Internet of Things equipment resource allocation method and system based on OpenHarmony

Abstract

The invention relates to the technical field of Internet of Things equipment, and discloses an OpenHarmony-based Internet of Things equipment resource allocation method and system, and the method comprises the steps: obtaining an equipment resource feature vector based on an OpenHarmony distributed network; calculating a resource dependency relationship between equipment nodes to obtain a resource sharing weight matrix; dividing the resource sharing weight matrix into a plurality of resource views, and performing feature extraction on the plurality of resource views to obtain an equipment resource state embedding matrix; performing resource allocation analysis of the dual-depth Q network on the device resource state embedding matrix to generate a global resource allocation strategy; generating an equipment dynamic scheduling sequence based on the equipment working state in each equipment cluster; according to the resource allocation method and system, the intra-region resource allocation operation and the cross-region resource request operation are constructed into a four-stage distributed transaction, the distributed resource allocation result is output, the overall stability of the system is improved, and the resource allocation problem of large-scale heterogeneous Internet of Things equipment is effectively solved.

Description

Technical Field

[0001] The present invention relates to the technical field of Internet of Things devices, and in particular to an OpenHarmony-based Internet of Things device resource allocation method and system. Background Art

[0002] With the widespread application of the OpenHarmony operating system in the field of the Internet of Things, the resource allocation problem of massive IoT devices has become increasingly prominent. When facing large-scale heterogeneous IoT devices, traditional centralized resource allocation methods have problems such as high communication overhead, high response delay, and high energy consumption, which makes it difficult to meet the real-time and reliability requirements of IoT applications.

[0003] Although the existing distributed resource allocation methods have alleviated the pressure of centralized allocation to a certain extent, they are still insufficient in dealing with complex resource dependencies between devices and dynamically changing resource requirements. Especially in the resource-constrained IoT environment, how to accurately capture the resource sharing mode between devices and make efficient resource allocation decisions based on real-time status has become a key issue that needs to be solved urgently. Summary of the invention

[0004] The present invention provides an Internet of Things device resource allocation method and system based on OpenHarmony, which improves the overall stability of the system and effectively solves the resource allocation problem of large-scale heterogeneous Internet of Things devices.

[0005] In a first aspect, the present invention provides an Internet of Things device resource allocation method based on OpenHarmony, and the Internet of Things device resource allocation method based on OpenHarmony comprises: Based on the OpenHarmony distributed network, the resource description matrix and resource interaction time series feature sequence are obtained, and the device resource feature vector is obtained through feature fusion calculation; Based on the device resource feature vector, the resource dependency relationship between device nodes is calculated to obtain a resource sharing weight matrix; Dividing the resource sharing weight matrix into multiple resource views according to resource types, and performing feature extraction on the multiple resource views to obtain a device resource state embedding matrix; Performing a resource allocation analysis of a dual-depth Q network on the device resource state embedding matrix to generate a global resource allocation strategy; Dividing the IoT devices into a plurality of device clusters according to the global resource allocation strategy, and generating a device dynamic scheduling sequence based on the working status of the devices in each of the device clusters; Based on the device dynamic scheduling sequence, the intra-region resource allocation operation and the cross-region resource request operation are constructed into a four-stage distributed transaction, and the distributed resource allocation result is output.

[0006] In a second aspect, the present invention provides an Internet of Things device resource allocation system based on OpenHarmony, and the Internet of Things device resource allocation system based on OpenHarmony includes: The acquisition module is used to obtain the resource description matrix and resource interaction time series feature sequence based on the OpenHarmony distributed network, and obtain the device resource feature vector through feature fusion calculation; A calculation module, used to calculate the resource dependency relationship between device nodes based on the device resource feature vector to obtain a resource sharing weight matrix; A partitioning module, used to partition the resource sharing weight matrix into multiple resource views according to resource types, and perform feature extraction on the multiple resource views to obtain a device resource status embedding matrix; An analysis module, configured to perform a resource allocation analysis of a dual-depth Q network on the device resource state embedding matrix to generate a global resource allocation strategy; A generation module, used to divide the IoT devices into a plurality of device clusters according to the global resource allocation strategy, and generate a device dynamic scheduling sequence based on the working status of the devices in each of the device clusters; The output module is used to construct the intra-region resource allocation operation and the cross-region resource request operation into a four-stage distributed transaction based on the device dynamic scheduling sequence, and output the distributed resource allocation result.

[0007] A third aspect of the present invention provides a computer device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the computer device executes the above-mentioned OpenHarmony-based Internet of Things device resource allocation method.

[0008] A fourth aspect of the present invention provides a computer-readable storage medium, which stores instructions. When the computer-readable storage medium is run on a computer, the computer executes the above-mentioned OpenHarmony-based Internet of Things device resource allocation method.

[0009] In the technical solution provided by the present invention, by constructing a multi-perspective spatiotemporal graph structure and a hybrid attention network, the device resource status and interaction characteristics are effectively extracted, the accuracy of resource status representation is improved, and a reliable data basis is provided for resource allocation decisions. A dynamic gated graph attention network is used to calculate the resource dependencies between devices, and the resource allocation strategy is optimized in combination with a dual deep Q learning algorithm, thereby improving the resource utilization and energy efficiency of the system. An intelligent sleep and wake-up mechanism based on a genetic algorithm is designed, which reduces the overall energy consumption of the system by optimizing the combination of device working states while ensuring the quality of service. A four-stage distributed transaction mechanism is proposed, which combines a two-stage locking protocol and a Paxos consensus algorithm to reduce the communication delay of the system and significantly improve the reliability of distributed resource allocation. Through the collaborative optimization of a multi-level resource allocation architecture, the overall stability of the system is improved, effectively solving the resource allocation problem of large-scale heterogeneous IoT devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0011] Figure 1 Schematic diagram of the steps of the method for allocating resources of IoT devices based on OpenHarmony in an embodiment of the present invention; Figure 2 It is a structural diagram of an Internet of Things device resource allocation system based on OpenHarmony in an embodiment of the present invention; Figure 3 It is a schematic block diagram of the structure of a computer device in an embodiment of the present invention. DETAILED DESCRIPTION

[0012] The embodiment of the present invention provides a method and system for allocating resources of IoT devices based on OpenHarmony. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0013] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 In an embodiment of the present invention, an embodiment of the method for allocating resources of IoT devices based on OpenHarmony includes: Step S1, based on the OpenHarmony distributed network, obtain the resource description matrix and the resource interaction time series feature sequence, and obtain the device resource feature vector through feature fusion calculation; It is understandable that the execution subject of the present invention may be an IoT device resource allocation system based on OpenHarmony, or a terminal or a server, which is not limited here. The embodiment of the present invention is described by taking a server as the execution subject as an example.

[0014] Specifically, the OpenHarmony distributed network is divided into multiple resource monitoring areas, and regional resource coordination nodes are configured in each resource monitoring area to solve the coordination problem of resource allocation between various device nodes in the distributed network. The regional coordination nodes effectively collect and coordinate the device resource information in each area to form a distributed resource monitoring network topology. Data collection is performed on the IoT device nodes in the distributed resource monitoring network topology to obtain the original resource status data set. The regional resource coordination node collects information from each device node and monitors the resource usage of the device, including data in multiple dimensions such as computing resources, storage, bandwidth, and energy consumption. The original resource status data set is normalized and mapped, and the data is converted into a standard resource description matrix through matrix conversion technology. At the same time, considering the resource interaction behavior between devices, the resource sharing historical data of the device nodes is sampled in time series. By setting the time window, the device identification pairs, shared resource types, and resource sharing amounts within t time windows are extracted to obtain the resource interaction time series feature sequence, which describes how devices share resources with each other in different time periods, thereby reflecting the collaboration mode and resource demand trend between devices. The resource description matrix is ​​processed by the feature encoder, and the feature encoder performs nonlinear transformation on the resource state features through a fully connected neural network to obtain the resource state feature subvector, capturing the fine-grained information of the resource state of each device. At the same time, for the resource interaction time series feature sequence, a sliding operation of the time window is performed to calculate the resource sharing frequency matrix and time interval matrix between the device pairs. These matrices reflect the sharing frequency and time interval between devices. Through this information, the dependency and interaction mode between devices can be inferred. Through the graph attention network, the sharing frequency matrix and time interval matrix are processed to generate the resource interaction feature subvector. The resource state feature subvector and the resource interaction feature subvector are input into the feature fusion network for fusion. The feature fusion network calculates the weights between different features through the multi-head attention mechanism to fully explore the correlation between resource state and resource interaction. The multi-head attention mechanism can focus on the features of multiple subspaces at the same time, thereby improving the model's expression ability and ability to process complex data. The fused weighted feature vector contains multi-dimensional information of resource state and resource interaction. This weighted feature vector is optimized through residual connection and layer normalization operations to obtain the device resource feature vector.

