly to a remote medical computing task intelligent scheduling and resource allocation technology based on an improved proximal policy optimization algorithm. BACKGROUND

[0002] In the modern intelligent medical system, edge computing task offloading technology is one of the core technologies to ensure the quality of remote medical services and the efficient operation of the system. With the rapid development of emerging technologies such as 5G communication, artificial intelligence, and the Internet of Things, the computing tasks in remote medical applications have shown explosive growth. The processing needs of these tasks have become more complex, and traditional task offloading strategies have gradually exposed their limitations, making it difficult to meet the strict requirements of real-time, reliability, and privacy protection in future intelligent medical systems.

[0003] Firstly, traditional task offloading strategies face the problem of low decision-making efficiency. Static strategies cannot be dynamically adjusted according to real-time network conditions, computing resource availability, and task characteristics, which can lead to unbalanced resource utilization. For example, in the remote image diagnosis scenario, a large amount of medical image data needs to be processed in real time, and static strategies may offload all tasks to edge servers, causing server overload; or process them all locally, causing terminal devices to respond slowly. Studies have shown that the performance gap between static strategies and optimal strategies can reach 30%-50%, seriously affecting the quality of medical services. Secondly, traditional strategies are difficult to meet the differentiated needs of medical tasks. Medical tasks have obvious priority differences, such as emergency tasks that need immediate response, and routine physical examination data that can be delayed. Static offloading strategies use a "one-size-fits-all" approach and cannot differentiate tasks based on their urgency, data sensitivity, and other characteristics. This rough management approach not only reduces system efficiency but also may cause urgent medical tasks to be delayed, posing a threat to patient safety. In addition, traditional strategies have serious shortcomings in privacy protection. Medical data contains a large amount of patient privacy information and is strictly protected by HIPAA, GDPR, and other regulations. Traditional offloading strategies mainly consider computing efficiency when making decisions, often ignoring the privacy attributes of data. For example, raw medical images containing patient identity information may be directly transmitted to public edge servers, increasing the risk of data leakage.

[0004] Traditional strategies exhibit significant limitations in adapting to network environments. Telemedicine application scenarios are complex and diverse, with network conditions varying dramatically, from high-speed connections in urban tertiary hospitals to unstable connections in remote areas. Static strategies are unable to perceive and adapt to these dynamic changes, resulting in either remote offloading during network congestion or local processing during smooth connections, wasting resources and degrading performance. In mobile healthcare scenarios, the network connectivity of ambulances and mobile medical units fluctuates frequently, making traditional strategies particularly incapable of effectively responding. Furthermore, traditional strategies lack intelligent learning and optimization capabilities. Medical task processing patterns exhibit certain regularities, such as the clustering of certain image analysis tasks during specific periods of time, or the low load on certain edge nodes at specific times. Traditional strategies are unable to learn these patterns from historical data or predict future task distribution, resulting in reactive decision-making and difficulty in achieving proactive optimization.

[0005] To address these challenges, AI-based intelligent task offloading has become a research hotspot in recent years. By incorporating machine learning, particularly reinforcement learning, the system can continuously optimize offloading strategies based on environmental feedback. However, most existing intelligent offloading solutions directly apply general reinforcement learning algorithms, failing to fully consider the specificities of medical scenarios. For example, traditional reinforcement learning algorithms may make high-risk decisions during the exploration phase, which is unacceptable in medical settings. The algorithms also have slow convergence rates, making it difficult to quickly adapt to changes in the medical environment. Furthermore, they lack support for multi-objective optimization and are unable to simultaneously address multiple dimensions, such as latency, energy consumption, and privacy.

