Data processing method and system for AI practical training computing power workstation

Abstract

The invention provides a data processing method and system for an AI practical training computing power workstation, and relates to the technical field of data processing, a four-camera matrix sensor group is built, practical training four-camera data acquisition is performed on practical training area information, practical training four-camera acquisition data is obtained, function analysis optimization and data feature processing are performed according to the practical training four-camera acquisition data, and the AI practical training computing power workstation is obtained. Obtaining practical training feature processing data; performing parallel mapping computing power processing state analysis on the practical training four-camera acquisition data and the practical training feature processing data to obtain computing power processing state judgment information; computing power resource allocation analysis is carried out according to the computing power processing state judgment information, computing power resource allocation is carried out according to the computing power resource allocation analysis information, and computing power resource allocation data is obtained. Multi-dimensional high-quality data collection, safety grading processing and dynamic precise computing power allocation can be achieved.

Description

Technical Field

[0001] This invention proposes a data processing method and system for an AI training computing power workstation, which relates to the field of data processing technology, specifically to the field of data processing technology for AI training computing power workstations. Background Technology

[0002] Current AI training largely relies on single or dual-camera devices to collect data, which suffers from incomplete field of view coverage and insufficient detail capture. Furthermore, the lack of targeted functional optimization and security-level processing for the collected data easily leads to poor data quality or privacy risks. Simultaneously, during multi-task parallel processing in training, computing resource allocation is mostly static, failing to perceive the real-time computing load status of different tasks. This results in redundant computing power being idle and scarce computing power not being replenished in a timely manner, thus wasting resources and affecting the efficiency of training tasks. The current computing power allocation does not incorporate differentiated management based on data security levels, potentially leading to security risks such as the misappropriation of computing power for sensitive or confidential data. Summary of the Invention

[0003] This invention provides a data processing method and system for an AI training computing workstation to solve the above-mentioned problems:

[0004] This invention proposes a data processing method and system for an AI training computing workstation, the method comprising:

[0005] S1. Build a four-camera matrix sensor group, collect training data from the training area using the four cameras, obtain training data collected by the four cameras, perform functional analysis and optimization and data feature processing based on the training data collected by the four cameras, and obtain training feature processing data.

[0006] S2. Analyze the computing power processing status by performing parallel mapping between the training four-camera acquisition data and the training feature processing data, and obtain computing power processing status determination information.

[0007] S3. Perform computing power resource allocation analysis based on the computing power processing status determination information, and allocate computing power resources based on the computing power resource allocation analysis information to obtain computing power resource allocation data.

[0008] Further, S1 includes:

[0009] Acquire training area information and build a four-camera matrix sensor group based on the training area information;

[0010] Acquire the preset training target data, and perform four-camera data acquisition on the preset training target data of the training area information based on the four-camera matrix sensor group to obtain the training four-camera acquisition data;

[0011] Feature acquisition data analysis was performed on the data collected by the four cameras in the training exercise to obtain feature acquisition analysis data;

[0012] Based on the feature acquisition and analysis data, the quad-camera acquisition function is adjusted to obtain quad-camera acquisition function adjustment data;

[0013] Based on the adjustment data of the four-camera acquisition function, adjust the four-camera data acquisition to obtain the updated data of the four-camera training.

[0014] The training data feature analysis and processing were performed on the updated data from the four cameras to obtain the training feature processing data.

[0015] Furthermore, the step of performing feature acquisition data analysis on the data collected by the four training cameras to obtain feature acquisition analysis data includes:

[0016] The training quad-camera data was divided into target quad-camera function data, obtaining main camera function data, ultra-wide-angle function data, telephoto function data and auxiliary function data;

[0017] The main camera function data, ultra-wide-angle function data, telephoto function data, and auxiliary function data are compared and analyzed with the preset training target data to obtain the four-camera function comparison data.

[0018] Based on the comparison data of the four camera functions, determine the unqualified function data;

[0019] The non-compliant functional data refers to the feature acquisition and analysis data.

[0020] Furthermore, the step of performing training data feature analysis on the updated data from the four training cameras to obtain training feature-processed data includes:

[0021] The training feature data is extracted from the updated data of the four training cameras to obtain the training feature extraction data.