[0015] Step S2: Calculate the resource dependency relationship between device nodes based on the device resource feature vector to obtain a resource sharing weight matrix; Specifically, a hierarchical resource allocation architecture is constructed based on the device resource feature vector. The architecture includes three levels: device layer, coordination layer, and management layer. In this architecture, the nodes of each layer represent resource allocation entities at different levels. The device layer contains physical devices, the coordination layer is responsible for resource coordination in local areas, and the management layer is responsible for global resource optimization and decision-making. By hierarchically encoding the resource status of each layer of nodes, a hierarchical resource status matrix is ​​obtained. The hierarchical resource status matrix is ​​processed by the gating unit in the dynamic gated graph attention network. The dynamic gated graph attention network can dynamically adjust the influence between each node according to the resource status and hierarchical structure of the node. In this process, the gating unit calculates a gating coefficient based on the characteristics of the current node and the characteristics of the neighboring nodes. The coefficient determines the intensity of the interaction between the nodes. Through this gating mechanism, a dynamic gated feature matrix is ​​obtained, which reflects the resource dependency relationship between nodes and can dynamically adjust the weight of their interaction according to the resource status of the nodes. The dynamic gated feature matrix is ​​input into the LeakyReLU activation function for nonlinear transformation. The LeakyReLU activation function can effectively solve the problem of ReLU function's gradient vanishing in the negative range, while retaining part of the information in the negative range, so as to better reflect the complex resource interaction pattern. The feature matrix after nonlinear transformation is calculated by the dot product operation of the node feature vector to calculate the attention score, which reflects the importance of resource sharing between nodes. Through the dot product operation, the original attention weight between each pair of nodes is obtained. The original attention weight is normalized by softmax, and the original attention weight is converted into a probability distribution, so that the relative importance between nodes can be quantified. The normalized weight is weightedly aggregated with the node features to obtain the updated node representation vector. The updated node representation vector integrates the resource characteristics of the node itself and the interaction information with the neighboring nodes. The resource consumption ratio and resource sharing frequency between device nodes are calculated based on the node representation vector. The resource consumption ratio reflects the resource demand of each device node in the resource allocation process, while the resource sharing frequency reflects the frequency of resource sharing between device nodes. Through these two indicators, a resource dependency metric between node pairs is constructed to obtain a dependency metric matrix. The dependency metric matrix is ​​input into the multi-head self-attention module, which calculates the attention weights of different heads through the query-key-value mechanism. The multi-head attention mechanism can focus on multiple different feature spaces at the same time, thereby enhancing the model's ability to model complex relationships. Each attention head will independently calculate an attention weight, and then the attention weights of multiple heads are spliced ​​and linearly transformed to obtain the final resource sharing weight matrix, which represents the degree of dependence between device nodes in resource sharing.

[0016] Step S3: Divide the resource sharing weight matrix into multiple resource views according to resource types, and perform feature extraction on the multiple resource views to obtain a device resource status embedding matrix; Specifically, the resource sharing weight matrix is ​​decomposed by resource type. The resource sharing weight matrix reflects the resource dependency between different devices, and different types of resources (such as CPU occupancy, memory usage, storage space occupancy, network bandwidth usage, and remaining power of the device) have different effects on the status and interdependence of the devices. The resource sharing weight matrix is ​​decomposed according to these different resource types to obtain multiple independent resource views, which respectively represent the status and interaction of the device under different resource types, and capture the role of various resources in device allocation. Each resource view is input into the graph convolution network for processing. The graph convolution network can effectively aggregate the neighborhood information between nodes. The graph convolution network performs convolution operations on each node, aggregates the information of its neighboring nodes, and obtains a node-level feature matrix, which effectively represents the status of each device node under a specific resource type and the interdependence with other device nodes. In order to improve the feature extraction capability, the node-level feature matrix is ​​extracted at multiple scales using dilated convolution. The dilated convolution captures a wider range of features by expanding the receptive field of the convolution kernel, thereby modeling the device status at different scales. The size of the convolution kernel is set to d, the hole rate is set to r, and the output feature dimension is set to f to obtain the time series feature sequence. These time series feature sequences reflect the changes in device status and can capture the resource allocation pattern and time-varying characteristics between devices. The time series feature sequence is input into the channel attention module for processing. The channel attention module dynamically adjusts the feature weights according to the importance of different channels, thereby highlighting the features that are more important for resource allocation. The time series feature sequence is processed by global average pooling and maximum pooling to obtain the channel attention weights. Global average pooling can extract global information from the entire feature map, while maximum pooling helps capture the most significant local features. Through these two pooling operations, more stable and representative channel weights are obtained. The channel attention weights are element-wise multiplied with the time series feature sequence to obtain the channel weighted time series features. Based on the channel weighted time series features, the time series correlation in the graph structure is calculated. The temporal dependency between device nodes is modeled to capture the trend of device resource status changes over time. The causal convolution structure is used to establish long-range dependencies. Causal convolution avoids the influence of future information on the current state by only using the current and past time steps in the time series, which conforms to the causal relationship of time series data. Through causal convolution, long-term resource sharing dependencies between modeled devices can be effectively captured when resource demand fluctuates greatly or when device states suddenly change. The time series dependency features are concatenated with the node-level feature matrix to form a device resource state embedding matrix.

[0017] Step S4: performing a resource allocation analysis of a dual-depth Q network on the device resource state embedding matrix to generate a global resource allocation strategy; Specifically, a resource state space and a resource allocation action space are constructed for each dimension of the device resource state embedding matrix. The resource state space reflects the various resource usage of the device, such as CPU, memory, storage space, network bandwidth, etc., while the resource allocation action space includes possible actions for resource allocation, such as resource sharing or task scheduling between devices. By encoding these states and actions, a state-action mapping matrix is ​​obtained to describe the relationship between different device resource states and corresponding actions. An online evaluation network and a target evaluation network are constructed for the state-action mapping matrix. The online evaluation network is used to evaluate the effect of the current strategy in real time, while the target evaluation network is used to evaluate the effect of the target strategy. The two networks continuously update each other to improve the accuracy of the allocation strategy. The core task of these two networks is to evaluate the value of different resource allocation strategies by extracting action value features, thereby providing a basis for subsequent decision-making. By training the dual-Q network evaluation model, the resource allocation strategy is continuously optimized to make it more accurately reflect the needs of device resource scheduling. Based on the dual-Q network evaluation model, the comprehensive reward score is calculated, and the optimization of the strategy is guided by the reward signal. The reward function can reflect the quality of the current strategy and promote the system to evolve towards the optimal strategy. The comprehensive reward score is normalized and input into the temporal difference learning unit. The discount factor γ is set to 0.95. The discount factor is used to balance the relationship between immediate rewards and long-term rewards in reinforcement learning. The learning target sequence is obtained by calculating the target Q value. The Huber loss is calculated for the learning target sequence and the output value of the online evaluation network. The Huber loss function is a loss function used in reinforcement learning. It can effectively deal with the fluctuation problem of the reward value and avoid gradient explosion or gradient disappearance through smoothing. After calculating the loss, the Adam optimizer is used to update the network parameters. The Adam optimizer is an adaptive learning rate optimization algorithm that can automatically adjust the learning rate according to the gradient changes of different parameters, thereby accelerating the convergence process. After multiple iterations and optimizations, the optimized network parameters are obtained. A cyclic experience replay buffer is constructed based on the optimized network parameters, and the state transition quadruple (including current state, action, reward and next state) is stored in the buffer. The experience replay buffer is a commonly used technology in reinforcement learning. By storing historical experience, it can break the correlation between data and improve the stability of learning. The state transition quadruple saved in the buffer provides sufficient data support for subsequent training. These quadruples are randomly extracted from the buffer and training data batches are generated so that the optimal policy can be learned in a more efficient way during each training process. During the training process, policy optimization iterations are performed on the training data batches. During this process, the parameters of the online evaluation network are updated to the target evaluation network through soft updates. The soft update operation allows the target network to make only small adjustments during each training, thereby avoiding training instability caused by large updates. Through continuous iterative optimization, the optimized dual-Q network model is obtained.The optimized double Q network model is used for resource allocation decision-making. In the process of strategy generation, the initial value of the exploration rate ε is set to 0.9, and the exploration rate is gradually updated according to the linear decay strategy. The exploration rate controls the balance between the agent exploring new strategies and utilizing existing strategies. A higher exploration rate helps to explore more possible resource allocation methods, while a lower exploration rate helps to utilize the learned strategies. Through the balance mechanism of exploration and utilization, the optimal strategy in the resource allocation of IoT devices is effectively found, and the global resource allocation strategy is generated.

