Multi-dimensional feature extraction and attention fusion edge resource calculation method and device

Abstract

The invention provides a multi-dimensional feature extraction and attention fusion edge resource calculation method and device, and the method comprises the steps: obtaining current scheduling data and time sequence data under a current historical time window when a current scheduling moment arrives; performing data preprocessing on the current scheduling data and the time sequence data under the current historical time window to obtain normalized time sequence data and scheduling data tensor; inputting the normalized time sequence data and the scheduling data tensor into a resource scheduling model, so that the resource scheduling model outputs a resource state prediction result of each host at the next moment and a task scheduling scheme at the next moment according to the normalized time sequence data and the scheduling data tensor; and updating a resource scheduling strategy according to the resource state prediction result of each host at the next moment and the task scheduling scheme at the next moment. According to the scheme of the invention, efficient fault prediction and fault-tolerant scheduling optimization are realized, and an efficient and adaptive solution is provided for dynamic fault tolerance in an edge computing environment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of edge computing, in particular to a multi-dimensional feature extraction and attention fusion edge resource computing method and device. BACKGROUND

[0002] With the explosive growth of Internet of Things devices, a large amount of sensing data emerges, and it is not feasible to send all data to the cloud backend for processing. Edge computing technology emerges as the times require to meet the demand for local data processing. However, the device resources in the edge computing environment are limited, which brings challenges to the reliability of the service. When task-intensive arrives or burst tasks arrive, it often leads to resource contention and system overload, causing application performance degradation and failure. Therefore, task scheduling in the edge computing environment and edge computing resource reliability guarantee mechanism have become the current research focus.

[0003] In the field of edge computing reliability guarantee, existing research focuses on solving the problems of resource limitation and fault tolerance, including traditional fault tolerance methods, unsupervised learning applications, deep reinforcement learning exploration, and integration with other technologies, each with its own progress and limitations. Traditional methods are difficult to efficiently and accurately cope with fault tolerance requirements in complex and variable edge computing environments. Existing task scheduling algorithms have high energy consumption of edge computing center hosts and fewer successfully processed tasks when scheduling tasks.

[0004] Therefore, how to realize efficient resource scheduling and dynamic fault tolerance mechanism in the resource-limited edge computing environment to ensure the high reliability and stability of the system has become a problem to be solved. SUMMARY

[0005] The present application provides a multi-dimensional feature extraction and attention fusion edge resource computing method and device to solve the defect that efficient resource scheduling and dynamic fault tolerance mechanism cannot be realized in the resource-limited edge computing environment in the prior art.

[0006] The present application provides a multi-dimensional feature extraction and attention fusion edge resource computing method, comprising the following steps: When the current scheduling time arrives, the current scheduling data and the time series data under the current historical time window are obtained; wherein the time series data includes the resource state data of each host in the edge computing under the current historical time window, and the resource state of the host includes CPU utilization, memory utilization, hard disk utilization and network utilization; The current scheduling data and the time series data under the current historical time window are preprocessed to obtain normalized time series data and scheduling data tensors; inputting the normalized time series data and the scheduling data tensor into a resource scheduling model, so that the resource scheduling model outputs a resource state prediction result of each host at a next time and a task scheduling scheme at the next time according to the normalized time series data and the scheduling data tensor; The resource scheduling model is a model pre-constructed and trained, and comprises a graph data preprocessing module, a graph network processing module, an LSTM and self-attention module, a residual module, an encoder, a state decoder, and a prototype decoder. According to the resource state prediction result of each host at the next time and the task scheduling scheme at the next time, a resource scheduling strategy is updated.

[0007] According to the resource scheduling model, the normalized time series data and the scheduling data tensor are used to output a resource state prediction result of each host at a next time and a task scheduling scheme at the next time. The graph data preprocessing module constructs a graph structure based on the normalized time series data and the scheduling data tensor, and outputs graph structure data. The graph network processing module performs multi-dimensional feature extraction on the graph structure data through the graph convolution network and the graph attention network, and outputs structure-enhanced features. The LSTM and self-attention module comprises an LSTM layer and a self-attention mechanism, and performs time series feature extraction on the structure-enhanced features by processing time series dependence through the LSTM layer and capturing long-term dependence through the self-attention mechanism, and outputs time series-enhanced features. The residual module receives output results of the LSTM layer and the self-attention mechanism, performs residual splicing, and outputs residual-enhanced features. The encoder performs nonlinear transformation and dimension reduction on the residual-enhanced features and the scheduling data tensor, and outputs latent state representation. The state decoder outputs the resource state prediction result of each host at the next time based on the latent state representation through linear transformation and a nonlinear activation function. The prototype decoder outputs the task scheduling scheme at the next time based on the latent state representation through linear transformation and a nonlinear activation function.

[0008] According to the resource scheduling model, the normalized time series data and the scheduling data tensor are used to output a resource state prediction result of each host at a next time and a task scheduling scheme at the next time. extract resource state data of each host in edge computing in a current scheduling interval from the normalized time series data as node features; determine a connection relationship between hosts according to the scheduling data tensor, construct an adjacency matrix, and remove a self-loop in the adjacency matrix; combine the node features and the adjacency matrix after removing the self-loop into graph structure data, and output the graph structure data.

[0009] According to the multi-dimensional feature extraction and attention fusion edge resource computing method provided by the application, the graph network processing module performs multi-dimensional feature extraction on the graph structure data through the graph convolution network and the graph attention network, and outputs structure enhanced features, including: The graph convolution network aggregates features of neighbor nodes by using the adjacency matrix, learns network local structure information, and obtains first structure enhanced features; The graph attention network assigns different weights to neighbors of each node, enhances the perception ability of important neighbors, and obtains second structure enhanced features; The first structure enhanced features and the second structure enhanced features are fused to obtain the structure enhanced features, and the structure enhanced features are output.

[0010] According to the multi-dimensional feature extraction and attention fusion edge resource computing method provided by the application, the LSTM and self-attention module process time series dependence through an LSTM layer, and capture long-term dependence relationship through a self-attention mechanism, perform time series feature extraction on the structure enhanced features, and output time series enhanced features, including: The structure enhanced features are stacked in time sequence to form sequence data; The LSTM layer is applied to process the sequence data, capture the change trend of host resources over time, and obtain time series encoding representation; The time series encoding is processed through the self-attention mechanism, the weights between time steps are dynamically adjusted, important information dependence between time steps is captured, and the time series enhanced features are output.

[0011] According to the multi-dimensional feature extraction and attention fusion edge resource computing method provided by the application, the residual module performs nonlinear transformation on the time series enhanced features to obtain transformed features, and adds the time series enhanced features and the transformed features through residual connection to output residual enhanced features, including: The time series encoding representation and the time series enhanced features are obtained; The time series encoding representation and the time series enhanced features are residual spliced to output the residual enhanced features.

[0012] According to the multi-dimensional feature extraction and attention fusion edge resource calculation method provided by the application, the encoder performs nonlinear transformation and dimension reduction on the residual enhancement feature and the scheduling data tensor, and outputs a latent state representation, which comprises: The residual enhancement feature is input into a linear transformation layer to obtain a linearly transformed residual enhancement feature, and the linearly transformed residual enhancement feature is added to the residual enhancement feature to obtain a time series enhancement representation. The time series enhancement representation and the scheduling data tensor are flattened into one-dimensional vectors and spliced to obtain a spliced feature. The encoder performs nonlinear compression on the spliced feature to obtain a latent state representation. The state decoder outputs the resource state prediction result of each host at the next time based on the latent state representation through linear transformation and nonlinear activation function, specifically comprising: The state decoder outputs the resource state prediction result of each host at the next time based on the latent state representation through two layers of linear transformation and nonlinear activation function. The prototype decoder outputs the task scheduling scheme at the next time based on the latent state representation through linear transformation and nonlinear activation function, specifically comprising: The prototype decoder outputs the task scheduling scheme at the next time based on the latent state representation through a multilayer perceptron.