[0022] The extracted training features are classified to obtain training feature category data;

[0023] Perform security feature analysis on the training feature category data to obtain security feature analysis data;

[0024] Based on the security feature analysis data, feature security extraction is performed on the training feature category data to obtain training feature category security extraction data.

[0025] The data extracted from the training feature categories is the training feature processing data.

[0026] Furthermore, the step of performing security feature analysis and processing on the training feature category data to obtain security feature analysis data includes:

[0027] Feature extraction of key information of preset encryption is performed on the training feature category data to obtain training encrypted feature extraction data;

[0028] Obtain the proportion of the extracted encryption features in the preset encryption key information to obtain the sensitivity coefficient of the training features;

[0029] The sensitivity coefficients of the training features are compared with preset sensitivity thresholds and preset confidentiality thresholds to obtain sensitivity level comparison results and confidentiality level comparison results.

[0030] The results of the sensitivity level comparison and the classification level comparison are the security feature analysis data.

[0031] Further, S2 includes:

[0032] The data collected by the four training cameras and the training feature processing data are mapped in parallel to obtain parallel mapped data.

[0033] Obtain computing power processing information for parallel mapping data;

[0034] Based on the computing power processing information, the computing power processing status of the parallel mapping data is determined to obtain computing power processing status determination information.

[0035] The computing power processing status determination information triggers the computing power processing allocation command.

[0036] Further, the step of determining the computing power processing status of the parallel mapping data based on the computing power processing information to obtain computing power processing status determination information includes:

[0037] Group the parallel mapping data to obtain multiple combinations of parallel mappings;

[0038] Obtain CPU core utilization data for parallel mapping data of each barracks mapping combination;

[0039] The ratio of the CPU core utilization rate data to the preset CPU core utilization rate data is obtained to obtain the computing power utilization coefficient.

[0040] The computing power utilization coefficient is compared with the preset computing power utilization threshold to obtain the computing power utilization comparison result;

[0041] When the computing power utilization comparison result is that the computing power utilization coefficient is greater than the preset computing power utilization threshold, a computing power processing status shortage judgment is made.

[0042] When the computing power utilization comparison result is that the computing power utilization coefficient is less than or equal to the preset computing power utilization threshold, a computing power processing state redundancy determination is performed.

[0043] Further, S3 includes:

[0044] When a computing power allocation command is triggered, the ratio of the computing power utilization coefficient to the preset computing power utilization threshold is obtained based on the computing power processing status shortage judgment information, and the computing power shortage coefficient is obtained.

[0045] The ratio of the computing power utilization coefficient to the preset computing power utilization threshold is obtained based on the computing power processing status redundancy determination information, thus obtaining the computing power redundancy coefficient.

[0046] Based on the computing power redundancy coefficient, the computing power shortage coefficient is analyzed to obtain computing power allocation analysis data.

[0047] Based on the computing power allocation analysis data, computing power resources are allocated to obtain computing power resource allocation data.

[0048] Further, the step of performing computing power allocation analysis on the computing power shortage coefficient based on the computing power redundancy coefficient to obtain computing power allocation analysis data includes:

[0049] The computing power redundancy coefficients of multiple parallel mapping combinations are sorted from largest to smallest to obtain a computing power redundancy sequence.

[0050] Obtain the difference between the computing power redundancy coefficient and the computing power shortage coefficient in each computing power redundancy sequence to obtain the computing power allocation difference;

[0051] Obtain the parallel mapping combination corresponding to the computing power redundancy coefficient with the smallest computing power allocation difference, and determine the computing power allocation source;

[0052] By allocating computing resources based on the computing redundancy coefficient of the computing power allocation source and the computing power shortage coefficient, computing power resource allocation data is obtained.

[0053] Furthermore, the system includes:

[0054] The quad-camera data processing module is used to build a quad-camera matrix sensor group, collect training quad-camera data on the training area, obtain training quad-camera collected data, perform functional analysis and optimization and data feature processing based on the training quad-camera collected data, and obtain training feature processing data.

[0055] The computing power status analysis module is used to perform parallel mapping of the data collected by the four training cameras and the training feature processing data to analyze the computing power processing status and obtain computing power processing status determination information.

[0056] The computing power allocation module is used to perform computing power resource allocation analysis based on the computing power processing status determination information, allocate computing power resources based on the computing power resource allocation analysis information, and obtain computing power resource allocation data.