[0018] Step S5: divide the IoT devices into multiple device clusters according to the global resource allocation strategy, and generate a device dynamic scheduling sequence based on the working status of the devices in each device cluster; Specifically, the global resource allocation strategy is analyzed for device characteristics. The similarity matrix between devices is calculated by resource utilization patterns and functional similarity to obtain a device association graph. The resource utilization pattern reveals the similarity of resource requirements between devices by analyzing the use of devices on different resource types; while functional similarity considers whether the tasks and functions performed by the devices are similar. In this way, the association between devices is represented in the form of a graph. Each node in the graph represents a device, and the edge weights between nodes represent the similarity between devices. The device association graph is input into the hierarchical clustering algorithm to cluster the devices. The hierarchical clustering algorithm is a clustering method based on the similarity between devices. Through a bottom-up iterative merging process, devices with high similarity are grouped into the same group. The clustering threshold is set to 0.75 to effectively control the size of the device cluster and the accuracy of clustering. When the similarity between devices is higher than the threshold, they will be merged into the same cluster to form multiple preliminary device clusters. The device status in each device cluster is binary encoded. The active state is encoded as 1 and the dormant state is encoded as 0. Through binary encoding, the device working state matrix is ​​obtained. Each row of the matrix represents a device and each column represents a certain state of the device. The device working state matrix is ​​input into the genetic algorithm encoder for chromosome construction. As an optimization method that simulates natural selection, the genetic algorithm continuously searches for the optimal solution through operations such as selection, crossover and mutation. In the chromosome construction stage, each row of the device state matrix represents a chromosome, and each gene in the chromosome represents the state of a device. Through genetic operations, the crossover probability is set to 0.8, and multiple device state combinations are exchanged to generate new device state combinations. Each device state combination is evaluated based on the fitness function of service quality and energy consumption. The evaluation criteria are based on the working efficiency of the device, resource consumption and the overall energy efficiency of the system. According to the evaluation results, individuals with high fitness are selected as the next generation population to obtain the optimized device state combination. The optimized device state combination is grouped according to the resource consumption level. By grouping each group of devices, the minimum active node number threshold in each group is calculated to ensure that each device cluster can meet the minimum active node requirement when performing tasks. This process helps determine the working mode sequence of the device cluster, ensuring that the devices in the cluster can perform tasks at a reasonable time sequence while maximizing the utilization of device resources. Based on the cluster working mode sequence, a dynamic scheduling schedule is constructed. The construction of the scheduling schedule takes into account the active time and sleep time of the equipment, and uses a sliding time window to set the activity cycle of the equipment. The time window mechanism helps to dynamically adjust the working status of the equipment so that it can be reasonably scheduled according to task requirements in different time periods to obtain an initial scheduling plan. The initial scheduling plan is input into the scheduling optimizer, and the scheduling timing is optimized through a dynamic programming algorithm.The dynamic programming algorithm can effectively find the global optimal scheduling strategy by decomposing the problem and solving the optimal substructure, thereby generating a dynamic scheduling sequence for the equipment.

[0019] Step S6: Based on the device dynamic scheduling sequence, the intra-region resource allocation operation and the cross-region resource request operation are constructed into a four-stage distributed transaction, and the distributed resource allocation result is output.

[0020] Specifically, the device dynamic scheduling sequence is divided into two types of resource allocation requests: intra-region resource allocation requests and cross-region resource allocation requests. Intra-region resource allocation requests involve the direct allocation of resources between devices, while cross-region resource allocation requests handle device resource requests that span multiple regions. To ensure efficient and orderly resource allocation, a resource scheduling priority sorting algorithm is used to queue these requests, resulting in a hierarchical resource request queue. During the sorting process, the priority is set based on factors such as the urgency of the device, the scarcity of resources, and the timeliness of the request, ensuring that high-priority requests can be processed first during actual execution. For each resource allocation operation in the hierarchical resource request queue, a transaction control block is constructed, which contains information such as a transaction identifier, resource type, requester identifier, and responder identifier. The transaction identifier is used to uniquely identify each resource allocation operation, the resource type indicates the type of resources involved in the operation, and the requester and responder identifiers correspond to the resource requester and resource provider, respectively. This information helps to clarify the operation objects and participants of each transaction, providing precise control for subsequent transaction processing and resource scheduling. By constructing these transaction control blocks, a transaction description set is obtained. The transaction description set is input into a four-phase commit protocol processor for processing. The four-phase commit protocol is a method for ensuring transaction consistency in a distributed system, and its process includes four stages: request initiation, resource reservation, execution commit, and completion confirmation. In the request initiation stage, the requester initiates a resource allocation request and waits for a response; in the resource reservation stage, the system allocates corresponding resources to the request according to the resource scheduling situation; in the execution commit stage, the resources are actually allocated and the task is executed; in the completion confirmation stage, the system confirms that the resource allocation has been completed and notifies the relevant parties. In accordance with the order of these four stages, the protocol processor ensures that the transaction can progress step by step according to the predetermined process and finally completes the resource allocation operation. To ensure the orderliness of transaction processing, a globally unique identifier is assigned to each transaction node in the distributed transaction. This identifier is used to distinguish different transaction nodes and ensure the uniqueness and integrity of transactions in the system. Through the vector clock algorithm, the temporal dependency relationship of transactions is maintained. The vector clock helps the system determine the sequence and dependency between transactions by recording the clock states of each transaction node, thus ensuring that transactions are processed in the correct order. With the help of the vector clock, an ordered transaction execution sequence is obtained. On the basis of ensuring the transaction order, a resource lock manager is constructed to effectively control concurrent access to competing resources. The role of the resource lock manager is to coordinate the access of multiple transactions to the same resource, avoiding conflicts or data inconsistencies caused by resource contention. For this purpose, a two-phase locking protocol is used to implement concurrent control. In the two-phase locking protocol, a transaction requests a resource lock at the beginning and will not release the resource lock until the transaction is processed and committed. In this way, it is ensured that only one transaction can access a certain resource at the same time, avoiding concurrent problems that occur when competing for resources.Through the two-phase locking protocol, the resource locking state table is obtained. Based on the resource locking state table, the resource allocation operation is performed in sequence according to the transaction priority order. In order to ensure the accuracy and traceability of resource allocation, the execution status of each operation is recorded through the resource scheduling log. After the resource allocation operation is completed, the consistency of the resource allocation result set is verified to ensure that the resource allocation status between different regions is consistent and the resource status of each node reaches a consensus in the distributed system. The Paxos consensus algorithm is adopted. The Paxos algorithm is a distributed consistency algorithm that can ensure consistency through the voting mechanism of the majority of nodes in an unreliable network environment. Through the Paxos algorithm, the resource allocation status is synchronized between regions to ensure the consistency of global resource allocation. After the global consistency state is reached, the state is input into the resource status updater, and the device resource status is updated through the atomic commit operation. The atomic commit operation ensures the atomicity of the update process, that is, all update operations are either all successful or all failed, thereby ensuring data consistency and system reliability. Through the above steps, the distributed resource allocation result is generated.

[0021] In the embodiment of the present invention, by constructing a multi-perspective spatiotemporal graph structure and a hybrid attention network, the device resource status and interaction characteristics are effectively extracted, the accuracy of resource status representation is improved, and a reliable data basis is provided for resource allocation decisions. A dynamic gated graph attention network is used to calculate the resource dependencies between devices, and the resource allocation strategy is optimized in combination with a dual deep Q learning algorithm, thereby improving the resource utilization and energy efficiency of the system. An intelligent sleep and wake-up mechanism based on a genetic algorithm is designed, which reduces the overall energy consumption of the system by optimizing the combination of device working states while ensuring the quality of service. A four-stage distributed transaction mechanism is proposed, which combines a two-stage locking protocol and a Paxos consensus algorithm to reduce the communication delay of the system and significantly improve the reliability of distributed resource allocation. Through the collaborative optimization of a multi-level resource allocation architecture, the overall stability of the system is improved, effectively solving the resource allocation problem of large-scale heterogeneous IoT devices.

[0022] In a specific embodiment, the process of executing step S1 may specifically include the following steps: Divide the OpenHarmony distributed network into multiple resource monitoring areas, configure regional resource coordination nodes for each resource monitoring area, and obtain the distributed resource monitoring network topology structure; Collect data from IoT device nodes in the distributed resource monitoring network topology to obtain an original resource status data set, and perform normalization mapping and matrix conversion on the original resource status data set to obtain a resource description matrix; Perform time series sampling on the resource sharing historical data of IoT device nodes, extract device identification pairs, shared resource types, and resource sharing amounts within t time windows, and obtain the resource interaction time series feature sequence; The resource description matrix is ​​input into the feature encoder, and the resource state features are nonlinearly transformed through a fully connected neural network to obtain a resource state feature sub-vector; Slide the time window of the resource interaction time series feature sequence, calculate the resource sharing frequency matrix and time interval matrix between device pairs, and generate the resource interaction feature sub-vector through the graph attention network; The resource status feature sub-vector and the resource interaction feature sub-vector are input into the feature fusion network, and the feature weights are calculated through the multi-head attention mechanism to obtain the weighted feature vector. The weighted feature vector is then subjected to residual connection and layer normalization operations to obtain the device resource feature vector.