[0013] According to the multi-dimensional feature extraction and attention fusion edge resource calculation method provided by the application, the method further comprises: Collecting historical scheduling data and historical time series data; Preprocessing the historical scheduling data and the historical time series data to obtain normalized historical scheduling data and converted historical scheduling data tensors; Based on the normalized historical scheduling data and the converted historical scheduling data tensors, a training set is constructed; An initial resource scheduling model is constructed, and the initial resource scheduling model is trained based on the training set to obtain a trained resource scheduling model.

[0014] According to the multi-dimensional feature extraction and attention fusion edge resource calculation method provided by the application, the method further comprises: The current task scheduling scheme is optimized by a scheduling optimization algorithm to obtain an optimized task scheduling scheme.

[0015] The application also provides a multi-dimensional feature extraction and attention fusion edge resource calculation device, comprising the following modules: An acquisition module is configured to acquire scheduling data at a current scheduling moment and timing data in a current historical time window, wherein the timing data comprises resource state data of each host in edge computing in the current historical time window, and the resource state of the host comprises CPU utilization, memory utilization, hard disk utilization and network utilization; A preprocessing module is configured to perform data preprocessing on the scheduling data at the current moment and the timing data in the current historical time window to obtain normalized timing data and scheduling data tensors; A prediction module is configured to input the normalized timing data and the scheduling data tensors into a resource scheduling model, so that the resource scheduling model outputs resource state prediction results of each host at a next moment and a task scheduling scheme at the next moment according to the normalized timing data and the scheduling data tensors; The resource scheduling model is a model pre-constructed and trained, and the resource scheduling model comprises a graph data preprocessing module, a graph network processing module, an LSTM and self-attention module, a residual module, an encoder, a state decoder and a prototype decoder, and the graph network processing module comprises a graph convolution network and a graph attention network. An updating module is configured to update a resource scheduling strategy according to the resource state prediction results of each host at the next moment and the task scheduling scheme at the next moment.

[0016] The application provides a multi-dimensional feature extraction and attention fusion edge resource calculation method and device, which can ensure that the model can make scheduling decisions based on the latest resource state and task allocation, and improve the response speed and adaptability of the system by obtaining the current scheduling data and the time series data under the current historical time window when the current scheduling moment comes. The current scheduling data and time series data are preprocessed to obtain normalized time series data and scheduling data tensors to adapt to the processing requirements of the model, ensure that the data can be effectively processed by the model, and improve the processing efficiency and stability of the model. Further, the normalized time series data and scheduling data tensors are input into the resource scheduling model, and the model outputs the resource state prediction results of each host at the next moment and the task scheduling scheme at the next moment according to the data. The resource scheduling model includes a graph data preprocessing module, a graph network processing module (including a graph convolution network and a graph attention network), an LSTM and self-attention module, a residual module, an encoder, a state decoder and a prototype decoder. Through the graph convolution network (GCN) and the graph attention network (GAT), the model can extract the spatial dependency between hosts, and enhance the multi-dimensional feature representation of the host resource state. Using the LSTM and self-attention mechanism, the model can capture the trend of the host resource over time, and enhance the modeling capability of the time series data. The residual module and the encoder can alleviate the gradient disappearance problem in the deep model training, and improve the training efficiency and stability of the model. The state decoder and the prototype decoder can predict the resource state and generate the task scheduling scheme, and improve the fault tolerance of the system. Further, according to the resource state prediction results of each host at the next moment and the task scheduling scheme at the next moment output by the model, the resource scheduling strategy is updated. The task allocation can be dynamically adjusted according to the prediction results to ensure efficient execution of the task in a resource-limited environment. By predicting potential failures and adjusting the task allocation in advance, the model can reduce resource contention and system overload, and improve the fault tolerance of the system. In summary, the scheme of the application realizes efficient fault prediction and fault-tolerant scheduling optimization, and provides an efficient and adaptive solution for dynamic fault tolerance in the edge computing environment. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the application or prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the application, and other drawings can also be obtained by those skilled in the art without any creative effort.

[0018] Figure 1 is a flowchart of the multi-dimensional feature extraction and attention fusion edge resource calculation method provided by the application.

[0019] Figure 2is the architecture schematic diagram of the resource scheduling model provided by the application.

[0020] Figure 3 is the structure schematic diagram of the edge resource computing device with multi-dimensional feature extraction and attention fusion provided by the application. DETAILED DESCRIPTION

[0021] In order to make the objects, technical solutions and advantages of the present application clearer, the technical solutions in the present application will be described clearly and completely below in combination with the drawings in the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the present application.

[0022] The multi-dimensional feature extraction and attention fusion edge resource computing method of the present application will be described below in combination with Figures 1-2

[0023] Figure 1 is the flowchart of the multi-dimensional feature extraction and attention fusion edge resource computing method provided by the application, as Figure 1 shown, the method comprises the following steps: Step 101, when the current scheduling time comes, the current scheduling data and the time series data under the current historical time window are obtained; The time series data comprises the resource state data of each host in the edge computing under the current historical time window, and the resource state of the host comprises CPU utilization, memory utilization, hard disk utilization and network utilization. Step 102, the current scheduling data and the time series data under the current historical time window are preprocessed to obtain normalized time series data and scheduling data tensors; Step 103, the normalized time series data and scheduling data tensors are input into the resource scheduling model, so that the resource scheduling model outputs the resource state prediction result of each host at the next time and the task scheduling scheme at the next time according to the normalized time series data and scheduling data tensors; The resource scheduling model is a model pre-constructed and trained, and the resource scheduling model comprises a graph data preprocessing module, a graph network processing module, an LSTM and self-attention module, a residual module, an encoder, a state decoder and a prototype decoder. The graph network processing module comprises a graph convolution network and a graph attention network. Step 104, the resource scheduling strategy is updated according to the resource state prediction result of each host at the next time and the task scheduling scheme at the next time.

[0024] ​In practical applications, the execution subject of the multi-dimensional feature extraction and attention fusion edge resource calculation method can be a multi-dimensional feature extraction and attention fusion edge resource calculation device. The multi-dimensional feature extraction and attention fusion edge resource calculation device can be implemented in various ways, such as through a computer program, for example, application software, etc.; or, for example, a chip, etc. It can also be implemented as a medium that stores a related computer program, such as a U disk, a cloud disk, etc.; or, it can also be implemented through an entity device that integrates or installs a related computer program, such as a server, a smart device, etc.

[0025] For example, the multi-dimensional feature extraction and attention fusion edge resource calculation device described above can be a proxy node of an edge computing center. The host node set included in the edge computing center is Each host node has differences in CPU core number, memory capacity, disk space, and energy supply.

[0026] Suppose there are p Internet of Things terminals, which need to send part of the business to the proxy node of the edge computing center when processing the business. The Internet of Things terminals send task processing requests to the proxy node in the edge computing center. After receiving the task processing request, the proxy node of the edge computing center schedules the task according to the resource state of each host in the edge computing center, with the optimization goal of the largest number of task processing and the smallest energy consumption of the edge computing center host.

[0027] In practice, the system running time is a finite interval The time axis is divided into equal length scheduling intervals The proxy node of the edge computing center updates the resource scheduling strategy according to the latest host state in the scheduling interval. It can be understood that the scheduling interval is a fixed time interval between two adjacent scheduling times.