[0057] The beneficial effects of this invention are as follows: This invention solves the technical problems of single data collection, static computing power allocation, and inefficient resource utilization in traditional AI training; it realizes the fully automated processing from data collection to computing power scheduling, which can adapt to the training needs of multiple scenarios without human intervention; it improves the overall operating efficiency of AI training, the comprehensiveness of data processing, and the utilization rate of computing power resources; and it reduces the operational errors, waste of computing power resources, and data security risks caused by human intervention. Attached Figure Description

[0058] Figure 1 A schematic diagram of a data processing method for an AI training computing workstation;

[0059] Figure 2 This is a schematic diagram of the four-camera monitoring system. Detailed Implementation

[0060] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0061] In one embodiment of the present invention, a data processing method and system for an AI training computing workstation is proposed, the method comprising:

[0062] S1. Build a four-camera matrix sensor group, collect training data from the training area using the four cameras, obtain training data collected by the four cameras, perform functional analysis and optimization and data feature processing based on the training data collected by the four cameras, and obtain training feature processing data.

[0063] S2. Analyze the computing power processing status by performing parallel mapping between the training four-camera acquisition data and the training feature processing data, and obtain computing power processing status determination information.

[0064] S3. Based on the computing power processing status determination information, perform computing power resource allocation analysis, and allocate computing power resources according to the computing power resource allocation analysis information to obtain computing power resource allocation data, such as... Figure 1 As shown.

[0065] The working principle and technical effects of the above technical solution are as follows: This method constructs a closed-loop process of data acquisition, status analysis, and resource allocation based on the core requirements of the AI ​​training computing power workstation. Multi-dimensional data acquisition of the training area is completed through a four-camera matrix sensor group. After functional optimization and feature security processing, the raw acquired data is transformed into usable data that meets the training requirements. The acquired data and processed data are associated through a parallel mapping method to accurately analyze the real-time status of computing power processing (scarcity / redundancy / reasonable idleness). Based on the status judgment results, targeted computing power resource allocation is triggered to achieve dynamic distribution of computing power among different training tasks.

[0066] This invention solves the technical problems of single data collection, static computing power allocation, and inefficient resource utilization in traditional AI training; it realizes the fully automated processing from data collection to computing power scheduling, which can adapt to the training needs of multiple scenarios without human intervention; it improves the overall operating efficiency of AI training, the comprehensiveness of data processing and the utilization rate of computing power resources; and it reduces the operational errors, waste of computing power resources and data security risks caused by human intervention.

[0067] In one embodiment of the present invention, S1 includes:

[0068] Obtain the training area information, and build a four-camera matrix sensor group based on the training area information, such as... Figure 2 As shown; a quad-camera matrix is ​​a hardware architecture and algorithm combination scheme for multiple cameras to work together. It is mainly used in smart terminals (such as mobile phones, drones, security cameras) or professional imaging equipment. The core is to achieve more comprehensive imaging capabilities than single or dual cameras by using four cameras with different functions, combined with algorithms such as image fusion, perspective stitching, and parameter coordination.

[0069] Acquire the preset training target data, and perform four-camera data acquisition on the preset training target data of the training area information based on the four-camera matrix sensor group to obtain the training four-camera acquisition data;

[0070] Feature acquisition data analysis was performed on the data collected by the four cameras in the training exercise to obtain feature acquisition analysis data;

[0071] The quad-camera acquisition function is adjusted based on the feature acquisition and analysis data to obtain quad-camera acquisition function adjustment data; the quad-camera acquisition function adjustment involves adjusting the function parameters of the unqualified function data until it is converted into qualified function data.

[0072] Based on the adjustment data of the four-camera acquisition function, adjust the four-camera data acquisition to obtain the updated data of the four-camera training.

[0073] The training data feature analysis and processing were performed on the updated data from the four cameras to obtain the training feature processing data.

