Artificial Intelligence Task Scheduling System and Method for Edge Computing

Abstract

This application relates to the field of computer science and discloses an artificial intelligence task scheduling system and method for edge computing. The system includes modules for semantic manifold kernel extraction, heterogeneous resource potential field construction, gradient tensor projection reconstruction, and streaming gradient optimization scheduling. The method utilizes a potential energy function to map heterogeneous network resources into a scalar potential field and calculates negative gradient vectors; it generates a semantic manifold kernel mask based on the semantic entropy density of feature channels, and combines it with an environmental pressure tensor mapped by the negative gradient magnitude to project and reconstruct the model weights to generate a dynamic inference model; it synthesizes a scheduling vector using the negative gradient vector and Lyapunov drift correction to determine the next-hop target node. This invention achieves adaptive model compression and network load balancing while preserving key semantic features through potential field gradient collaborative optimization of the model structure and transmission path, solving the problem of balancing transmission efficiency and inference accuracy in edge computing scenarios.

Description

Technical Field

[0001] This invention relates to the field of computer technology, specifically to an artificial intelligence task scheduling system and method for edge computing. Background Technology

[0002] With the rapid development of IoT technology and deep learning algorithms, deploying artificial intelligence applications at the edge has become a trend. Edge computing, by moving computing tasks from the cloud to the network edge, can effectively reduce data transmission latency and alleviate core network bandwidth pressure. However, edge network environments are typically composed of heterogeneous devices with varying computing power, energy supply, and communication bandwidth, and deep neural network models often have a large number of parameters and computational complexity. This poses a significant challenge to efficient task scheduling and model transmission in resource-constrained edge environments.

[0003] Current edge intelligence task scheduling technologies suffer from the following limitations: First, existing solutions typically treat AI models as fixed-structure data blocks for transmission or scheduling. Traditional model compression techniques are mostly completed statically before deployment, lacking the ability to dynamically adjust based on real-time network pressure and the semantic features of input data. This static approach prevents the flexible removal of redundant parameters from the model to reduce transmission overhead when network bandwidth is limited or computing resources are scarce, or the inability to effectively identify and retain key semantic features during compression. This results in a significant decrease in inference accuracy, making it difficult to achieve a balance between transmission efficiency and task performance.

[0004] Secondly, edge networks have complex topologies and highly dynamic node states, involving multi-dimensional physical attributes such as computing power, remaining energy, and chip temperature. Existing scheduling algorithms often make decisions based on single indicators or static rules, or use computationally complex global optimization algorithms, making it difficult to reflect the distribution trends of multi-dimensional resources in the network in real time and intuitively. This limitation makes it easy for tasks to be scheduled to nodes that seem to have sufficient computing power but may be about to fail due to high temperature throttling or insufficient power, failing to form an effective resource optimization orientation and resulting in uneven overall network resource utilization.

[0005] Furthermore, existing greedy scheduling strategies tend to send tasks to the nodes with the most abundant resources, ignoring the backlog in the task queues of those nodes. In multi-task concurrent scenarios, this strategy can easily lead to localized congestion due to excessive concentration of traffic to specific high-performance nodes, causing a surge in queuing latency. The lack of a long-term queue stability control mechanism makes it difficult for the scheduling system to guarantee the continuity of task processing and low latency requirements when facing sudden traffic surges. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides an artificial intelligence task scheduling system and method for edge computing, which solves the problem that existing technologies struggle to balance transmission efficiency and task performance.

[0007] To achieve the above objectives, the present invention provides the following technical solution: The first aspect of this invention provides an artificial intelligence task scheduling system for edge computing, deployed in a heterogeneous computing network environment including edge terminal devices, edge gateway nodes, and cloud servers. This system maps the heterogeneous computing network environment as a scalar potential field, guiding task flow and dynamically adjusting model structure by calculating the gradient of the potential field. The system mainly includes a semantic manifold kernel extraction module, a heterogeneous resource potential field construction module, a gradient tensor projection reconstruction module, and a streaming gradient optimization scheduling module.

[0008] The semantic manifold kernel extraction module is configured to receive the data stream of the AI ​​task to be processed and obtain intermediate layer feature maps through convolution operations. This module calculates the semantic entropy density of each channel in the intermediate layer feature map, which reflects the information carrying capacity of the feature channel. Based on the semantic entropy density, the module generates a semantic manifold kernel mask, which is used to mark the positions of channels carrying key semantic information in the model.

[0009] The heterogeneous resource potential field construction module is configured to collect resource status data of each node in the network, including remaining computing power, available communication bandwidth, remaining energy, and chip temperature. This module uses a potential energy function to construct a scalar potential energy field covering the network topology space, where nodes with high load, high temperature, or low energy correspond to high potential energy regions, and vice versa. The module calculates the negative gradient vector of the scalar potential energy field at the current spatial coordinates. The direction of this vector indicates the transmission direction for resource optimization, and its magnitude characterizes the resource pressure intensity of the current environment on the task.

[0010] The gradient tensor projection reconstruction module is configured to receive negative gradient vectors and semantic manifold kernel masks. This module establishes a mapping between physical environment pressure and neural network weights, converting the magnitude of the negative gradient vector into an environment pressure tensor and generating an environment-tolerant mask accordingly. The module performs a union operation on the semantic manifold kernel mask and the environment-tolerant mask to generate a retained mask, which is then used to project onto the original AI model weight tensor. This operation removes redundant weights suppressed by environment pressure while forcibly retaining the key weights marked by the semantic manifold kernel mask, generating dynamic inference model weights adapted to the current network environment.