[0023] Specifically, the OpenHarmony distributed network is divided into multiple resource monitoring areas, and a regional resource coordination node is configured for each resource monitoring area to obtain the topological structure of the distributed resource monitoring network. The division of resource monitoring areas is determined based on the geographical location, functional similarity and concentration of resource requirements between devices. Each resource monitoring area is managed by one or more regional resource coordination nodes, which are responsible for monitoring the device resource status, scheduling and coordinating resource usage in the area. In this way, the devices in the OpenHarmony network are effectively organized in different monitoring areas, and the resource management of each area is performed by a designated coordination node, forming a distributed resource monitoring network topological structure. Data collection is performed on the IoT device nodes in the distributed resource monitoring network topological structure to obtain the original resource status data set, including information such as the CPU usage, memory occupancy, network bandwidth usage, storage space usage, and remaining power of each device. The original data is presented as a data set of multiple dimensions, and each item of the data set represents a certain resource status of the device. The original resource status data set is normalized and mapped and converted into a resource description matrix. The construction process of the resource description matrix includes normalizing each resource type so that the values ​​of different resource types have the same dimension and eliminate the scale differences of different resources. The formula for normalization mapping is expressed as: ; in, Indicates the device No. The original value of the resource, It is The average value of the resource, It is The standard deviation of the resource, is the normalized value. The resource sharing historical data of IoT device nodes is sampled in time series. By sampling the historical usage of device resources, the device identification pairs, shared resource types, and resource sharing amounts within t time windows are extracted to obtain the resource interaction time series feature sequence. The time window size is set to , record the resource sharing situation between devices in each time window, and extract the following information: In the time window, the device identification pair ( ), shared resource type ( ) and the amount of shared resources ( ). These data reflect the resource interaction between devices in a specific time period. By sliding the time window, the interaction features of each moment are gradually generated to form time series data. In order to process these resource interaction features, the resource description matrix is ​​input into the feature encoder, and the resource state features are nonlinearly transformed through a fully connected neural network. The fully connected neural network is expressed as: ; in, It is The output feature vector of the layer, It is The weight matrix of the layer, is the bias term, is the activation function. Through the nonlinear transformation of the fully connected layer, the high-order features of the resource state are extracted to obtain the resource state feature sub-vector. At the same time, the time window of the resource interaction time series feature sequence is slid to calculate the resource sharing frequency matrix and time interval matrix between the device pairs. The resource sharing frequency matrix represents the frequency of resource sharing between devices over a period of time, while the time interval matrix reflects the time interval of resource interaction between devices. These two matrices are processed by the graph attention network to generate a resource interaction feature sub-vector. The graph attention network can mine the potential relationship between devices through weighted calculation of the connections between devices and generate a more accurate resource interaction feature sub-vector. The update formula of the graph attention network is expressed as: ; in, Is a node The updated feature vector of Representation Node The set of neighbor nodes of Is a node and nodes The attention coefficient between is the weight matrix. Attention coefficient It is calculated through the correlation between node features, so as to assign different weights to each node. The resource state feature subvector and the resource interaction feature subvector are input into the feature fusion network, and the feature weights are calculated through the multi-head attention mechanism to obtain the weighted feature vector. The multi-head attention mechanism captures richer feature information by performing parallel calculations on multiple attention heads. Through this mechanism, information is integrated between multiple features to obtain a more comprehensive weighted feature vector. The calculation formula of the multi-head attention mechanism is: ; in, Represents query, key and value respectively, is the output of each attention head, is the output weight matrix. Through the weighted calculation of multi-head attention, the weighted feature vector is obtained. The weighted feature vector is subjected to residual connection and layer normalization operations to obtain the device resource feature vector. The purpose of residual connection is to alleviate the gradient vanishing problem in the deep layer of the neural network, and to help information flow smoothly through the network layer by directly adding the input and output. Layer normalization helps to stabilize the training process and improve the convergence speed. Through the above steps, the device resource feature vector is obtained, which describes the resource status and resource interaction behavior of each device.

[0024] In a specific embodiment, the process of executing step S2 may specifically include the following steps: Based on the device resource feature vector, a three-layer resource allocation architecture consisting of the device layer, coordination layer, and management layer is constructed. The resource status of each layer node is encoded hierarchically to obtain a hierarchical resource status matrix. The hierarchical resource state matrix is ​​processed by the gating unit of the dynamic gated graph attention network, the gating coefficients between nodes are calculated, and the dynamic gated feature matrix is ​​obtained; The dynamic gated feature matrix is ​​input into the LeakyReLU activation function for nonlinear transformation, and the attention score is calculated by the dot product operation of the node feature vector to obtain the original attention weight; The original attention weights are normalized by softmax and weighted aggregated with the node features to obtain the updated node representation vector; Based on the node representation vector, the resource consumption ratio and resource sharing frequency between device nodes are calculated, and the resource dependency measurement index between node pairs is constructed to obtain the dependency measurement matrix. The dependency metric matrix is ​​input into the multi-head self-attention module, and the attention weights of different heads are calculated through the query-key-value mechanism to obtain the multi-head attention features. The multi-head attention features are concatenated and linearly transformed to obtain the resource sharing weight matrix.

[0025] Specifically, the resource allocation system is divided into three layers: device layer, coordination layer, and management layer. The hierarchical architecture design can effectively organize and manage the resource allocation of devices, reduce the complexity of the system, and improve the efficiency of resource scheduling. In this three-layer architecture, the device layer represents each specific IoT device, the coordination layer is responsible for the coordination and resource scheduling between devices, and the management layer is responsible for the overall resource allocation decision. The nodes of each layer correspond to different levels of resource management and scheduling functions, and the resource status of each layer also has different encoding methods, reflecting the hierarchical management method. The node resource status of each layer is hierarchically encoded to obtain a hierarchical resource status matrix. According to the resource management requirements of different levels, the device resources of each layer are independently encoded and represented. The resource status of each node in the device layer, coordination layer, and management layer will be represented in a separate matrix for subsequent processing and calculation. For the hierarchical resource status matrix, the dynamic gated graph attention network (DG-GAT) is used for processing. DG-GAT is a graph neural network that can dynamically adjust the information propagation mode between nodes. By introducing gating units, the network can flexibly control the intensity and direction of information transmission between nodes. Through the dynamic gating mechanism, the gating coefficients between each node are calculated according to the resource requirements and interaction relationships of each node. These gating coefficients can adjust the influence between nodes, thereby obtaining a dynamic gating feature matrix. The calculation formula of the gating coefficient is expressed as: ; in, Representation Node and nodes The gating coefficient between and The nodes are and nodes The characteristic vector of Indicates a connection operation. and are the weights and biases of the gating unit, is a sigmoid activation function, which is used to map the output to the range of (0,1). The dynamic gate feature matrix is ​​input into the LeakyReLU activation function for nonlinear transformation to enhance the nonlinear expression ability of the network so as to better capture the complex nonlinear relationship between the device resource states. The form of the LeakyReLU activation function is: ; in, is the input value, is a small constant, set to 0.01. Through this transformation of the activation function, the negative gradient information can be maintained, thereby avoiding the "dead neuron" problem during neural network training. After the nonlinear transformation, the obtained node feature vector is more expressive and can adapt to more complex resource scheduling and allocation problems. By calculating the dot product operation of the same amount of node features, the attention score between nodes is calculated. Evaluate the importance of the interaction between device nodes. The formula for the dot product operation is: ; in, Representation Node and nodes The attention score between and The nodes are and nodes The characteristic vector of is a trainable weight matrix. By calculating the attention score, we can evaluate which nodes have stronger resource sharing and dependency. We normalize the original attention weights by softmax and convert all attention scores into probability distributions so that the sum of the attention weights of all nodes is 1. The formula of the Softmax function is: ; in, Is a node and nodes The original attention weights between Representation Node The neighbor node set of each node. Through this process, the attention weights of all nodes will be normalized so that they can reflect the relative importance of different nodes in the subsequent weighted aggregation process. The features of the nodes are updated through weighted aggregation operations. The features of the neighbor nodes of each node are weighted summed according to the normalized attention weights to obtain the updated node representation vector. The update formula is: ; in, is the updated node The characteristic vector of is the normalized attention weight, is a weight matrix. Through weighted aggregation operation, the updated node representation vector is obtained. Based on the updated node representation vector, the resource consumption ratio and resource sharing frequency between device nodes are calculated to construct the resource dependency measurement index between node pairs. The formula of resource dependency measurement is expressed as: ; in, Indicates the device and equipment The resource dependency measure between Yes Equipment and equipment The amount of resource sharing between Yes Equipment and equipment The resource consumption between them. The dependency metric matrix is ​​input into the multi-head self-attention module for processing. The multi-head self-attention module calculates the attention weights of different heads through the query-key-value mechanism. For each pair of device nodes, their feature representations are used as input, and different attention weights are calculated through multiple independent attention heads. The features of these different heads are spliced ​​together and linearly transformed to obtain the final resource sharing weight matrix. The calculation formula of the multi-head self-attention module is: ; in, and They are query, key and value, is the output of each attention head, Concat represents the concatenation operation, is the output weight matrix. Through this multi-head mechanism, the multi-level resource sharing and dependency relationships between device nodes are captured, and finally the resource sharing weight matrix is ​​obtained.