[0028] Suppose that host node failure is mainly caused by resource contention, and the failure probability of the host node is positively correlated with the resource utilization rate. Define the dynamic threshold function to represent the utilization rate threshold of the host at time for resource . When , it is marked as a resource contention failure. The threshold is dynamically adjusted according to historical load and node performance. For example, it can be set according to the utilization rate average of the past window. This paper assumes that the node power supply is stable, and the failure recovery only requires resource release or task migration, so power interruption and other non-recoverable failures can be ignored.

[0029] In terms of workload model, a bag-of-tasks workload model is adopted to simulate the edge computing scenario. At the beginning of each scheduling interval (when the scheduling time arrives), the IoT devices generate a batch of independent tasks that can be parallelized and send them to the proxy nodes in the edge computing in batches. The task attributes include computing resource requirements, SLO deadlines, etc. The tasks are abstracted as container instances to achieve resource isolation and rapid deployment. The SLO constraint is the time limit for the task from arriving at the proxy node in the edge computing to completing processing.

[0030] At the beginning of each scheduling interval , the proxy nodes in the edge computing center generate task scheduling decisions , where is the number of active tasks, represents the task assigned to the host . For new tasks, is the initial scheduling. For existing tasks, if , migration from the original host to the new host is triggered. In terms of state prediction, based on the historical state sequence , the time series data under the current historical time window is used to predict the next state . Wherein, contains the resource state data of each host in the edge computing.

[0031] Specifically, step 101 comprises: at the arrival of the current scheduling time, obtaining the current scheduling data and the time series data under the current historical time window.

[0032] Wherein, the current scheduling data includes the task scheduling decision matrix generated by the proxy node at the arrival of the current scheduling time.

[0033] Wherein, the time series data includes the resource state data of each host in the edge computing under the current historical time window, and the resource state of the host includes CPU utilization, memory utilization, hard disk utilization and network utilization.

[0034] In combination with the above description, the proxy nodes in the edge computing center update the resource scheduling strategy according to the latest host state within the scheduling interval. The scheduling interval is a fixed time interval between two adjacent scheduling times. In practice, at the arrival of the current scheduling time, the current scheduling data and the time series data under the current historical time window are obtained to update the resource scheduling strategy according to the current scheduling data and the time series data under the current historical time window.

[0035] It can be understood that by obtaining the current scheduling data and the timing data under the current historical time window when the current scheduling moment comes, it can be ensured that the model can make scheduling decisions based on the latest resource status and task allocation, thereby improving the response speed and adaptability of the system.

[0036] Further, the step 102 comprises: performing data preprocessing on the current scheduling data and the timing data under the current historical time window to obtain normalized timing data and scheduling data tensors.

[0037] In actual application, the current scheduling data and the timing data under the current historical time window are preprocessed and converted into a format that can be processed by the model, thereby enhancing the feature expression capability. For example, in terms of timing data, the timing data includes multiple time series, which are the sequences of CPU utilization, memory utilization, hard disk utilization and network utilization of each host. The format of the sequences of CPU utilization, memory utilization, hard disk utilization and network utilization of each host is [T, 3×N], where T is the time step and N is the number of hosts. In terms of scheduling data, the allocation relationship matrix of containers and hosts has a format of [T, N×N].

[0038] In terms of timing data standardization, the time series is cut into a fixed length window, and the first frame data is filled when the length is insufficient to ensure the consistency of the input dimension. Each feature dimension is normalized to .

[0039] wherein, represents a training data set, represents the maximum value of each feature dimension of the training data, represents a minimum value (such as 1e-8) to prevent division by zero.

[0040] In terms of scheduling data encoding, the scheduling decision matrix is converted into a tensor as a static feature of the model input. The scheduling data represents the task allocation relationship between hosts in the form of an adjacency matrix, denoted as , wherein N is the number of hosts.

[0041] In addition, in one example, the abnormal threshold calculation is also performed in the data preprocessing stage. Specifically, in terms of abnormal threshold calculation, the abnormal threshold of each feature dimension is dynamically generated based on the percentile of the training data, and the following formula is used to calculate: wherein, represents a training data set, is a percentile (such as 98), which represents taking a higher value of the data distribution as a reference. is an adjustment coefficient (such as 0.99999) for amplifying the threshold to reduce false positives.

[0042] It can be understood that the current scheduling data and timing data are preprocessed to obtain normalized timing data and scheduling data tensors to adapt to the processing requirements of the model, ensure that the data can be effectively processed by the model, and improve the processing efficiency and stability of the model.

[0043] Further, the step 103 comprises: inputting the normalized timing data and scheduling data tensors into the resource scheduling model, so that the resource scheduling model outputs the resource state prediction result of each host at the next moment and the task scheduling scheme at the next moment according to the normalized timing data and scheduling data tensors.

[0044] Figure 2 is an architecture schematic diagram of the resource scheduling model provided by the application, as shown in Figure 2 The resource scheduling model comprises a graph data preprocessing module, a graph network processing module, an LSTM and self-attention module, a residual module, an encoder, a state decoder and a prototype decoder, and the graph network processing module comprises a graph convolution network and a graph attention network.

[0045] As shown in Figure 2 The graph data preprocessing module comprises obtaining a node number, creating a node pair, removing a self-loop and generating a graph network. The module converts the normalized timing data and scheduling data tensors into a graph structure representation, preparing for subsequent graph convolution network and graph attention network processing. By creating a node pair to construct an adjacency relationship and removing a self-loop to ensure the rationality of the graph network.

[0046] The graph network processing module comprises a graph convolution network (GCN) and a graph attention network (GAT). The GCN network extracts node features through an adjacency matrix and learns local structure information of the network. The GAT network introduces an attention mechanism to assign differentiated weights to neighbor nodes and enhance the perception ability of important neighbors. This module can extract structured information between nodes and enhance the node feature representation ability using adjacency information.

[0047] The LSTM and self-attention module comprises an LSTM layer and a self-attention mechanism. The LSTM layer handles timing dependencies. The self-attention mechanism realizes dynamic weighting of key time steps and improves long-term dependency capture ability. This module realizes the function of jointly modeling the timing evolution characteristics of the graph structure.

[0048] The residual module comprises a residual connection and a residual mapping. The residual connection establishes a direct path from the input to the output. The residual mapping learns the difference between the input and the output. This module alleviates the gradient vanishing in deep model training, allows information to bypass the deep layer from the previous layer, and improves the learning efficiency.

[0049] The encoder module includes a flatten operation, a fully connected layer, and a LeakyReLU. The flatten operation converts the graph structure output into a vector. The fully connected layer and the LeakyReLU perform nonlinear transformation and dimension reduction on the graph convolution output. The module compresses the graph embedding into a latent state representation, retaining key information.

[0050] The state decoder module includes a linear layer, a LeakyReLU activation function, a linear layer, and a Sigmoid activation function. The module restores the predicted state output from the embedding state, which is used to predict the state of the next time window of the node.

[0051] The prototype decoder module includes a linear layer, a LeakyReLU activation function, a linear layer, a LeakyReLU activation function, a linear layer, and a Sigmoid activation function. The module is used to calculate the distance between the input and each prototype to determine its abnormality or belonging category.

