Brain-like ai architecture based edge side computing power acceleration method and device and computer equipment

Abstract

The application relates to a side power acceleration method and device based on a brain-like AI architecture, computer equipment, a computer readable storage medium and a computer program product. The method comprises the following steps: classifying heterogeneous data collected by a side device according to a preset strategy to obtain heterogeneous sub-data of different categories; combining the power resource data collected by the brain-like AI architecture and the characteristics of the heterogeneous sub-data to construct a path evaluation index strategy and match the optimal processing path of each sub-data; after processing each sub-data slice, the real-time and importance are sorted; a scheduling strategy is determined according to the current power resource state, and the data slices are sequentially transmitted to the corresponding processing unit; the power allocation resource is determined according to the processing unit load and the data slice demand, and the processing task is executed. The application ensures that real-time tasks can be responded in time, optimizes the allocation of power resources, and avoids the problem of low processing efficiency caused by uneven load or unreasonable resource allocation of the processing unit.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method, apparatus, computer device, computer-readable storage medium, and computer program product for accelerating edge computing power based on a brain-like AI architecture. Background Technology

[0002] With the development of artificial intelligence and edge computing, edge computing acceleration methods based on brain-inspired AI (Artificial Intelligence) architectures have gradually become a research focus in order to improve data processing speed and reduce transmission latency. Brain-inspired AI architectures mimic the structure and function of the human brain, achieving complex computational tasks by simulating the connections between neurons. This architecture not only enhances the computing power of edge devices (side devices) but also enables operation under low power consumption conditions.

[0003] In related technologies, edge computing acceleration methods based on neuromorphic AI architectures typically employ a unified, fixed data processing path, processing all types of data according to the same process. When encountering complex, large-volume heterogeneous data, the fixed processing path leads to severe delays in data transmission and processing. Furthermore, a large amount of unnecessary computation consumes limited edge computing resources, preventing edge devices from efficiently responding to real-time tasks. This can easily result in low data processing efficiency due to uneven load or unreasonable resource allocation. Summary of the Invention

[0004] Therefore, it is necessary to provide a method, device, computer equipment, computer-readable storage medium, and computer program product based on a neuromorphic AI architecture to address the above-mentioned technical problems and solve the problem of low processing efficiency caused by uneven load or unreasonable resource allocation of processing units.

[0005] Firstly, this application provides a method for accelerating edge computing power based on a brain-like AI architecture, including:

[0006] Based on a preset classification strategy, heterogeneous data collected by edge devices is classified to obtain multiple heterogeneous sub-data corresponding to different categories.

[0007] Based on a pre-built path evaluation index strategy, each type of heterogeneous sub-data is matched with multiple preset candidate processing paths to determine the data processing target path corresponding to each type of heterogeneous sub-data; the pre-built path evaluation index strategy is used to indicate the path quality evaluation index and the corresponding index weight.

[0008] According to the adaptation of each type of heterogeneous sub-data to its corresponding data processing target path, each type of heterogeneous sub-data is segmented to obtain multiple target data segments corresponding to each type of heterogeneous sub-data. Based on the data real-time performance and data importance of each type of heterogeneous sub-data, the target data segments corresponding to each type of heterogeneous sub-data are prioritized to obtain the data sorting result.

[0009] Based on the data sorting results and the current computing power resource data of the edge device, a sharding scheduling strategy is determined, and the target data is sharded and transmitted sequentially to the corresponding processing unit based on the sharding scheduling strategy; the current computing power resource data is the computing power resource data of the edge device collected based on the neuromorphic AI architecture;

[0010] For the target data shard transmitted to the processing unit, based on the current load status of the processing unit and the processing requirements of the target data shard, computing power allocation resources are determined, and processing tasks for the target data shard are executed based on the computing power allocation resources.

[0011] In one embodiment, the step of segmenting each type of heterogeneous sub-data according to the adaptation between each type of heterogeneous sub-data and its corresponding data processing target path, to obtain multiple target data segments corresponding to each type of heterogeneous sub-data, includes:

[0012] For any type of heterogeneous sub-data, extract the feature dimensions of the heterogeneous sub-data and the capacity parameters of each node in the corresponding data processing target path;

[0013] Based on the path fit degree corresponding to any type of heterogeneous sub-data, a sharding priority sequence corresponding to any type of heterogeneous sub-data is obtained; the path fit degree includes multiple parameters, namely, the path fit degree between each feature dimension of any type of heterogeneous sub-data and the data processing target path corresponding to any type of heterogeneous sub-data.

[0014] Based on the sharding priority sequence, the feature dimensions of the heterogeneous sub-data are grouped according to the number of nodes in the data processing target path corresponding to the heterogeneous sub-data, resulting in multiple feature dimension groups of the heterogeneous sub-data.

[0015] Based on the information entropy density of each feature dimension group and the capacity parameters of each node in the data processing target path, the granularity coefficient corresponding to each feature dimension group is determined.

[0016] Based on the granularity coefficient corresponding to each of the feature dimension groups, the heterogeneous sub-data is sharded to obtain multiple target data shards corresponding to the heterogeneous sub-data.

[0017] In one embodiment, the step of partitioning the heterogeneous sub-data according to the partitioning granularity coefficient corresponding to each of the feature dimension groups to obtain multiple target data partitions corresponding to the heterogeneous sub-data includes:

[0018] Based on the granularity coefficient of each feature dimension group, the heterogeneous sub-data portion corresponding to each feature dimension group in any type of heterogeneous sub-data is horizontally sharded to obtain multiple initial data shards.

[0019] The number of initial data shards is adjusted based on the preset load threshold of each node in the data processing target path to obtain the adjusted number of shards.

[0020] Based on the adjusted number of shards, the initial data shards are split or merged to generate multiple target data shards corresponding to any type of heterogeneous sub-data.

[0021] In one embodiment, the path adaptability calculation formula is as follows:

[0022]

[0023] Among them, F d,k f represents the path fit between the d-th feature dimension and the k-th data processing target path; d,m c is the value of the m-th parameter in the d-th feature dimension; k,m The capacity parameter value for the m-th node in the target path of the k-th data processing line is given; q is the number of parameter items; g d,m h is the weight coefficient of the m-th parameter in the d-th feature dimension; k,m The influence coefficient of the capacity parameter of the m-th node in the target path of the k-th data processing.

[0024] In one embodiment, the current load status includes CPU utilization, memory usage, and task queue length, and the processing requirements include computational complexity, storage requirements, and processing latency thresholds;

[0025] The process of determining computing power allocation resources based on the current load status of the processing unit and the processing requirements of the target data shard includes:

[0026] Based on the CPU utilization, memory usage, and task queue length of the processing unit, and combined with the computational complexity, storage requirements, and processing latency threshold of the target data shard, a load demand mapping matrix is ​​constructed.

[0027] Based on the matching degree between each processing unit and each target data shard in the load demand mapping matrix, the resource demand gap coefficient of each processing unit for each target data shard is determined; the resource demand gap coefficient is used to quantify the degree to which the processing unit meets the resource demand of the target data shard.

[0028] Based on the resource demand gap coefficient, each target data segment is filtered to obtain target processing units that meet the first preset minimum resource demand gap coefficient, and the target processing units are used as candidate binding units to form an initial binding relationship group.

[0029] Based on the initial binding relationship group, a resource demand load correlation function is constructed between each target data shard and its corresponding candidate binding unit; the resource demand load correlation function is used to quantify the degree of impact of the load change of the candidate binding unit on meeting the resource demand of the corresponding target data shard;

[0030] Based on the resource demand load correlation function, combined with the load changes of the candidate binding units and the processing progress of the target data shards, the computing power allocation resources are determined.

[0031] In one embodiment, determining the computing power allocation resources based on the resource demand load correlation function, combined with the load changes of the candidate binding units and the processing progress of the target data shard, includes:

[0032] Based on the resource demand load correlation function, the computing resources between the target data shard and the candidate binding unit in each initial binding relationship group are analyzed to generate a candidate scheme group; each scheme in the candidate scheme group contains a set of computing power allocation values ​​of the candidate binding unit to the target data shard;

[0033] For each scheme in the candidate scheme group, a scheme suitability index is determined based on the load balancing degree of the candidate binding unit and the demand satisfaction degree of the target data sharding;

[0034] The schemes are sorted in descending order of their suitability index, and the scheme with the largest suitability index is selected as the optimal computing power allocation scheme. The optimal computing power allocation scheme is used to output the computing power allocation amount of each processing unit to the target data shard.

[0035] The computing power allocation amount of the optimal computing power allocation scheme is adjusted based on the load changes of the candidate binding units and the processing progress of the target data shards to obtain the computing power allocation resources.

[0036] In one embodiment, the resource demand load correlation function is:

[0037]

[0038] Among them, F s,u (x1,x2,x3) is the resource demand load correlation function between the target data fragment s and the candidate binding unit u; x1, x2, and x3 are the changes in CPU utilization, memory utilization, and task queue length, respectively; k1, k2, and k3 are the influence weight coefficients.

[0039] Secondly, this application also provides an edge computing acceleration device based on a brain-like AI architecture, comprising:

[0040] The classification module is used to classify heterogeneous data collected by edge devices based on a preset classification strategy, and obtain multiple heterogeneous sub-data corresponding to different categories.

[0041] The path planning module is used to match each type of heterogeneous sub-data with multiple preset candidate processing paths based on a pre-built path evaluation index strategy, and determine the data processing target path corresponding to each type of heterogeneous sub-data; the pre-built path evaluation index strategy is used to indicate the path quality evaluation index and the corresponding index weight.