[0026] In a specific embodiment, the process of executing step S3 may specifically include the following steps: The resource sharing weight matrix is ​​decomposed into resource types according to CPU usage, memory usage, storage space usage, network bandwidth usage, and remaining power of the device to obtain multiple resource views; Each resource view is input into the graph convolutional network, and the node neighborhood information is aggregated to obtain the node-level feature matrix. Multi-scale feature extraction is performed on the node-level feature matrix through dilated convolution. The convolution kernel size is set to d, the dilation rate is set to r, and the output feature dimension is set to f to obtain the temporal feature sequence. The temporal feature sequence is input into the channel attention module for global average pooling and maximum pooling to obtain the channel attention weight, and the channel attention weight is multiplied element by element with the temporal feature sequence to obtain the channel weighted temporal feature; The timing correlation of the graph structure is calculated based on the channel-weighted timing features. The long-range dependency relationship is established through the causal convolution structure to obtain the timing dependency features. The timing dependency features are then concatenated with the node-level feature matrix to generate the device resource status embedding matrix.

[0027] Specifically, consider the multi-dimensional resource usage of each device. IoT devices involve multiple resource types at the same time, including CPU usage, memory usage, storage space usage, network bandwidth usage, and remaining device power. These resource types have their own independent attributes and complex dependencies between them. Decompose the resource sharing weight matrix according to these resource types, and each resource type corresponds to a resource view. For example, for a device For example, the elements of the resource sharing weight matrix are Indicates the device With equipment The weight of shared resources between devices is decomposed according to different resource types such as the device's CPU occupancy rate and memory usage rate to obtain multiple independent resource views. These resource views respectively represent the sharing of devices on different resources. Input each resource view into the graph convolutional network. Graph convolutional network is a deep learning method that can effectively process graph structured data. Each node in the resource view represents an IoT device, and the edges between nodes represent the resource sharing relationship between devices. The graph convolutional network aggregates the information of neighboring nodes through the edge connection information between nodes, thereby learning the global resource characteristics of the device. Take each resource view as input, and perform aggregation operations on the neighborhood information of the node through the graph convolutional network. The calculation formula of the graph convolutional layer is: ; in, Is a node The updated feature vector, Representation Node The set of neighbor nodes of is a trainable weight matrix, Neighbor node The characteristic vector of is the bias term, is the activation function. Through this aggregation operation, the node The feature vector contains not only the device The resource status of the network itself is also integrated with the resource information of its neighboring devices. After the graph convolution operation is completed, the node-level feature matrix is ​​extracted at multiple scales through dilated convolution. Dilated convolution is a type of extended convolution that effectively expands the receptive field while maintaining computational efficiency, thereby capturing longer-distance dependencies. The key to dilated convolution is to set the size of the convolution kernel. and void ratio , where the dilation rate determines the spacing between adjacent elements in the convolution operation. The formula for dilated convolution is: ; in, represents the output eigenvalue, is the input eigenvalue, is the weight of the convolution kernel, is the void ratio, is the size of the convolution kernel. Through the atrous convolution operation, each node of the node-level feature matrix will obtain richer temporal information from its neighboring nodes. The temporal feature sequence is input into the channel attention module. The channel attention mechanism enables the network to focus on important feature channels by assigning different weights to the features of each channel. The channel attention module processes the temporal feature sequence through global average pooling and maximum pooling operations. Global average pooling obtains the global information of each channel by calculating the average value of each feature channel; while maximum pooling extracts the most significant features by calculating the maximum value of each feature channel. The operations of global average pooling and maximum pooling are expressed as: ; in, represents the length of the feature sequence, It is Through these pooling operations, the global information and the most significant information of each feature channel are obtained. The channel attention weight is multiplied element by element with the time series feature sequence to obtain the channel weighted time series feature. Through the weighted operation, the network can better focus on important feature channels and suppress unimportant channels, thereby improving the model's ability to process time series features. The weighted operation is expressed as: ; in, is the channel attention weight, is the time series feature, is a weighted feature. Based on the channel weighted timing features, the timing correlation of the graph structure is calculated. In the graph structure, the timing dependency of device nodes is reflected by factors such as the resource sharing frequency and resource consumption between nodes. In order to capture these timing dependencies, a causal convolution structure is adopted, which is a convolution method that can effectively capture the dependencies between the previous and next time steps in the sequence. Causal convolution relies only on the current or previous time step through convolution operations without leaking future information. The calculation formula of causal convolution is: ; in, Represents the output at the current moment, is the input at the current time and before, is the weight of the convolution kernel, is the size of the convolution kernel. Through causal convolution, the changes in device resource status over time and long-range dependencies are effectively captured. The time-dependent features are concatenated with the node-level feature matrix. Information from different sources is integrated to generate a device resource status embedding matrix that fully represents the resource status of the device and its time-series changes. The formula for feature concatenation is: ; in, is the concatenated feature vector, is the feature vector of the node, It is a timing dependent feature.

[0028] In a specific embodiment, the process of executing step S4 may specifically include the following steps: Construct resource state space and resource allocation action space for each dimension in the device resource state embedding matrix, and perform state encoding to obtain a state-action mapping matrix; An online evaluation network and a target evaluation network are constructed for the state-action mapping matrix, and action value features are extracted to obtain a double Q network evaluation model. The comprehensive reward score is calculated based on the dual Q network evaluation model to obtain the standardized reward value, and the standardized reward value is input into the temporal difference learning unit. The discount factor γ is set to 0.95, the target Q value is calculated, and the learning target sequence is obtained. The Huber loss is calculated for the learning target sequence and the output value of the online evaluation network, and the network parameters are updated through the Adam optimizer to obtain the optimized network parameters; Based on the optimized network parameters, a cyclic experience replay buffer is constructed, and the state transition quadruple is stored in the buffer to obtain a training data batch; The strategy optimization iteration is performed on the training data batches, and the parameters of the online evaluation network are soft-updated to the target evaluation network to obtain the optimized dual-Q network model. The optimized dual-Q network model is used for resource allocation decisions. The initial value of the exploration rate ε is set to 0.9 and updated according to the linear attenuation strategy to generate a global resource allocation strategy.

