Resource and task aware visual processing edge adaptive decision-making method

Abstract

The invention belongs to the technical field of artificial intelligence and computer vision, particularly relates to a visual processing edge adaptive decision-making method for resource and task perception, and aims to solve the problem of scheduling mismatch caused by resource dynamic change and task demand diversity in visual task processing in an edge computing environment. The method comprises the following steps: collecting multi-dimensional resource state data of edge nodes in real time to form a resource state vector with high time resolution; analyzing the visual task request, and constructing a quantifiable task feature vector; and establishing a resource-task association mapping model based on a dynamic weight distribution mechanism. The method also supports cross-edge domain collaborative decision, and processes a pipeline dynamic reconstruction and security isolation mechanism. According to the technical scheme, the fluctuation of the resource utilization rate is reduced to 15% or below, the average task processing delay is reduced to 60%, the scheduling satisfaction degree is improved by 40% or above, and the self-adaptability and the service quality guarantee capability of the edge vision system are remarkably enhanced.

Description

Technical Field

[0001] This application belongs to the field of artificial intelligence and computer vision technology, specifically relating to a visual processing edge adaptive decision-making method for resource and task perception. Background Technology

[0002] With the continuous evolution of edge computing technology, its core role in industrial intelligence, real-time information processing, and distributed system control is becoming increasingly prominent. As a key hub connecting terminal devices and cloud data centers, edge nodes undertake important functions such as low-latency response, local autonomous decision-making, and multi-source information fusion processing. Especially in complex and dynamic environments, the edge side needs to efficiently analyze and schedule sensed data in real time to support the stable operation of advanced process control and information planning tasks, which places stringent requirements on resource utilization efficiency, task response capabilities, and system adaptability.

[0003] Edge computing architectures for visual processing are gradually becoming the technological foundation for scenarios such as intelligent manufacturing, unmanned inspection, and intelligent monitoring. Their core objective is to achieve localized analysis of image data and a closed-loop system for real-time decision-making. Such systems typically need to perform feature extraction, object recognition, and behavior assessment on high-throughput video streams under limited computing power. They also need to coordinate multi-dimensional resource status, including network bandwidth, storage capacity, and computing load, to complete task priority scheduling and processing flow optimization, thereby ensuring the quality of service for critical applications.

[0004] Existing technologies for achieving collaborative resource and task perception in visual processing mechanisms generally face multiple bottlenecks. First, resource allocation strategies are mostly based on static thresholds or periodic monitoring, lacking the dynamic perception capability for real-time load fluctuations at edge nodes, leading to frequent idle or overloaded computing resources. Second, task scheduling models often consider single indicators such as latency or energy consumption in isolation, failing to establish a correlation between task criticality and system resource status, making it difficult to achieve adaptive distribution and execution path reconstruction for heterogeneous tasks. Furthermore, when handling sudden visual events, existing architectures rely on preset rules to trigger resource reallocation, unable to make proactive decisions based on contextual semantics and operational status, resulting in both response delays and resource mismatches. These problems are particularly prominent in high-concurrency, real-time-critical industrial control scenarios, severely restricting the robustness and autonomous decision-making capabilities of edge vision systems. Therefore, a resource and task-aware adaptive decision-making method for edge visual processing is urgently needed. Summary of the Invention

[0005] The purpose of this invention is to provide a resource and task-aware visual processing edge adaptive decision-making method that can effectively solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: a resource and task-aware visual processing edge adaptive decision-making method, comprising the following specific steps: Step (1) Real-time acquisition of multi-dimensional resource status data of edge nodes: Through the resource monitoring agent module deployed in the edge computing device, the CPU utilization rate, memory occupancy rate, network bandwidth availability, storage space balance and temperature sensor readings of the current node are periodically collected to form a resource status vector containing 5 dimensions. The sampling period is 100 milliseconds to ensure high temporal resolution perception of dynamic changes in resources; Step (2) Acquiring and parsing the visual task features to be processed: receiving Visual task requests from front-end sensing devices are analyzed, including task type, input data volume, expected processing latency limit, computational complexity level, and task priority label. A task feature matrix is ​​constructed, and the unstructured task description is transformed into a quantifiable feature vector through a semantic encoder. Step (3) Establish a resource-task association mapping model: Based on the resource state vector and task feature vector obtained in steps (1) and (2), a resource adaptability scoring function is constructed using a dynamic weight allocation mechanism. This function includes a resource demand matching item, a load balancing penalty item, and a task urgency incentive item. The three coefficients are adjusted online based on historical scheduling feedback. Output the comprehensive adaptation score of each task under the current resource conditions; Step (4) Execute adaptive task scheduling and resource reallocation: Based on the adaptation score generated in step (3), prioritize all tasks to be processed, trigger the resource reallocation strategy, if the comprehensive score is lower than the preset threshold, start the computing power migration process, unload the task to the nearby edge node or cloud, if the resource is locally overloaded, activate the lightweight model replacement mechanism, and use the compressed neural network model to replace the original model to reduce the computing load; Step (5) Implement context-aware decision optimization: Introduce the operation status assessment module, combine environmental context information, including time series trends, task bursts Based on the system stability index and system stability indicators, a multilayer perceptron prediction model is constructed to predict the fluctuation of resource demand and task arrival patterns within the next 200 milliseconds. Based on the prediction results, the resource reservation strategy and cache preloading mechanism are adjusted in advance to realize the decision-making paradigm shift from passive response to active regulation. Step (6) Closed-loop feedback and model self-learning: The results of each scheduling execution, including task completion delay, resource utilization rate change, energy consumption increment and user satisfaction score, are used as feedback signals to input to the reinforcement learning agent module. This module uses a deep deterministic strategy gradient algorithm to continuously optimize the parameter weights in the resource-task mapping model and form a closed-loop adaptive optimization mechanism.