[0042] The sharding and sorting module is used to shard each type of heterogeneous sub-data according to the adaptation of each type of heterogeneous sub-data to its corresponding data processing target path, thereby obtaining multiple target data shards corresponding to each type of heterogeneous sub-data. The module also sorts the target data shards corresponding to each type of heterogeneous sub-data based on the data real-time performance and data importance of each type of heterogeneous sub-data, thereby obtaining the data sorting result.

[0043] The scheduling and transmission module is used to determine a sharding scheduling strategy based on the data sorting result and the current computing power resource data of the edge device, and to transmit the target data in shards to the corresponding processing units in sequence based on the sharding scheduling strategy; the current computing power resource data is the computing power resource data of the edge device collected based on the neuromorphic AI architecture;

[0044] The computing power allocation and task execution module is used to determine the computing power allocation resources for the target data shards transmitted to the processing unit based on the current load status of the processing unit and the processing requirements of the target data shards, and to execute the processing tasks for the target data shards based on the computing power allocation resources.

[0045] Thirdly, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, implements the steps of the method described above.

[0046] Fourthly, this application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements the steps of the above-described method.

[0047] Fifthly, this application also provides a computer program product. The computer program product includes a computer program that, when executed by a processor, implements the steps of the above-described method.

[0048] The aforementioned edge computing acceleration method, device, computer equipment, computer-readable storage medium, and computer program product based on a neuromorphic AI architecture classifies heterogeneous data collected by edge devices using a preset classification strategy to obtain multiple heterogeneous sub-data categories. Based on a pre-built path evaluation index strategy, each category of heterogeneous sub-data is matched with multiple preset candidate processing paths to determine the target data processing path for each category. The pre-built path evaluation index strategy indicates the path quality evaluation index and its corresponding weight. Each category of heterogeneous sub-data is then segmented according to its compatibility with its corresponding target data processing path to obtain the desired data for each category. Each data segment corresponds to multiple target data segments. These segments are prioritized based on the real-time performance and importance of each type of heterogeneous sub-data, resulting in a data sorting result. Based on the sorting result and the current computing power resources of the edge devices, a segmentation scheduling strategy is determined, and the target data segments are sequentially transmitted to their corresponding processing units. The current computing power resources data are collected from the edge devices using a brain-like AI architecture. For each target data segment transmitted to the processing unit, computing power allocation is determined based on the current load of the processing unit and the processing requirements of the target data segment. Processing tasks for the target data segments are then executed based on the allocated computing power resources.

[0049] Thus, by classifying the acquired heterogeneous data and considering the differences in characteristics among different data, this approach changes the previous method of using a uniform and fixed processing path for all data. The classified heterogeneous sub-data, combined with the collected computing resource data, can be matched to find the optimal target data processing path. This solves the problem of related technologies not considering the dynamic changes in computing resources of side devices and avoids unnecessary computation consuming limited computing resources under a fixed path. Furthermore, by using the optimal data processing path to segment the heterogeneous sub-data and prioritizing it based on real-time performance and importance, and determining the segmentation scheduling strategy based on current computing resources, high-priority data segmentation can be processed first, ensuring timely response to real-time tasks. Finally, computing resources are allocated based on the processing unit load and data segmentation processing requirements, dynamically adapting to the processing unit's status and data needs, further optimizing the allocation of computing resources and avoiding low processing efficiency caused by uneven load or unreasonable resource allocation. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 This is a flowchart illustrating an edge computing acceleration method based on a brain-like AI architecture in one embodiment.

[0052] Figure 2 This is a flowchart illustrating the steps for obtaining multiple target data fragments corresponding to each type of heterogeneous sub-data in one embodiment.

[0053] Figure 3 This is a flowchart illustrating the steps for determining computing power allocation resources in one embodiment;

[0054] Figure 4 This is a flowchart illustrating the steps for determining computing power allocation resources in another embodiment;

[0055] Figure 5 This is a flowchart illustrating an edge computing acceleration method based on a brain-like AI architecture in another embodiment.

[0056] Figure 6 This is a structural block diagram of an edge computing acceleration device based on a brain-like AI architecture in one embodiment;

[0057] Figure 7 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0059] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0060] In one embodiment, such as Figure 1 As shown, an edge computing acceleration method based on a neuromorphic AI architecture is provided. This embodiment demonstrates that this method can be applied to computer devices. It is understood that the computer device can be a terminal, a server, or a system including both a terminal and a server. It should be noted that in this application, the edge device can also be referred to as a side-side device. In this embodiment, the computer device is specifically described as an edge computing acceleration system. The method includes steps S110 to S150:

[0061] Step S110: Based on a preset classification strategy, the heterogeneous data collected by the edge device is classified to obtain multiple heterogeneous sub-data corresponding to different categories.

[0062] Among them, edge devices can be edge devices in the power grid scenario, and heterogeneous data can be heterogeneous power data.

[0063] In some embodiments, the edge computing acceleration system has a pre-defined classification strategy that covers multiple dimensions of classification criteria, including data source, data structure complexity, and data real-time requirements. When the system receives heterogeneous data (including voltage and current waveform data, device temperature data, and device status image data) collected by edge devices (such as smart meters, transformers, and infrared temperature measuring devices in the power grid), it performs a classification operation using the pre-defined strategy. During the classification process, key features of the heterogeneous data are first extracted, such as data size, data format (structured, semi-structured, unstructured), and data update frequency. Then, based on the pre-defined classification strategy, the data is divided into different categories, ultimately forming multiple categories of heterogeneous sub-data.

[0064] In one embodiment, taking a side-side scenario of a power grid as an example, the heterogeneous data collected by the side-side devices includes voltage and current time-series data uploaded by smart meters every 5 minutes (structured, medium data volume, high real-time requirements), thermal imaging images of the equipment taken by infrared temperature measuring devices (unstructured, large data volume, medium real-time requirements), and vibration frequency values ​​transmitted back by equipment status monitoring sensors (structured, small data volume, extremely high real-time requirements). The side-side computing acceleration system, based on a preset classification strategy, categorizes this heterogeneous data into three types of heterogeneous sub-data: time-series data, image data, and high-frequency numerical data.

[0065] Step S120: Based on the pre-built path evaluation index strategy, each type of heterogeneous sub-data is matched with multiple preset candidate processing paths to determine the data processing target path corresponding to each type of heterogeneous sub-data.

[0066] Among them, the pre-built path evaluation index strategy is used to indicate the path quality evaluation index and the corresponding index weight.

[0067] Among them, path evaluation index strategies can be constructed based on the computing power resource data and feature information in heterogeneous sub-data collected from edge devices using a brain-like AI architecture.

[0068] In some embodiments, the edge computing acceleration system collects real-time computing resource data from edge devices based on a neuromorphic AI architecture, such as CPU utilization, memory usage, and network bandwidth. Simultaneously, it extracts feature information from various heterogeneous sub-data, such as data dimension, data correlation, and data change rate. First, a path evaluation index strategy is constructed based on this information. This strategy includes multiple evaluation dimensions (path quality evaluation indicators), such as processing latency, computing power utilization, and data transmission stability. Then, this strategy is matched with multiple preset candidate processing paths. By quantitatively analyzing the performance of each candidate path across various evaluation dimensions, the optimal data processing target path corresponding to each type of heterogeneous sub-data is selected.

[0069] Furthermore, when matching this strategy with multiple preset candidate processing paths, the formula for calculating the path matching degree is as follows:

[0070]

[0071] Among them, M i,j The value represents the matching degree between the i-th heterogeneous sub-data and the j-th candidate processing path; n is the number of evaluation dimensions; x i,k b is the feature value of the i-th type of heterogeneous subdata in the k-th evaluation dimension; j,k w is the baseline value for the j-th candidate processing path in the k-th evaluation dimension; k Let be the weight of the k-th evaluation dimension, and When Mi,j The closer the value is to 1, the higher the matching degree, and the corresponding path is the optimal data processing target path. Furthermore, max(x) is introduced. i,k ,b j,k Using 0.1 as the denominator effectively avoids calculation errors caused by eigenvalues ​​or benchmarks being zero. It also adapts to eigenvalues ​​and benchmarks of different magnitudes, ensuring stable calculations under various data conditions. Furthermore, by using relative differences rather than absolute differences in the calculations, the impact of differences in data magnitude across different dimensions is eliminated, making the evaluation results from different dimensions comparable and thus more accurately determining the optimal data processing target path.

[0072] Continuing with the above embodiments, the neuromorphic AI architecture collects data showing that the current CPU utilization of the edge device is 60%, memory usage is 45%, and network bandwidth is 20Mbps. For heterogeneous sub-data of time-series data, the characteristics are: data dimension 100, strong data correlation, and moderate data change rate; for heterogeneous sub-data of image data, the characteristics are: data dimension 500, weak data correlation, and low data change rate; for heterogeneous sub-data of high-frequency numerical data, the characteristics are: data dimension 10, moderate data correlation, and high data change rate. In the constructed path evaluation index strategy, the processing latency weight is 0.4, the computing power utilization weight is 0.3, and the data transmission stability weight is 0.3. Preset candidate processing paths include a local fast processing path, an edge node collaborative processing path, and a cloud-assisted processing path. Through matching analysis, the heterogeneous sub-data of time series data adopts the edge node collaborative processing path (low processing latency and moderate computing power consumption), the image data adopts the cloud-assisted processing path (which can reduce the local computing power burden and is suitable for large data volume processing), and the high-frequency numerical data adopts the local fast processing path (high real-time requirements and minimum local processing latency). Based on this, the optimal data processing target path for each is obtained.

[0073] Step S130: According to the adaptation of each type of heterogeneous sub-data to its corresponding data processing target path, each type of heterogeneous sub-data is segmented to obtain multiple target data segments corresponding to each type of heterogeneous sub-data. Based on the data real-time performance and data importance of each type of heterogeneous sub-data, the target data segments corresponding to each type of heterogeneous sub-data are prioritized to obtain the data sorting result.