[0029] Specifically, considering that each dimension in the device resource state embedding matrix represents a different aspect of the device resources, such as CPU occupancy, memory usage, storage space occupancy, network bandwidth usage, and device remaining power, each resource type is regarded as an independent resource state space. On this basis, a resource allocation action space is constructed, and the action space represents possible resource scheduling decisions. For example, for a certain device, resource scheduling actions include limiting the device's CPU usage within a certain range, or reallocating storage resources among certain devices. In order to effectively model these resource states and actions, state encoding is used to represent the device's resource state in a discrete, numerical vector form. The resource state of each device is represented by a multidimensional vector, in which each dimension represents the state of a specific resource. The correspondence between these state information and the action space constitutes a state-action mapping matrix, and each element of the matrix represents the reward value for performing an action under a specific state. Set the state space to , the action space is , then the state-action mapping matrix It is expressed as: ; Is to perform an action After receiving the instant reward, is the discount factor, Represents the maximum value of all actions in the next state. The double Q network model is constructed using the state-action mapping matrix. The double Q network (DoubleQ-learning) is used to solve the problem of overestimation in the traditional Q-learning method when estimating the action value. The double Q network includes two Q networks: one is the online evaluation network, which is used to generate action estimates under the current strategy; the other is the target evaluation network, which is used to calculate the target Q value. The online evaluation network estimates the possible action value in each state through the current strategy, while the target evaluation network is used to generate the target Q value based on historical data. The Q value of the online evaluation network is set to , the Q value of the target evaluation network is , then the update formula of the double Q network is: ; In order to calculate the reward score and update the network parameters, the comprehensive reward score is calculated based on the double Q network evaluation model. The reward score is used to measure the value of a state-action pair in the current environment and is defined as: ; in, is the immediate reward based on the current state and action, Is to perform an action The cost required, is a weight parameter used to balance rewards and costs. In the dual-Q network, the comprehensive reward score is standardized to ensure that the reward value is within a reasonable range. The standardized reward value is expressed as: ; in, and are the mean and standard deviation of the reward values, respectively, which are used to scale the reward values ​​to a standard range. The standardized reward values ​​are input into the temporal difference learning unit, which is a reinforcement learning method used to evaluate and optimize strategies. Set the discount factor , then the calculation formula of the target Q value is: ; On this basis, the learning target sequence is calculated and compared with the output value of the online evaluation network. In order to minimize this gap, the Huber loss function is used to calculate the error between the target Q value and the output of the online evaluation network. The Huber loss function effectively reduces the impact of outliers and is defined as: ; in, is a hyperparameter used to control the sensitivity of the loss. The parameters of the network are updated through the Adam optimizer. Adam is an optimization algorithm that can adaptively adjust the learning rate and is suitable for large-scale data training in deep learning. The Adam optimizer updates the parameters through the following formula: ; in, are the parameters of the network, and are the estimates of momentum and gradient, respectively, is the learning rate, is a small constant used to avoid division by zero errors. In this process, a cyclic experience replay buffer is constructed based on the optimized network parameters. The cyclic experience replay buffer is used to store state transition quads, i.e., the action taken in each state, the immediate reward, the next state, and the flag of whether it is terminated. The use of the replay buffer helps improve the stability and convergence speed of training. The stored state transition quad is expressed as: ; in, is the current state, is the action taken, It is the reward obtained. is the next state, and done is the mark of whether it is terminated. A batch of training data is sampled from the playback buffer for optimization to obtain a batch of training data. During the optimization process, policy optimization iterations are performed to ensure consistency between the target network and the online network by soft-updating the parameters of the online evaluation network to the target evaluation network. The soft update operation is implemented with a small update step size to avoid drastic changes in the parameters of the target network. A global resource allocation strategy is generated for resource allocation decisions through the optimized dual Q network model. In order to enhance the exploration ability of the model, the exploration rate is set The initial value is 0.9 and is updated according to the linear decay strategy , thus balancing the relationship between exploration and utilization. After the above optimization and update, a global resource allocation strategy is finally obtained.

[0030] In a specific embodiment, the process of executing step S5 may specifically include the following steps: Perform device feature analysis on the global resource allocation strategy, calculate the similarity matrix between devices through resource utilization patterns and functional similarity, and obtain the device association graph; The device association graph is input into the hierarchical clustering algorithm, and the initial device clusters are obtained through a bottom-up iterative merging process with a clustering threshold of 0.75. Binary encode each device state in the initial device cluster, encode the active state as 1, and encode the dormant state as 0, generate a device working state matrix, and input the device working state matrix into the genetic algorithm encoder for chromosome construction to obtain the initial population; Perform genetic operations on the initial population, set the crossover probability to 0.8, and perform individual evaluation and selection through the fitness function based on service quality and energy consumption to obtain the optimized device state combination; The optimized device status combinations are grouped according to resource consumption levels, and the minimum active node number threshold is calculated for each group of devices to obtain the cluster working mode sequence; A dynamic scheduling schedule is constructed for the cluster working mode sequence. The active time and sleep time of the equipment are set through a sliding time window to obtain the initial scheduling plan. The initial scheduling plan is input into the scheduling optimizer. The scheduling timing is optimized through a dynamic programming algorithm to generate a dynamic scheduling sequence for the equipment.

[0031] Specifically, the global resource allocation strategy is analyzed for device characteristics, and the similarity matrix between devices is calculated based on the resource utilization patterns of the devices and the functional similarity between them. The resource utilization patterns of the devices are obtained by monitoring various resource indicators of the devices, such as CPU, memory, storage, bandwidth, and power. For each pair of devices, the Euclidean distance of their resource utilization patterns is calculated, which is defined as: ; in, Indicates the device and equipment The resource utilization pattern distance between Indicates the device In the The usage of resource types. is the number of resource types. The functional similarity of devices is calculated based on the working functions and service types of the devices. and equipment If they perform similar tasks (such as processing similar types of data or providing similar services), the similarity between them is high. Based on these calculations, a similarity matrix between devices is constructed, which is used to construct a device association graph, in which the edge weights of each device node and other devices represent the similarity between them. The device association graph is input into the hierarchical clustering algorithm for analysis, and a bottom-up hierarchical clustering method is used to form device clusters by gradually merging similar device nodes. During the clustering process, a threshold of 0.75 is set to determine whether the similarity between devices is large enough to decide whether to merge them into one cluster. The hierarchical clustering algorithm determines the merging order by calculating the similarity between devices and generates preliminary device clusters. Assuming that each device is initialized as a separate cluster, the goal of clustering is to eventually obtain a device cluster by merging clusters with high similarity. In each merging process, the similarity of devices in the cluster will continue to increase until the preset similarity threshold is reached. After obtaining the initial device cluster, each device in the cluster is binary-encoded to identify the working status of the device. The active state of the device is encoded as 1, and the dormant state is encoded as 0 to generate a device working status matrix. In this matrix, rows represent devices, columns represent different time periods, and a matrix element value of 1 indicates that the device is active during that time period, and a value of 0 indicates that the device is dormant. In time period The working status is expressed as: ; After the working status of all devices is matrixed, it is input into the genetic algorithm encoder. Genetic algorithm is an optimization algorithm that simulates natural selection and genetic mechanisms, generating the optimal solution through continuous selection, crossover and mutation. In the encoder, the device working status matrix is ​​converted into chromosomes, where the working status of each device is part of the chromosome. The genetic algorithm generates a new population by performing selection, crossover and mutation operations. The crossover probability is set to 0.8, which means that there is an 80% probability that each pair of parent individuals will crossover to generate a new generation of individuals. The fitness function is used to evaluate the quality of each individual. The fitness function is designed based on quality of service (QoS) and energy consumption. The goal is to maximize the quality of service and reduce energy consumption. Assume that the fitness function is: ; wherein, and are the weight coefficients of quality of service and energy consumption respectively, QoS represents the evaluation value of quality of service, and Energy represents the energy consumption of the device. After completing the fitness evaluation and selection, an optimized device state combination is obtained. According to the resource consumption level of the devices, the optimized device state combination is grouped, and each group of devices corresponds to a cluster. To effectively manage the resources of each cluster, the threshold of the minimum number of active nodes in each group of devices is calculated, that is, how many devices must be in the active state when the cluster is running. It is obtained by calculating the number of 1s in each column of the device working state matrix in each cluster. If the number of active devices in a certain period is less than the threshold, the working mode of the cluster needs to be adjusted. Based on the working mode of the cluster, a dynamic scheduling time table is constructed. The scheduling time table determines the active and dormant states of the devices in different time periods. By sliding the time window, the active time and dormant time of each device in different time periods are set to obtain an initial scheduling scheme. Assume that the size of the time window is , and its state in each time period is adjusted according to the working requirements of the device. The initial scheduling scheme is expressed as: ; wherein, represents the working state of device i in a certain time period. The initial scheduling scheme is input into the scheduling optimizer, and the dynamic programming algorithm is used to optimize the scheduling time sequence. The dynamic programming algorithm optimizes the resource allocation of the devices in a recursive manner, ensures the reasonable utilization of resources, and reduces the energy consumption while ensuring the quality of service. During the optimization process, dynamic programming calculates the optimal scheduling strategy by analyzing the active and dormant states of each device in different time periods. Through optimization, the dynamic scheduling sequence of the devices is finally generated.

[0032] In a specific embodiment, the process of executing step S6 may specifically include the following steps: Divide the device dynamic scheduling sequence into in-region resource allocation requests and cross-region resource allocation requests, and perform queue sorting through the resource scheduling priority sorting algorithm to obtain a hierarchical resource request queue; Construct a transaction control block TCB for each resource allocation operation in the hierarchical resource request queue, including a transaction identifier, a resource type, a requester identifier, and a responder identifier, to obtain a transaction description set; Input the transaction description set into the four-phase commit protocol processor, and construct a transaction state transition rule in the order of request initiation, resource reservation, execution commit, and completion confirmation to obtain a distributed transaction execution framework; Assign a globally unique identifier TID to the transaction node in the distributed transaction execution framework, maintain the transaction timing dependency through the vector clock algorithm, obtain the transaction ordered execution sequence, and build a resource lock manager based on the transaction ordered execution sequence. Perform concurrency control on competing resources through a two-phase locking protocol to obtain a resource locking status table. The resource locking state table is used to perform resource allocation operations in the order of transaction priority, and the execution status of each operation is recorded through the resource scheduling log to obtain a resource allocation result set; The resource allocation result set is verified for consistency, and the resource allocation status is synchronized between regions through the Paxos consensus algorithm to obtain the global consistency status. The global consistency status is input into the resource status updater, and the device resource status is updated through atomic commit operations to generate distributed resource allocation results.