[0007] Preferably, in step (1), the resource monitoring agent module is integrated into the kernel layer of the edge operating system. It directly reads the hardware register status through the system call interface, avoiding the delay caused by user-mode polling, and ensuring that the measurement error of the CPU utilization and memory occupancy is less than 1%. The network bandwidth monitoring adopts traffic mirroring and packet capture technology, combined with the sliding window algorithm to calculate the instantaneous throughput. The storage space monitoring distinguishes between hot data area and cold data area, and records the read / write frequency and fragmentation degree respectively. The temperature sensor is deployed in the hot area of ​​the chip. The sampled data is filtered by Kalman to eliminate noise interference. The final resource status vector is updated every 100 milliseconds and written to the shared memory buffer for subsequent module calls.

[0008] Preferably, in step (2), the visual task request is initiated by an industrial camera, drone, or mobile inspection terminal. The task types include four categories: target detection, behavior recognition, anomaly alarm, and image enhancement. The input data volume is calculated based on the video resolution and frame rate. The upper limit of the expected processing latency is defined by the application layer. The computational complexity level is divided according to the number of floating-point operations of the called model. The task priority label is represented by a 3-bit binary code, corresponding to safety criticality, business importance, and time sensitivity, respectively. The semantic encoder uses a pre-trained bidirectional long short-term memory network to embed the task description text and outputs a 128-dimensional feature vector. This vector and the structured parameters together constitute the task feature matrix, ensuring the comparability and unified modeling basis between heterogeneous tasks.

[0009] Preferably, the resource suitability scoring function in step (3) is defined as follows: ,in To determine the resource demand matching degree, cosine similarity is used to calculate the degree of similarity between the resource distribution required by the task and the current available resource distribution. This is a load balancing penalty term, calculated based on the deviation of the current node's load standard deviation from the mean of its neighboring nodes. The task urgency incentive term is obtained by normalizing the product of the task priority label and the remaining tolerable latency, with a weighting coefficient. , , The initial values ​​are set to 0.4, 0.3, and 0.3, and are dynamically adjusted by the reinforcement learning agent module based on the scheduling success rate and resource waste rate. The update cycle is a gradient update performed after every 100 scheduling decisions are completed.

[0010] Preferably, in step (4), the priority reordering adopts a non-dominated sorting strategy, takes the adaptation score as the main sorting basis, and combines the task arrival time to carry out fairness constraints to prevent low-scoring tasks from being starved for a long time. The computing power migration process selects the optimal forwarding path based on the topological connectivity graph between edge nodes, taking into account link delay and the remaining capacity of intermediate nodes. The lightweight model replacement mechanism has three built-in compression models, which correspond to 70%, 50% and 30% of the parameters of the original model, respectively. The replacement decision is triggered by whether the current central processor utilization rate exceeds 85% for three consecutive cycles. The model switching process is accompanied by the output confidence threshold adjustment to ensure that the accuracy decrease does not exceed the preset tolerance.

[0011] Preferably, in step (5), the operational status assessment module performs a status sampling every 50 milliseconds. The time series trend is extracted by exponential smoothing to determine the direction of growth or decay. The task burst index is obtained by normalizing the standard deviation of the number of new tasks per unit time. The system stability index is calculated by weighting voltage fluctuation, temperature gradient and task drop rate. The multilayer perceptron prediction model includes an input layer, two hidden layers and an output layer, with 64, 32 and 16 neurons respectively. The activation function is a modified linear unit. The training data comes from historical operation logs. The prediction results are used to dynamically adjust the weight coefficient prediction in step (3) and the resource reservation ratio in step (4) to achieve forward-looking regulation across cycles.

[0012] Preferably, in step (6), the reinforcement learning agent module runs in the edge coordinator controller. The state space is formed by concatenating the resource state vector and the task feature matrix. The action space is defined as the fine-tuning direction of three weight coefficients, and the reward function... ,in To determine the success rate of the task. This represents the improvement in energy efficiency ratio. The resource idle rate is represented by a coefficient. , , The values ​​are 0.5, 0.3, and 0.2 respectively. The deep deterministic policy gradient algorithm uses an experience replay mechanism to store historical state-action-reward tuples with a batch size of 32. The target network is updated softly once every 1000 steps to ensure stable convergence during the learning process.

[0013] Preferably, it also includes a cross-edge domain collaborative decision-making mechanism, in which a master node is set up in a cluster composed of multiple edge nodes. The master node regularly collects resource status summaries and task load reports of each member node to build a global resource view. When any node triggers a resource crisis warning, the master node starts a cross-domain task redirection protocol, selects the optimal receiving node based on the global view, and coordinates the reservation of communication resources to achieve cluster-level load balancing and fault tolerance. The communication overhead of this mechanism is controlled within 5% of the total bandwidth and the synchronization delay is less than 10 milliseconds.

[0014] Preferably, it also includes a dynamic reconstruction function for the visual processing pipeline, which presets multiple processing flow templates for different task types, including single-stage detection flow, two-stage recognition flow and multimodal fusion flow. The scheduling decision module automatically matches the optimal flow according to the task characteristics, and dynamically enables or disables specific processing stages based on resource feedback during runtime. For example, it turns off the high-resolution reconstruction module when computing power is tight, and enables the data compression encoding module when bandwidth is limited, to ensure that the processing flow and resource conditions are aligned in real time.