[0074] In some embodiments, the edge computing acceleration system segments each type of heterogeneous sub-data according to a defined data processing target path, dividing it into multiple target data fragments. Simultaneously, the system prioritizes the target data fragments corresponding to each type of heterogeneous sub-data based on their real-time performance and importance. Data real-time performance is measured by the freshness of the data's timestamp, while data importance is measured by the data's impact on core business operations, ultimately yielding the data ranking result.

[0075] Continuing with the above embodiments, based on the obtained high-frequency numerical data heterogeneous sub-data, it is segmented according to the local fast processing path, dividing it into 10 target data segments. Each segment contains vibration frequency values ​​for 10 consecutive seconds. Since this type of data has extremely high real-time requirements and is directly related to whether the equipment will malfunction, all 10 target data segments are set to the highest priority. For time-series data heterogeneous sub-data, it is segmented into 5 target data segments. Based on the data's temporal order, the 2 most recently acquired segments are given high priority, and the remaining 3 are given medium priority. For image data heterogeneous sub-data, it is segmented into 8 target data segments. The 3 segments containing key parts of the equipment are given medium priority, and the remaining 5 are given low priority, resulting in the corresponding data sorting results.

[0076] Step S140: Based on the data sorting results and the current computing power resource data of the edge devices, determine the sharding scheduling strategy, and transmit the target data shards to the corresponding processing units in sequence based on the sharding scheduling strategy.

[0077] The current computing power resource data is the computing power resource data of edge devices collected based on the brain-like AI architecture.

[0078] In some embodiments, the edge computing acceleration system formulates a sharding scheduling strategy based on the data sorting results and the current computing resource data of the edge devices. This strategy comprehensively considers prioritizing the scheduling of high-priority data shards and the current computing resource status to avoid computing overload caused by centralized scheduling of high-priority shards. Subsequently, the system transmits the target data shards sequentially to the corresponding processing units, such as local CPUs, edge servers, and cloud processing centers, according to the sharding scheduling strategy.

[0079] Continuing with the above embodiment, the current computing resources of the edge device are: CPU utilization reduced to 50%, memory usage at 40%, and network bandwidth at 25Mbps. In the data sorting results, high-frequency numerical data fragments have the highest priority, followed by time-series data fragments, and image data fragments have the lowest priority. The fragmentation scheduling strategy determines to first transmit 10 target data fragments of high-frequency numerical data to the local CPU for processing. During transmission, when the local CPU utilization reaches 70%, transmission is paused, and transmission of 2 high-priority time-series data fragments begins to the edge server. Once the local CPU utilization drops below 60%, the remaining high-frequency numerical data fragments continue to be transmitted, and other target data fragments are transmitted according to a similar logic.

[0080] Step S150: For the target data fragment transmitted to the processing unit, based on the current load status of the processing unit and the processing requirements of the target data fragment, determine the computing power allocation resources, and execute the processing task for the target data fragment based on the computing power allocation resources.

[0081] In some embodiments, the edge computing acceleration system can monitor the current load status of the processing unit in real time, such as CPU utilization, memory usage, and task queue length, for the target data shards transmitted to the processing unit. Simultaneously, it determines the computing power allocation resources, including the number of allocated CPU cores, memory size, and computation time, based on the processing requirements of the target data shards, such as computational complexity, storage requirements, and processing latency thresholds. Then, the system executes the processing tasks for the target data shards according to these allocated computing power resources.

[0082] Continuing with the above embodiments, for high-frequency numerical data fragments transmitted to the local CPU, the edge computing acceleration system allocates one CPU core and 50MB of memory to each high-frequency numerical data fragment, with a computation time of 0.5 seconds, ensuring that processing requirements are met without overloading the local CPU. For time-series data fragments transmitted to the edge server, based on the edge server's load and fragment processing requirements, two virtual CPU cores and 100MB of memory are allocated to each fragment to execute processing tasks.

[0083] In the aforementioned edge computing acceleration method based on a neuromorphic AI architecture, heterogeneous data collected by edge devices is classified using a pre-defined classification strategy to obtain multiple heterogeneous sub-data types corresponding to different categories. Based on a pre-constructed path evaluation index strategy, each type of heterogeneous sub-data is matched with multiple pre-defined candidate processing paths to determine the target data processing path for each type of heterogeneous sub-data. The pre-constructed path evaluation index strategy is used to indicate the path quality evaluation index and its corresponding weight. According to the compatibility between each type of heterogeneous sub-data and its corresponding target data processing path, each type of heterogeneous sub-data is segmented to obtain multiple target data segments corresponding to each type of heterogeneous sub-data. The system prioritizes the target data fragments corresponding to each type of heterogeneous sub-data based on their real-time performance and importance, resulting in a data sorting result. Based on the data sorting result and the current computing power resource data of the edge devices, a fragmentation scheduling strategy is determined, and the target data fragments are sequentially transmitted to the corresponding processing units according to the fragmentation scheduling strategy. The current computing power resource data is the computing power resource data of the edge devices collected based on the neuromorphic AI architecture. For the target data fragments transmitted to the processing units, computing power allocation resources are determined based on the current load status of the processing units and the processing requirements of the target data fragments, and processing tasks for the target data fragments are executed based on the allocated computing power resources.

[0084] Thus, by classifying the acquired heterogeneous data and considering the differences in characteristics among different data, this approach changes the previous method of using a uniform and fixed processing path for all data. The classified heterogeneous sub-data, combined with the collected computing resource data, can be matched to find the optimal target data processing path. This solves the problem of related technologies not considering the dynamic changes in computing resources of side devices and avoids unnecessary computation consuming limited computing resources under a fixed path. Furthermore, by using the optimal data processing path to segment the heterogeneous sub-data and prioritizing it based on real-time performance and importance, and determining the segmentation scheduling strategy based on current computing resources, high-priority data segmentation can be processed first, ensuring timely response to real-time tasks. Finally, computing resources are allocated based on the processing unit load and data segmentation processing requirements, dynamically adapting to the processing unit's status and data needs, further optimizing the allocation of computing resources and avoiding low processing efficiency caused by uneven load or unreasonable resource allocation.

[0085] In some embodiments, such as Figure 2 As shown, in step S130, according to the adaptation of each type of heterogeneous sub-data to its corresponding data processing target path, each type of heterogeneous sub-data is segmented to obtain multiple target data segments corresponding to each type of heterogeneous sub-data, including steps S1301 to S1305:

[0086] Step S1301: For any type of heterogeneous sub-data, extract the feature dimensions of any type of heterogeneous sub-data and the capacity parameters of each node in the corresponding data processing target path.

[0087] In some embodiments, the edge computing acceleration system first extracts the feature dimensions for each type of heterogeneous sub-data, including data size, data update frequency, data complexity, and real-time requirement level. Then, it obtains the capacity parameters of each node in the data processing target path corresponding to that type of heterogeneous sub-data, such as the maximum data storage capacity of the node, the data processing capacity per unit time, and the upper limit of inter-node transmission bandwidth. Finally, it associates and maps the extracted feature dimensions with the capacity parameters of each node to construct a feature-capacity mapping set. This set clearly presents the correspondence between the features of heterogeneous sub-data and the capacity of path nodes.

[0088] In one embodiment, for heterogeneous sub-data of high-frequency numerical data, its characteristic dimensions are small data volume (10KB / minute), high data update frequency (100 times / second), low data complexity (single numerical type), and extremely high real-time requirement level (Level 1). Its data processing target path is a local fast processing path, involving a local CPU. The capacity parameters of this node are a maximum data storage capacity of 500KB, a data processing volume of 500 times / second per unit time, and unlimited transmission bandwidth within the node (local processing). The edge computing acceleration system correlates this information to construct a feature capacity mapping set for this type of heterogeneous sub-data. Similarly, corresponding feature capacity mapping sets are constructed for heterogeneous sub-data of time-series data (target path is edge node collaborative processing path) and image data (target path is cloud-assisted processing path).

[0089] Step S1302: Based on the path adaptability corresponding to any type of heterogeneous sub-data, obtain the sharding priority sequence corresponding to any type of heterogeneous sub-data.

[0090] The path fit includes multiple factors, namely, the path fit of each feature dimension of any type of heterogeneous sub-data with the data processing target path corresponding to any type of heterogeneous sub-data.

[0091] In some embodiments, the edge computing acceleration system calculates the path fit degree between each feature dimension and its corresponding data processing target path based on the feature capacity mapping set. This path fit degree measures the degree of matching between the feature dimension and the capacity of the path nodes; a higher fit degree indicates that the data under that feature dimension is more suitable for processing via that path. Then, the system sorts the feature dimensions in descending order of their path fit degrees to obtain a sharding priority sequence. This sequence determines the priority processing order of the sub-data portions corresponding to each feature dimension during sharding.

[0092] Furthermore, the formula for calculating path adaptability is as follows:

[0093]

[0094] Among them, F d,k f represents the path fit between the d-th feature dimension and the k-th data processing target path; d,m c is the value of the m-th parameter in the d-th feature dimension; k,m The capacity parameter value for the m-th node in the target path of the k-th data processing line; q is the number of parameter items; g d,m h is the weight coefficient of the m-th parameter in the d-th feature dimension; k,m The influence coefficient of the capacity parameter of the m-th node in the target path of the k-th data processing.

[0095] This formula calculates the absolute difference between feature dimensions and node capacity using Euclidean distance, while introducing a sigmoid function to strengthen the influence weight of key parameters. This preserves the accuracy of difference quantification and highlights the adaptation contribution of core features, making the adaptation results more closely match real-world scenarios. Furthermore, the introduction of weighting and influence coefficients allows for dynamic adjustment based on power grid business scenarios (such as equipment monitoring and load forecasting), improving the formula's applicability to various scenarios.