[0011] The streaming gradient optimization scheduling module is configured to receive negative gradient vectors and calculate Lyapunov drift corrections based on the network queue backlog status to balance resource optimization and queue stability. This module synthesizes a scheduling vector based on the negative gradient vector and the Lyapunov drift correction, determines the next-hop target node for task transmission, and sends the dynamic inference model weights to that target node.

[0012] A second aspect of the present invention provides an artificial intelligence task scheduling method for edge computing, which achieves edge intelligent task scheduling by collaboratively optimizing model structure and transmission path.

[0013] This method first receives the AI ​​task data to be processed, and extracts features from the task data to obtain intermediate layer feature maps. Based on the activation distribution of the intermediate layer feature maps, the channel-level semantic entropy density is calculated to quantify the feature importance of each channel. By comparing the semantic entropy density with a dynamic statistical threshold, a semantic manifold kernel mask identifying key feature channels is generated.

[0014] This method collects resource status data of each node in an edge computing network, including computing power, bandwidth, energy, and temperature. It uses a potential energy function to map the discrete resource status of network nodes into a continuous scalar potential energy field. The negative gradient vector of the potential energy field at the current location of the task to be scheduled is calculated, thereby obtaining the guiding direction pointing to resource-rich areas and the gradient magnitude quantifying the resource scarcity at the current location.

[0015] This method further maps the negative gradient vector of the potential field to an environmental pressure tensor and generates a corresponding environmental tolerance mask. The environmental tolerance mask is then logically joined with the semantic manifold kernel mask to generate the final preserved mask. This preserved mask is used to project and reconstruct the weight tensor of the AI ​​model, eliminating parameters with low semantic importance and suppressed by environmental pressure, thus generating a sparse dynamic inference model.

[0016] This method plans transmission paths based on a gradient descent strategy with Lyapunov drift correction. A comprehensive scheduling vector is calculated based on the resource optimization direction indicated by the negative gradient vector of the potential field, combined with the drift correction derived from the network queue backlog state. The projection of the comprehensive scheduling vector onto the link directions of each neighbor node is calculated, and the neighbor node with the largest projection value is selected as the next-hop target node, to which the dynamic inference model is transmitted. When the current node is determined to be a local potential minimum point, the model inference task is executed locally.

[0017] This invention provides an artificial intelligence task scheduling system and method for edge computing. It has the following beneficial effects: 1. This invention achieves adaptive dynamic adjustment of the model structure to the network environment through a joint projection mechanism of semantic manifold kernel mask and environment-tolerant mask. The system uses channel-level semantic entropy density to identify and lock the weight parameters carrying key information, ensuring that they are not compressed during transmission; at the same time, it maps the resource pressure of the network environment to gradient magnitude, and accordingly performs sparsity pruning on non-critical weights to different degrees. This mechanism effectively reduces the model's transmission bandwidth consumption and energy consumption while maximizing the preservation of the model's feature extraction capabilities, solving the problem of balancing transmission efficiency and inference accuracy in edge computing scenarios.

[0018] 2. This invention constructs a multi-dimensional resource potential field encompassing computing power, bandwidth, energy, and temperature, mapping the discrete heterogeneous network node states into a continuous scalar field space. By calculating the negative gradient vector of the potential field function, the system can directly obtain the optimal transmission direction pointing to resource-rich regions. Compared to traditional rule-based or heuristic scheduling algorithms, the field-theory gradient-based optimization method has lower computational complexity and can guide task flows to automatically avoid congested nodes with high load, high temperature, or low energy, significantly improving the resource utilization balance in heterogeneous edge network environments.

[0019] 3. This invention introduces a Lyapunov drift correction strategy during the gradient optimization process, effectively ensuring the long-term stability of the network queue. The scheduling module does not solely rely on the guidance of physical resource gradients, but instead uses the task queue backlog state of network nodes as a correction term to synthesize the final scheduling vector. This drift penalty control avoids excessive concentration of tasks on resource-rich but heavily queued local nodes, preventing network congestion and thus achieving an optimized balance between throughput and end-to-end latency in dynamically changing edge network environments. Attached Figure Description

[0020] Figure 1 This is a structural diagram of the present invention; Figure 2 This is a flowchart of the present invention.

[0021] Among them, 110 is the semantic manifold kernel extraction module; 120 is the heterogeneous resource potential field construction module; 130 is the gradient tensor projection reconstruction module; and 140 is the streaming gradient optimization scheduling module. Detailed Implementation

[0022] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Example: Please see the appendix Figure 1 This invention provides an AI task scheduling system for edge computing. This system is deployed in a heterogeneous computing network environment comprising edge terminal devices, edge gateway nodes, and cloud servers. The edge terminal devices include, but are not limited to, surveillance cameras, industrial sensors, or mobile terminals; the edge gateway nodes possess computing and forwarding capabilities; and the cloud servers possess high computing power resources. The system is configured to map the heterogeneous computing network environment as a virtual potential field on a Riemannian manifold space and to process the computational tasks to be processed as fluid entities with variable structural characteristics.