[0052] Specifically, in one possible implementation, the resource scheduling model in the above step 103 outputs the resource state prediction result of each host at the next time and the task scheduling scheme at the next time according to the normalized time series data and the scheduling data tensor, including: Step 1031, a graph data preprocessing module constructs a graph structure based on the normalized time series data and the scheduling data tensor, and outputs graph structure data; Step 1032, a graph network processing module extracts multi-dimensional features from the graph structure data through a graph convolution network and a graph attention network, and outputs structure-enhanced features; Step 1033, the LSTM and self-attention module includes an LSTM layer and a self-attention mechanism. The LSTM and self-attention module processes time series dependence through the LSTM layer, captures long-term dependence through the self-attention mechanism, extracts time series features from the structure-enhanced features, and outputs time series-enhanced features; Step 1034, a residual module receives the output results of the LSTM layer and the self-attention mechanism, and performs residual concatenation to output residual-enhanced features; Step 1035, an encoder performs nonlinear transformation and dimension reduction on the residual-enhanced features and the scheduling data tensor, and outputs a latent state representation; Step 1036, a state decoder outputs the resource state prediction result of each host at the next time based on the latent state representation through linear transformation and nonlinear activation function; Step 1037, a prototype decoder outputs the task scheduling scheme at the next time based on the latent state representation through linear transformation and nonlinear activation function.

[0053] Optionally, in one example, the above step 1031 includes: extract resource state data of each host in edge computing in a current scheduling interval as node features from the normalized time series data; determine a connection relationship between hosts according to the scheduling data tensor, construct an adjacency matrix, and remove a self-loop in the adjacency matrix; combine the node features and the adjacency matrix after removing the self-loop into graph structure data, and output the graph structure data.

[0054] In actual application, the graph data preprocessing module aims to provide structural prior information, i.e., the connection topology structure between hosts, for the graph network processing module, and represent the connection topology structure as an adjacency matrix . indicates the connection relationship between nodes (without self-loop); is a degree matrix, recording the number of neighbors of each node; indicates that there is an edge connection between node and node , otherwise 0. This structure avoids redundant self-connection, which is conducive to enhancing the modeling ability of node interaction and improving the robustness of spatial dependence modeling. The expressions of the adjacency matrix and the degree matrix are as follows: Optionally, in an example, the above step 1032 includes: The graph convolution network aggregates the features of neighbor nodes using the adjacency matrix, learns the local structure information of the network, and obtains first structure-enhanced features; The graph attention network assigns different weights to the neighbors of each node, enhances the perception ability of important neighbors, and obtains second structure-enhanced features; fuse the first structure-enhanced features and the second structure-enhanced features to obtain structure-enhanced features, and output the structure-enhanced features.

[0055] In combination with the above description, the graph network processing module includes a graph convolution network (GCN) and a graph attention network (GAT).

[0056] In actual application, the graph convolution network aggregates neighbor node features, and the graph convolution network is composed of two layers of graph convolution layers stacked, which are updated according to the following formula.

[0057] The output of each layer as the input of the next layer . indicates the trainable weight matrix of the layer; is the input feature of the layer, indicates the first layer, and the input is the node feature ; indicates the second layer, and the input is the output of the previous layer ; is the normalized adjacency matrix . It can prevent the numerical instability caused by the large difference of node degrees during aggregation, and ensure that the neighbor information can affect the node features evenly. After two layers of graph convolution, the first structure-enhanced feature is obtained, which is used to fuse with the results of the GAT module.

[0058] Combining Figure 2 , the graph attention network is composed of two layers of graph attention layers, each of which adopts the following attention calculation mechanism.

[0059] where denotes two nodes in the graph. , i.e. is the neighbor of . denotes the neighbor node set of node . denotes the node feature, and the dimension is . denotes the weight matrix, which linearly transforms the feature of each node to a new space. denotes the feature concatenation operation. denotes the feature of node after linear mapping transformation. denotes the attention vector, which maps the concatenated feature to a scalar attention score. denotes the attention score, which represents the relevance of to . denotes the exponential function, which is used for softmax operation to make the attention weight positive and normalized. denotes the normalized attention weight, which represents the attention degree of node when aggregating the neighbor features.

[0060] The weighted aggregation of node features is shown in the following formula: where denotes the number of GAT layers, and the final output is used as the input of the time series modeling module. denotes the nonlinear activation function. GAT performs weighted summation on the features of neighbor nodes through the attention mechanism. The representation of node in the layer is obtained by linear transformation of its neighbor features and attention weight The weighted sum is activated to dynamically model the importance of different neighbors. The second structure-enhanced feature is output by two-layer attention computation for fusion with the GCN layer result.

[0061] Further, the first structure-enhanced feature and the second structure-enhanced feature are obtained at each time step to obtain a structure-enhanced feature, and the expression is as follows: Optionally, in an example, the above step 1033 includes: Stack the structure-enhanced features in chronological order to form sequence data; Apply an LSTM layer to process the sequence data to capture the trend of host resources over time, and obtain a time-series encoding representation; Process the time-series encoding through a self-attention mechanism to dynamically adjust the weights between time steps and capture important information dependencies across time steps, and output a time-series enhanced feature.

[0062] Specifically, to model the trend of host resources over time, the LSTM and self-attention module use LSTM (Long Short-Term Memory Network) to encode the structure embedding sequence extracted by GCN and GAT.

[0063] Stack the structure-enhanced features in chronological order to form sequence data, and the expression of the sequence data is as follows: Further, LSTM can effectively capture long-term dependencies in sequences, and the time-series feature of each host node is obtained at time steps of the graph structure representation , and the time-series features of all nodes are stacked into a sequence by time steps. To adapt to the input format of PyTorch LSTM, the batch dimension is expanded to make the input , and then it is sent to the LSTM layer to obtain the output of the LSTM layer: Further, the output is obtained by removing the batch dimension to obtain a time-series encoding representation : For example, the output of the last time step of the time-series encoding representation is taken as the time-series feature representation of the host : To further model the global dependency between time steps, the model introduces a self-attention mechanism. This mechanism dynamically adjusts the weight distribution of key dimensions by calculating the feature correlation between time steps. represents the time series feature representation after attention enhancement, which is used to capture important information dependencies across time steps. This module can dynamically identify and strengthen key time segments, improving the model's ability to perceive abnormalities or emergencies. The expression of the time series enhancement feature is as follows: Optionally, in an example, the above step 1034 includes: obtaining the time series encoding representation and the time series enhancement feature; concatenating the time series encoding representation and the time series enhancement feature in residual, and outputting the residual enhancement feature.

[0064] Specifically, to enhance the model's ability to model key events in time series, the residual module introduces a double residual connection structure based on the time series feature encoding. First, the time series encoding representation and the time series enhancement feature are obtained. Further, the time series encoding representation and the time series enhancement feature are concatenated in residual, and the residual enhancement feature is output: Optionally, in an example, the above step 1035 includes: inputting the residual enhancement feature into a linear transformation layer to obtain a linearly transformed residual enhancement feature, and adding the linearly transformed residual enhancement feature to the residual enhancement feature to obtain a time series enhancement representation; flattening the time series enhancement representation and the scheduling data tensor into one-dimensional vectors and concatenating them to obtain a concatenated feature; the encoder compresses the concatenated feature non-linearly to obtain a latent state representation; The above step 1036 specifically includes: The state decoder outputs the predicted resource state of each host at the next time based on the latent state representation through two linear transformations and a nonlinear activation function; The above step 1037 specifically includes: The prototype decoder outputs the task scheduling scheme at the next time based on the latent state representation through a multi-layer perceptron.

[0065] Specifically, after extracting the spatio-temporal features, the model uses an encoder-decoder structure to reconstruct the system state and compress it to a low-dimensional representation in the latent space for anomaly detection and reconstruction. The residual enhancement feature The input is added to itself and passed through a linear transformation layer to form the final temporal augmented representation: where, are the residual fully connected layer weights.