[0015] Preferably, it also includes a security isolation mechanism, which establishes virtualized containers or lightweight sandboxes in the task execution environment. Each task runs in an independent security domain to prevent malicious code or abnormal behavior from affecting other tasks or the system kernel. The resource monitoring module monitors the resource usage of each container at the same time. Once an out-of-bounds behavior is detected, the isolation policy is immediately triggered and the security audit log is reported to ensure the overall security and reliability of the edge system.

[0016] Compared with the prior art, the present invention has the following beneficial effects: It achieves high-frequency, low-error dynamic perception of multi-dimensional resource status of edge nodes, breaks through the technical limitation of traditional static threshold monitoring methods that cannot respond to rapid load fluctuations, reduces the fluctuation range of resource utilization to below 15%, and significantly reduces the frequency of idle computing power and overload phenomena. A quantitative correlation mapping model between task characteristics and resource status was constructed. Through a dynamic weight scoring mechanism, the deep integration of task criticality and system carrying capacity was achieved, which solved the resource mismatch problem caused by the isolated optimization of a single indicator in the existing scheduling strategy. The overall satisfaction of task scheduling was improved by more than 40%. By introducing a context-aware, proactive decision-making optimization mechanism and utilizing an operational situation prediction model to adjust resource allocation strategies in advance, the traditional technology-dependent passive response mode based on preset rules is changed. This improves the system's response speed to sudden visual events by 2 times and reduces the average task processing latency to 60% of the original latency. A closed-loop self-learning scheduling optimization framework was established. By continuously iterating the resource-task mapping parameters through a reinforcement learning agent module, the system has the ability to continuously improve decision quality as the running time increases. After 1000 scheduling iterations, the resource waste rate decreased by 50% and the energy efficiency ratio increased by 35%. It supports cross-edge domain collaboration and dynamic reconfiguration of processing pipelines, enhancing the system's robustness and flexibility in complex industrial scenarios. It can maintain a task completion rate of over 95% under high concurrency and strong real-time conditions, meeting the stringent service quality requirements of advanced process control and information planning. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall technical architecture of the visual processing edge adaptive decision-making method for resource and task awareness proposed in this invention. Figure 2 This is a schematic diagram of the core principle framework of resource-task association mapping and adaptive scheduling decision in this invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0019] Please refer to Figure 1 and Figure 2In the above-mentioned resource and task-aware visual processing edge adaptive decision-making method, step (1) involves real-time collection of multi-dimensional resource status data of edge nodes. Through a resource monitoring agent module deployed in the edge computing device, the CPU utilization, memory usage, network bandwidth availability, storage space availability, and temperature sensor readings of the current node are periodically collected to form a resource status vector containing five dimensions. The sampling period is 100 milliseconds, ensuring high temporal resolution perception of dynamic resource changes. Specifically, the resource monitoring agent module is integrated into the kernel layer of the edge operating system. It directly reads the hardware register status through the system call interface, avoiding the delay caused by user-mode polling, and ensuring that the measurement error of CPU utilization and memory usage is less than 1%. This collection mechanism runs in the operating system kernel space, bypassing the conventional process scheduling and interrupt handling path, and directly accesses the underlying status registers of the processor performance monitoring unit (PMU) and memory management unit (MMU), thereby achieving microsecond-level response and sub-millisecond-level synchronization accuracy. CPU utilization is calculated based on the ratio of each core's running cycles to idle cycles, with a sampling window length of 100 milliseconds. A moving average filter is used to suppress instantaneous jitter. Memory usage monitoring covers both physical and virtual memory, distinguishes between user space and kernel space memory allocation, and tracks memory fragmentation in real time. The fragmentation index is calculated using the distribution of free blocks in the Buddy System. Network bandwidth monitoring employs traffic mirroring and packet capture technologies. Port mirroring is set up at the data link layer, copying all data packets entering and leaving the network card to a dedicated monitoring channel. Instantaneous throughput is calculated using a sliding window algorithm with a window size of 50 milliseconds and a step size of 10 milliseconds. It supports the classification and statistics of multi-protocol traffic such as TCP, UDP, and QUIC, and identifies the proportion of bursty data streams and long-lived connection traffic. Storage space monitoring distinguishes between hot and cold data areas. Hot data areas correspond to frequently read and written data blocks, while cold data areas correspond to archived data accessed infrequently. The read / write frequency, IOPS (Input / Output Operations Per Second), and fragmentation level are recorded for each area. A Logical Block Address (LBA) mapping table is used to analyze the storage layout, and a B+ tree index structure accelerates metadata lookup. Temperature sensors are deployed in hot areas of the chip, including near the CPU core, GPU core, and power management unit. High-precision digital sensors (such as the DS18B20) are used, with a sampling frequency of once every 100 milliseconds. Raw data undergoes Kalman filtering to eliminate noise interference. The filter state transition matrix is ​​modeled as a first-order inertial system, with the process noise covariance set to 0.01 and the observation noise covariance set to 0.1 to ensure the dynamic response and stability of temperature readings.The resulting resource status vector is encapsulated in a structured data format (such as Protobuf) and contains five fields: CPU utilization (0-100%, precision 0.1%), memory usage (0-100%, precision 0.1%), available network bandwidth (Mbps), available storage space (GB), and highest temperature (°C). It is updated every 100 milliseconds and written to a shared memory buffer. This buffer adopts a circular queue structure with a capacity of 10 historical states, which can be read by subsequent modules as needed. Access permissions are constrained by the operating system's memory protection mechanism, allowing only the scheduling decision module and the operational status assessment module to perform read-only mapping through the mmap system call.