[0074] The working principle and technical effects of the above technical solution are as follows: Based on the spatial layout and environmental conditions of the training area, a four-camera matrix sensor group consisting of four functionally differentiated cameras is constructed to ensure full-view acquisition coverage required for the training; according to the preset acquisition objectives (such as panoramic coverage, detail capture, etc.), data of the training area is synchronously acquired through the four-camera matrix to obtain initial training four-camera acquisition data; feature acquisition analysis is performed on the acquired data to identify unqualified items in the four-camera functions that do not meet the preset objectives; camera parameters (such as focal length, exposure, angle of view, etc.) are adjusted for the unqualified functional data, and optimized training four-camera updated data is obtained by re-acquiring; finally, feature extraction, classification, and security processing are performed on the updated data to form training feature processing data that meets the training requirements of the training model. The entire process follows an iterative optimization logic of construction, acquisition, analysis, adjustment, re-acquisition, and processing to ensure that data quality is gradually improved.

[0075] This method addresses the technical problems of incomplete perspective, lack of detail, and limited functionality in traditional single- or dual-camera data acquisition, as well as the lack of security processing for the acquired data. It achieves multi-dimensional, high-quality acquisition and secure and compliant processing of training data, ensuring that the data fully covers training needs while meeting security control requirements. It improves the integrity, accuracy, and security of training data, reduces the risks of poor model training results and data leakage due to poor data quality, and also reduces the manual cost of secondary data processing.

[0076] In one embodiment of the present invention, the step of performing feature acquisition data analysis on the training quad-camera data to obtain feature acquisition analysis data includes:

[0077] The training quad-camera data was divided into target quad-camera function data, obtaining main camera function data, ultra-wide-angle function data, telephoto function data and auxiliary function data;

[0078] The main camera function data, ultra-wide-angle function data, telephoto function data, and auxiliary function data are compared and analyzed with the preset training target data to obtain the four-camera function comparison data.

[0079] Based on the comparison data of the four camera functions, determine the unqualified function data;

[0080] The non-compliant functional data refers to the feature acquisition and analysis data.

[0081] The working principle and technical effect of the above technical solution are as follows: Based on the functional positioning of each camera in the quad-camera matrix (main camera, ultra-wide-angle, telephoto, and auxiliary camera), the data collected by the four cameras in the training exercise is functionally divided, clarifying the data types corresponding to each type of camera. The data collected for each functional category is then compared one by one with preset target data (e.g., the main camera needs to clearly capture the main subject, the ultra-wide-angle needs to completely cover the panorama, etc.) to analyze whether each type of data meets the preset standards. Data items that do not meet the standards are filtered out based on the comparison results; these are considered unqualified functional data and are used as feature collection and analysis data. Through the logic of functional division, item-by-item comparison, and unqualified identification, the functional shortcomings in the quad-camera data collection process are accurately located.

[0082] This method solves the technical problems of ambiguous positioning of data acquisition function in four-camera arrays and inability to accurately identify the acquisition defects of individual cameras; it enables accurate evaluation and defect location of the acquisition function of each camera in the four-camera array, and clarifies the specific optimization direction; it improves the pertinence and efficiency of adjusting the four-camera acquisition function, avoids the waste of resources caused by blind adjustment; and it reduces the risk of affecting the quality of training data due to the failure to detect acquisition function defects in time.

[0083] In one embodiment of the present invention, the step of performing training data feature analysis processing on the updated data from the four training cameras to obtain training feature processing data includes:

[0084] The training feature data is extracted from the updated data of the four training cameras to obtain the training feature extraction data.

[0085] The training feature extraction data is classified to obtain training feature category data; the training feature category data includes panoramic data, detail data, face tracking data, and first-person view shooting data, etc.

[0086] Perform security feature analysis on the training feature category data to obtain security feature analysis data;

[0087] Based on the security feature analysis data, feature security extraction is performed on the training feature category data to obtain training feature category security extraction data.

[0088] The data extracted from the training feature categories is the training feature processing data.

[0089] The working principle and technical effects of the above technical solution are as follows: Core features (such as target contours, motion trajectories, environmental parameters, etc.) related to the training task are extracted from the updated data of the four training cameras to obtain training feature extraction data; the extracted feature data is classified into training feature categories such as panoramic data, detailed data, face tracking data, and first-person perspective shooting data according to training needs; security feature analysis is performed on each category of data to identify the privacy information or sensitive content contained therein; based on the security analysis results, the feature category data undergoes security processing such as desensitization and encryption to extract feature data that retains both core training value and meets security requirements, i.e., securely extracted training feature category data. This approach balances the training value and security compliance of the data, achieving a complete chain of processing from extraction, classification, security analysis, to secure extraction.