[0024] The AI ​​task scheduling system for edge computing mainly includes: a semantic manifold kernel extraction module 110, a heterogeneous resource potential field construction module 120, a gradient tensor projection reconstruction module 130, and a streaming gradient optimization scheduling module 140. The modules are connected via data communication interfaces or a system bus to achieve data interaction and control command transmission.

[0025] The semantic manifold kernel extraction module 110 is configured to receive the data stream of the artificial intelligence task to be processed. This module integrates a feature extraction unit for performing convolution operations on the input data to obtain intermediate layer feature maps. The semantic manifold kernel extraction module 110 is also configured to calculate the activation probability distribution and semantic entropy density of each channel in the intermediate layer feature map, and generate a semantic manifold kernel mask based on the semantic entropy density. This semantic manifold kernel mask is used to identify rigid subspaces in the weight parameters of the artificial intelligence model corresponding to high-information features; these subspaces maintain structural integrity in subsequent processing.

[0026] The heterogeneous resource potential field construction module 120 is configured to communicate with each computing node in the network environment and collect resource status data of each node in real time. The resource status data includes the node's remaining computing power, available communication bandwidth, remaining energy, and current chip temperature. The heterogeneous resource potential field construction module 120 is configured to construct a scalar potential energy field covering the network topology space using a preset potential energy function based on the resource status data, and calculate the negative gradient vector of this potential energy field at arbitrary spatial coordinates. The negative gradient vector contains direction information and magnitude information, representing the transmission direction for resource optimization and the pressure intensity of the environment on the task, respectively.

[0027] The gradient tensor projection reconstruction module 130 connects the semantic manifold kernel extraction module 110 and the heterogeneous resource potential field construction module 120. The gradient tensor projection reconstruction module 130 is configured to receive the negative gradient vector and the semantic manifold kernel mask. This module uses a nonlinear mapping algorithm to convert the magnitude of the negative gradient vector into an environmental pressure tensor consistent with the weight dimension of the artificial intelligence model, and generates an environmental tolerance mask accordingly. The gradient tensor projection reconstruction module 130 is further configured to perform a union operation on the semantic manifold kernel mask and the environmental tolerance mask to generate a final retained mask. This module uses the retained mask to project the original artificial intelligence model weight tensor, removing redundant weight parameters and generating dynamic inference model weights.

[0028] The streaming gradient optimization scheduling module 140 connects the heterogeneous resource potential field construction module 120 and the gradient tensor projection reconstruction module 130. The streaming gradient optimization scheduling module 140 is configured to receive the negative gradient vector and calculate the Lyapunov drift correction based on the network queue backlog state. This module is configured to solve the motion equation of the task in the network space based on the corrected gradient direction, determine the next-hop target node for task transmission, and send the dynamic inference model weights to the target node.

[0029] Each module in the AI ​​task scheduling system for edge computing can be implemented by computer program instructions running on a computing device, or by hardware circuits such as application-specific integrated circuits (ASICs) and field-programmable gate arrays (FPGAs).

[0030] See attached document Figure 2 , Figure 2 This is a schematic diagram illustrating the workflow of an AI task scheduling method for edge computing according to an embodiment of the present invention. The method mainly includes four stages: task semantic analysis, resource potential field construction, dynamic model reconstruction, and gradient flow scheduling.

[0031] Step S100: Receive the artificial intelligence task data to be processed, extract features from the task data to obtain an intermediate layer feature map, calculate the channel-level semantic entropy density based on the intermediate layer feature map, and then generate a semantic manifold kernel mask.

[0032] In step S100, for the video stream or image data collected by the edge terminal, the system uses a preset lightweight convolutional neural network to obtain the first... Feature map tensors of layers For the first feature map For each channel, calculate the normalized probability distribution of its pixel activation values. Based on this distribution, calculate the semantic entropy density of that channel. The semantic entropy density The numerical value is proportional to the amount of information contained in that feature channel.

[0033] The system calculates the statistical distribution of semantic entropy density for all channels in the current layer and constructs a dynamic filtering threshold. The calculated semantic entropy density With this screening threshold A comparison is performed. When the semantic entropy density of a channel is greater than the filtering threshold, the model weight corresponding to that channel is determined to belong to the semantic manifold kernel feature. A semantic manifold kernel mask is then generated based on this. The mask is a binary vector, with the channel index positions that meet the conditions marked as 1 and the remaining positions marked as 0.

[0034] Step S200: Collect resource status data of each node in the edge computing network, map the network environment into a continuous scalar potential energy field, and calculate the negative gradient vector of the potential energy field at the current position of the task to be scheduled.

[0035] In step S200, the system acquires the nodes in the network topology. state vector The state vector This includes the normalized remaining computing power factor, remaining energy factor, available bandwidth factor, and temperature factor.

[0036] Based on the state vectors of each node, a potential energy function covering the network space is constructed. This potential energy function defines any position in space. exist The resource value at any given moment. The construction principle is: the larger the normalized remaining computing power and remaining energy values ​​of a node, the smaller the potential energy function value corresponding to that location; the larger the normalized temperature and task queue backlog values ​​of a node, the larger the potential energy function value corresponding to that location.

[0037] The system has a certain potential energy function. Perform spatial differentiation to calculate the negative gradient vector. :

[0038] The negative gradient vector The direction indicates the direction of the fastest decrease in potential energy, and its magnitude... The numerical value corresponds to the intensity of environmental resource pressure at the current spatial location.