[0020] In the above method, step (2) involves acquiring and parsing the visual task features to be processed, receiving visual task requests from front-end sensing devices, parsing their task type, input data volume, expected processing latency limit, computational complexity level, and task priority label, constructing a task feature matrix, and converting the unstructured task description into a quantifiable feature vector through a semantic encoder. Specifically, the visual task request is initiated by an industrial camera, drone, or mobile inspection terminal. The task types include four categories: target detection, behavior recognition, anomaly alarm, and image enhancement. The input data volume is calculated based on the video resolution and frame rate. For example, a 1080p@30fps video stream corresponds to a data rate of approximately 1.5Gbps. The expected processing latency limit is defined by the application layer. For example, security monitoring tasks require an end-to-end latency of less than 200 milliseconds, while data analysis tasks can tolerate a latency of more than 500 milliseconds. The computational complexity level is divided into three levels based on the number of floating-point operations (FLOPs) of the called model: low (<1GFLOPs), medium (1~5GFLOPs), and high (>5GFLOPs), used to estimate the load pressure of the task on the central processing unit or GPU. The task priority label is represented by a 3-bit binary code, corresponding to safety criticality, business importance, and time sensitivity. The first bit represents safety criticality (1 indicates that it involves personal or equipment safety, such as fire alarm), the second bit represents business importance (1 indicates that it is related to the core production process, such as quality inspection), and the third bit represents time sensitivity (1 indicates that it must be completed within a limited time, such as real-time tracking). This forms 8 priority combinations, with the highest priority being "111" and the lowest being "000". The semantic encoder uses a pre-trained bidirectional long short-term memory (Bi-LSTM) network to embed the task description text. The input text is a natural language description provided by the task initiator, such as "detection of abnormal pedestrian gathering at the east gate entrance of the factory area". The encoder first decomposes it into a sequence of words through a word segmentation module. Each word is mapped to a 300-dimensional vector by a pre-trained word vector model (such as Word2Vec or GloVe), and then input into the bidirectional LSTM network. The forward LSTM captures the contextual dependencies from left to right, and the backward LSTM captures the semantic associations from right to left. Finally, the hidden states in the two directions are concatenated to output a 128-dimensional feature vector. This feature vector, together with the aforementioned structured parameters (task type, data volume, latency limit, complexity level, priority label), constitutes the task feature matrix. The matrix dimension is N×M, where N is the number of tasks to be processed, and M is the total number of feature dimensions (128+5=133). The feature vectors of all tasks are normalized and arranged in rows to form a unified input space, ensuring comparability and a unified modeling basis among heterogeneous tasks.Task requests are received using a lightweight communication protocol based on RESTful API, with task metadata encapsulated in JSON format and transmitted via HTTPS encryption. Nginx is deployed on the server side as a reverse proxy to achieve load balancing and request queue management. The maximum queue length is 200 tasks. If this is exceeded, a flow control mechanism is triggered, rejecting new requests and returning a 503 status code.

[0021] In the above method, step (3) establishes a resource-task association mapping model. Based on the resource state vector and task feature vector obtained in steps (1) and (2), a resource suitability scoring function is constructed using a dynamic weight allocation mechanism. This function includes a resource demand matching item, a load balancing penalty item, and a task urgency incentive item. The three coefficients are adjusted online based on historical scheduling feedback, and the comprehensive suitability score of each task under the current resource conditions is output. Specifically, the resource suitability scoring function is defined as: in, To determine the resource demand matching degree, the cosine similarity is used to calculate the closeness between the resource distribution required by the task and the current available resource distribution. The resource distribution required by the task is derived from the computational complexity level, expected latency, and input data volume in the task feature matrix, and mapped to a 5-dimensional vector: CPU demand intensity (low, medium, and high correspond to 0.3, 0.6, and 0.9 respectively), memory demand intensity (similarly), network bandwidth demand (Mbps), storage read / write intensity (IOPS), and expected heat load (temperature increment). The current available resource distribution is directly taken from the resource state vector generated in step (1), and normalized to the [0,1] interval. The cosine similarity calculation formula is: The load balancing penalty term is calculated based on the deviation of the current node's load standard deviation from the mean of its neighboring nodes. The system maintains a topology connectivity graph of edge nodes. Each node maintains a resource state summary of all nodes in its one-hop neighborhood. The summary contains a weighted average of CPU utilization, memory usage, and network load, with weights of 0.5, 0.3, and 0.2, respectively. The load vector of the current node is a combination of these three metrics. Its Euclidean distance from the neighborhood mean vector is calculated and then divided by the load standard deviation of the neighboring nodes to obtain the normalized deviation, which serves as the load balancing penalty term. The value of is in the range of [0,∞), and the larger the value, the more unbalanced the load. This is the task urgency incentive term, obtained by normalizing the product of the task priority label and the remaining tolerable latency. The 3-bit binary code of the priority label is converted to a decimal value (0~7). The remaining tolerable latency is the upper limit of the expected processing latency of the task minus the waiting time. The product of the two is mapped to the [0,1] interval through min-max normalization. Weighting coefficients. , , The initial values ​​are set to 0.4, 0.3, and 0.3, and are dynamically adjusted by the reinforcement learning agent module based on scheduling success rate and resource waste rate. The update cycle is one gradient update after every 100 scheduling decisions. The output S of the scoring function is a real number representing the overall suitability score of the task executed on the current node; a higher score indicates better suitability for local processing. This model is loaded as a dynamic library in the scheduling decision module, supporting hot updates of weight parameters without requiring a service restart.