[0066] The temporal augmented representation is then flattened into a one-dimensional vector and concatenated with the scheduling data tensor to obtain the concatenated features: The encoder consists of two linear mappings with LeakyReLU activations to non-linearly compress the high-dimensional concatenated feature vector to obtain the latent representation . denote the fully connected layer weights; are the corresponding bias terms; denote the vector concatenation; is the final output latent state representation: Further, the state decoder module is used to restore the current system state (CPU, memory, disk usage) from the latent representation , i.e., to predict the normalized value of each indicator. The first layer is a non-linear transformation with LeakyReLU activation. The second layer is mapped and activated by Sigmoid to output the predicted state. denote the first linear layer parameters. denote the second linear layer parameters. Where, denote the normalized system state (CPU, memory, disk utilization). and The expressions are as follows: Further, to realize the discrimination of abnormal categories, the prototype decoder generates prototype vectors from the latent representation through a multi-layer perceptron (MLP). Figure 2 are the weights and biases of the first layer, and the relevant expressions are as follows: where, , , are the latent state representations of the three fully connected layers, respectively; , , are the weights of the three fully connected layers, respectively; , , are the bias terms corresponding to the three fully connected layers, respectively; and is an activation function, represents a Sigmoid activation function; is the latent state representation output by the encoder module.

[0067] It can be understood that the normalized time series data and the scheduling data tensor are input into the resource scheduling model, and the model outputs the resource state prediction results of each host at the next moment and the task scheduling scheme at the next moment according to the data. The resource scheduling model includes a graph data preprocessing module, a graph network processing module (including a graph convolution network and a graph attention network), an LSTM and a self-attention module, a residual module, an encoder, a state decoder and a prototype decoder. Through the graph convolution network (GCN) and the graph attention network (GAT), the model can extract the spatial dependency between hosts, enhance the multi-dimensional feature representation of the host resource state. Using the LSTM and the self-attention mechanism, the model can capture the trend of the host resource over time, and enhance the modeling ability of the time series data. The residual module and the encoder can alleviate the gradient vanishing problem in the training of the deep model, improve the training efficiency and stability of the model. The state decoder and the prototype decoder can predict the resource state and generate the task scheduling scheme, and improve the fault tolerance of the system.

[0068] Further, step 104 includes updating the resource scheduling strategy according to the resource state prediction results of each host at the next moment and the task scheduling scheme at the next moment.

[0069] It can be understood that the resource scheduling strategy is updated according to the resource state prediction results of each host at the next moment and the task scheduling scheme at the next moment output by the model. The task allocation can be dynamically adjusted according to the prediction results to ensure efficient execution of tasks in a resource-limited environment. By predicting potential failures and adjusting the task allocation in advance, the model can reduce resource contention and system overload, and improve the fault tolerance of the system In addition, in practice, it is necessary to pre-construct and train the resource scheduling model. Optionally, in a possible implementation, the method further includes: collecting historical scheduling data and historical time series data; preprocessing the historical scheduling data and the historical time series data to obtain normalized historical scheduling data and converted historical scheduling data tensors; construct a training set based on the normalized historical scheduling data and the converted historical scheduling data tensor; construct an initial resource scheduling model, and train the initial resource scheduling model based on the training set to obtain a trained resource scheduling model.

[0070] In this application, in order to improve the performance of the resource scheduling model, a model training algorithm is designed. The model training algorithm constructs a robust fault tolerance recovery mechanism by jointly optimizing the state prediction ability and the anomaly classification ability of the model. The model training algorithm mainly includes: spatio-temporal coding of historical state features and scheduling information, state and prototype decoding, loss function calculation and parameter optimization, dynamic prototype updating and adaptive adjustment of its update factor.

[0071] Specifically, the input of the model training algorithm is an initial resource scheduling model, historical time window host state features , a scheduling decision matrix , a category prototype set , a dynamic anomaly threshold ; the output of the model training algorithm is a trained resource scheduling model.

[0072] Exemplarily, the model training algorithm includes the following steps: Step 1, traverse the training set to obtain historical state input and the current scheduling matrix .

[0073] Step 2, input the historical state input and the current scheduling matrix into the current resource scheduling model.

[0074] Step 3, output the resource state predicted by the current resource scheduling model at the next time the embedding vector set of the current host the current prototype set .

[0075] Specifically, in the training process, the historical time series feature window (such as CPU, memory utilization, etc.) and the current scheduling decision matrix are input as input, pass through the spatio-temporal feature extraction module, output the next time resource state prediction value and the embedding representation set of each host . The embedding vector is matched with the prototype set , thereby constructing a loss function. Wherein represents a normal class prototype, represents an abnormal class prototype.

[0076] It can be understood that the space-time feature extraction module can extract spatial features and temporal features. In combination Figure 2 , the space-time feature extraction module includes a graph network processing module, an LSTM and self-attention module, and a residual module.

[0077] Step 4, calculate the mean square error between the predicted state and the real state, denoted as state reconstruction loss , the expression is as follows: In practice, the state reconstruction loss is used to measure the mean square error between the predicted state and the real state , and is used to measure the prediction accuracy. Wherein, is the number of hosts.

[0078] Step 5, initialize the triplet loss The cumulative value is zero.

[0079] Step 6, get each host : Step 61, judge whether its predicted state exceeds the threshold .

[0080] Step 62, according to the judgment result, the sample is regarded as “normal class” or “abnormal class”.

[0081] Step 63, select the corresponding prototype to construct a triplet: If it exceeds, it is regarded as an abnormal class, and the nearest abnormal prototype is selected as a positive class, and the normal prototype is selected as a negative class; Otherwise, it is regarded as a normal class, and the normal prototype is selected as a positive class, and the farthest abnormal prototype is selected as a negative class.

[0082] Specifically, the predicted state and the resource threshold are used for abnormal judgment to determine the category to which the sample belongs (step 61-step 62). The determination result is used to construct the positive class and negative class prototype selection strategy in the triplet loss. For each anchor sample , the positive class prototype corresponds to the prototype vector of the anchor class, and the negative class prototype is the prototype vector of other classes.

[0083] Specifically, judge whether its predicted state exceeds the threshold , according to the judgment result, the sample is regarded as “normal class” or “abnormal class”, which can be represented by the following formula: The triplet loss is defined as follows: For optimizing the structure of embedding space, the anchor feature vector is pushed close to the positive class prototype and far away from the negative class prototype (step 63). Wherein is the margin, which controls the minimum interval. is the positive class prototype, which represents the prototype vector of the class to which the anchor sample belongs. The smaller the positive class loss (the mean square error of the anchor and p⁺), the closer the distance between the anchor feature and the positive class prototype. is the negative class prototype, which represents the prototype vector of other classes. The larger the negative class loss, the farther the distance between the anchor feature and the negative class prototype.

[0084] Step 64, if the current triple satisfies the update condition (positive class loss ≤ all negative class minimum loss), update the class prototype, the expression is as follows: Specifically, when the positive class loss is the smallest, the corresponding class prototype is updated with a decay factor. Wherein is the update weight (PROTO_UPDATE_FACTOR), which is a decay factor, determining the influence weight of the new feature on the prototype when updating. represents the feature vector of the current sample. is the vector of the th class prototype at time step .

[0085] Step 7, calculate the total loss.

[0086] Specifically, the weighted average triple loss is calculated to enhance the abnormal class discrimination ability, and the expression of the weighted average triple loss is as follows: Wherein is the embedding feature vector of the th host.

[0087] Further, the total loss of the overall optimization target is constructed by combining the above state reconstruction loss and the triple loss . .