[0096] Continuing with the above embodiments, for the feature capacity mapping set of heterogeneous sub-data of high-frequency numerical data, the path fit degree for each feature dimension is calculated. The fit degree between high data update frequency and high local CPU data processing volume per unit time is 0.95; the fit degree between extremely high real-time requirements and low local processing latency is 0.98; the fit degree between small data volume and local CPU storage capacity is 0.8; and the fit degree between low data complexity and local CPU processing power is 0.85. Sorting the fit degree from high to low, the sharding priority sequence is: Real-time requirement level > Data update frequency > Data complexity > Data volume.

[0097] Step S1303: Based on the sharding priority sequence, group the feature dimensions of any type of heterogeneous sub-data according to the number of nodes in the data processing target path corresponding to any type of heterogeneous sub-data, to obtain multiple feature dimension groups of any type of heterogeneous sub-data.

[0098] In some embodiments, the edge computing acceleration system groups the feature dimensions of any type of heterogeneous sub-data according to a sharding priority sequence and the number of nodes in the data processing target path corresponding to any type of heterogeneous sub-data. During grouping, the feature dimensions are sequentially assigned to groups equal to the number of nodes, in descending order of priority, ensuring that the feature dimensions in each group are relatively balanced in priority and that each group corresponds to one node in the path. This grouping allows different nodes to process the sub-data shards corresponding to the feature dimensions of different groups in parallel, thereby improving processing efficiency.

[0099] In one embodiment, the sharding priority sequence for heterogeneous sub-data of time-series data is: data correlation > data update frequency > data volume > data complexity. The target data processing path is an edge node collaborative processing path, involving two nodes (local edge node and adjacent edge node). The edge computing acceleration system divides the feature dimensions into two groups based on the number of nodes (2). The first group includes data correlation and data update frequency (the top two with higher priority), and the second group includes data volume and data complexity (the bottom two with lower priority). Each group corresponds to one node.

[0100] Step S1304: Based on the information entropy density of each feature dimension group and the capacity parameters of each node in the data processing target path, determine the granularity coefficient corresponding to each feature dimension group.

[0101] In some embodiments, the edge computing acceleration system first calculates the information entropy density of each feature dimension group. This information entropy density reflects the degree of information disorder and concentration within the feature dimension group. The higher the density, the more concentrated the information within the group, requiring finer sharding during processing. Then, combining the capacity parameters (such as processing capacity per unit time, storage capacity, etc.) of the corresponding nodes in the data processing target path, the sharding granularity coefficient for each feature dimension group is determined. This sharding granularity coefficient determines the shard size of the sub-data portion corresponding to that feature dimension group; the larger the coefficient, the larger the shard; the smaller the coefficient, the smaller the shard.

[0102] Furthermore, the formula for calculating information entropy density is as follows:

[0103]

[0104] Among them, E t Let E(G) be the information entropy density of the t-th feature dimension group. t ) is the feature dimension group G t The information entropy; Len(d) is the cardinality of the value space of feature dimension d, that is, the number of possible values ​​of this dimension.

[0105] The formula for calculating the particle size coefficient is as follows:

[0106]

[0107] Among them, G t C is the granularity coefficient of the t-th feature dimension group; t,max P represents the maximum capacity parameter (e.g., maximum storage) of the node corresponding to the t-th group; t T represents the average priority of the feature dimensions within the t-th group; tLet be the unit-time processing threshold for the node corresponding to the t-th feature dimension group. This formula, through the product of information entropy density and the logarithm of the node's maximum capacity, can simultaneously reflect the information concentration of the feature group (such as the "critical area proportion" in power grid image data) and the node's carrying capacity limit, ensuring a dual match between the granularity of the segmentation and the information complexity and equipment capacity.

[0108] In one embodiment, the feature dimensions of the heterogeneous sub-data of image data are divided into two groups (corresponding to two nodes in the cloud-assisted processing path). The first group includes image resolution and the proportion of key regions, with an information entropy density of 0.8 (relatively concentrated information). The corresponding node has a processing capacity of 10 images / second and a storage capacity of 10GB. The second group includes image acquisition time and data format, with an information entropy density of 0.4 (relatively dispersed information). The corresponding node has a processing capacity of 20 images / second and a storage capacity of 20GB. The edge computing acceleration system calculates the granularity coefficient of the first group to be 0.3 (smaller granularity, facilitating fine processing) and the granularity coefficient of the second group to be 0.6 (larger granularity, improving processing efficiency).

[0109] Step S1305: Based on the sharding granularity coefficient corresponding to each feature dimension group, shard any type of heterogeneous sub-data to obtain multiple target data shards corresponding to any type of heterogeneous sub-data.

[0110] In some embodiments, the edge computing acceleration system performs sharding processing on any type of heterogeneous sub-data according to the sharding granularity coefficient of each feature dimension group, to obtain multiple target data shards corresponding to any type of heterogeneous sub-data.

[0111] The technical solution of this embodiment involves extracting the feature dimensions of any type of heterogeneous sub-data and the capacity parameters of each node in the corresponding data processing target path for any given type of heterogeneous sub-data; obtaining a sharding priority sequence for any type of heterogeneous sub-data based on the path fit degree; the path fit degree includes multiple parameters, namely the path fit degree between each feature dimension of any type of heterogeneous sub-data and the corresponding data processing target path; grouping each feature dimension of any type of heterogeneous sub-data according to the number of nodes in the corresponding data processing target path based on the sharding priority sequence, resulting in multiple feature dimension groups for any type of heterogeneous sub-data; determining the sharding granularity coefficient for each feature dimension group based on the information entropy density of each feature dimension group and the capacity parameters of each node in the data processing target path; and sharding the heterogeneous sub-data according to the sharding granularity coefficient for each feature dimension group, resulting in multiple target data shards for any type of heterogeneous data.

[0112] Thus, by first establishing a mapping relationship between features and capacity, the compatibility between data and paths is clarified; then, a priority sequence is obtained through path compatibility to ensure that important features are processed first; grouping by the number of nodes provides the basis for parallel processing; and the granularity of sharding is determined based on information entropy density and node capacity to match the shard size with the processing capacity. This achieves refined sharding based on heterogeneous sub-data features and processing path node capacity, improving the targeting and efficiency of sharding processing and reducing processing delays and resource waste caused by unreasonable sharding.

[0113] In some embodiments, step S1305 involves partitioning any type of heterogeneous subdata according to the partitioning granularity coefficient corresponding to each feature dimension group, to obtain multiple target data partitions corresponding to any type of heterogeneous data, including steps S13051 to S13053:

[0114] Step S13051: Based on the granularity coefficient corresponding to each feature dimension group, perform horizontal sharding on the heterogeneous sub-data portion corresponding to each feature dimension group in any type of heterogeneous sub-data to obtain multiple initial data shards.

[0115] In some embodiments, the edge computing acceleration system performs horizontal sharding on the heterogeneous sub-data portion corresponding to each feature dimension group, based on a sharding granularity coefficient. Horizontal sharding refers to dividing a continuous sub-data portion into multiple independent initial data shards according to dimensions such as the order of data records or time series, with each shard retaining the complete feature attributes of that feature dimension group. During the sharding process, the size of each initial data shard is ensured to match the sharding granularity coefficient; the larger the coefficient, the more data a single initial shard contains.

[0116] Furthermore, the initial data sharding calculation formula is as follows:

[0117]

[0118] Where, N t,init S represents the initial number of slices for the t-th feature dimension group; t γ represents the total data volume of the sub-data portion corresponding to the t-th feature dimension group; t G represents the feature importance coefficient of this group (range 0-1); t This is the particle size coefficient for fragmentation; represents the average data size of a single standard slice, and [·] is the rounding up function.

[0119] In one embodiment, the first feature dimension group (including image resolution and key region proportion) of the heterogeneous sub-data of image data corresponds to 100 thermal imaging images of the device, with a granularity coefficient of 0.3. Based on this coefficient, horizontal sharding yields 15 initial data shards, each containing approximately 6-7 images (the total data volume is evenly distributed). The second feature dimension group (including image acquisition time and data format) also corresponds to related auxiliary information of these 100 images, with a granularity coefficient of 0.6. Horizontal sharding yields 8 initial data shards, each containing auxiliary information of 12-13 images.

[0120] Step S13052: Adjust the number of initial data shards based on the preset load threshold of each node in the data processing target path to obtain the adjusted number of shards.

[0121] In some embodiments, the edge computing acceleration system acquires a load threshold preset for each node in the target data processing path. This threshold includes the maximum number of concurrent data shards processed by the node and the maximum amount of data that can be received per unit time. The initial number of data shards is compared with the load threshold. If the initial number of shards exceeds the load threshold, the number of shards needs to be increased to reduce the processing pressure on a single node; if the initial number of shards is much lower than the load threshold, the number of shards can be reduced to improve processing efficiency, ultimately resulting in an adjusted number of shards.

[0122] Furthermore, the adjusted formula for calculating the number of shards is as follows:

[0123]

[0124] Where, N t,adj θ is the number of slices after adjustment for the t-th feature dimension group; t κ is the load threshold of the node corresponding to the t-th feature dimension group; κ is the adjustment coefficient (range 1-2); tanh(·) is the hyperbolic tangent function, used to smooth the adjustment amplitude.