[0022] In the above method, step (4) involves performing adaptive task scheduling and resource reallocation. Based on the adaptation score generated in step (3), all tasks to be processed are prioritized and reordered, triggering a resource reallocation strategy. If the overall score is lower than a preset threshold, a computing power migration process is initiated to unload the tasks to nearby edge nodes or the cloud. If resources are locally overloaded, a lightweight model replacement mechanism is activated, using a compressed neural network model to replace the original model to reduce the computational load. Specifically, the priority reordering adopts a non-dominated sorting strategy, using the adaptation score as the main sorting basis and combining it with the task arrival time for fairness constraints to prevent low-scoring tasks from being starved for a long time. The non-dominated sorting is based on the Pareto optimality principle, treating tasks as a solution set in a multi-objective optimization. The objective is to maximize the adaptation score and minimize the waiting time. First, the set of tasks with the highest adaptation score (the first non-dominated layer) is selected, and then the remaining tasks are selected again, and so on, forming multiple priority levels. Within the same level, tasks are arranged in ascending order of arrival time. The computing power migration process selects the optimal forwarding path based on the topological connectivity graph between edge nodes. The topological graph is represented as an undirected graph, where nodes are edge devices and edges are communication links. The edge weight is a weighted function of link latency and the remaining capacity of intermediate nodes. Latency is measured by the round-trip time of ICMP probe packets, and the remaining capacity is calculated by the resource digest broadcast by neighboring nodes. The weighting coefficients are 0.7 for latency and 0.3 for capacity. Path selection uses an improved Dijkstra algorithm, supporting multi-hop forwarding with a maximum hop limit of 3 to avoid excessive latency introduced by cross-domain transmission. Once the path is determined, the system sends a task migration request to the target node via the gRPC protocol, including task data, model parameters, and context information. After the target node verifies resource availability, it returns an ACK, and the source node begins data transmission, employing a chunked transmission and verification mechanism to ensure integrity. The lightweight model replacement mechanism incorporates three compressed models, corresponding to 70%, 50%, and 30% of the parameters of the original model, respectively. Compression methods include channel pruning, knowledge distillation, and quantization (INT8). Model versions are stored in a local model repository and indexed by task type. The replacement decision is triggered by whether the current CPU utilization exceeds 85% for three consecutive cycles. The detection logic is implemented in the resource monitoring agent module, which checks the utilization every 100 milliseconds. If it exceeds the threshold three times consecutively, an overload alarm is sent to the scheduling decision module. The model switching process is accompanied by a linked adjustment of the output confidence threshold. The confidence threshold for the original model is 0.5. When switching to a 70% model, it is increased to 0.55, to a 50% model to 0.6, and to a 30% model to 0.7, to compensate for accuracy loss and ensure that the sum of the false positive rate and false negative rate of the final output does not exceed a preset tolerance (e.g., 5%). The resource reallocation strategy also includes dynamic memory reclamation and cache cleanup. When the memory utilization exceeds 90%, the LRU (Least Recently Used) algorithm is triggered to clear the intermediate cache data of historical tasks.

[0023] In the above method, step (5) implements context-aware decision optimization, introduces an operational status assessment module, and combines environmental context information, including time series trends, task suddenness index and system stability indicators, to construct a multilayer perceptron prediction model to predict resource demand fluctuations and task arrival patterns within the next 200 milliseconds. Based on the prediction results, resource reservation strategies and cache preloading mechanisms are adjusted in advance to realize a shift in decision paradigm from passive response to active control. Specifically, the operational status assessment module performs a state sampling every 50 milliseconds. The time series trend is extracted by exponential smoothing to determine the direction of growth or decay. The smoothing coefficient α is set to 0.3. The resource utilization rate of the past 10 sampling points is weighted and averaged to calculate the trend slope. If the slope is greater than 0.05, it is determined to be an upward trend, and if it is less than -0.05, it is a downward trend. The task suddenness index is obtained by normalizing the standard deviation of the number of new tasks per unit time. The statistical window is 1 second. The sequence standard deviation of the number of new tasks every 50 milliseconds is calculated and then divided by the historical mean to obtain the relative volatility, which is normalized to the [0,1] interval. A value greater than 0.8 is determined to be a sudden state. The system stability index is calculated by weighting voltage fluctuation, temperature gradient, and task drop rate. Voltage fluctuation is taken from the ADC sampling data of the power management unit, and the peak-to-peak value within 100 milliseconds is calculated. The temperature gradient is the difference between the highest and lowest temperatures divided by the device size (unit: °C / cm). The task drop rate is the proportion of tasks rejected or timed out in the past second out of the total number of tasks. The weights of the three factors are 0.3, 0.4, and 0.3, respectively, and the sum constitutes the stability score. The lower the score, the more unstable the system. The multilayer perceptron prediction model includes an input layer, two hidden layers, and an output layer, with 64, 32, and 16 neurons, respectively. The input layer receives the aforementioned contextual features (trend direction, burst index, stability score, historical resource vector, and task arrival rate). The hidden layer uses the modified linear unit (ReLU) activation function, and the output layer uses linear activation. It predicts the mean and variance of resource requirements for the next 200 milliseconds, as well as the Poisson distribution parameter λ of task arrival. The model training data comes from historical operation logs, which are stored in Avro format and include timestamps, resource status, task features, scheduling actions, and execution results. These logs are updated incrementally daily and trained offline using the TensorFlow framework. The model version number is bound to the training timestamp, and canary releases are supported. Prediction results are used to dynamically adjust the weight coefficient estimates in step (3). For example, when resource demand is predicted to increase and task bursts are high, the weight coefficients are temporarily increased. (Urgency incentive item) reduced to 0.5, decrease Up to 0.3, prioritize high-priority tasks; at the same time, adjust the resource reservation ratio in step (4) to reserve CPU cores and memory page frames for high-priority tasks in advance. The reservation ratio is set according to the 95th percentile of the predicted load.