[0006] In the design of edge computing architecture, the existing medical task offloading scheme has adopted a multi-layer architecture, but still has the following characteristics and deficiencies: (1) Lack of fine-grained decomposition mechanism of medical tasks: existing schemes usually make offloading decisions as a whole, ignoring the decomposability of medical tasks. For example, medical image analysis tasks can be decomposed into image preprocessing, feature extraction, lesion recognition, and other subtasks, and different subtasks have different requirements for computing resources and privacy protection. Coarse-grained task processing cannot fully utilize the advantages of distributed computing resources. (2) Weak coordination mechanism between edge nodes: although multiple edge nodes are deployed, edge nodes in existing schemes often run independently and lack effective coordination mechanisms. When a certain edge node is overloaded, it cannot dynamically migrate tasks to nearby idle nodes, causing local resource bottlenecks and low overall efficiency. This "information island" phenomenon is particularly prominent in scenarios where medical resources are unevenly distributed. (3) Lack of understanding of the characteristics of medical data: existing architectures fail to fully consider the special properties of medical data, such as temporal and spatial correlation, multi-modal features, and quality differences. For example, continuous physiological monitoring data has strong temporal correlation, and adjacent time data can be incrementally processed. However, existing schemes often process each time data independently, resulting in a lot of redundant computation. (4) Lack of differentiated quality of service guarantee: medical tasks have strict and differentiated requirements for service quality, such as sub-second response for emergency imaging and minute-level delay for health examination data. Existing architectures use a unified QoS standard, which cannot provide differentiated performance guarantees for different priority medical tasks, which may result in key tasks not being prioritized. SUMMARY

[0007] To address the challenges of traditional task offloading strategies in modern telemedicine scenarios, such as decision-making efficiency, differentiated services, privacy protection, and environmental adaptability, the present invention proposes an edge computing telemedicine task offloading method based on an improved PPO algorithm. By constructing a three-layer task offloading architecture and combining an improved PPO-MED algorithm, intelligent offloading decisions for medical tasks are achieved. The technical solutions provided by the present invention are as follows:

[0008] The edge computing telemedicine task offloading method based on the improved PPO algorithm includes the following steps:

[0009] Step 1, perform resource assessment and environment initialization, the medical image terminal detects local computing resources and establishes a secure connection with the edge server, the PPO-MED algorithm model is initialized, and the pre-trained policy network and value network are loaded;

[0010] Step 2, perform medical image offloading task processing based on the PPO-MED algorithm, and select the optimal action from the LCO, ESO, and EIPO offloading modes;

[0011] Step 3: PPO-MED algorithm optimization, through the reward function to update the policy network and value network.

[0012] Preferably, step 2 specifically includes the following steps:

[0013] Step 2.1: After receiving the medical imaging data, extract the task state vector ,in represents the task feature vector; It is the local resource status; Edge server status; For network conditions;

[0014] Step 2.2: Calculate the probability distribution of the unloading action based on the current state: ,in is the deep representation of state features, and is the output layer parameter of the policy network, softmax is the activation function, and the algorithm selects the optimal action from the three offloading modes: LCO, ESO and EIPO.

[0015] Preferably, when the EIPO mode is selected in step 2.2, intelligent preprocessing is performed locally, and the feature extraction function The original image Convert to compressed feature representation ,in For different types of feature extractors, is the corresponding weight; the preprocessed data is transmitted to the edge server. After receiving the task, the edge server calls the medical AI model for diagnostic analysis, generates the diagnostic results, and returns the diagnostic results to the terminal.

[0016] Preferably, step 3 specifically includes the following steps:

[0017] Step 3.1, calculate the comprehensive reward value: ,in is the actual delay, is the reference delay threshold; is energy consumption, is the maximum allowable energy consumption; Score for privacy protection; It is an indicator of diagnostic accuracy;

[0018] Step 3.2, update the policy network using the PPO clipping objective function:

[0019]

[0020] in is the importance sampling ratio, which represents the action probability ratio of the new strategy to the old strategy, is the advantage function, which represents the advantage of taking a specific action relative to the average level. A positive value indicates that the action is better than the average, and a negative value indicates that it is worse than the average. is the cropping parameter, is the expected value;

[0021] Step 3.3, update the value network by minimizing the mean square error:

[0022]

[0023] The learning target is calculated by Monte Carlo return or temporal difference method, which represents the expected cumulative reward that can be obtained by executing the strategy from the current state. The value network minimizes the predicted value. and target value The mean square error between them is gradually improved to improve the accuracy of the assessment of long-term returns under different conditions.

[0024] Preferably, the value network parameters in step 3.3 are Regularly synchronize from the main network through soft updates, and update the parameters Expressed as ,in is the update rate.

[0025] Preferably, step 3 further comprises the following steps:

[0026] In step 3.4, the experience replay buffer is maintained to store historical interaction data. During each training cycle, multiple small batches of data are randomly sampled from the buffer for training to improve sample utilization efficiency. The buffer adopts a priority sampling mechanism, giving higher sampling weights to samples with larger prediction errors.