[0125] Continuing with the above embodiments, the preset load threshold for the processing nodes corresponding to the first feature dimension group of image data is 10, which is the maximum number of concurrent processing fragments. The initial number of data fragments is 15. If this threshold is exceeded, the edge computing acceleration system adjusts it to 20 (an increase of 5). The preset load threshold for the processing nodes corresponding to the second feature dimension group is 8, which is the maximum number of concurrent processing fragments. The initial number of data fragments is 8, which is consistent with the threshold and does not require adjustment. The adjusted number of fragments remains 8.

[0126] Step S13053: Based on the adjusted number of shards, split or merge the initial data shards to generate multiple target data shards corresponding to any type of heterogeneous sub-data.

[0127] In some embodiments, the edge computing acceleration system performs secondary splitting or merging of the initial data shards based on the adjusted number of shards. When the adjusted number is greater than the initial number, some of the larger initial shards are split a second time, and the resulting sub-shards retain the feature information of the original shards. When the adjusted number is less than the initial number, multiple smaller initial shards are merged into a new shard, and the merged shard contains the complete data of the original multiple shards. The resulting multiple target data shards must meet the load threshold requirements of each node, and each target data shard must contain the corresponding data fragments of all feature dimension groups.

[0128] Continuing with the above embodiment, the number of slices after adjustment for the first feature dimension group of the image data class is 20, while the initial number is 15. Five initial slices need to be split twice, each split into two, resulting in 20 target data slices. The number of slices after adjustment for the second feature dimension group is 8, consistent with the initial number, requiring no splitting or merging. The edge computing acceleration system combines the 20 target data slices from the first group with the 8 target data slices from the second group according to their correspondence, ultimately obtaining 20 target data slices containing complete image information and auxiliary information (8 of which contain complete corresponding segments from both groups, and 12 contain copied information of the segments split from the first group and the corresponding segments from the second group).

[0129] The technical solution of this embodiment involves horizontally sharding the heterogeneous sub-data portion corresponding to each feature dimension group in any type of heterogeneous sub-data based on the sharding granularity coefficient corresponding to each feature dimension group, thereby obtaining multiple initial data shards; adjusting the number of shards of the initial data shards based on the preset load threshold of each node in the data processing target path, thereby obtaining the adjusted number of shards; and splitting or merging the initial data shards based on the adjusted number of shards to generate multiple target data shards corresponding to any type of heterogeneous sub-data.

[0130] In this way, horizontal sharding ensures the integrity of data within feature dimension groups, and adjustments based on load thresholds prevent node overload or resource idleness. Secondary splitting and merging ensure the rationality of target data sharding and data correlation, thereby achieving refined processing from initial sharding to target data sharding that meets node load requirements.

[0131] In some embodiments, the current load status includes CPU utilization, memory usage, and task queue length, and the processing requirements include computational complexity, storage requirements, and processing latency thresholds; such as Figure 3 As shown, step S150, based on the current load status of the processing unit and the processing requirements of the target data shards, determines the allocation of computing power resources, including steps S1501 to S1505:

[0132] Step S1501: Based on the CPU utilization rate, memory utilization rate and task queue length of the processing unit, and combined with the computational complexity, storage requirements and processing latency threshold of the target data sharding, construct a load demand mapping matrix.

[0133] In some embodiments, the edge computing acceleration system first collects real-time status data (current load status) of the processing units, including CPU utilization, memory usage, and task queue length, to reflect the current load of the processing units. Simultaneously, it extracts key attributes of the target data shards to determine processing requirements, namely computational complexity (measuring the amount of computation required to process the shard), storage requirements (the memory space required for processing), and processing latency threshold (the maximum allowed time to complete the processing). These data are then correlated and integrated to construct a load-demand mapping matrix. The rows of this matrix represent processing units, the columns represent target data shards, and the elements in the matrix reflect the load-demand matching relationship between the corresponding processing unit and the target data shard.

[0134] Furthermore, the formula for calculating the elements of the load demand mapping matrix is ​​as follows:

[0135]

[0136] Among them, M u,s The mapping value between processing unit u and target data fragment s; CPU u CPU utilization rate of processing unit u; C s The computational complexity of partitioning the target data into s pieces; Mem u Memory usage of processing unit u; S s Storage requirements for the target data s s; Queue u T is the length of the task queue for processing unit u; s The processing latency threshold for the target data segment s.

[0137] In one embodiment, taking two processing units as an example, processing unit A has a CPU utilization rate of 60%, a memory utilization rate of 55%, and a task queue length of 8; processing unit B has a CPU utilization rate of 30%, a memory utilization rate of 40%, and a task queue length of 3. The target data shards include two target data shards (shard 1 and shard 2) of high-frequency numerical data, with computational complexities of 0.8 and 0.7 respectively, storage requirements of 60MB and 50MB respectively, and a processing latency threshold of 1s for both; and one target data shard (shard 3) of time-series data, with a computational complexity of 0.5, a storage requirement of 80MB, and a processing latency threshold of 3s. In the load demand mapping matrix constructed by the edge computing acceleration system, the mapping value between processing unit A and shard 1 is 0.65 (indicating a medium matching degree), the mapping value between processing unit A and shard 2 is 0.7, and the mapping value between processing unit A and shard 3 is 0.5; the mapping value between processing unit B and shard 1 is 0.85, the mapping value between processing unit B and shard 2 is 0.9, and the mapping value between processing unit B and shard 3 is 0.75.

[0138] Step S1502: Based on the matching degree between each processing unit and each target data shard in the load demand mapping matrix, determine the resource demand gap coefficient of each processing unit for each target data shard.

[0139] Among them, the resource demand gap coefficient is used to quantify the degree of gap between the processing unit and the resource demand of the target data shard.

[0140] In some embodiments, the edge computing acceleration system calculates the resource requirement gap coefficient for each processing unit to each target data shard based on the load demand mapping matrix. This coefficient quantifies the degree to which the processing unit meets the resource requirements of the target data shard. The smaller the coefficient, the better the existing resources of the processing unit can meet the requirements of the target data shard; the larger the coefficient, the larger the gap, requiring more additional resources to process the target data shard. The calculation also comprehensively considers the differences between the various load indicators of the processing unit and the various requirement indicators of the target data shard.

[0141] Furthermore, the formula for calculating the resource demand gap coefficient is as follows:

[0142]

[0143] Among them, G u,s The resource requirement gap coefficient for processing unit u on target data fragment s.

[0144] Continuing with the above embodiments, for processing unit A and shard 1, the resource demand gap coefficient is calculated based on the information in the load demand mapping matrix. There is a gap between the CPU utilization of processing unit A (60%) and the computational complexity of shard 1 (0.8), the memory utilization of 55% and the storage requirement of 60MB, and the task queue length of 8 and the processing latency threshold of 1s. The overall calculated resource demand gap coefficient is 0.35. Processing unit B and shard 1 have a high degree of matching across their metrics, resulting in a resource demand gap coefficient of 0.15.

[0145] Step S1503: Based on the resource demand gap coefficient, each target data segment is filtered to obtain the target processing unit that meets the preset minimum resource demand gap coefficient, and the target processing unit is used as a candidate binding unit to form an initial binding relationship group.

[0146] In some embodiments, the edge computing acceleration system pre-defines a preset number (e.g., 2 or 3) to represent the maximum number of candidate binding units selected for each target data shard. For each target data shard, the resource requirement gap coefficients of all processing units are sorted, and the top preset number of processing units with the smallest gap coefficients are selected as the target processing units for that target data shard, i.e., candidate binding units. The target data shards are associated with the corresponding candidate binding units to form an initial binding relationship group, providing a basis for subsequent resource allocation.

[0147] Continuing with the above embodiments, with a preset quantity of 2, for shard 1, the resource demand gap coefficients of each processing unit are ranked as follows: processing unit B (0.15) < processing unit A (0.35). Therefore, processing unit B and processing unit A are selected as candidate binding units for shard 1, forming an initial binding relationship group (shard 1 - processing unit B, shard 1 - processing unit A). For shard 2, the gap coefficient of processing unit B is 0.12, and the gap coefficient of processing unit A is 0.28, forming an initial binding relationship group (shard 2 - processing unit B, shard 2 - processing unit A).

[0148] Step S1504: Based on the initial binding relationship group, construct the resource demand load association function between each target data shard and the corresponding candidate binding unit.

[0149] Among them, the resource demand load correlation function is used to quantify the impact of load changes of candidate binding units on the resource demand to meet the corresponding target data shards.

[0150] In some embodiments, the edge computing acceleration system constructs a resource requirement load correlation function for each target data shard and its corresponding candidate binding unit based on an initial binding relationship group. This function quantifies the impact of load changes in candidate binding units (such as increased CPU utilization, increased task queue length, etc.) on meeting the resource requirements of the corresponding target data shard. When the load of a candidate binding unit increases, a large change in the function value indicates that the load change has a significant impact on the satisfaction of the resource requirements of the corresponding target data shard; conversely, a smaller impact indicates a smaller impact.

[0151] Furthermore, the formula for calculating the resource demand load correlation function is as follows:

[0152]

[0153] Among them, F s,u (x1,x2,x3) is the resource demand load correlation function between the target data s and the candidate binding unit u; x1, x2, and x3 are the changes in CPU utilization, memory utilization, and task queue length, respectively; k1, k2, and k3 are the influence weight coefficients, and k1+k2+k3=1.

[0154] Continuing with the above embodiment, for the initial binding relationship between shard 1 and processing unit B, the constructed resource demand load correlation function is F(x) = 0.2x1 + 0.3x2 + 0.5x3, where x1 represents the change in CPU utilization, x2 represents the change in memory utilization, and x3 represents the change in task queue length. This function shows that the change in task queue length has the greatest impact on meeting the resource requirements of shard 1, while the change in CPU utilization has the least impact.