[0024] In the above method, step (6), closed-loop feedback and model self-learning, takes the results of each scheduling execution, including task completion delay, resource utilization change, energy consumption increment and user satisfaction score, as feedback signals and inputs them to the reinforcement learning agent module. This module uses the deep deterministic policy gradient algorithm to continuously optimize the parameter weights in the resource-task mapping model, forming a closed-loop adaptive optimization mechanism. Specifically, the reinforcement learning agent module runs in the edge coordinator controller. The state space is composed of resource state vector and task feature matrix, with a dimension of 5+133=138. The action space is defined as the fine-tuning direction of 3 weight coefficients. Each coefficient can perform three actions: "+Δ", "-Δ" or "hold". Δ is set to 0.05. Therefore, the action space has 27 discrete combinations, but it is modeled as a continuous action output by the deep deterministic policy gradient (DDPG) algorithm. The policy network outputs a 3-dimensional vector, representing , , The adjustment amount is limited to the range [-0.1, 0.1]. The reward function is defined as: in, The task success rate is the proportion of tasks that are completed on time and meet the accuracy requirements out of the total number of scheduled tasks. The energy efficiency ratio improvement value is defined as the task processing volume per unit energy consumption. The baseline value is taken from the historical average value under the initial configuration, and the improvement value is the current value minus the baseline value. This is the resource idle rate, a weighted average of the idle rates of the CPU, memory, and network, with weights of 0.4, 0.3, and 0.3 respectively. (Coefficient) , , The values ​​are 0.5, 0.3, and 0.2 respectively, ensuring that the reward function prioritizes task success, then improves energy efficiency, and finally reduces waste. The deep deterministic policy gradient algorithm uses an experience replay mechanism to store historical state-action-reward tuples. The replay buffer capacity is 10,000 entries, and a FIFO strategy is used to discard old data. The batch size is 32, used for updates during each training iteration. Both the Actor and Critic networks are fully connected neural networks. The Actor network structure is 138-64-32-3, and the Critic network is a concatenation of state and action inputs (138+3-64-32-1). The Adam optimizer is used, with a learning rate of 0.001. The target network is soft-updated every 1000 steps, with an update coefficient τ of 0.005, ensuring stable convergence during the learning process. During training, a policy evaluation is performed every 100 scheduling decisions. If the evaluated reward value increases by less than 1% for five consecutive times, model saving and deployment are triggered, and the new weights are written to the configuration center and loaded in real-time by the scheduling decision module.

[0025] Furthermore, this invention also includes a cross-edge domain collaborative decision-making mechanism. A master node is established in a cluster composed of multiple edge nodes. The master node periodically collects resource status summaries and task load reports from each member node to construct a global resource view. When any node triggers a resource crisis warning, the master node initiates a cross-domain task redirection protocol, selects the optimal receiving node based on the global view, and coordinates communication resource reservation to achieve cluster-level load balancing and fault tolerance. Specifically, the master node is deployed on the cluster's central server, using ZooKeeper for high-availability election to avoid single points of failure. Member nodes report a resource status summary to the master node every 200 milliseconds. The summary includes CPU utilization, memory usage, network load, available storage space and temperature, as well as the current queue length and average waiting time of pending tasks. Data is sent in batches via UDP protocol with a compression rate of 50%, and communication overhead is controlled within 5% of the total bandwidth. The master node maintains a global resource view table, storing the status of all nodes in a two-dimensional matrix, and calculates the reachability and transmission cost between nodes using a topology connectivity graph. When the CPU utilization of any node exceeds 95% for five consecutive cycles, or the memory usage exceeds 98%, or the temperature exceeds 85°C, a resource crisis warning is triggered, and the master node initiates a cross-domain task redirection protocol. The protocol first filters candidate nodes with a load rate below 70% and good connectivity to the source node, sorting them in ascending order of transmission latency, and selecting the top three as candidates. Then, it calculates the expected load increment after each candidate node takes over the task, selecting the node with the smallest increment that meets resource constraints as the optimal node to take over. The master node sends a resource reservation request to this node, reserving CPU cores and memory resources for 1.5 times the expected task execution time. Simultaneously, it returns a redirection instruction to the source node, which migrates the task context and data to the target node and releases local resources after the migration is complete. The entire synchronization latency is less than 10 milliseconds, ensuring seamless task switching.