[0027] In step 3.5, an adaptive exploration strategy is introduced to dynamically adjust the exploration level based on the training stage and task type. In the early stages of training, a higher action entropy regularization coefficient is used to encourage the strategy to explore different offloading solutions. As training progresses, the action entropy regularization coefficient is gradually reduced to allow the strategy to converge to the optimal solution. For emergency medical tasks, a greedy strategy is directly used to ensure service reliability.

[0028] In step 3.6, the model performance is continuously optimized through the incremental learning mechanism, and the elastic weight consolidation technology is used to maintain the memory of the learned knowledge when learning new task patterns.

[0029] Preferably, when multiple medical terminals exist, the PPO-MED is extended to a multi-agent version MA-PPO-MED, each terminal shares a global value network but maintains an independent policy network, and a cooperative reward function adds a fairness term, realizing implicit coordination between terminals through centralized training and distributed execution.

[0030] Preferably, it also includes a task priority management mechanism, which dynamically schedules medical tasks according to their urgency by maintaining a priority queue, and adopts a conservative strategy for high-priority tasks to ensure the reliability of task execution.

[0031] Preferably, it also includes an exception handling and recovery mechanism that triggers failover when an exception is detected, sets a task timeout threshold, and automatically reschedules a timeout task.

[0032] The edge computing remote medical task offloading system based on the improved PPO algorithm includes a medical image acquisition terminal, a PPO-MED intelligent decision module and an edge server cluster. The system adopts a distributed deployment architecture, the medical image acquisition terminal is deployed in primary medical institutions, and the edge server cluster is deployed in regional medical centers. The system dynamically selects the optimal offloading mode through the PPO-MED intelligent decision module, and the offloading modes include local computing offloading mode LCO, edge server offloading mode ESO and edge intelligent preprocessing offloading mode EIPO.

[0033] Compared with the prior art, the beneficial effects achieved by the present application are:

[0034] (1) Intelligent adaptive decision-making capability: Traditional remote medical task offloading schemes usually adopt fixed offloading strategies, which cannot be adjusted adaptively according to dynamic changes in network environment and task characteristics, resulting in serious performance degradation during network congestion or task peak periods. The improved PPO-MED algorithm in the present application realizes intelligent decision-making, which can dynamically select the optimal offloading strategy according to real-time system state, task urgency and resource availability, significantly improving the adaptability and service quality of the system in complex environments.

[0035] (2) Dual optimization of computing efficiency and privacy protection: Existing edge computing schemes often use simple binary offloading modes (local processing or edge processing), which are difficult to balance between computing efficiency and privacy protection. The edge intelligent preprocessing offloading mode (EIPO) proposed in the present application intelligently preprocesses and extracts features locally, significantly reducing data transmission volume and network overhead, and protecting patient privacy through data desensitization technology, realizing dual optimization of computing efficiency and privacy protection.

[0036] (3) Targeted optimization for medical scenarios: Traditional reinforcement learning algorithms lack targeted design for task characteristics when applied in medical scenarios, making it difficult to meet the high reliability requirements of medical services. The PPO-MED algorithm of the present application introduces a multi-objective reward function, considering four dimensions of delay, energy consumption, privacy protection and reliability, and dynamically adjusts the exploration strategy according to the urgency of medical tasks, ensuring stable and reliable services for urgent medical tasks, while optimizing the long-term performance of the system.

[0037] (4) Improvement of large-scale data processing capability: Existing telemedicine systems face serious bandwidth bottlenecks and delay problems when processing large-scale medical image data, affecting the real-time nature of diagnosis. The present application, through the cooperative work of the three-layer offloading architecture, combined with local lightweight processing, intelligent feature extraction and edge deep analysis, can compress the original medical data by several orders of magnitude, significantly reducing the transmission delay while ensuring the accuracy of diagnosis, especially suitable for remote diagnosis applications under the limited bandwidth conditions of primary medical institutions. BRIEF DESCRIPTION OF DRAWINGS

[0038] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, together with the embodiments of the present application, to explain the present application, and do not constitute a limitation on the present application. In the drawings:

[0039] Figure 1 is a method flowchart of the present application;

[0040] Figure 2 is a system architecture schematic diagram of the present application;

[0041] Figure 3 is an average task completion delay comparison of various methods including the present application under different task scales. DETAILED DESCRIPTION

[0042] The technical solutions in the embodiments of the present application will be described in detail below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the present application.