[0155] Step S1505: Based on the resource demand load correlation function, combined with the load changes of candidate binding units and the processing progress of target data shards, determine the computing power allocation resources.

[0156] In some embodiments, the edge computing acceleration system monitors in real time the load changes of candidate bound units (e.g., CPU utilization increasing from 30% to 40%) and the processing progress of the target data shard (e.g., 30% of the processing has been completed). This real-time data is analyzed in conjunction with a load correlation function to adjust the computing resources allocated to the target data shard, thereby determining the computing resource allocation.

[0157] In this embodiment, the current load status includes CPU utilization, memory usage, and task queue length, while the processing requirements include computational complexity, storage requirements, and processing latency thresholds. A load requirement mapping matrix is ​​constructed based on the CPU utilization, memory usage, and task queue length of the processing unit, combined with the computational complexity, storage requirements, and processing latency thresholds of the target data shard. Based on the matching degree between each processing unit and each target data shard in the load requirement mapping matrix, the resource requirement gap coefficient for each processing unit to each target data shard is determined. The resource requirement gap coefficient is used to quantify the degree to which the processing unit fails to meet the resource requirements of the target data shard. Based on the resource demand gap coefficient, each target data shard is screened to obtain target processing units that meet the first preset minimum resource demand gap coefficient. These target processing units are then used as candidate binding units to form an initial binding relationship group. Based on the initial binding relationship group, a resource demand load correlation function is constructed between each target data shard and its corresponding candidate binding unit. The resource demand load correlation function is used to quantify the impact of load changes in candidate binding units on meeting the resource demands of the corresponding target data shards. Based on the resource demand load correlation function, and combined with the load changes of candidate binding units and the processing progress of the target data shards, the computing power allocation resources are determined.

[0158] In this way, by constructing mapping relationships, determining gap coefficients, screening candidate units, establishing correlation functions, and finally determining computing power allocation resources, we can dynamically adapt to changes in the load of processing units and the processing progress of data shards, ensuring that data shards can receive sufficient computing power support while meeting latency requirements, making the allocation of computing power resources more reasonable and efficient.

[0159] In some embodiments, such as Figure 4 As shown, step S1505, based on the resource demand load correlation function, combined with the load changes of candidate binding units and the processing progress of target data shards, determines the computing power allocation resources, including steps S15051 to S15054:

[0160] Step S15051: Analyze the computing resources between the target data shards and candidate binding units in each initial binding relationship group based on the resource demand load correlation function to generate candidate scheme groups.

[0161] Each scheme in the candidate scheme group contains a set of candidate binding units for the computing power allocation value of the target data shard.

[0162] In some embodiments, the edge computing acceleration system can analyze the computing resources between data shards and candidate binding units in each initial binding relationship group based on the resource demand load correlation function to obtain computing power allocation values, and form a candidate scheme group based on the computing power allocation values; each scheme in the candidate scheme group contains a set of computing power allocation values ​​of candidate binding units to the target data shard.

[0163] In some embodiments, the edge computing acceleration system substitutes the relevant data of the target data shards and candidate binding units in each initial binding relationship group into the resource demand load correlation function. The function calculates the computing power allocation value that satisfies the resource requirements of the target data shards, including the number of CPU cores, memory allocation, and computation time. The system combines these computing power allocation values ​​to form candidate scheme groups. Each scheme corresponds to a complete computing power allocation strategy, covering the computing power allocation between all target data shards and their corresponding candidate binding units.

[0164] In one embodiment, the initial binding relationship group includes (Shard 1 - Processing Unit B, Shard 1 - Processing Unit A) and (Shard 2 - Processing Unit B, Shard 2 - Processing Unit A). For Shard 1 and Processing Unit B, the computing power allocation value calculated by substituting into the resource demand load correlation function is 2 CPU cores and 60MB memory; the computing power allocation value for Shard 1 and Processing Unit A is 3 CPU cores and 60MB memory; the computing power allocation value for Shard 2 and Processing Unit B is 1 CPU core and 50MB memory; and the computing power allocation value for Shard 2 and Processing Unit A is 2 CPU cores and 50MB memory. These are combined to form a candidate scheme group, one of which is (Shard 1 - Processing Unit B: 2 cores 60MB, Shard 2 - Processing Unit B: 1 core 50MB), and another is (Shard 1 - Processing Unit A: 3 cores 60MB, Shard 2 - Processing Unit B: 1 core 50MB), etc.

[0165] Step S15052: For each scheme in the candidate scheme group, determine the scheme adaptability index based on the load balancing degree of the candidate binding unit and the demand satisfaction degree of the target data sharding.

[0166] In some embodiments, the edge computing acceleration system calculates the load balancing degree of candidate binding units and the target data shard demand satisfaction degree for each scheme in the candidate scheme group. The load balancing degree measures the evenness of load distribution among processing units; a value closer to 1 indicates a more balanced load. The target data shard demand satisfaction degree measures the extent to which the computing power allocation in the scheme meets the requirements of the target data shard; a value closer to 1 indicates a higher satisfaction degree. Based on these two indicators, a scheme suitability index is calculated using a specific formula, which comprehensively reflects the merits of the scheme.

[0167] Continuing with the above embodiments, for the candidate scheme (shard 1 - processing unit B: 2 cores 60MB, shard 2 - processing unit B: 1 core 50MB), the load balancing degree of processing unit B is 0.8 (because only this unit undertakes the task, the load is concentrated), and the demand satisfaction degrees of shard 1 and shard 2 are 0.9 and 0.95 respectively, resulting in a scheme adaptability index of 0.85. For the candidate scheme (shard 1 - processing unit B: 2 cores 60MB, shard 2 - processing unit A: 2 cores 50MB), the load balancing degree of processing units A and B is 0.9, the shard demand satisfaction degree is 0.9 for both, and the scheme adaptability index is 0.92.

[0168] Step S15053: Sort the scheme adaptability index in descending order and select the scheme with the largest adaptability index as the optimal computing power allocation scheme.

[0169] The optimal computing power allocation scheme is used to output the computing power allocation amount for each processing unit to the target data shard.

[0170] In some embodiments, the edge computing acceleration system sorts all candidate schemes by their scheme fit index in descending order, i.e., from largest to smallest. The scheme with the highest scheme fit index is selected as the optimal computing power allocation scheme. This scheme achieves the best balance between load balancing and demand satisfaction, determining the final processing unit and corresponding computing power allocation for each target data shard.

[0171] Continuing with the above embodiments, the suitability index of each scheme in the candidate scheme group is ranked as follows: Scheme C (0.92) > Scheme A (0.85) > Scheme B (0.8). Therefore, Scheme C is selected as the optimal computing power allocation scheme. This scheme determines that shard 1 is processed by processing unit B, allocated with 2 CPU cores and 60MB of memory; shard 2 is processed by processing unit A, allocated with 2 CPU cores and 50MB of memory.

[0172] Step S15054: Adjust the computing power allocation amount of the optimal computing power allocation scheme based on the load changes of the candidate binding units and the processing progress of the target data shards to obtain the computing power allocation resources.

[0173] In some embodiments, the edge computing acceleration system tracks in real time the load changes of candidate bound units (e.g., increased CPU utilization, increased memory usage) and the processing progress of target data shards (e.g., 50% of the processing has been completed). Based on this real-time information, it determines whether the optimal computing power allocation scheme can still meet the demand. If the load change results in insufficient originally allocated computing power, or the processing progress lags behind expectations, the computing power allocation is adjusted, such as increasing the number of CPU cores or extending the computation time, to ultimately obtain the allocated computing power resources.

[0174] Furthermore, the formula for calculating the computing power allocation adjustment coefficient is as follows:

[0175]

[0176] Where K is the computing power allocation adjustment coefficient; ΔL is the load change rate (the difference between the current load and the initial load); and ΔP is the processing progress rate (the ratio of the current progress to the expected progress). The final adjusted computing power allocation is the original allocation multiplied by (1+K).

[0177] Continuing with the above embodiment, during the execution of the optimal computing power allocation scheme, the CPU utilization rate of processing unit B increased from 40% to 55% (load change), and the processing progress of shard 1 was 40% (lagging behind the expected 50%). After analysis of the edge computing power acceleration system, the CPU core allocation of shard 1 was increased from 2 to 3, while the memory remained unchanged at 60MB, as the final computing power allocation resource.

[0178] The technical solution of this embodiment analyzes the computing resources between the target data shard and candidate binding units in each initial binding relationship group based on a resource demand load correlation function to generate a candidate scheme group. Each scheme in the candidate scheme group contains a set of computing power allocation values ​​of candidate binding units to the target data shard. For each scheme in the candidate scheme group, a scheme suitability index is determined based on the load balancing degree of the candidate binding units and the demand satisfaction degree of the target data shard. The scheme suitability indices are sorted in descending order, and the scheme with the largest scheme suitability index is selected as the optimal computing power allocation scheme. The optimal computing power allocation scheme is used to output the computing power allocation amount of each processing unit to the target data shard. The computing power allocation amount of the optimal computing power allocation scheme is adjusted based on the load changes of the candidate binding units and the processing progress of the target data shard to obtain the computing power allocation resources.

[0179] In this way, the optimal computing power allocation scheme can be selected from multiple candidate schemes and dynamically adjusted according to real-time conditions, ensuring that the allocation of computing power resources always matches the load of processing units and the needs of data sharding, making computing power allocation more accurate and flexible, achieving load balancing between processing units and fully meeting the processing needs of data sharding.

[0180] In another embodiment, such as Figure 5 The diagram illustrates a process flow of an edge computing acceleration method based on a neuromorphic AI architecture, comprising the following steps:

[0181] Step S502: Based on a preset classification strategy, the heterogeneous data collected by the edge device is classified to obtain multiple heterogeneous sub-data corresponding to different categories.