[0088] Wherein, is the balance hyperparameter, with a fixed value of 0.01, used to balance the proportion of reconstruction accuracy and prototype distribution optimization. The value is fixed during training and does not change with the number of rounds.

[0089] Step 8, if prototype update occurs in this round of training, perform exponential decay on the prototype update factor.

[0090] Specifically, the dynamic attenuation update factor : wherein, denotes the update factor used at the th update; denotes the update factor at the previous update; is a constant less than 1, used to control the gradual attenuation of .

[0091] Step 9, return the trained resource scheduling model.

[0092] In addition, the present application also sets up a scheduling optimization algorithm. The optimization process is independently executed in the inference stage, does not update the model parameters, and is only used for the correction and enhancement of the current scheduling scheme. The scheduling optimization algorithm takes the historical time window features and the current scheduling matrix as inputs, combines the model structure, resource threshold , maximum iteration number , continuous non-change round threshold , and executes the scheduling strategy adjustment process based on the differentiable model structure.

[0093] Optionally, in a possible implementation, the method further includes: optimizing the current task scheduling scheme by the scheduling optimization algorithm to obtain an optimized task scheduling scheme.

[0094] Specifically, the input of the task scheduling scheme is the current historical time window , the current scheduling matrix , the model structure model, the host resource usage threshold , the maximum iteration number , and the termination count threshold .

[0095] Illustratively, the scheduling optimization algorithm includes the following steps: Step 1, convert the current scheduling matrix into a trainable tensor , and set it as the current optimal scheme .

[0096] Step 2, initialize the AdamW optimizer and the cosine annealing learning rate scheduler.

[0097] Specifically, the input parameters are input and initialized through steps 1 and 2. The current scheduling matrix is converted into a trainable tensor , and the initial optimal scheduling scheme Initialize AdamW optimizer with cosine annealing learning rate scheduler, set the "unchanged epoch counter" to 0.

[0098] Step 3, enter the maximum number of iterations optimization loop: Step 30, concatenate the scheduling variables with the feature sequence input model.

[0099] Step 31, extract node spatial features through graph structure encoding.

[0100] Step 32, perform temporal modeling and attention mechanism enhanced representation.

[0101] Step 33, encoder output fusion representation; decoder generates predicted resource state and failure prototype P.

[0102] In practical applications, feature fusion of the model is performed through steps 30 to 33. Specifically, the and are concatenated as model inputs. In each iteration, the model processes the input data in the following steps. First, graph structure encoding. The host connection relationship is encoded through a graph convolutional network (GCN) and a graph attention network (GAT) to capture spatial dependencies between nodes. Second, temporal encoding. The windowed temporal data is processed through an LSTM layer to capture time dependencies, output hidden states, and enhance feature representation through residual connection and self-attention mechanism. Finally, encoding and decoding. The encoder maps the spatio-temporal features and scheduling data concatenated to the latent space, and the decoder generates the predicted resource state and failure prototype .

[0103] Step 34, calculate the optimization loss .

[0104] Step 35, if the predicted state is normal (i.e. ), terminate optimization immediately.

[0105] In practical applications, steps 34 and 35 are executed, and the optimization target and loss function. Specifically, the objective is to minimize the amount of abnormality in the predicted state that exceeds the threshold, and the optimization loss is constructed: where is the threshold for each dimension. If , it indicates that the current schedule has eliminated all abnormalities, and the iteration is terminated early and the optimal solution is returned.

[0106] Step 36, sequentially perform emptying gradient, back propagation, updating parameters, updating learning rate; Step 37, continuously schedule matrix projection (one-hot encoding): project the softmax output into one-hot format scheduling matrix; Step 38, compare the current scheduling scheme with the last round optimal scheme: If the current scheduling scheme is the same as the last round, increase the "unchanged round number" count; If the current scheduling scheme is different from the last round, reset the "unchanged round number" count to zero; Step 39, if the "unchanged round number" exceeds the set threshold , or the current state is completely normal, terminate early.

[0107] Specifically, by performing steps 36 to 39, gradient optimization and scheduling projection are performed. Specifically, in the back propagation and parameter updating step, gradient descent is performed using the AdamW optimizer, combined with the cosine annealing learning rate scheduler to dynamically adjust the learning rate, and through gradient clipping to ensure training stability (step 36). The initial learning rate is set to 0.8, and the cosine annealing learning rate scheduling period is set to . After each iteration, the scheduling matrix is projected back to the one-hot space to ensure that the final scheduling scheme is executable. Through the above optimization process, the final optimized scheduling scheme is obtained: In the one-hot encoding projection step, the optimized continuous value scheduling matrix is converted to a probability distribution through softmax, and then projected into one-hot encoding. Only the host allocation with the maximum probability is retained to ensure the discreteness and executability of the decision (step 37). In the convergence judgment step, if the continuous round scheduling matrix does not change ( ), it is considered that the optimization converges; otherwise, update and reset the count (steps 38 to 39) Step 4, return the current optimal scheduling scheme as the final optimization result.

[0108] The edge resource computation method with multidimensional feature extraction and attention fusion provided by this invention, by acquiring the current scheduling data and time-series data under the current historical time window when the current scheduling moment arrives, ensures that the model can make scheduling decisions based on the latest resource status and task allocation, thereby improving the system's response speed and adaptability. Preprocessing the current scheduling data and time-series data yields normalized time-series data and scheduling data tensors to suit the model's processing requirements, ensuring that the data can be effectively processed by the model, thus improving the model's processing efficiency and stability. Further, the normalized time-series data and scheduling data tensors are input into the resource scheduling model, which outputs the predicted resource status of each host at the next moment and the task scheduling scheme for the next moment based on this data. The resource scheduling model includes a graph data preprocessing module, a graph network processing module (including a graph convolutional network and a graph attention network), an LSTM and self-attention module, a residual module, an encoder, a state decoder, and a prototype decoder. Through the graph convolutional network (GCN) and graph attention network (GAT), the model can extract the spatial dependencies between hosts, enhancing the multidimensional feature representation of host resource status. By employing LSTM and self-attention mechanisms, the model can capture the changing trends of host resources over time, enhancing its ability to model time-series data. Residual modules and encoders alleviate the vanishing gradient problem in deep model training, improving training efficiency and stability. State decoders and prototype decoders predict resource states and generate task scheduling schemes, improving system fault tolerance. Furthermore, the resource scheduling strategy is updated based on the model's output of the next time step's predicted resource states of each host and the next time step's task scheduling scheme. Task allocation can be dynamically adjusted based on prediction results, ensuring efficient task execution in resource-constrained environments. By predicting potential faults and adjusting task allocation in advance, the model can reduce resource contention and system overload, improving system fault tolerance. In summary, the proposed solution achieves efficient fault prediction and fault-tolerant scheduling optimization, providing an efficient and adaptive solution for dynamic fault tolerance in edge computing environments.

[0109] The edge resource computing device for multidimensional feature extraction and attention fusion provided by the present invention is described below. The edge resource computing device for multidimensional feature extraction and attention fusion described below can be referred to in correspondence with the edge resource computing method for multidimensional feature extraction and attention fusion described above.