[0043] In order to make the above-mentioned purposes, features and effects of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0044] Example 1: An edge computing telemedicine task offloading method based on an improved PPO algorithm. To achieve intelligent task offloading decisions, the PPO-MED algorithm models the medical task offloading problem as a Markov decision process and optimizes the offloading strategy through deep reinforcement learning. The algorithm's state space includes multidimensional information such as task type, urgency, data volume, local and edge resource status, and network conditions. The action space is defined as the choice of three offloading modes: LCO, ESO, or EIPO.

[0045] The core of the PPO-MED algorithm is the design of a multi-objective reward function:

[0046]

[0047] in Reflects the delay in task completion, represents the energy consumption, Measuring the degree of privacy protection, Represents the reliability of task execution. Weight coefficient 、 、 and Dynamically adjust according to specific medical scenarios.

[0048] The algorithm adopts an improved strategy update mechanism and introduces adaptive clipping parameters:

[0049]

[0050] Importance Sampling Ratio Represents the action probability ratio of the new strategy to the old strategy. This ratio measures the magnitude of the strategy update; advantage function Indicates the advantage of taking a specific action relative to the average level, a positive value indicates that the action is better than the average, and a negative value indicates that it is worse than the average; for urgent tasks, adaptive clipping parameters Smaller, limiting the strategy update range to ensure stability; for routine tasks, allowing a larger exploration space, Dynamic adjustments are made based on the urgency of the task to ensure that high-priority medical tasks receive stable and reliable services.

[0051] The PPO-MED algorithm's execution process is as follows: the system first constructs a policy network and a value network. Both utilize a multi-layer, fully connected neural network architecture, using as the activation function. The policy network's output layer generates a probability distribution over three unloading modes using a function, while the value network outputs a single numerical value representing the estimated state value. During the initialization phase, key hyperparameters must be set, including the learning rate, batch size, and number of training rounds. A large-capacity experience replay buffer, D, is also established to store historical interaction data.

[0052] When the medical system starts running, the algorithm enters the online data collection and decision-making phase. Whenever a new medical task arrives, the system first extracts multi-dimensional feature information from the task request, including task attributes such as task type, urgency level, data size, privacy level, and current system state information such as local CPU utilization, available memory, edge server load, network bandwidth and latency, which together form the current state vector . Based on this state vector, the strategy network calculates the probability distribution of the three offloading modes . For urgent medical tasks, the algorithm automatically switches to a deterministic strategy, directly selecting the action with the highest probability , ensuring the reliability of task execution; for regular tasks, the algorithm performs random sampling according to the probability distribution, retaining a certain degree of exploration to discover better strategies.

[0053] After the task is executed according to the selected offloading mode, the system records the execution results in detail, including actual completion time, energy consumption, whether it is successfully completed, number of retries and other key indicators. Based on this feedback information, the algorithm calculates the comprehensive reward value R, which considers delay, energy consumption, privacy protection and reliability in four dimensions. Subsequently, the system stores the experience tuple composed of state , action , reward and next state in the buffer D, providing training data for subsequent policy optimization.

[0054] When sufficient experience data has accumulated in the buffer, the algorithm enters the policy optimization phase. This phase first uses the Generalized Advantage Estimation (GAE) method to calculate the advantage function value at each time, which reflects the degree of superiority or inferiority of taking a particular action relative to the average level. The calculation formula is:

[0055]

[0056] where is the time difference error. Then, the algorithm randomly samples multiple mini-batches from the buffer for batch training.

[0057] In the update process of the strategy network, the algorithm first calculates the importance sampling ratio of the new and old strategies:

[0058]

[0059] This ratio measures the magnitude of policy updates. Then apply the adaptive clipping mechanism to calculate the loss function , where the clipping parameter According to the task urgency dynamic adjustment, ensure that the more conservative update strategy is adopted for urgent tasks. In order to maintain moderate exploration, the algorithm also adds an entropy regularization term , which encourages the strategy to maintain a certain randomness. The final strategy loss is , using the Adam optimizer to update the policy network parameters by gradient descent.

[0060] At the same time, the value network also needs to be updated to provide more accurate state value estimates. The algorithm calculates the target value:

[0061]

[0062] Where is the discounted cumulative reward from the current time. By minimizing the mean square error between the predicted value and the target value , the value network parameters φ are constantly optimized. This dual network structure enables the algorithm to simultaneously learn the optimal policy and accurate value estimates.