[0039] Step S300: Map the negative gradient vector of the potential energy field to the environmental pressure tensor, and combine it with the semantic manifold kernel mask to project and reconstruct the weight tensor of the artificial intelligence model to generate a dynamic inference model.

[0040] In step S300, the system performs a mapping from the potential field parameters to the neural network weight parameters. First, a Sigmoid-type nonlinear function is used. The magnitude of the negative gradient vector Converted to scalar stress values, and expanded to be compatible with the model's first scalar stress value via tensor broadcasting. Environmental pressure tensor with consistent layer weight dimensions .

[0041] The system calculates the weights of the original model. The absolute value of the environmental tolerance mask is calculated and compared with the environmental pressure tensor to generate an environmental tolerance mask. When the absolute value of a weight is lower than the value at the corresponding position in the environmental pressure tensor, that position is marked as 0 in the environmental tolerance mask, indicating that the weight parameter is pruned; otherwise, it is marked as 1.

[0042] The system uses the semantic manifold kernel mask generated in step S100. With the aforementioned environment tolerance mask Perform a logical OR operation to generate the final reserved mask. :

[0043] Finally, the original weights are multiplied element-wise using the final preserved mask to obtain the reconstructed dynamic inference model weights. This step ensures that the weight parameters marked as 1 by the semantic manifold kernel mask are always preserved regardless of the environmental pressure tensor.

[0044] Step S400: Based on the gradient descent algorithm with Lyapunov drift correction, calculate the transmission path of the task and schedule the reconstructed dynamic inference model to the next hop node.

[0045] In step S400, the system introduces the Lyapunov optimization method to maintain the stability of the network queue. The Lyapunov function is defined. It is half the sum of the squares of the task queue lengths of all nodes in the network. Calculate the Lyapunov drift term, which is in the negative direction of the queue length gradient.

[0046] The system calculates the final scheduling vector. This vector is formed by the negative gradient vector of the potential field. The system performs a weighted synthesis with the Lyapunov drift correction term. It iterates through all neighboring nodes of the current node, calculating the difference vector between the position vector of each neighboring node in the network topology space and the current position vector. This difference vector is then compared with the scheduling vector. The system calculates the cosine similarity and selects the neighbor node with the highest cosine similarity value as the next-hop target node. After determining the target node, the system uses the dynamic reasoning model generated in step S300. Send to the target node.

[0047] To more clearly illustrate the technical solution of the present invention, the above step S100 will be described in detail below: Step S101: Input the task data to be processed into the preset feature extraction network to obtain the intermediate feature map tensor of the target neural network layer.

[0048] In step S101, the feature extraction network can employ a pre-trained lightweight convolutional neural network structure (such as MobileNetV3 or ResNet-18). The system receives raw input data. Through the operations of convolutional layers, normalization layers, and activation functions, the output is the first... Feature map tensors of layers ,in The number of feature channels, For feature map height, The width of the feature map. Each element in this feature map tensor Representing the The spatial location of each channel The activation response value at that location.

[0049] Step S102: Perform spatial dimension probability normalization processing on each feature channel in the intermediate feature map tensor.

[0050] In step S102, in order to measure the information distribution characteristics of different channels at a uniform scale, the system performs probability mapping on the activation values ​​within each feature channel. For the first... Each feature channel is used to calculate its spatial location. Normalized probability at point The calculation uses the spatial softmax function:

[0051] The above calculations guarantee that the sum of the probabilities of all spatial locations within the channel is 1. This probability distribution reflects the... The degree of spatial concentration or dispersion of the response of each feature channel to the input data.

[0052] Step S103: Based on the normalized probability, calculate the semantic entropy density of each feature channel to quantify the information carrying capacity of the feature channel.

[0053] In step S103, the semantic entropy density of the channel is defined using Shannon entropy. The calculation formula is:

[0054] in, It is a very small positive number (for example, its range is 1). to Semantic entropy density is used to maintain the stability of numerical computation. The numerical value is positively correlated with the texture detail and semantic richness contained in the feature channel.

[0055] Step S104: Calculate the dynamic filtering threshold based on the semantic entropy density statistical distribution of all channels in the current layer, and generate the semantic manifold kernel mask.

[0056] In step S104, in order to adapt to the differences in semantic complexity of different input data (such as video frames from different scenes), the system calculates the mean of the semantic entropy density of the current layer in real time. and standard deviation :

[0057] Based on the above statistics, a dynamic screening threshold is constructed. :

[0058] in, The semantic preservation coefficient is set to a range of values. .

[0059] The system will calculate the semantic entropy density of each channel. With dynamic filtering threshold Compare and generate a binary semantic manifold kernel mask vector. :

[0060] The semantic manifold kernel mask vector The element "1" indicates that the convolution kernel parameter corresponding to this channel belongs to the "rigid skeleton" of the model. In subsequent model reconstruction steps, regardless of changes in environmental pressure, the weight parameters of this part are forcibly retained and no pruning operation is performed.

[0061] To more clearly illustrate the technical solution of the present invention, the above step S200 will be described in detail below: Step S201: Periodically collect the physical resource parameters of each node in the edge computing network, and normalize the physical resource parameters to generate a node state vector.

[0062] In step S201, the system obtains the network information of the first... Nodes (denoted as ) )exist Real-time operational data at any given moment. The physical resource parameters include: remaining computing power. Remaining battery energy Current chip temperature at the node and the current task queue length .