[0110] Figure 2 This is a schematic diagram of the edge resource computing device for multi-dimensional feature extraction and attention fusion provided by the present invention, as shown below. Figure 3 Figure 3 As shown, the edge resource computing device for multidimensional feature extraction and attention fusion includes: an acquisition module 41, a preprocessing module 42, a processing module 43, and an update module 44; The acquisition module 41 is configured to acquire scheduling data at a current scheduling moment and time sequence data in a current historical time window; the time sequence data comprises resource state data of each host in the edge computing in the current historical time window, and the resource state of the host comprises CPU utilization, memory utilization, hard disk utilization, and network utilization; The preprocessing module 42 is configured to perform data preprocessing on the scheduling data at the current scheduling moment and the time sequence data in the current historical time window, to obtain normalized time sequence data and scheduling data tensors; The prediction module 43 is configured to input the normalized time sequence data and the scheduling data tensors into a resource scheduling model, so that the resource scheduling model outputs resource state prediction results of each host at a next moment and a task scheduling scheme at the next moment according to the normalized time sequence data and the scheduling data tensors; The resource scheduling model is a model that is pre-constructed and trained, and the resource scheduling model comprises a graph data preprocessing module, a graph network processing module, an LSTM and self-attention module, a residual module, an encoder, a state decoder, and a prototype decoder; the graph network processing module comprises a graph convolution network and a graph attention network. The update module 44 is configured to update a resource scheduling strategy according to the resource state prediction results of each host at the next moment and the task scheduling scheme at the next moment.

[0111] Optionally, in a possible implementation, the resource scheduling model comprises a graph data preprocessing module, a graph network processing module, an LSTM and self-attention module, a residual module, an encoder, a state decoder, and a prototype decoder; the graph network processing module comprises a graph convolution network and a graph attention network. The prediction module 43 is specifically configured to: input the normalized time sequence data and the scheduling data tensors into the graph data preprocessing module; The graph data preprocessing module is configured to construct a graph structure based on the normalized time sequence data and the scheduling data tensors, and output graph structure data; The graph network processing module is configured to perform multi-dimensional feature extraction on the graph structure data through the graph convolution network and the graph attention network, and output structure-enhanced features; The LSTM and self-attention module comprises an LSTM layer and a self-attention mechanism; the LSTM and self-attention module is configured to process time sequence dependence through the LSTM layer, capture long-term dependence through the self-attention mechanism, perform time sequence feature extraction on the structure-enhanced features, and output time sequence-enhanced features; The residual module is configured to receive output results of the LSTM layer and the self-attention mechanism, and perform residual splicing, to output residual-enhanced features; an encoder configured to perform nonlinear transformation and dimensionality reduction on the residual enhancement feature and the scheduling data tensor, and output a latent state representation; a state decoder configured to output a predicted result of a resource state of each host at a next time based on the latent state representation, through a linear transformation and a nonlinear activation function; a prototype decoder configured to output a task scheduling scheme at the next time based on the latent state representation, through a linear transformation and a nonlinear activation function.

[0112] Optionally, in an example, the graph data preprocessing module is specifically configured to: extract resource state data of each host in edge computing within a current scheduling interval from the normalized time series data as node features; determine a connection relationship between the hosts according to the scheduling data tensor, construct an adjacency matrix, and remove a self-loop in the adjacency matrix; combine the node features and the adjacency matrix with the self-loop removed into graph structure data, and output the graph structure data.

[0113] Optionally, in an example, the graph network processing module is specifically configured to: a graph convolution network applies the adjacency matrix to aggregate features of neighbor nodes, learns local structure information of the network, and obtains first structure enhancement features; a graph attention network assigns different weights to neighbors of each node, enhances a perception ability for important neighbors, and obtains second structure enhancement features; fuses the first structure enhancement features and the second structure enhancement features to obtain structure enhancement features, and outputs the structure enhancement features.

[0114] Optionally, in an example, the LSTM and self-attention module is specifically configured to: stacks the structure enhancement features in a time sequence to form sequence data; applies an LSTM layer to process the sequence data, captures a change trend of the host resources over time, and obtains time series encoding representations; processes the time series encoding through a self-attention mechanism, dynamically adjusts weights between time steps, captures important information dependencies across time steps, and outputs time series enhancement features.

[0115] Optionally, in an example, the residual module is specifically configured to: obtains the time series encoding representations and the time series enhancement features; performs residual concatenation on the time series encoding representations and the time series enhancement features, and outputs residual enhancement features.

[0116] Optionally, in an example, the encoder is specifically configured to: The residual enhancement features are input into the linear transformation layer to obtain the linearly transformed residual enhancement features. The linearly transformed residual enhancement features are then added to the residual enhancement features to obtain the temporal enhancement representation. The temporal augmentation representation and scheduling data tensor are flattened into one-dimensional vectors and concatenated to obtain concatenated features; The encoder performs nonlinear compression on the spliced ​​features to obtain the latent state representation; The aforementioned state decoder is specifically used for: The state decoder is based on the latent state representation and outputs the resource state prediction results of each host in the next time step through two layers of linear transformation and nonlinear activation function. The aforementioned prototype decoder is specifically used for: The prototype decoder, based on the latent state representation, outputs the task scheduling scheme for the next time step through a multilayer perceptron.

[0117] Optionally, in one possible implementation, the aforementioned edge resource computing device for multidimensional feature extraction and attention fusion further includes: a training module; the training module is specifically used for: Collect historical scheduling data and historical time-series data; The historical scheduling data and historical time-series data are preprocessed to obtain normalized historical scheduling data and transformed historical scheduling data tensors. A training set is constructed based on normalized historical scheduling data and the transformed historical scheduling data tensor. An initial resource scheduling model is constructed, and the initial resource scheduling model is trained based on the training set to obtain the trained resource scheduling model.

[0118] Optionally, in one possible implementation, the aforementioned edge resource computing device for multidimensional feature extraction and attention fusion further includes: an optimization module; the optimization module is used to: The current task scheduling scheme is optimized by using a scheduling optimization algorithm to obtain an optimized task scheduling scheme.

[0119] In the edge resource computing device with multidimensional feature extraction and attention fusion provided by this invention, the acquisition module acquires the current scheduling data and time-series data under the current historical time window when the current scheduling moment arrives. This ensures that the model can make scheduling decisions based on the latest resource status and task allocation, improving the system's response speed and adaptability. The preprocessing module preprocesses the current scheduling data and time-series data to obtain normalized time-series data and scheduling data tensors to adapt to the model's processing requirements, ensuring that the data can be effectively processed by the model and improving the model's processing efficiency and stability. Furthermore, the prediction module inputs the normalized time-series data and scheduling data tensors into the resource scheduling model. Based on this data, the model outputs the resource status prediction results of each host at the next moment and the task scheduling scheme for the next moment. The resource scheduling model includes a graph data preprocessing module, a graph network processing module (including a graph convolutional network and a graph attention network), an LSTM and self-attention module, a residual module, an encoder, a state decoder, and a prototype decoder. Through the graph convolutional network (GCN) and graph attention network (GAT), the model can extract the spatial dependencies between hosts, enhancing the multidimensional feature representation of host resource status. By employing LSTM and self-attention mechanisms, the model can capture the changing trends of host resources over time, enhancing its ability to model time-series data. The residual module and encoder alleviate the vanishing gradient problem in deep model training, improving training efficiency and stability. The state decoder and prototype decoder predict resource states and generate task scheduling schemes, improving the system's fault tolerance. Furthermore, the update module updates the resource scheduling strategy based on the model's output predictions of each host's resource state at the next time step and the task scheduling scheme at the next time step. It can dynamically adjust task allocation based on prediction results, ensuring efficient task execution in resource-constrained environments. By predicting potential faults and adjusting task allocation in advance, the model can reduce resource contention and system overload, improving the system's fault tolerance. In summary, the proposed solution achieves efficient fault prediction and fault-tolerant scheduling optimization, providing an efficient and adaptive solution for dynamic fault tolerance in edge computing environments.