[0063] During the entire running process, the algorithm also contains various dynamic adjustment mechanisms. The exploration rate gradually decreases as the training progresses, and the entropy regularization coefficient decreases according to the exponential decay law, so that the strategy gradually transitions from active exploration in the early stage to stable utilization in the later stage. The learning rate is adjusted using the cosine annealing strategy, maintaining a high learning rate in the early stage of training to accelerate convergence, and reducing the learning rate in the later stage for fine tuning. The system also monitors key performance indicators in real time, including average task delay, success rate, resource utilization, etc. When performance anomalies are detected, the policy rollback mechanism is triggered to restore to the previous stable version.

[0064] In actual deployment, the PPO-MED algorithm is closely integrated with the three-layer offloading architecture. The algorithm continuously monitors the system running state and dynamically optimizes offloading decisions based on historical experience and real-time feedback. For example, when network congestion is detected, the algorithm will increase the proportion of EIPO mode usage, reducing data transmission through local preprocessing; when the edge server load is low, the algorithm will preferentially select the ESO mode to fully utilize edge computing resources.

[0065] The system also introduces a task priority queue management mechanism. Medical tasks are divided into multiple priorities according to urgency, such as emergency tasks, routine examinations, health monitoring, etc. High-priority tasks can preempt low-priority tasks for resources, ensuring that urgent medical needs are responded to in a timely manner. At the same time, the system maintains a record of task execution history for algorithm training and performance optimization.

[0066] To ensure the security of medical data, the system establishes encrypted transmission channels between each layer, using end-to-end encryption to ensure data transmission security. On edge servers, data processing is performed in isolated secure containers, and temporary data is immediately cleared after processing to prevent data leakage.

[0067] The present invention specifically comprises the following steps:

[0068] Step 1: Perform resource evaluation and environment initialization. The medical imaging terminal detects local computing resources and establishes a secure connection with the edge server. Initialize the PPO-MED algorithm model and load the pre-trained policy network. and value network ,in and are the network parameters respectively.

[0069] Step 2: Medical imaging task processing. The specific steps are as follows:

[0070] Step 2.1, task reception and feature extraction. When the medical device collects the patient's image data, the system constructs the task state vector ,in Represents the task feature vector, which includes information such as image type, data volume, and urgency; It is the local resource status; Edge server status; For network conditions.

[0071] Step 2.2, intelligent offloading decision. The PPO-MED algorithm calculates the probability distribution of offloading actions based on the current state: ,in is the deep representation of state features, and is the policy network output layer parameter. The algorithm selects the optimal action from the three offloading modes: LCO, ESO and EIPO .

[0072] Step 2.3, execute the unloading strategy. When the EIPO mode is selected, the system performs intelligent preprocessing locally, through the feature extraction function The original image Convert to compressed feature representation ,in For different types of feature extractors, The preprocessed data is transmitted to the edge server for in-depth analysis.

[0073] Step 2.4: Edge computing and result generation. After receiving the task, the edge server calls the medical AI model. Perform diagnostic analysis and generate diagnostic results The system packages the analysis results, visual annotations, and diagnostic suggestions and returns them to the terminal.

[0074] Step 3, PPO-MED algorithm optimization mechanism, the specific steps are as follows:

[0075] Step 3.1, reward function calculation. After the task is completed, the system calculates the comprehensive reward value: ,in is the actual delay, is the reference delay threshold; is energy consumption, is the maximum allowable energy consumption; Score for privacy protection; An indicator of diagnostic accuracy.

[0076] Step 3.2, policy network update. Update the policy parameters using the PPO clipping objective function:

[0077]

[0078] in is the importance sampling ratio, is the advantage function estimate, is the cropping parameter.

[0079] Step 3.3, value network update. Update the value network by minimizing the mean square error:

[0080]

[0081] In this process, the value network continuously optimizes its ability to estimate the value of the state through supervised learning. Specifically, The learning target of the value function is calculated by Monte Carlo return or temporal difference method, which represents the expected cumulative reward that can be obtained by executing the strategy from the current state. The value network minimizes the predicted value. and target value The mean square error between them is gradually improved to improve the accuracy of the assessment of long-term benefits under different system states.

[0082] In order to improve training stability, the system adopts target network technology, which is to maintain a target value network with slower parameter updates. , whose parameters are regularly synchronized from the main network via soft updates: , where τ is the update rate (usually 0.01). This design effectively avoids numerical oscillations during training and ensures smooth convergence of value estimation.