[0182] Step S504: Based on the pre-built path evaluation index strategy, match each type of heterogeneous sub-data with multiple preset candidate processing paths to determine the data processing target path corresponding to each type of heterogeneous sub-data.

[0183] Step S506: For any type of heterogeneous sub-data, extract the feature dimensions of any type of heterogeneous sub-data and the capacity parameters of each node in the corresponding data processing target path.

[0184] Step S508: Based on the path adaptability corresponding to any type of heterogeneous sub-data, obtain the sharding priority sequence corresponding to any type of heterogeneous sub-data.

[0185] Step S510: Based on the sharding priority sequence, group the feature dimensions of any type of heterogeneous sub-data according to the number of nodes in the data processing target path corresponding to any type of heterogeneous sub-data, to obtain multiple feature dimension groups of any type of heterogeneous sub-data.

[0186] Step S512: Based on the information entropy density of each feature dimension group and the capacity parameters of each node in the data processing target path, determine the granularity coefficient corresponding to each feature dimension group.

[0187] Step S514: Based on the granularity coefficient corresponding to each feature dimension group, perform horizontal sharding on the heterogeneous sub-data portion corresponding to each feature dimension group in any type of heterogeneous sub-data to obtain multiple initial data shards.

[0188] Step S516: Adjust the number of initial data shards based on the preset load threshold of each node in the data processing target path to obtain the adjusted number of shards.

[0189] Step S518: Split or merge the initial data fragments based on the adjusted number of fragments to generate multiple target data fragments corresponding to any type of heterogeneous sub-data.

[0190] Step S520: Based on the data sorting results and the current computing power resource data of the edge devices, determine the sharding scheduling strategy, and transmit the target data shards sequentially to the corresponding processing units based on the sharding scheduling strategy.

[0191] Step S522: Based on the CPU utilization rate, memory utilization rate and task queue length of the processing unit, and combined with the computational complexity, storage requirements and processing latency threshold of the target data sharding, construct a load demand mapping matrix.

[0192] Step S524: Based on the matching degree between each processing unit and each target data shard in the load demand mapping matrix, determine the resource demand gap coefficient of each processing unit for each target data shard.

[0193] Step S526: Based on the resource demand gap coefficient, each target data segment is filtered to obtain the target processing unit that meets the preset minimum resource demand gap coefficient, and the target processing unit is used as a candidate binding unit to form an initial binding relationship group.

[0194] Step S528: Based on the initial binding relationship group, construct the resource demand load association function between each target data shard and the corresponding candidate binding unit.

[0195] Step S530: Analyze the computing resources between the target data shards and candidate binding units in each initial binding relationship group based on the resource demand load correlation function to generate candidate scheme groups.

[0196] Step S532: For each scheme in the candidate scheme group, determine the scheme adaptability index based on the load balancing degree of the candidate binding unit and the demand satisfaction degree of the target data sharding.

[0197] Step S534: Sort the scheme adaptability index in descending order and select the scheme with the largest adaptability index as the optimal computing power allocation scheme.

[0198] Step S536: Adjust the computing power allocation of the optimal computing power allocation scheme based on the load changes of the candidate binding units and the processing progress of the target data shards to obtain computing power allocation resources, and execute the processing task for the target data shards based on the computing power allocation resources.

[0199] It should be noted that the specific limitations of the above steps can be found in the above description of the specific limitations of an edge computing acceleration method based on a brain-like AI architecture, and will not be repeated here.

[0200] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0201] Based on the same inventive concept, this application also provides a brain-inspired AI architecture-based edge computing acceleration device for implementing the aforementioned brain-inspired AI architecture-based edge computing acceleration method. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the brain-inspired AI architecture-based edge computing acceleration device provided below can be found in the limitations of the brain-inspired AI architecture-based edge computing acceleration method described above, and will not be repeated here.

[0202] In one exemplary embodiment, such as Figure 6 As shown, an edge computing acceleration device based on a neuromorphic AI architecture is provided, including: a classification module 610, a path planning module 620, a sharding and sorting module 630, a scheduling and transmission module 640, and a computing power allocation and task execution module 650, wherein:

[0203] The classification module 610 is used to classify the heterogeneous data collected by the edge device based on a preset classification strategy to obtain multiple heterogeneous sub-data corresponding to different categories.

[0204] The path planning module 620 is used to match each type of heterogeneous sub-data with multiple preset candidate processing paths based on a pre-built path evaluation index strategy, and determine the data processing target path corresponding to each type of heterogeneous sub-data; the pre-built path evaluation index strategy is used to indicate the path quality evaluation index and the corresponding index weight.

[0205] The sharding and sorting module 630 is used to shard each type of heterogeneous sub-data according to the adaptation of each type of heterogeneous sub-data to its corresponding data processing target path, thereby obtaining multiple target data shards corresponding to each type of heterogeneous sub-data, and prioritizing the target data shards corresponding to each type of heterogeneous sub-data based on the data real-time performance and data importance of each type of heterogeneous sub-data, thereby obtaining the data sorting result.

[0206] The scheduling and transmission module 640 is used to determine a sharding scheduling strategy based on the data sorting result and the current computing power resource data of the edge device, and to transmit the target data in shards to the corresponding processing units in sequence based on the sharding scheduling strategy; the current computing power resource data is the computing power resource data of the edge device collected based on the neuromorphic AI architecture.

[0207] The computing power allocation and task execution module 650 is used to determine computing power allocation resources for the target data shards transmitted to the processing unit based on the current load status of the processing unit and the processing requirements of the target data shards, and to execute processing tasks for the target data shards based on the computing power allocation resources.

[0208] In one embodiment, the sharding and sorting module 630 is specifically configured to: extract the feature dimensions of any type of heterogeneous sub-data and the capacity parameters of each node in the corresponding data processing target path for any type of heterogeneous sub-data; obtain a sharding priority sequence for any type of heterogeneous sub-data based on the path fit degree corresponding to the any type of heterogeneous sub-data; the path fit degree includes multiple parameters, namely, the path fit degree between each feature dimension of the any type of heterogeneous sub-data and the data processing target path corresponding to the any type of heterogeneous sub-data; group each feature dimension of the any type of heterogeneous sub-data according to the number of nodes in the data processing target path corresponding to the any type of heterogeneous sub-data based on the sharding priority sequence, to obtain multiple feature dimension groups of the any type of heterogeneous sub-data; determine the sharding granularity coefficient corresponding to each feature dimension group based on the information entropy density of each feature dimension group and the capacity parameters of each node in the data processing target path; and shard the any type of heterogeneous sub-data according to the sharding granularity coefficient corresponding to each feature dimension group, to obtain multiple target data shards corresponding to the any type of heterogeneous sub-data.

[0209] In one embodiment, the sharding and sorting module 630 is specifically used to horizontally shard the heterogeneous sub-data portion corresponding to each feature dimension group in any type of heterogeneous sub-data based on the sharding granularity coefficient corresponding to each feature dimension group, to obtain multiple initial data shards; adjust the number of shards of the initial data shards based on the preset load threshold of each node in the data processing target path, to obtain the adjusted number of shards; and split or merge the initial data shards based on the adjusted number of shards to generate multiple target data shards corresponding to any type of heterogeneous sub-data.

[0210] In one embodiment, the path adaptability calculation formula is as follows:

[0211]

[0212] Among them, F d,k f represents the path fit between the d-th feature dimension and the k-th data processing target path; d,m c is the value of the m-th parameter in the d-th feature dimension; k,m The capacity parameter value for the m-th node in the target path of the k-th data processing line is given; q is the number of parameter items; g d,m h is the weight coefficient of the m-th parameter in the d-th feature dimension; k,m The influence coefficient of the capacity parameter of the m-th node in the target path of the k-th data processing.

[0213] In one embodiment, the current load status includes CPU utilization, memory usage, and task queue length, and the processing requirements include computational complexity, storage requirements, and processing latency thresholds. The computing power allocation and task execution module 650 is specifically used to construct a load requirement mapping matrix based on the CPU utilization, memory usage, and task queue length of the processing unit, combined with the computational complexity, storage requirements, and processing latency thresholds of the target data shard. Based on the matching degree between each processing unit and each target data shard in the load requirement mapping matrix, a resource requirement gap coefficient is determined for each processing unit to each target data shard. The resource requirement gap coefficient is used to quantify the processing unit's ability to meet the requirements of the target data shard. The resource demand gap is assessed; based on the resource demand gap coefficient, each target data shard is screened to obtain target processing units that meet the first preset minimum resource demand gap coefficient, and these target processing units are used as candidate binding units to form an initial binding relationship group; based on the initial binding relationship group, a resource demand load correlation function is constructed between each target data shard and its corresponding candidate binding unit; the resource demand load correlation function is used to quantify the impact of the load change of the candidate binding unit on meeting the resource demand of the corresponding target data shard; based on the resource demand load correlation function, combined with the load change of the candidate binding unit and the processing progress of the target data shard, the computing power allocation resources are determined.

[0214] In one embodiment, the computing power allocation and task execution module 650 is specifically used to analyze the computing power resources between the target data shard and the candidate binding unit in each initial binding relationship group based on the resource demand load correlation function, and generate a candidate scheme group; each scheme in the candidate scheme group includes a set of computing power allocation values ​​of the candidate binding unit to the target data shard; for each scheme in the candidate scheme group, a scheme suitability index is determined based on the load balancing degree of the candidate binding unit and the demand satisfaction degree of the target data shard; the scheme suitability index is sorted in descending order, and the scheme with the largest scheme suitability index is selected as the optimal computing power allocation scheme; the optimal computing power allocation scheme is used to output the computing power allocation amount of each processing unit to the target data shard; the computing power allocation amount of the optimal computing power allocation scheme is adjusted based on the load change of the candidate binding unit and the processing progress of the target data shard to obtain the computing power allocation resources.