[0120] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-dimensional feature extraction and attention fusion edge resource calculation method, characterized in that, The method comprises: when the current scheduling moment arrives, obtaining current scheduling data and time series data under a current historical time window; wherein the time series data comprises resource state data of each host in edge computing under the current historical time window, and the resource state of the host comprises CPU utilization, memory utilization, hard disk utilization and network utilization; performing data preprocessing on the current scheduling data and the time series data under the current historical time window to obtain normalized time series data and scheduling data tensors; inputting the normalized time series data and the scheduling data tensors into a resource scheduling model, so that the resource scheduling model outputs resource state prediction results of each host at the next moment and a task scheduling scheme at the next moment according to the normalized time series data and the scheduling data tensors; wherein the resource scheduling model is a model pre-constructed and trained, and the resource scheduling model comprises a graph data preprocessing module, a graph network processing module, an LSTM and self-attention module, a residual module, an encoder, a state decoder and a prototype decoder, and the graph network processing module comprises a graph convolution network and a graph attention network; updating a resource scheduling strategy according to the resource state prediction results of each host at the next moment and the task scheduling scheme at the next moment.

2. The multi-dimensional feature extraction and attention fusion edge resource computing method according to claim 1, characterized in that, The resource scheduling model outputs resource state prediction results of each host at the next moment and a task scheduling scheme at the next moment according to the normalized time series data and the scheduling data tensors, which comprises: the graph data preprocessing module constructs a graph structure based on the normalized time series data and the scheduling data tensors, and outputs graph structure data; the graph network processing module extracts multi-dimensional features from the graph structure data through the graph convolution network and the graph attention network, and outputs structure-enhanced features; the LSTM and self-attention module comprises an LSTM layer and a self-attention mechanism, and performs time series feature extraction on the structure-enhanced features by processing time series dependence through the LSTM layer and capturing long-term dependence through the self-attention mechanism, and outputs time series-enhanced features; the residual module receives output results of the LSTM layer and the self-attention mechanism, performs residual splicing, and outputs residual-enhanced features; the encoder performs nonlinear transformation and dimension reduction on the residual-enhanced features and the scheduling data tensors, and outputs latent state representation; the state decoder outputs the resource state prediction results of each host at the next moment based on the latent state representation through linear transformation and a nonlinear activation function; the prototype decoder outputs the task scheduling scheme at the next moment based on the latent state representation through linear transformation and a nonlinear activation function.

3. The multi-dimensional feature extraction and attention fusion edge resource computing method according to claim 2, characterized in that, The graph data preprocessing module constructs a graph structure based on the normalized time series data and the scheduling data tensors, and outputs graph structure data, which comprises: extracting resource state data of each host in edge computing within a current scheduling interval as node features from the normalized time series data; determine a connection relationship between hosts according to the scheduling data tensor, construct an adjacency matrix, and remove a self-loop in the adjacency matrix; combine the node features and the adjacency matrix after removing the self-loop into graph structure data, and output the graph structure data.

4. The multi-dimensional feature extraction and attention fusion edge resource computing method according to claim 3, characterized in that, The graph network processing module performs multi-dimensional feature extraction on the graph structure data through the graph convolution network and the graph attention network, and outputs structure enhancement features, including: The graph convolution network aggregates features of neighbor nodes using the adjacency matrix, learns local structure information of the network, and obtains first structure enhancement features; The graph attention network assigns different weights to neighbors of each node, enhances the perception ability of important neighbors, and obtains second structure enhancement features; The first structure enhancement features and the second structure enhancement features are fused to obtain the structure enhancement features, and the structure enhancement features are output.

5. The multi-dimensional feature extraction and attention fusion edge resource computing method according to claim 2, characterized in that, The LSTM and self-attention module process temporal dependencies through an LSTM layer and capture long-term dependencies through a self-attention mechanism, and perform temporal feature extraction on the structure enhancement features to output temporal enhancement features, including: The structure enhancement features are stacked in chronological order to form sequence data; An LSTM layer is applied to process the sequence data to capture the trend of host resources over time, and obtain temporal encoding representations; The temporal encoding is processed through a self-attention mechanism to dynamically adjust the weights between time steps and capture important information dependencies across time steps, and output temporal enhancement features.

6. The multi-dimensional feature extraction and attention fusion edge resource computing method according to claim 5, characterized in that, The residual module performs nonlinear transformation on the temporal enhancement features to obtain transformed features, and adds the temporal enhancement features and the transformed features through a residual connection to output residual enhancement features, including: The temporal encoding representations and the temporal enhancement features are obtained; The temporal encoding representations and the temporal enhancement features are spliced through a residual connection to output the residual enhancement features.

7. The multi-dimensional feature extraction and attention fusion edge resource computing method according to claim 2, characterized in that, The encoder performs nonlinear transformation and dimensionality reduction on the residual enhancement features and the scheduling data tensor to output latent state representations, including: The residual enhancement features are input into a linear transformation layer to obtain linearly transformed residual enhancement features, and the linearly transformed residual enhancement features and the residual enhancement features are added to obtain temporal enhancement representations; The temporal enhancement representations and the scheduling data tensor are flattened into one-dimensional vectors and spliced to obtain spliced features; The encoder performs nonlinear compression on the spliced features to obtain latent state representations; The state decoder outputs the predicted results of the resource states of the hosts at the next time based on the latent state representations through linear transformation and nonlinear activation functions, specifically including: The state decoder outputs the predicted results of the resource states of the hosts at the next time based on the latent state representations through two layers of linear transformation and nonlinear activation functions; The prototype decoder outputs the task scheduling scheme at the next time based on the latent state representations through linear transformation and nonlinear activation functions, specifically including: The prototype decoder outputs the task scheduling scheme at the next time based on the latent state representations through a multilayer perceptron.

8. The multi-dimensional feature extraction and attention fusion edge resource computing method according to any one of claims 1-7, characterized in that, The method further includes: Collect historical scheduling data and historical timing data; Preprocess the historical scheduling data and the historical timing data to obtain normalized historical scheduling data and converted historical scheduling data tensors; Based on the normalized historical scheduling data and the converted historical scheduling data tensors, construct a training set; Construct an initial resource scheduling model and train the initial resource scheduling model based on the training set to obtain a trained resource scheduling model.

9. The multi-dimensional feature extraction and attention fusion edge resource computing method according to any one of claims 1-7, characterized in that, The method further comprises: Optimize the current task scheduling scheme by a scheduling optimization algorithm to obtain an optimized task scheduling scheme.

10. An edge resource computing device for multi-dimensional feature extraction and attention fusion, comprising: The device comprises: An acquisition module configured to, when a current scheduling time arrives, acquire current scheduling data and timing data under a current historical time window; wherein the timing data comprises resource state data of each host in edge computing under the current historical time window, and the resource state of the host comprises CPU utilization, memory utilization, hard disk utilization, and network utilization; A preprocessing module configured to preprocess the current scheduling data and the timing data under the current historical time window to obtain normalized timing data and scheduling data tensors; A prediction module configured to input the normalized timing data and the scheduling data tensors into a resource scheduling model, so that the resource scheduling model outputs resource state prediction results of each host at a next time and a task scheduling scheme at the next time according to the normalized timing data and the scheduling data tensors; The resource scheduling model is a model constructed and trained in advance, and comprises a graph data preprocessing module, a graph network processing module, an LSTM and self-attention module, a residual module, an encoder, a state decoder, and a prototype decoder; the graph network processing module comprises a graph convolution network and a graph attention network; An updating module configured to update a resource scheduling strategy according to the resource state prediction results of each host at the next time and the task scheduling scheme at the next time.

Images