[0083] Furthermore, accurate estimation of the value network is crucial for computing the advantage function It is crucial because Reflects the degree of advantage of taking a specific action relative to the average level. Accurate value estimation can help the policy network better identify which actions are truly superior to the baseline level, guiding the policy to improve in a more optimal direction, which is of great significance to the optimization of medical task offloading decisions.

[0084] Step 3.4, Experience Replay and Batch Optimization. The system maintains an experience replay buffer Buffer, storing historical interaction data . In each training cycle, multiple small batches of data are randomly sampled from the buffer for training, improving sample utilization efficiency. In specific implementation, the buffer uses a priority sampling mechanism, giving higher sampling weight to samples with larger prediction error: , where is the time difference error, and a is the priority index.

[0085] Step 3.5, Exploration and Utilization Balance Mechanism. The algorithm introduces an adaptive exploration strategy, dynamically adjusting the exploration degree according to the training stage and task type. For the early training stage, a higher action entropy regularization coefficient β_entropy is used to encourage the strategy to explore different offloading schemes; as training progresses, β_entropy is gradually reduced, allowing the strategy to converge to the optimal solution. For urgent medical tasks, a greedy strategy is directly used to ensure service reliability.

[0086] Step 3.6, Multi-Agent Collaborative Optimization. When there are multiple medical terminals in the system, PPO-MED is extended to a multi-agent version MA-PPO-MED. Each terminal shares a global value network but maintains an independent policy network . Through centralized training and distributed execution, implicit coordination between terminals is achieved, avoiding resource competition. The collaborative reward function adds a fairness term: to promote load balancing.

[0087] Step 3.7, Online Adaptation and Incremental Learning. The system continuously collects real-time running data after deployment, constantly optimizing model performance through incremental learning mechanisms. The Elastic Weight Consolidation (EWC) technique is used to maintain memory of learned knowledge when learning new task patterns: , where is the Fisher information matrix diagonal element, is the optimal parameter of the old task.

[0088] The system also has some system operation guarantee mechanisms as follows:

[0089] 1. Task priority management. The system maintains a priority queue , dynamically dispatches according to the urgency of the medical task. For high priority tasks, the system adopts a conservative strategy , ensuring the reliability of task execution.

[0090] 2. Exception handling and recovery. When a system exception is detected, the failover mechanism is triggered. Set the task timeout threshold. , timeout tasks are automatically rescheduled. The system uses the state transition probability , predict possible failure scenarios and formulate response strategies in advance.

[0091] Through the above-mentioned implementation methods, the present invention realizes the intelligent offloading of medical imaging diagnosis tasks, effectively reduces system delay and resource consumption while ensuring the quality of diagnosis, and provides an efficient and reliable technical solution for telemedicine applications.

[0092] Figure 3 The comparison of the average task completion delay of five task offloading algorithms under different task scales is shown, namely the PPO-MED algorithm of the present invention, traditional PPO, greedy algorithm, polling scheduling and random allocation strategy. The horizontal axis represents the number of tasks and the vertical axis represents the average task completion delay, both of which use logarithmic scales. Figure 3 It can be seen that as the task scale increases, the PPO-MED algorithm, i.e., the present invention, shows significant advantages. Especially in large-scale task scenarios, the delay increases very slowly and always remains at the lowest level. 5 When , the average delay of PPO-MED is still at a low level, while the delay of traditional PPO has increased significantly, the delay of polling scheduling has increased significantly, and the random allocation strategy performs the worst.

[0093] This outstanding performance is primarily due to the PPO-MED algorithm's three-layer intelligent offloading architecture and adaptive decision-making mechanism. PPO-MED dynamically optimizes task allocation strategies through deep reinforcement learning, intelligently selecting local processing (LCO), edge processing (ESO), or intelligent pre-processing (EIPO) based on task characteristics, system load, and network status, effectively avoiding resource bottlenecks. The innovative EIPO mode, in particular, significantly reduces data transmission volume and network overhead through local pre-processing.

[0094] In contrast, traditional PPO uses a fixed binary offloading strategy, lacking targeted optimization for medical task characteristics and experiencing a sharp decline in performance under high load. While the greedy algorithm performs reasonably well for small-scale tasks, its short-sighted decision-making mechanism leads to severe load imbalance as the task scale increases. Round-robin scheduling, due to its mechanical task allocation, incurs significant scheduling overhead and queuing delays. Random allocation strategies are completely unable to adapt to dynamically changing system environments and perform the worst.