[0063] The system processes the above parameters using the rated maximum values ​​of each node to construct the node state vector. :

[0064] The calculation formulas for each item are as follows:

[0065] in, and They are nodes Maximum computing power and maximum battery capacity; This is the critical temperature for thermal throttling of the chip. The ambient reference temperature; This sets the upper limit for the queue buffer capacity. This step ensures that parameters with different physical dimensions are mapped to... Within the unified range.

[0066] Step S202: Determine the coordinate position of each node in the network topology space, and construct a scalar potential energy function based on the node state vector.

[0067] In step S202, the system determines each node spatial coordinates The spatial coordinates can be geographic coordinates based on the physical location of the node, or virtual network coordinates generated by multidimensional scaling analysis (MDS) based on network communication latency.

[0068] Construct a scalar potential energy function in a metric space with defined coordinates. The function value represents the spatial location. The resource pressure potential energy at a given location is the sum of the local potential energies generated by all nodes in the network. The specific functional form is as follows:

[0069] in, For nodes The source potential energy intensity is positively correlated with the node's load level and negatively correlated with the node's available resources. It is defined as:

[0070] In the formula, These are preset weighting coefficients for positive real numbers; It is a non-zero small constant. Here, the radial basis function describes the decay of nodal potential energy with spatial distance:

[0071] in, Indicates spatial location With node coordinates The Euclidean distance between them This is the potential field smoothing parameter, and its value is usually set to the average distance between adjacent nodes in the network.

[0072] Step S203: Perform spatial differentiation on the scalar potential energy function to calculate the negative gradient vector of the potential field at the current position of the task to be scheduled.

[0073] In step S203, the system calculates the potential energy function. At the current node location of the task The negative gradient vector at point According to the rules of vector differentiation, the calculation process is as follows:

[0074] The negative gradient vector This is the composite vector of the potential energy thrust generated by each neighboring node. Physically, this vector points in the direction of the fastest decrease in the potential energy function value, i.e., away from regions with high temperature, high load, and low remaining resources, and towards regions with low temperature, low load, and high remaining resources. The magnitude of this vector is... The numerical value quantifies the intensity of environmental resource pressure on the current location.

[0075] The specific implementation of step S300 includes the following: Step S301: Receive the negative gradient vector of the potential field output by the heterogeneous resource potential field construction module, and calculate the scalar environmental pressure value using a nonlinear mapping function.

[0076] In step S301, the system obtains the negative gradient vector of the potential field at the current position of the task. The magnitude of this vector The numerical value is positively correlated with the resource load of the current network node. The system uses the Sigmoid function to construct a nonlinear mapping relationship to calculate the scalar environmental pressure value. :

[0077] in, The Euclidean norm of the negative gradient vector; This is the pressure bias threshold. When the gradient magnitude is less than this threshold, the output pressure value approaches 0. To control the sensitivity of pressure values ​​to gradient changes in response to the slope coefficient; This is the maximum pruning threshold coefficient, and its unit is consistent with the numerical unit of the model weights. It is usually set as a multiple of the standard deviation of the model weights (e.g., 1.0 to 3.0 times). This formula ensures that the calculated scalar environmental pressure value is accurate. Always within the range Inside.

[0078] Step S302: Expand the scalar environmental pressure value into an environmental pressure tensor that is consistent with the weight dimension of the target neural network layer.

[0079] In step S302, for the first The layered artificial intelligence model, whose original weight tensor is represented as: Its dimension is denoted as The system constructs a dimension and The same full 1 tensor, and the scalar environmental pressure value Multiplying this tensor by the total 1s generates the environmental pressure tensor. The value of each element in this environmental pressure tensor is equal to... This enables numerical broadcasting from scalar to tensor space.

[0080] Step S303: Compare the original weight tensor with the environmental pressure tensor element by element to generate an environmental tolerance mask.

[0081] In step S303, the system calculates each weight parameter in the original weight tensor. absolute value The absolute value is then compared with the corresponding value at the position in the environmental pressure tensor. Compare and generate a binarized environment-tolerant mask. :

[0082] This step filters parameters based on weight magnitude. When the absolute value of a weight parameter is lower than the threshold corresponding to the current environmental pressure, the parameter is marked as 0 in the environmental tolerance mask, indicating that it is a redundant parameter that can be pruned.

[0083] Step S304: Combine the semantic manifold kernel mask generated by the semantic manifold kernel extraction module to calculate the final dynamic projection mask.

[0084] In step S304, the system obtains the semantic manifold kernel mask. .because The length is equal to the number of output channels. For a one-dimensional vector, the system first performs a dimension expansion operation on it, changing its dimension from... Adjusted to Subsequently, it was extended to include environmentally tolerable masks via tensor broadcasting. Same dimensions .

[0085] The system performs a logical OR operation on the extended semantic manifold kernel mask and the environment-tolerant mask to generate the final dynamic projection mask. :

[0086] This logical operation establishes a hybrid retention mechanism: as long as the channel to which a certain weight parameter belongs is determined to be a semantic manifold kernel (i.e., the semantic mask is 1), or the magnitude of the weight parameter itself is greater than the environmental pressure threshold (i.e., the environmental mask is 1), the weight parameter is marked as 1 in the final mask.

[0087] Step S305: The original weight tensor is projected and reconstructed using the dynamic projection mask to generate a sparse dynamic inference model.