[0090] This method addresses the technical issues of inaccurate feature extraction and chaotic classification of training data, as well as the failure to consider data security risks. It achieves accurate feature extraction, standardized classification, and secure and compliant processing of training data, ensuring that the data meets the needs of model training without leaking privacy or sensitive information. It improves the usability, standardization, and security of training data, reduces the low model training efficiency caused by inaccurate feature extraction and the compliance risks caused by data security issues, and also reduces the cost of feature data processing.

[0091] In one embodiment of the present invention, the step of performing security feature analysis and processing on the training feature category data to obtain security feature analysis data includes:

[0092] Feature extraction of key information of preset encryption is performed on the training feature category data to obtain training encrypted feature extraction data;

[0093] Obtain the proportion of the extracted encryption features in the preset encryption key information to obtain the sensitivity coefficient of the training features;

[0094] The sensitivity coefficients of the training features are compared with preset sensitivity thresholds and preset confidentiality thresholds to obtain sensitivity level comparison results and confidentiality level comparison results.

[0095] The results of the sensitivity level comparison and the classification level comparison are the security feature analysis data.

[0096] The working principle and technical effect of the above technical solution are as follows: Predefined key encrypted information (such as facial information, classified identifiers, privacy data fields, etc.) is identified; feature data containing this key information is extracted from the training feature category data to obtain training encrypted feature extraction data; the proportion of the extracted key encrypted information feature data to the total amount of predefined key encrypted information is calculated to obtain the training feature sensitivity coefficient, quantifying the sensitivity of the data; the sensitivity coefficient is compared with predefined sensitivity thresholds and confidentiality thresholds to determine whether the data reaches a sensitive or confidentiality level, forming sensitivity level comparison results and confidentiality level comparison results, which are used as security feature analysis data. Through the logic of key information extraction, sensitivity coefficient calculation, and threshold comparison, accurate determination of data security level is achieved.

[0097] This method addresses the technical problem of insufficient targeted security processing due to the lack of quantitative standards and ambiguous judgment results in determining the security level of training data. It enables quantitative assessment of the sensitivity of training data and accurate classification of security levels. It improves the targetedness and accuracy of data security processing, ensuring that data of different security levels receive appropriate protection measures. It reduces the problem of over-protection or under-protection caused by misjudgment of security levels, thus ensuring data security while avoiding the impact of over-processing on the value of data training.

[0098] In one embodiment of the present invention, S2 includes:

[0099] The data collected by the four training cameras and the training feature processing data are mapped in parallel to obtain parallel mapped data.

[0100] Obtain computing power processing information for parallel mapping data;

[0101] Based on the computing power processing information, the computing power processing status of the parallel mapping data is determined to obtain computing power processing status determination information.

[0102] The computing power processing status determination information triggers the computing power processing allocation command.

[0103] The working principle and technical effect of the above technical solution are as follows: Parallel mapping processing is performed on the data collected by the four training cameras and the training feature processing data. This involves decomposing the two types of data into multiple parallel-computable subtasks and allocating them to different computing units (such as CPU cores) to form parallel mapped data. Then, computing power processing information, including the load status of each computing unit and task processing progress, is collected from the parallel mapped data. Based on the collected computing power processing information, the current computing power processing status is determined (scarce / redundant / reasonably idle). According to the status determination result, corresponding computing power allocation instructions are triggered. If the status is scarce, a computing power replenishment instruction is triggered; if it is redundant, a computing power distribution instruction is triggered. The entire process, through the logic of parallel mapping, information collection, status determination, and instruction triggering, achieves real-time perception of computing power status and precise triggering of allocation instructions.

[0104] This method solves the technical problems of traditional computing power scheduling, such as the inability to perceive the computing power processing status in real time and the lag in triggering allocation instructions. It realizes real-time monitoring of computing power processing status and automatic triggering of allocation instructions, ensuring that computing power scheduling responds to the needs of training tasks in a timely manner. It improves the real-time performance and accuracy of computing power scheduling. It reduces problems such as task processing delays and waste of computing power resources caused by untimely perception of computing power status, and ensures the smooth operation of training tasks.