[0095] Overall, the PPO-MED algorithm of the present application exhibits excellent performance advantages under various task scales, fully demonstrating its practical value and technical advancement in the remote medical edge computing scenario.

[0096] Embodiment 2: An edge computing remote medical task offloading system based on an improved PPO algorithm, including a medical image acquisition terminal, a PPO-MED intelligent decision module, and an edge server cluster. The system adopts a distributed deployment architecture, as shown in Figure 2 The medical image acquisition terminal is deployed in primary medical institutions, and the edge server cluster is deployed in regional medical centers. According to the characteristics of medical tasks and the state of system resources, the PPO-MED intelligent decision module dynamically selects the optimal offloading mode, including local computation offloading mode (LCO), edge server offloading mode (ESO), and edge intelligent preprocessing offloading mode (EIPO).

[0097] The local computation offloading mode mainly handles lightweight tasks and privacy-sensitive data. When the medical terminal device receives a task request, it first evaluates the computational complexity and data sensitivity of the task. For lightweight tasks such as basic physiological parameter computation and simple image preprocessing, the system directly processes them locally. The specific process is as follows: the terminal device receives raw medical data such as electrocardiogram signals and blood pressure data, and executes basic algorithms such as signal filtering, feature extraction, and threshold judgment through the local processor. The processing results can be directly fed back to the user or used as input for subsequent processing. The local computation offloading layer is also responsible for managing local cache, storing commonly used medical data templates and processing results, and reducing repeated computation. When a task is detected to have high privacy sensitivity, the system preferentially selects local processing to avoid network transmission of sensitive data.

[0098] The edge server offloading mode is responsible for handling medium complexity tasks and latency-sensitive applications. When a task exceeds the local processing capacity but requires fast response, the system offloads the task to the edge server. The offloading process is as follows: the terminal device first establishes a secure connection with the edge server, and then sends the task data and processing requirements in a package. After receiving the task, the edge server invokes the corresponding medical AI model for processing, such as medical image analysis, real-time video encoding and decoding, and multi-modal data fusion. During processing, the edge server can cooperate with other edge nodes to achieve load balancing and resource sharing. After processing is completed, the results are returned to the terminal device through an optimized transmission protocol. To ensure service quality, the system monitors the resource utilization and network state of the edge server in real time, and dynamically adjusts the task allocation strategy.

[0099] The edge intelligence preprocessing offloading mode is an innovative design of the present application, which establishes an intelligent preprocessing mechanism between the local and edge servers. The workflow of this layer is as follows: firstly, the original medical data is intelligently preprocessed locally, including data dimension reduction, feature extraction and privacy desensitization. For example, for high-resolution medical images, the system extracts key feature vectors locally, compressing the data volume from tens of MB to several KB; for data containing patient identity information, the system performs de-identification processing. The preprocessed lightweight data is then offloaded to the edge server for deep analysis. This way not only reduces network transmission overhead, but also protects patient privacy. After the edge server completes the analysis based on the preprocessed data, it can request specific parts of the original data as needed, realizing on-demand transmission.

[0100] The above merely describes the preferred embodiments of the present application and is not intended to limit the present application. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art can modify the technical solutions described in the foregoing embodiments or make equivalent replacements to some technical features. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. The edge computing telemedicine task offloading method based on the improved PPO algorithm is characterized by: The following steps are involved: Step 1: Perform resource evaluation and environment initialization. The medical imaging terminal detects local computing resources and establishes a secure connection with the edge server. The PPO-MED algorithm model is initialized and the pre-trained policy network and value network are loaded. Step 2: Process the medical image offloading task based on the PPO-MED algorithm and select the optimal action from the three offloading modes: LCO, ESO, and EIPO. Step 3: PPO-MED algorithm optimization, through the reward function to update the policy network and value network.

2. The edge computing telemedicine task offloading method based on the improved PPO algorithm according to claim 1 is characterized in that: Step 2 specifically includes the following steps: Step 2.1: After receiving the medical imaging data, extract the task state vector ,in represents the task feature vector; It is the local resource status; Edge server status; For network conditions; Step 2.2: Calculate the probability distribution of the unloading action based on the current state: ,in is the deep representation of state features, and is the output layer parameter of the policy network, softmax is the activation function, and the algorithm selects the optimal action from the three offloading modes: LCO, ESO and EIPO.