[0088] In step S305, the final dynamic projection mask is applied to the original weights using the Hadamard product:

[0089] Calculation result This refers to the reconstructed dynamic weights. For positions where the mask is set to 0, the system encodes and stores the weight tensor using either Compressed Sparse Row (CSR) or Compressed Sparse Column (CSC) format. The reconstructed model data packet is then sent to the streaming gradient optimization scheduling module.

[0090] The specific implementation of step S400 includes the following: Step S401: Construct the Lyapunov function for network queue stability and calculate the local queue difference vector of the current node relative to its neighboring nodes.

[0091] In step S401, the system establishes a queue constraint model based on Lyapunov optimization theory. A quadratic Lyapunov function is defined. It is half the sum of the squares of the task queue backlogs of all nodes in the network:

[0092] in, For nodes In the time slot The actual length of the task queue.

[0093] The system obtains the current node single-hop physical neighbor node set For each neighbor node Calculate the difference in queue backlog. Based on this difference, a local queue difference vector is constructed. :

[0094] in, and These are the spatial coordinate vectors of the current node and its neighboring nodes, respectively. This formula shows that... The direction points to the weighted geometric center where the queue backlog is significantly lower than the current node's neighboring nodes.

[0095] Step S402: Introduce drift penalty control parameters, normalize and weight the potential field negative gradient vector and the local queue difference vector to generate a comprehensive optimization guidance vector.

[0096] In step S402, the system receives the negative gradient vector of the potential field output by the heterogeneous resource potential field construction module. To eliminate the influence of differences in physical dimensions on the synthesis direction, the system first... and Perform L2 norm normalization on each vector to obtain the unit vector. and .

[0097] System construction of comprehensive optimization guidance vector The calculation formula is as follows:

[0098] in, This is a drift penalty control parameter, and its value range is... .

[0099] When the setting is large When the value is set, the scheduling strategy focuses on optimizing physical resources (computing power, energy); when a smaller value is set... At this value, the scheduling strategy focuses on using queue differential balancing to balance network load. By adjusting... This value achieves a dynamic balance between resource utilization and network throughput stability.

[0100] Step S403: Based on the comprehensive optimization guidance vector, perform maximum projection optimization in the discrete set of neighbor nodes to determine the next-hop target node for task transmission.

[0101] In step S403, the system maps the guidance vectors in the continuous space to the physical links. The current node is traversed. Each neighbor node Calculate the link unit vector pointing to this neighbor node. :

[0102] Calculate the comprehensive optimization guidance vector Projected scalar values ​​on each link unit vector : The system compares the projected scalar values ​​of all neighboring nodes. If at least one neighboring node has a projected scalar value greater than zero, the neighboring node with the largest projected scalar value is selected as the best next-hop node. :

[0103] If the projected scalar values ​​of all neighboring nodes are less than or equal to zero, it indicates that the current node is at a local minimum point of the potential energy field or queue pressure. The system determines that the current node is a computation execution node, terminates the task scheduling process, and triggers the local model inference service.

[0104] Step S404: The reconstructed dynamic inference model and related task metadata are encapsulated and sent to the determined next-hop target node via the network interface.

[0105] In step S404, the system adjusts the weights of the sparse dynamic inference model generated in Example 5. Perform serialization and encapsulation. The packet header contains metadata information: Model structure descriptor: defines the dimensions and data type of the tensor; Sparse index table: used by the receiver to reconstruct the weight matrix using the CSR or CSC algorithm; Manifold kernel checksum: Used to verify the integrity of semantic manifold kernel data.

[0106] The system sends the encapsulated data packet to the next-hop target node determined in step S403 via a network communication interface (such as a TCP / IP socket or RDMA interface). After receiving the data, if the target node determines that it is the computation execution node, it deserializes the model and performs inference; otherwise, it becomes the new current node and repeats steps S401 to S404.

[0107] This invention is implemented based on an electronic device, which serves as a computing node in a heterogeneous edge computing network and is configured to execute the heterogeneous edge network streaming gradient optimization method based on semantic manifold kernels described in the foregoing embodiments.

[0108] The electronic device includes a processor, memory, communication interface, input / output interface, and physical status monitoring unit that are interconnected via a system bus.

[0109] The processor can specifically be a central processing unit (CPU), graphics processing unit (GPU), neural network processor (NPU), or field-programmable gate array (FPGA). Given that this invention relates to a large number of feature map tensor operations and differential geometry calculations, the processor integrates a vector computation unit or a tensor core and supports a single-instruction multiple-data (SIMD) instruction set. The processor is configured to call instructions from memory to perform mathematical operations including convolutional feature extraction, spatial softmax probability normalization, Hadamard product masking, and potential field gradient vector calculation.

[0110] The memory includes non-volatile memory and volatile memory. The non-volatile memory stores the operating system, device drivers, and computer program instructions for implementing the method of this invention. The volatile memory provides a high-speed data buffer for temporarily storing intermediate feature map tensors generated by the semantic manifold kernel extraction module, environmental pressure tensors generated by the heterogeneous resource potential field construction module, and sparse model weight data to be transmitted. The system bus is responsible for transmitting control and data signals between the processor, memory, and various peripheral interfaces.