[0033] Specifically, the dynamic scheduling sequence of the device is obtained from the dynamic changes of the device resources. These scheduling sequences record the working status, resource requirements and adjustment requirements of the device. According to this dynamic scheduling information, it is divided into two types of requests: intra-regional resource allocation request and cross-regional resource allocation request. Intra-regional resource allocation request refers to the resource scheduling and adjustment between devices in the same physical area, while cross-regional resource allocation request involves resource allocation between multiple areas. For example, if a device requests access to network bandwidth or storage resources, and these resources are not sufficient to meet the demand in the area where the device is located, a cross-regional request will be generated. In order to process these requests more efficiently, a resource scheduling priority sorting algorithm is adopted. The algorithm sorts the requests according to the urgency of the request, the importance of the resources and the remaining available resources, and gives priority to high-priority requests. After sorting, all resource requests form a hierarchical resource request queue, in which the priority of each request is explicitly marked. For each resource allocation operation in the hierarchical resource request queue, a transaction control block (TCB) is constructed. The transaction control block is a data structure used to describe the basic information of a resource allocation operation. The TCB contains the following key information: transaction identifier (TID), resource type, requester identifier and responder identifier. For example, in the process of cross-region resource allocation, TID will uniquely identify a resource allocation request, resource type indicates the requested resource (such as CPU, memory, bandwidth, etc.), and requester identifier and responder identifier indicate the device that initiates the request and the device that provides the resource, respectively. With this information, a transaction description set is generated. Each element in the transaction description set is a transaction control block. After the transaction description is input into the four-phase commit protocol processor, the system constructs transaction state transition rules in the order of request initiation, resource reservation, execution submission, and completion confirmation. The four-phase commit protocol is a distributed transaction protocol that ensures the consistency and reliability of resource scheduling in a distributed environment. The four phases are: request initiation phase, resource reservation phase, execution submission phase, and completion confirmation phase. In the request initiation phase, the requester device initiates a resource request; in the resource reservation phase, the responder device reserves resources and prepares for execution; in the execution submission phase, the actual allocation operation of the resource begins; in the completion confirmation phase, the resource allocation result is confirmed and the transaction is completed. In order to ensure that these transactions can be executed in the correct order, a globally unique identifier is assigned to each transaction node, and the vector clock algorithm is used to maintain the timing dependency of the transaction. The vector clock algorithm tracks the order of transaction execution by assigning a vector identifier to each transaction node to ensure the order of transactions. Whenever the state of a transaction changes, its corresponding clock value is updated to ensure that the dependencies between transactions are accurately maintained. Based on the timing dependencies of transactions, the sequence of orderly execution of transactions is derived, and a resource lock manager is constructed. The resource lock manager is used to perform concurrency control on competing resources to prevent multiple transactions from modifying the same resource state at the same time.Through the two-phase locking protocol, concurrent access is effectively controlled during the resource scheduling process. The basic principle of the two-phase locking protocol is that during the execution of a transaction, a lock needs to be requested in the locking phase and released in the release phase, ensuring that access to resources is exclusively owned by one transaction at any time. When executing resource allocation operations, resource requests are processed in order of priority of transactions. The execution status of resource allocation operations will be recorded in the resource scheduling log for subsequent rollback operations and state recovery. The resource scheduling log not only records the allocation of resources, but also includes the execution time, status, and success and failure flags of each operation. These log information helps the system to roll back and recover operations when errors or failures occur. The resource allocation result set is verified for consistency. Through the Paxos consensus algorithm, the resource allocation status is synchronized between regions to ensure that all devices in a distributed environment maintain consistency in access and use of resources. The Paxos algorithm ensures consensus among nodes in multiple regions through a multi-round voting mechanism, thereby avoiding resource conflicts or failures caused by inconsistent information between nodes. The global consistency state is input into the resource status updater, and the resource status of the device is updated through an atomic commit operation. The atomic commit operation ensures the atomicity of resource status update, that is, either the resource status of all devices is successfully updated, or it rolls back to the state before the update when an error occurs, ensuring the consistency and reliability of the system. After the above steps, the distributed resource allocation result is finally obtained.

[0034] The above describes the IoT device resource allocation method based on OpenHarmony in the embodiment of the present invention. The following describes the IoT device resource allocation system based on OpenHarmony in the embodiment of the present invention. Figure 2 In an embodiment of the present invention, an embodiment of an Internet of Things device resource allocation system based on OpenHarmony includes: The acquisition module is used to obtain the resource description matrix and resource interaction time series feature sequence based on the OpenHarmony distributed network, and obtain the device resource feature vector through feature fusion calculation; A calculation module, used to calculate the resource dependency relationship between device nodes based on the device resource feature vector, and obtain a resource sharing weight matrix; A partitioning module is used to partition the resource sharing weight matrix into multiple resource views according to resource types, and perform feature extraction on the multiple resource views to obtain a device resource state embedding matrix; An analysis module, used to perform a resource allocation analysis of a dual-depth Q network on the device resource state embedding matrix and generate a global resource allocation strategy; A generation module is used to divide IoT devices into multiple device clusters according to the global resource allocation strategy, and generate a device dynamic scheduling sequence based on the working status of the devices in each device cluster; The output module is used to construct the intra-region resource allocation operation and the cross-region resource request operation into a four-stage distributed transaction based on the device dynamic scheduling sequence, and output the distributed resource allocation result.

[0035] Through the collaborative cooperation of the above components, by constructing a multi-perspective spatiotemporal graph structure and a hybrid attention network, the device resource status and interaction characteristics are effectively extracted, the accuracy of resource status representation is improved, and a reliable data basis is provided for resource allocation decisions. The dynamic gated graph attention network is used to calculate the resource dependencies between devices, and the resource allocation strategy is optimized in combination with the dual deep Q learning algorithm, which improves the resource utilization and energy efficiency of the system. An intelligent sleep and wake-up mechanism based on genetic algorithm is designed, which reduces the overall energy consumption of the system by optimizing the combination of device working states while ensuring the service quality. A four-stage distributed transaction mechanism is proposed, which combines the two-stage locking protocol and the Paxos consensus algorithm to reduce the communication delay of the system and significantly improve the reliability of distributed resource allocation. Through the collaborative optimization of the multi-level resource allocation architecture, the overall stability of the system is improved, and the resource allocation problem of large-scale heterogeneous IoT devices is effectively solved.

[0036] Reference Figure 3 In an embodiment of the present invention, a computer device is also provided. The computer device may be a server, and its internal structure may be as follows: Figure 3 As shown. The computer device includes a processor, a memory, a display screen, an input device, a network interface and a database connected through a system bus. Among them, the processor designed by the computer is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the above method is implemented.

[0037] Those skilled in the art will understand that Figure 3 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied.

[0038] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the above method is implemented. It can be understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0039] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided by the present invention and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double-speed data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM.

[0040] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0041] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the whole or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.

[0042] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for allocating IoT device resources based on OpenHarmony, characterized in that: The method comprises: Based on the OpenHarmony distributed network, the resource description matrix and resource interaction time series feature sequence are obtained, and the device resource feature vector is obtained through feature fusion calculation; Based on the device resource feature vector, the resource dependency relationship between device nodes is calculated to obtain a resource sharing weight matrix; Dividing the resource sharing weight matrix into multiple resource views according to resource types, and performing feature extraction on the multiple resource views to obtain a device resource state embedding matrix; Performing a resource allocation analysis of a dual-depth Q network on the device resource state embedding matrix to generate a global resource allocation strategy; Dividing the IoT devices into a plurality of device clusters according to the global resource allocation strategy, and generating a device dynamic scheduling sequence based on the working status of the devices in each of the device clusters; Based on the device dynamic scheduling sequence, the intra-region resource allocation operation and the cross-region resource request operation are constructed into a four-stage distributed transaction, and the distributed resource allocation result is output.

2. The method for allocating IoT device resources based on OpenHarmony according to claim 1, characterized in that: The method of obtaining a resource description matrix and a resource interaction time series feature sequence based on the OpenHarmony distributed network and obtaining a device resource feature vector through feature fusion calculation includes: Divide the OpenHarmony distributed network into multiple resource monitoring areas, configure a regional resource coordination node for each resource monitoring area, and obtain a distributed resource monitoring network topology structure; Performing data collection on the IoT device nodes in the distributed resource monitoring network topology to obtain an original resource status data set, and performing normalization mapping and matrix conversion on the original resource status data set to obtain a resource description matrix; Performing time series sampling on the resource sharing historical data of the IoT device nodes, extracting device identification pairs, shared resource types, and resource sharing amounts within t time windows, and obtaining a resource interaction time series feature sequence; Input the resource description matrix into a feature encoder, perform nonlinear transformation on the resource state features through a fully connected neural network, and obtain a resource state feature subvector; Sliding the resource interaction time series feature sequence in a time window, calculating the resource sharing frequency matrix and the time interval matrix between the device pairs, and generating a resource interaction feature sub-vector through a graph attention network; The resource state feature sub-vector and the resource interaction feature sub-vector are input into a feature fusion network, feature weights are calculated through a multi-head attention mechanism to obtain a weighted feature vector, and residual connection and layer normalization operations are performed on the weighted feature vector to obtain a device resource feature vector.