3. The edge computing telemedicine task offloading method based on the improved PPO algorithm according to claim 2 is characterized in that: When the EIPO mode is selected in step 2.2, intelligent preprocessing is performed locally, and the feature extraction function The original image Convert to compressed feature representation ,in For different types of feature extractors, is the corresponding weight; The preprocessed data is transmitted to the edge server. After receiving the task, the edge server calls the medical AI model for diagnostic analysis, generates diagnostic results, and returns the diagnostic results to the terminal.

4. The edge computing telemedicine task offloading method based on the improved PPO algorithm according to claim 1 is characterized in that: Step 3 specifically includes the following steps: Step 3.1, calculate the comprehensive reward value: ,in is the actual delay, is the reference delay threshold; is energy consumption, is the maximum allowable energy consumption; Score for privacy protection; It is an indicator of diagnostic accuracy; Step 3.2, update the policy network using the PPO clipping objective function: ； in is the importance sampling ratio, which represents the action probability ratio of the new strategy to the old strategy, is the advantage function, which represents the advantage of taking a specific action relative to the average level. A positive value indicates that the action is better than the average, and a negative value indicates that it is worse than the average. is the cropping parameter, is the expected value; Step 3.3, update the value network by minimizing the mean square error: ； The learning target is calculated by Monte Carlo return or temporal difference method, which represents the expected cumulative reward that can be obtained by executing the strategy from the current state. The value network minimizes the predicted value. and target value The mean square error between them is gradually improved to improve the accuracy of the assessment of long-term returns under different conditions.

5. The edge computing telemedicine task offloading method based on the improved PPO algorithm according to claim 4 is characterized in that: Value network parameters in step 3.3 Regularly synchronize from the main network through soft updates, and update the parameters Expressed as ,in is the update rate.

6. The edge computing telemedicine task offloading method based on the improved PPO algorithm according to claim 4 is characterized in that: Step 3 further includes the following steps: In step 3.4, the experience replay buffer is maintained to store historical interaction data. During each training cycle, multiple small batches of data are randomly sampled from the buffer for training to improve sample utilization efficiency. The buffer adopts a priority sampling mechanism, giving higher sampling weights to samples with larger prediction errors. In step 3.5, an adaptive exploration strategy is introduced to dynamically adjust the exploration level based on the training stage and task type. In the early stages of training, a higher action entropy regularization coefficient is used to encourage the strategy to explore different offloading solutions. As training progresses, the action entropy regularization coefficient is gradually reduced to allow the strategy to converge to the optimal solution. For emergency medical tasks, a greedy strategy is directly used to ensure service reliability. In step 3.6, the model performance is continuously optimized through the incremental learning mechanism, and the elastic weight consolidation technology is used to maintain the memory of the learned knowledge when learning new task patterns.

7. The edge computing telemedicine task offloading method based on the improved PPO algorithm according to claim 6 is characterized in that: When there are multiple medical terminals, PPO-MED is extended to the multi-agent version MA-PPO-MED. Each terminal shares the global value network but maintains an independent policy network. The collaborative reward function adds a fairness term and achieves implicit coordination between terminals through centralized training and distributed execution.

8. The edge computing telemedicine task offloading method based on the improved PPO algorithm according to claim 1 is characterized in that: It also includes a task priority management mechanism, which maintains a priority queue and dynamically schedules medical tasks according to their urgency. For high-priority tasks, a conservative strategy is adopted to ensure the reliability of task execution.

9. The edge computing telemedicine task offloading method based on the improved PPO algorithm according to claim 1 is characterized in that: It also includes an exception handling and recovery mechanism. When an exception is detected, a failover is triggered, a task timeout threshold is set, and timed-out tasks are automatically rescheduled.

10. An edge computing telemedicine task offloading system based on an improved PPO algorithm, used to execute the edge computing telemedicine task offloading method based on an improved PPO algorithm according to any one of claims 1 to 9, characterized in that: The system includes medical image acquisition terminals, PPO-MED intelligent decision-making modules and edge server clusters. The system adopts a distributed deployment architecture. Medical image acquisition terminals are deployed in primary medical institutions, and edge server clusters are deployed in regional medical centers. The system dynamically selects the optimal offloading mode through the PPO-MED intelligent decision-making module. The offloading modes include local computing offloading mode LCO, edge server offloading mode ESO and edge intelligent preprocessing offloading mode EIPO.

Images