[0215] In one embodiment, the resource demand load correlation function is:

[0216]

[0217] Among them, F s,u(x1,x2,x3) is the resource demand load correlation function between the target data fragment s and the candidate binding unit u; x1, x2, and x3 are the changes in CPU utilization, memory utilization, and task queue length, respectively; k1, k2, and k3 are the influence weight coefficients.

[0218] The modules in the aforementioned edge computing acceleration device based on a neuromorphic AI architecture can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0219] In one exemplary embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 7 As shown, the computer device includes a processor, memory, input / output interfaces, a communication interface, a display unit, and an input device. The processor, memory, and input / output interfaces are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interfaces. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The input / output interfaces are used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, Near Field Communication (NFC), or other technologies. When the computer program is executed by the processor, it implements an edge computing acceleration method based on a brain-like AI architecture. The display unit is used to form a visually visible image and can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0220] Those skilled in the art will understand that Figure 7 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0221] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0222] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0223] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0224] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0225] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0226] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0227] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for accelerating edge computing power based on a brain-like AI architecture, characterized in that, The method includes: Based on a preset classification strategy, heterogeneous data collected by edge devices is classified to obtain multiple heterogeneous sub-data corresponding to different categories. Based on a pre-built path evaluation index strategy, each type of heterogeneous sub-data is matched with multiple preset candidate processing paths to determine the data processing target path corresponding to each type of heterogeneous sub-data; the pre-built path evaluation index strategy is used to indicate the path quality evaluation index and the corresponding index weight. Based on the adaptation of each type of heterogeneous sub-data to its corresponding data processing target path, each type of heterogeneous sub-data is segmented to obtain multiple target data segments corresponding to each type of heterogeneous sub-data. The target data segments corresponding to each type of heterogeneous sub-data are then prioritized based on the data real-time performance and data importance, resulting in a data ranking result. This includes: for any type of heterogeneous sub-data, extracting the feature dimensions of that type of heterogeneous sub-data and the capacity parameters of each node in the corresponding data processing target path; and obtaining the segment priority sequence corresponding to that type of heterogeneous sub-data based on the path adaptation degree. The path adaptation degree includes multiple parameters, namely... Each feature dimension of any type of heterogeneous sub-data is assigned a path fit with the target data processing path corresponding to that type of heterogeneous sub-data. Based on the sharding priority sequence, each feature dimension of the heterogeneous sub-data is grouped according to the number of nodes in the target data processing path corresponding to that type of heterogeneous sub-data, resulting in multiple feature dimension groups for that type of heterogeneous sub-data. Based on the information entropy density of each feature dimension group and the capacity parameter of each node in the target data processing path, a sharding granularity coefficient corresponding to each feature dimension group is determined. According to the sharding granularity coefficient corresponding to each feature dimension group, the heterogeneous sub-data is sharded to obtain multiple target data shards corresponding to that type of heterogeneous data. Based on the data sorting results and the current computing power resource data of the edge device, a sharding scheduling strategy is determined, and the target data is sharded and transmitted sequentially to the corresponding processing unit based on the sharding scheduling strategy; the current computing power resource data is the computing power resource data of the edge device collected based on the neuromorphic AI architecture; For the target data shard transmitted to the processing unit, based on the current load status of the processing unit and the processing requirements of the target data shard, computing power allocation resources are determined, and processing tasks for the target data shard are executed based on the computing power allocation resources.

2. The method according to claim 1, characterized in that, The step of partitioning any type of heterogeneous sub-data according to the partitioning granularity coefficient corresponding to each of the feature dimension groups to obtain multiple target data partitions corresponding to the any type of heterogeneous sub-data includes: Based on the granularity coefficient of each feature dimension group, the heterogeneous sub-data portion corresponding to each feature dimension group in any type of heterogeneous sub-data is horizontally sharded to obtain multiple initial data shards. The number of initial data shards is adjusted based on the preset load threshold of each node in the data processing target path to obtain the adjusted number of shards. Based on the adjusted number of shards, the initial data shards are split or merged to generate multiple target data shards corresponding to any type of heterogeneous sub-data.

3. The method according to claim 1, characterized in that, The formula for calculating path adaptability is as follows: ； in, Indicates the first The feature dimension and the first Path fit of the target path for data processing; For the first The first feature dimension Item parameter value; For the first Data processing target path number The capacity parameter values ​​for each node; The number of parameter items; For the first The first feature dimension Weighting coefficients of the item parameters; For the first Data processing target path number The influence coefficient of the capacity parameter of each node.

4. The method according to claim 1, characterized in that, The current load status includes CPU utilization, memory usage, and task queue length; the processing requirements include computational complexity, storage requirements, and processing latency threshold. The process of determining computing power allocation resources based on the current load status of the processing unit and the processing requirements of the target data shard includes: Based on the CPU utilization, memory usage, and task queue length of the processing unit, and combined with the computational complexity, storage requirements, and processing latency threshold of the target data shard, a load demand mapping matrix is ​​constructed. Based on the matching degree between each processing unit and each target data shard in the load demand mapping matrix, the resource demand gap coefficient of each processing unit for each target data shard is determined; the resource demand gap coefficient is used to quantify the degree to which the processing unit meets the resource demand of the target data shard. Based on the resource demand gap coefficient, each target data segment is filtered to obtain target processing units that meet the first preset minimum resource demand gap coefficient, and the target processing units are used as candidate binding units to form an initial binding relationship group. Based on the initial binding relationship group, a resource demand load correlation function is constructed between each target data shard and its corresponding candidate binding unit; the resource demand load correlation function is used to quantify the degree of impact of the load change of the candidate binding unit on meeting the resource demand of the corresponding target data shard; Based on the resource demand load correlation function, combined with the load changes of the candidate binding units and the processing progress of the target data shards, the computing power allocation resources are determined.

5. The method according to claim 4, characterized in that, The determination of computing power allocation resources based on the resource demand load correlation function, combined with the load changes of the candidate binding units and the processing progress of the target data shards, includes: Based on the resource demand load correlation function, the computing resources between the target data shard and the candidate binding unit in each initial binding relationship group are analyzed to generate a candidate scheme group; each scheme in the candidate scheme group contains a set of computing power allocation values ​​of the candidate binding unit to the target data shard; For each scheme in the candidate scheme group, a scheme suitability index is determined based on the load balancing degree of the candidate binding unit and the demand satisfaction degree of the target data sharding; The schemes are sorted in descending order of their suitability index, and the scheme with the largest suitability index is selected as the optimal computing power allocation scheme. The optimal computing power allocation scheme is used to output the computing power allocation amount of each processing unit to the target data shard. The computing power allocation amount of the optimal computing power allocation scheme is adjusted based on the load changes of the candidate binding units and the processing progress of the target data shards to obtain the computing power allocation resources.

6. The method according to claim 4, characterized in that, The resource demand load correlation function is: ； in, Slice the target data With candidate binding unit Resource demand load correlation function; These represent the changes in CPU utilization, memory usage, and task queue length, respectively. To influence the weighting coefficients.

7. An edge computing acceleration device based on a brain-like AI architecture, characterized in that, The device includes: The classification module is used to classify heterogeneous data collected by edge devices based on a preset classification strategy, and obtain multiple heterogeneous sub-data corresponding to different categories. The path planning module is used to match each type of heterogeneous sub-data with multiple preset candidate processing paths based on a pre-built path evaluation index strategy, and determine the data processing target path corresponding to each type of heterogeneous sub-data; the pre-built path evaluation index strategy is used to indicate the path quality evaluation index and the corresponding index weight. The sharding and sorting module is used to shard each type of heterogeneous sub-data according to the adaptation of each type of heterogeneous sub-data to its corresponding data processing target path, thereby obtaining multiple target data shards corresponding to each type of heterogeneous sub-data. The module also sorts the target data shards corresponding to each type of heterogeneous sub-data based on the data real-time performance and data importance of each type of heterogeneous sub-data, thereby obtaining the data sorting result. The sharding and sorting module is specifically used to: extract the feature dimensions of any type of heterogeneous sub-data and the capacity parameters of each node in the corresponding data processing target path for any given type of heterogeneous sub-data; obtain a sharding priority sequence for any type of heterogeneous sub-data based on the path fit degree corresponding to the given type of heterogeneous sub-data; the path fit degree includes multiple parameters, namely, the path fit degree between each feature dimension of any type of heterogeneous sub-data and the corresponding data processing target path; based on the sharding priority sequence, group each feature dimension of any type of heterogeneous sub-data according to the number of nodes in the corresponding data processing target path for any given type of heterogeneous sub-data to obtain multiple feature dimension groups for any type of heterogeneous sub-data; determine the sharding granularity coefficient corresponding to each feature dimension group based on the information entropy density of each feature dimension group and the capacity parameters of each node in the data processing target path; and shard the given type of heterogeneous sub-data according to the sharding granularity coefficient corresponding to each feature dimension group to obtain multiple target data shards corresponding to the given type of heterogeneous sub-data. The scheduling and transmission module is used to determine a sharding scheduling strategy based on the data sorting result and the current computing power resource data of the edge device, and to transmit the target data in shards to the corresponding processing units in sequence based on the sharding scheduling strategy; the current computing power resource data is the computing power resource data of the edge device collected based on the neuromorphic AI architecture; The computing power allocation and task execution module is used to determine the computing power allocation resources for the target data shards transmitted to the processing unit based on the current load status of the processing unit and the processing requirements of the target data shards, and to execute the processing tasks for the target data shards based on the computing power allocation resources.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images