3. The method for allocating IoT device resources based on OpenHarmony according to claim 2, characterized in that: The calculating the resource dependency relationship between the device nodes based on the device resource feature vector to obtain a resource sharing weight matrix includes: Based on the device resource feature vector, a three-layer resource allocation architecture of a device layer, a coordination layer, and a management layer is constructed, and the resource status of each layer node is hierarchically encoded to obtain a hierarchical resource status matrix; The hierarchical resource state matrix is ​​processed by a gating unit of a dynamic gated graph attention network, and the gating coefficients between nodes are calculated to obtain a dynamic gated feature matrix; The dynamic gated feature matrix is ​​input into the LeakyReLU activation function for nonlinear transformation, and the attention score is calculated by the dot product operation of the node feature vector to obtain the original attention weight; The original attention weight is subjected to softmax normalization processing, and weighted aggregation is performed with the node feature to obtain an updated node representation vector; Based on the node representation vector, the resource consumption ratio and resource sharing frequency between the device nodes are calculated, and the resource dependency measurement index between the node pairs is constructed to obtain a dependency measurement matrix; The dependency metric matrix is ​​input into the multi-head self-attention module, and the attention weights of different heads are calculated through the query-key-value mechanism to obtain the multi-head attention features, and the multi-head attention features are concatenated and linearly transformed to obtain the resource sharing weight matrix.

4. The method for allocating IoT device resources based on OpenHarmony according to claim 3, characterized in that: The resource sharing weight matrix is ​​divided into a plurality of resource views according to resource types, and features are extracted from the plurality of resource views to obtain a device resource state embedding matrix, including: Decomposing the resource sharing weight matrix by resource type according to CPU occupancy, memory occupancy, storage space occupancy, network bandwidth occupancy and device remaining power to obtain multiple resource views; Input each resource view into the graph convolutional network, perform aggregation operation on the node neighborhood information to obtain a node-level feature matrix, and perform multi-scale feature extraction on the node-level feature matrix through hole convolution, set the convolution kernel size to d, the hole rate to r, and the output feature dimension to f, to obtain a temporal feature sequence; Input the temporal feature sequence into the channel attention module for global average pooling and maximum pooling to obtain the channel attention weight, and multiply the channel attention weight by the temporal feature sequence element by element to obtain the channel weighted temporal feature; The timing correlation of the graph structure is calculated based on the channel weighted timing features, and a long-range dependency relationship is established through a causal convolution structure to obtain a timing dependency feature. The timing dependency feature is then feature-concatenated with a node-level feature matrix to generate a device resource status embedding matrix.

5. The method for allocating IoT device resources based on OpenHarmony according to claim 4, characterized in that: The performing of a dual-depth Q network resource allocation analysis on the device resource state embedding matrix to generate a global resource allocation strategy includes: Constructing a resource state space and a resource allocation action space for each dimension in the device resource state embedding matrix, and performing state encoding to obtain a state-action mapping matrix; An online evaluation network and a target evaluation network are constructed for the state-action mapping matrix, and action value features are extracted to obtain a double Q network evaluation model; Calculate the comprehensive reward score based on the dual Q network evaluation model to obtain a standardized reward value, input the standardized reward value into the temporal difference learning unit, set the discount factor γ to 0.95, calculate the target Q value, and obtain the learning target sequence; Calculating Huber loss for the learning target sequence and the output value of the online evaluation network, updating the network parameters through the Adam optimizer, and obtaining optimized network parameters; Constructing a cyclic experience replay buffer based on the optimized network parameters, and storing the state transition quadruple in the buffer to obtain a training data batch; The strategy optimization iteration is performed on the training data batch, and the parameters of the online evaluation network are soft-updated to the target evaluation network to obtain an optimized dual-Q network model. The optimized dual-Q network model is used for resource allocation decision-making, and the initial value of the exploration rate ε is set to 0.9 and updated according to the linear attenuation strategy to generate a global resource allocation strategy.

6. The method for allocating IoT device resources based on OpenHarmony according to claim 5, characterized in that: The method of dividing the IoT devices into a plurality of device clusters according to the global resource allocation strategy and generating a device dynamic scheduling sequence based on the working status of the devices in each of the device clusters includes: Performing device feature analysis on the global resource allocation strategy, calculating a similarity matrix between devices through resource utilization patterns and functional similarities, and obtaining a device association graph; The device association graph is input into a hierarchical clustering algorithm, and an initial device cluster is obtained by setting a clustering threshold of 0.75 through a bottom-up iterative merging process; Binary encoding is performed on each device state in the initial device cluster, the active state is encoded as 1, and the dormant state is encoded as 0, to generate a device working state matrix, and the device working state matrix is ​​input into a genetic algorithm encoder for chromosome construction to obtain an initial population; Performing genetic operations on the initial population, setting the crossover probability to 0.8, and performing individual evaluation and selection through a fitness function based on service quality and energy consumption to obtain an optimized device state combination; The optimized device state combinations are grouped according to resource consumption levels, and a minimum active node number threshold is calculated for each group of devices to obtain a cluster working mode sequence; A dynamic scheduling schedule is constructed for the cluster working mode sequence, and the active time and sleep time of the equipment are set through a sliding time window to obtain an initial scheduling plan. The initial scheduling plan is input into a scheduling optimizer, and the scheduling timing is optimized through a dynamic programming algorithm to generate a dynamic scheduling sequence for the equipment.

7. The method for allocating IoT device resources based on OpenHarmony according to claim 6, characterized in that: Based on the device dynamic scheduling sequence, the intra-region resource allocation operation and the cross-region resource request operation are constructed into a four-stage distributed transaction, and the distributed resource allocation result is output, including: Dividing the device dynamic scheduling sequence into intra-regional resource allocation requests and cross-regional resource allocation requests, sorting the queues by a resource scheduling priority sorting algorithm, and obtaining a hierarchical resource request queue; Constructing a transaction control block TCB for each resource allocation operation in the hierarchical resource request queue, including a transaction identifier, a resource type, a requester identifier and a responder identifier, to obtain a transaction description set; Input the transaction description set into the four-phase commit protocol processor, construct transaction state transition rules in the order of request initiation, resource reservation, execution submission and completion confirmation, and obtain a distributed transaction execution framework; Assigning a globally unique identifier TID to a transaction node in the distributed transaction execution framework, maintaining transaction timing dependencies through a vector clock algorithm, obtaining a transaction ordered execution sequence, and building a resource lock manager based on the transaction ordered execution sequence, performing concurrency control on competing resources through a two-phase locking protocol, and obtaining a resource locking state table; The resource locking state table is used to perform resource allocation operations according to the transaction priority order, and the execution status of each operation is recorded through the resource scheduling log to obtain a resource allocation result set; The resource allocation result set is verified for consistency, and the resource allocation status is synchronized between regions through the Paxos consensus algorithm to obtain a global consistency state, and the global consistency state is input into the resource status updater, and the device resource status is updated through an atomic commit operation to generate a distributed resource allocation result.

8. An IoT device resource allocation system based on OpenHarmony, characterized in that: Used to execute the OpenHarmony-based IoT device resource allocation method according to any one of claims 1 to 7, the system comprising: The acquisition module is used to obtain the resource description matrix and resource interaction time series feature sequence based on the OpenHarmony distributed network, and obtain the device resource feature vector through feature fusion calculation; A calculation module, used to calculate the resource dependency relationship between device nodes based on the device resource feature vector to obtain a resource sharing weight matrix; A partitioning module, used to partition the resource sharing weight matrix into multiple resource views according to resource types, and perform feature extraction on the multiple resource views to obtain a device resource status embedding matrix; An analysis module, configured to perform a resource allocation analysis of a dual-depth Q network on the device resource state embedding matrix to generate a global resource allocation strategy; A generation module, used to divide the IoT devices into a plurality of device clusters according to the global resource allocation strategy, and generate a device dynamic scheduling sequence based on the working status of the devices in each of the device clusters; The output module is used to construct the intra-region resource allocation operation and the cross-region resource request operation into a four-stage distributed transaction based on the device dynamic scheduling sequence, and output the distributed resource allocation result.

9. A computer device, characterized in that: It includes a memory and a processor, the memory stores a computer program that can be run on the processor, and is characterized in that when the processor executes the computer program, it implements the Internet of Things device resource allocation method based on OpenHarmony as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the processor executes the method for allocating IoT device resources based on OpenHarmony as claimed in any one of claims 1 to 7.

Images