[0111] The communication interface is used to establish network connections between electronic devices and other edge nodes and cloud servers. This interface supports the TCP / IP protocol stack and Remote Direct Data Access (RDMA) technology, and can parse custom-encapsulated data packets containing model structure descriptors, sparse index tables, and manifold kernel checksums. During streaming gradient optimization, the communication interface is responsible for frequently receiving resource state vectors broadcast by neighboring nodes and sending the reconstructed model data according to the processor's scheduling instructions.

[0112] The physical state monitoring unit is connected to the processor and is used to collect real-time physical operating parameters of the electronic equipment, and convert the physical signals into digital signals to provide to the heterogeneous resource Riemann potential energy field construction module. This unit specifically includes: A thermistor array, connected to the system bus via an analog-to-digital converter (ADC), reads the junction temperature data of the processor chip and power devices in real time to provide the temperature variable in the formula. ; The power management integrated circuit (PMIC) communicates with the processor via I2C or SMBus bus to read the voltage, current, and coulomb counter values ​​from the battery register to provide information on remaining energy. ; The system load monitor is implemented as a performance counter interface at the operating system kernel level (such as the / proc filesystem interface in Linux systems). It is used to count the number of ready task processes and the frequency of interrupt requests in real time, providing a variable indicating the current task queue length. and remaining computing power variables .

[0113] The input / output interface is used to connect external data acquisition devices, such as cameras, lidar, or various IoT sensors, to obtain raw task data to be processed.

[0114] An embodiment of the present invention also provides a computer-readable storage medium storing computer program instructions thereon. The computer-readable storage medium may be a non-transitory storage medium, such as a hard disk, optical disk, flash memory, or embedded multimedia card (eMMC). When the computer program instructions are read and executed by the logic circuitry of a computer or processor, the computer or processor performs the steps of the method described in any of the foregoing embodiments, specifically including: The intermediate feature map tensor is obtained using a feature extraction network, and the semantic entropy density of the feature channels is calculated. A semantic manifold kernel mask is generated based on the semantic entropy density, and a heterogeneous resource Riemann potential field is constructed by combining the data collected by the physical state monitoring unit. Calculate the negative gradient vector of the potential field and map it to the environmental pressure tensor, then combine the semantic manifold kernel mask to generate a dynamic projection mask. The weights of the original model are reconstructed by sparsification, and the next-hop target node is calculated based on the potential field gradient and the queue difference vector.

[0115] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An artificial intelligence task scheduling system for edge computing, deployed in a heterogeneous computing network environment comprising edge terminal devices, edge gateway nodes, and cloud servers, characterized in that, The system includes: The semantic manifold kernel extraction module (110) is configured to access the data stream of the artificial intelligence task to be processed, perform convolution operation on the input data to obtain the intermediate layer feature map, calculate the semantic entropy density of each channel in the intermediate layer feature map, and generate a semantic manifold kernel mask based on the semantic entropy density. The heterogeneous resource potential field construction module (120) is configured to collect resource status data of each node, construct a scalar potential field covering the network topology space using the potential energy function, and calculate the negative gradient vector of the scalar potential field at any spatial coordinate. The negative gradient vector contains direction information and magnitude information. A gradient tensor projection reconstruction module (130) is configured to receive the negative gradient vector and the semantic manifold kernel mask, convert the magnitude of the negative gradient vector into an environmental pressure tensor and generate an environmental tolerance mask accordingly, perform a union operation between the semantic manifold kernel mask and the environmental tolerance mask to generate a preserved mask, and use the preserved mask to project the original artificial intelligence model weight tensor to generate dynamic inference model weights; and The streaming gradient optimization scheduling module (140) is configured to receive the negative gradient vector, calculate the Lyapunov drift correction based on the network queue backlog state, solve the motion equation of the task in the network space according to the corrected gradient direction, determine the next hop target node for task transmission, and send the dynamic inference model weights to the target node. 2.The edge computing oriented artificial intelligence task scheduling system according to claim 1, characterized in that, The semantic manifold kernel extraction module (110) is specifically configured as follows: The feature map tensor of the l-th layer is obtained using a convolutional neural network. The normalized probability distribution of the pixel activation value of the c-th channel in the feature map is calculated, and the semantic entropy density of the channel is calculated based on the distribution. Calculate the statistical distribution of semantic entropy density of all channels in the current layer, and construct a dynamic filtering threshold; The semantic entropy density is compared with the dynamic filtering threshold. When the semantic entropy density of a certain channel is greater than the dynamic filtering threshold, the corresponding channel index position is marked as 1 in the semantic manifold kernel mask; otherwise, it is marked as 0. 3.The edge computing oriented artificial intelligence task scheduling system according to claim 1, wherein, The heterogeneous resource potential field construction module (120) is specifically configured as follows: Obtain the state vector of each node in the network topology. The state vector includes the remaining computing power factor, remaining energy factor, available bandwidth factor, and temperature factor after normalization. Construct a potential energy function covering the network space. This potential energy function defines the resource cost at any location in the space. The larger the normalized remaining computing power and remaining energy of a node, the smaller the potential energy function value corresponding to that location. The larger the normalized temperature and task queue backlog of a node, the larger the potential energy function value corresponding to that location. The potential energy function is spatially differentiated to calculate the negative gradient vector, the direction of which indicates the direction of the fastest decrease in potential energy. 4.The edge computing oriented artificial intelligence task scheduling system according to claim 1, wherein, The gradient tensor projection reconstruction module (130) is specifically configured as follows: The magnitude of the negative gradient vector is converted into a scalar pressure value using a nonlinear mapping algorithm, and then expanded into the environmental pressure tensor with the same dimension as the model weights through tensor broadcasting. Calculate the absolute value of the weight tensor of the original artificial intelligence model and compare it with the environmental pressure tensor. When the absolute value of the weight is lower than the value at the corresponding position in the environmental pressure tensor, mark it as 0 in the environmental tolerance mask; otherwise, mark it as 1. The semantic manifold kernel mask and the environment tolerance mask are logically ORed to generate the reserved mask. The original weights are then multiplied element-wise using the reserved mask to obtain the dynamic inference model weights.