[0026] Furthermore, this invention also includes a dynamic reconstruction function for a visual processing pipeline. Multiple processing flow templates are preset for different task types, including single-stage detection, two-stage recognition, and multimodal fusion flows. The scheduling decision module automatically matches the optimal flow based on task characteristics and dynamically enables or disables specific processing stages based on resource feedback during runtime. For example, it disables the high-resolution reconstruction module when computing power is limited and enables the data compression encoding module when bandwidth is restricted, ensuring real-time alignment between the processing flow and resource conditions. Specifically, the single-stage detection flow is suitable for target detection tasks, including image preprocessing, YOLO model inference, and post-processing (NMS); the two-stage recognition flow is suitable for behavior recognition, including optical flow computation, two-stream network inference, and temporal fusion; the multimodal fusion flow is suitable for complex scene understanding, fusing visible light, infrared, and radar data, including modal alignment, feature-level fusion, and decision-level fusion. The flow template is defined in the form of a DAG (Directed Acyclic Graph), where each node is a processing operator and edges represent data flow dependencies. During task allocation, the scheduling decision module matches the most suitable template based on the task type and feature vector and instantiates it as an execution plan. During runtime, the resource monitoring module continuously reports the current resource status. If the CPU utilization exceeds 80%, the high-resolution reconstruction module (such as SRGAN) is automatically disabled, and bilinear interpolation is used instead. If the network bandwidth is below the threshold, the data compression encoding module (such as H.265) is enabled, with a compression ratio set to 20:1. If memory is scarce, the intermediate feature maps are quantized from FP32 to FP16 for storage. All reconstruction operations are implemented through a dynamic graph scheduling engine, supporting operator replacement and branch jumps, with a reconstruction decision latency of less than 5 milliseconds.

[0027] Furthermore, this invention includes a security isolation mechanism. Virtualized containers or lightweight sandboxes are established within the task execution environment, with each task running in an independent security domain. This prevents malicious code or abnormal behavior from affecting other tasks or the system kernel. The resource monitoring module simultaneously monitors the resource usage of each container. Upon detecting out-of-bounds behavior, an isolation policy is immediately triggered, and a security audit log is reported, ensuring the overall security and reliability of the edge system. Specifically, security isolation employs Kata Containers technology. Each task runs in an independent lightweight virtual machine with its own kernel and root file system. Memory and CPU resources are hard isolated via cgroups, the network is isolated via a virtual bridge, and storage is mounted via a tmpfs temporary file system, prohibiting access to host devices. The resource monitoring module periodically queries the resource usage of each container through the libvirt interface. Monitoring metrics include CPU utilization, memory RSS, disk I / O, and network throughput. Thresholds are set as follows: CPU usage exceeding 120% of its allocated quota for 100 consecutive milliseconds, memory usage exceeding 110%, or abnormal network traffic surges (exceeding three times the standard deviation of the mean). Once an out-of-bounds behavior is detected, an isolation strategy is immediately implemented: first, the container process is paused; then, a memory snapshot is saved for evidence collection; finally, the container is terminated and resources are released. At the same time, an alarm log is sent to the security audit module. The log contains the container ID, out-of-bounds type, timestamp, and context information. It is forwarded to the central log server via the Syslog protocol and retained for no less than 180 days.

[0028] To verify the technical effectiveness of this invention, a specific application example was constructed: 10 edge computing nodes were deployed in a smart manufacturing park. Each node was equipped with an 8-core ARM processor, 16GB of memory, a 10Gbps network interface, and a 2TB SSD, running a Linux 5.10 kernel. Front-end sensing devices included 50 industrial cameras (1080p@30fps), 10 drones (4K@25fps), and 20 mobile inspection terminals. The system ran continuously for 30 days, processing 1.2 million visual tasks with an average task arrival rate of 400 tasks / minute. Experimental data showed that the fluctuation range of resource utilization decreased from 40% in traditional static scheduling to 12.7%, the overall task scheduling satisfaction increased by 43.6%, the response speed to emergencies increased by 2.1 times, the average processing latency decreased from 380 milliseconds to 228 milliseconds, and after 1000 scheduling iterations, the resource waste rate decreased from the initial 35% to 17.5%, and the energy efficiency ratio improved by 36.2%. In high-concurrency scenarios (peak 1000 times / minute), the system maintains a 95.4% task completion rate, meeting the stringent service quality requirements of advanced process control.

[0029] 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 illustrative of the principles of 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 resource- and task-aware visual processing edge adaptive decision-making method, comprising the following steps: Real-time collection of multi-dimensional resource status data of edge nodes; the resource monitoring agent module deployed in the edge computing device periodically obtains the status data of the current node to form a resource status vector. The system acquires and parses visual task requests to be processed, extracts task type, input data volume, expected processing latency limit, computational complexity level, and task priority label, constructs a task feature matrix, and transforms unstructured task descriptions into quantized feature vectors through a semantic encoder. A resource-task association mapping model is established. Based on the resource state vector and task feature vector, a resource adaptability scoring function is constructed using a dynamic weight allocation mechanism. The resource adaptability scoring function is adjusted online according to scheduling execution feedback and outputs the comprehensive adaptability score of each task under the current resource conditions. Perform adaptive task scheduling and resource reallocation, and reorder the priority of all pending tasks based on the comprehensive adaptation score to trigger the resource reallocation strategy. To implement context-aware decision optimization, an operational status assessment module is introduced, and a multi-layer perceptron prediction model is constructed by combining time series trends, task contingency index, and system stability indicators. Closed-loop feedback and model self-learning are used to input the results of each scheduling execution into the reinforcement learning agent module. The reinforcement learning agent module uses a deep deterministic policy gradient algorithm to continuously optimize the parameter weights in the resource-task mapping model, forming a closed-loop adaptive optimization mechanism.