5. The AI ​​task scheduling system for edge computing according to claim 1, characterized in that, The streaming gradient optimization scheduling module (140) is specifically configured as follows: Define the Lyapunov function as half the sum of the squares of the task queue lengths of all nodes in the network, and calculate the Lyapunov drift term in the negative direction of the queue length gradient. Calculate the scheduling vector, which is a weighted composite of the negative gradient vector and the Lyapunov drift term; Calculate the difference vector between the position vector of each neighbor node in the network topology space and the current position vector, and calculate the cosine similarity between the difference vector and the scheduling vector. The neighbor node with the highest cosine similarity value is selected as the next-hop target node.

6. An artificial intelligence task scheduling method for edge computing, characterized in that, The method includes the following steps: The system receives AI task data to be processed, extracts features from the task data to obtain an intermediate layer feature map, calculates the channel-level semantic entropy density based on the intermediate layer feature map, and then generates a semantic manifold kernel mask. Collect resource status data of each node in the edge computing network, use the potential energy function to map the network environment into a scalar potential energy field, and calculate the negative gradient vector of the potential energy field at the current position of the task to be scheduled. The negative gradient vector of the potential energy field is mapped to the environmental pressure tensor and an environmental tolerance mask is generated. The semantic manifold kernel mask is combined to generate a retention mask. The retention mask is used to project and reconstruct the weight tensor of the artificial intelligence model to generate a dynamic inference model. Based on the gradient descent algorithm with Lyapunov drift correction, the scheduling vector is calculated according to the negative gradient vector of the potential field and the network queue backlog state to determine the next-hop target node for task transmission, and the dynamic inference model is sent to the next-hop target node.

7. The artificial intelligence task scheduling method for edge computing according to claim 6, characterized in that, The step of calculating the channel-level semantic entropy density based on the intermediate layer feature map, and then generating the semantic manifold kernel mask, includes: For each feature channel in the intermediate layer feature map tensor, the spatial dimension probability is normalized using the spatial Softmax function to obtain the normalized probability. Based on the normalized probability, the semantic entropy density of each feature channel is calculated using the Shannon entropy formula. Calculate the dynamic filtering threshold based on the semantic entropy density statistical distribution of all channels in the current layer. ,in The mean, Standard deviation, These are semantic preservation coefficients; The semantic entropy density of each channel is compared with the dynamic filtering threshold to generate a binary semantic manifold kernel mask.

8. The artificial intelligence task scheduling method for edge computing according to claim 6, characterized in that, The method of mapping the network environment to a scalar potential field using a potential energy function includes: Based on the resource status data, the source potential energy intensity of each node is constructed. The source potential energy intensity is positively correlated with the current chip temperature and current task queue length of the node, and negatively correlated with the remaining computing power and remaining energy of the node. The radial basis function kernel is used to describe the decay of nodal potential energy with spatial distance, and the scalar potential energy function is constructed by superimposing the local potential energy of each node. Perform spatial differentiation on the scalar potential energy function to calculate the negative gradient vector of the potential energy field. The magnitude of this vector corresponds to the environmental resource pressure intensity at the current spatial location.

9. The artificial intelligence task scheduling method for edge computing according to claim 6, characterized in that, The step of projecting and reconstructing the weight tensor of the artificial intelligence model using the preserved mask to generate a dynamic inference model includes: The magnitude of the negative gradient vector of the potential energy field is converted into a scalar pressure value using a Sigmoid-type nonlinear function, and then broadcast as the environmental pressure tensor with the same dimension as the model weights. The absolute value of the original weight tensor is compared element by element with the environmental pressure tensor to generate an environmental tolerance mask. Perform a logical OR operation on the semantic manifold kernel mask and the environment-tolerant mask to generate the final preserved mask. This logical OR operation ensures that the weight parameters marked as 1 by the semantic manifold kernel mask are preserved. The preserved mask is applied to the original weight tensor using the Hadamard product to obtain the reconstructed dynamic inference model weights.

10. The artificial intelligence task scheduling method for edge computing according to claim 6, characterized in that, The determination of the next-hop target node for task transmission includes: Construct a Lyapunov function for network queue stability and compute the local queue difference vector of the current node relative to its neighbors; By introducing drift penalty control parameters, the negative gradient vector of the potential energy field and the local queue difference vector are normalized and weighted to generate a comprehensive optimization guidance vector. Calculate the projection scalar value of the integrated optimization guidance vector onto the link unit vector pointing to each neighbor node, and select the neighbor node with the largest projection scalar value as the next hop target node; If the projected scalar values ​​of all neighboring nodes are less than or equal to zero, then the current node is determined to be the computation execution node and local model inference is executed.

Images