2. According to the method described in claim 1, the resource monitoring agent module is integrated into the kernel layer of the edge operating system, and directly reads the hardware register status through the system call interface to avoid the latency caused by user-mode polling. The measurement of CPU utilization and memory occupancy is based on the acquisition of the status registers of the underlying performance counter and memory management unit. The network bandwidth monitoring adopts traffic mirroring and packet capture technology combined with the sliding window algorithm to calculate the instantaneous throughput. The storage space monitoring distinguishes between hot data areas and cold data areas and records the read / write frequency and fragmentation degree respectively. The temperature sensor is deployed in the hot area of ​​the chip. The sampled data is filtered by Kalman to eliminate noise interference. The final resource status vector is written to the shared memory buffer for subsequent module calls.

3. The method according to claim 1, wherein the visual task request is initiated by an industrial camera, drone or mobile inspection terminal, the task type includes target detection, behavior recognition, anomaly alarm and image enhancement, the input data volume is derived from the video resolution and frame rate, the upper limit of the expected processing latency is defined by the application layer, the computational complexity level is divided according to the number of floating-point operations of the called model, the task priority label uses multi-bit binary encoding to represent safety criticality, business importance and time sensitivity, the semantic encoder uses a pre-trained bidirectional long short-term memory network to embed the task description text and output a fixed-dimensional feature vector, which together with the structured parameters constitute the task feature matrix.

4. According to the method described in claim 1, the resource demand matching term of the resource suitability scoring function calculates the closeness between the resource distribution required by the task and the current available resource distribution through cosine similarity calculation. The resource distribution required by the task is mapped to a multi-dimensional vector by the computational complexity level, expected latency, and input data magnitude in the task features. The current available resource distribution is taken from the resource state vector and normalized to a unified interval. The load balancing penalty term is calculated based on the deviation of the current node's load standard deviation from the mean of neighboring nodes. The task urgency incentive term is obtained by normalizing the product of the task priority label and the remaining tolerable latency. The weight coefficient is dynamically adjusted by the reinforcement learning agent module according to the scheduling success rate and resource waste rate.

5. According to the method described in claim 1, the priority reordering adopts a non-dominated sorting strategy, uses the adaptation score as the sorting basis, and combines the task arrival time to impose fairness constraints to prevent low-scoring tasks from being starved for a long time. The computing power migration process selects the optimal forwarding path based on the topological connectivity graph between edge nodes, taking into account link latency and the remaining capacity of intermediate nodes. The lightweight model replacement mechanism has multiple built-in compression models, each corresponding to different parameter ratios of the original model. The replacement decision is triggered by the number of cycles in which the central processing unit utilization exceeds the threshold. The model switching process is accompanied by a linkage adjustment of the output confidence threshold to compensate for accuracy loss.

6. According to the method described in claim 1, the operational status assessment module performs status sampling at a fixed frequency, the time series trend is extracted by exponential smoothing to determine the direction of growth or decay, the task burst index is obtained by normalizing the standard deviation of the number of new tasks per unit time, the system stability index is calculated by weighting voltage fluctuation, temperature gradient and task drop rate, the multilayer perceptron prediction model includes an input layer, a hidden layer and an output layer, the activation function is a modified linear unit, the training data comes from historical operation logs, and the prediction results are used to dynamically adjust the weight coefficients in the resource suitability scoring function and the resource reservation ratio in task scheduling.

7. According to the method described in claim 1, the reinforcement learning agent module runs in the edge coordinator controller, the state space is composed of the resource state vector and the task feature matrix, the action space is defined as the fine-tuning direction of multiple weight coefficients, the reward function is composed of the task success rate, the energy efficiency ratio improvement value and the resource idle rate weighted, the deep deterministic policy gradient algorithm uses an experience replay mechanism to store historical state-action-reward tuples, batch updates the policy network and the value network, and the target network periodically synchronizes parameters in a soft update manner to ensure stable convergence of the learning process.

8. The method according to claim 1 further includes a cross-edge domain collaborative decision-making mechanism, wherein a master control node is set up in a cluster composed of multiple edge nodes, the master control node periodically collects resource status summaries and task load reports of each member node, constructs a global resource view, and when any node triggers a resource crisis warning, the master control node initiates a cross-domain task redirection protocol, selects the optimal receiving node according to the global view, and coordinates the reservation of communication resources to achieve cluster-level load balancing and fault tolerance, with communication overhead controlled within a limited proportion of the total bandwidth and synchronization delay less than a preset threshold.

9. The method according to claim 1 further includes a dynamic reconstruction function for the visual processing pipeline, which presets multiple processing flow templates for different task types, including a single-stage detection flow, a two-stage recognition flow, and a multi-modal fusion flow. The scheduling decision module automatically matches the optimal flow according to the task characteristics and dynamically enables or disables specific processing stages based on resource feedback during runtime. When computing power is tight, the high-resolution reconstruction module is turned off, and when bandwidth is limited, the data compression encoding module is enabled to ensure that the processing flow and resource conditions are aligned in real time.

10. The method according to claim 1 further includes a security isolation mechanism, which establishes virtualized containers or lightweight sandboxes in the task execution environment, with each task running in an independent security domain. The resource monitoring module simultaneously monitors the resource usage of each container, and immediately triggers an isolation policy and reports the security audit log once an out-of-bounds behavior is detected. Memory and CPU resources are hard isolated through control groups, the network is isolated through virtual bridges, and storage is mounted through a temporary file system, prohibiting access to host devices.

Images