[0105] In one embodiment of the present invention, the step of determining the computing power processing status of parallel mapping data based on the computing power processing information to obtain computing power processing status determination information includes:

[0106] Group the parallel mapping data to obtain multiple combinations of parallel mappings;

[0107] Obtain CPU core utilization data for parallel mapping data of each barracks mapping combination;

[0108] The ratio of the CPU core utilization rate data to the preset CPU core utilization rate data is obtained to obtain the computing power utilization coefficient.

[0109] The computing power utilization coefficient is compared with the preset computing power utilization threshold to obtain the computing power utilization comparison result;

[0110] When the computing power utilization comparison result is that the computing power utilization coefficient is greater than the preset computing power utilization threshold, a computing power processing status shortage judgment is made.

[0111] When the computing power utilization comparison result is that the computing power utilization coefficient is less than or equal to the preset computing power utilization threshold, a computing power processing state redundancy determination is performed.

[0112] When the computing power utilization comparison result is that the computing power utilization coefficient is greater than the preset reasonable idle threshold and less than or equal to the preset computing power utilization threshold, it is determined that the computing power processing state is reasonably idle.

[0113] The working principle and technical effect of the above technical solution are as follows: Parallel mapping data is grouped according to dimensions such as task type and data security level to form multiple parallel mapping combinations, ensuring consistency in computing power requirements for each group. Then, CPU core utilization data corresponding to each parallel mapping combination is collected to reflect the actual utilization of computing resources by that group of tasks. By calculating the ratio of the actual CPU core utilization rate to the preset CPU core utilization data, a computing power utilization coefficient is obtained, quantifying the computing power utilization degree of each group of tasks. The computing power utilization coefficient is compared with preset reasonable idle thresholds and computing power utilization thresholds, respectively, to classify three states: a coefficient greater than the computing power utilization threshold indicates computing power shortage; a coefficient between the reasonable idle threshold and the computing power utilization threshold indicates reasonable computing power idleness; and a coefficient less than or equal to the reasonable idle threshold indicates computing power redundancy. This process, through the logic of grouping, utilization rate collection, coefficient calculation, and threshold comparison, achieves refined determination of computing power status.

[0114] This invention solves the technical problem that the determination of computing power processing status is based on a single dimension and is coarsely divided, which cannot accurately reflect the differences in computing power requirements of different task groups. It realizes a refined determination of computing power status for different parallel mapping combinations, distinguishing between three states: scarce, reasonably idle, and redundant. It improves the accuracy and refinement of computing power status determination, reduces the problem of unreasonable computing power allocation caused by coarse status determination, avoids unnecessary allocation of tasks in reasonably idle states, and ensures the stable utilization of computing power resources.

[0115] In one embodiment of the present invention, S3 includes:

[0116] When a computing power allocation command is triggered, the ratio of the computing power utilization coefficient to the preset computing power utilization threshold is obtained based on the computing power processing status shortage judgment information, and the computing power shortage coefficient is obtained.

[0117] The ratio of the computing power utilization coefficient to the preset computing power utilization threshold is obtained based on the computing power processing status redundancy determination information, thus obtaining the computing power redundancy coefficient.

[0118] Based on the computing power redundancy coefficient, the computing power shortage coefficient is analyzed to obtain computing power allocation analysis data.

[0119] Based on the computing power allocation analysis data, computing power resources are allocated to obtain computing power resource allocation data.

[0120] The working principle and technical effect of the above technical solution are as follows: When a computing power allocation command is triggered, for parallel mapping combinations with scarce computing power, the ratio of their computing power utilization coefficient to a preset computing power utilization threshold is calculated to obtain a computing power scarcity coefficient, quantifying the degree of scarcity; for parallel mapping combinations with redundant computing power, the ratio of their computing power utilization coefficient to a preset computing power utilization threshold is calculated to obtain a computing power redundancy coefficient, quantifying the degree of redundancy; based on the correspondence between the computing power redundancy coefficient and the computing power scarcity coefficient, the matching degree between the computing power resources available for allocation in each redundant combination and the computing power resources required by the scarce combination is analyzed; according to the matching degree analysis results, the idle computing power resources of the redundant combination are allocated to the scarce combination, forming computing power resource allocation data. Following the logic of coefficient calculation, matching analysis, and resource allocation, the precise flow of computing power resources is achieved.

[0121] This method solves the technical problems of uneven resource allocation, ineffective utilization of redundant computing power, and failure to replenish scarce computing power in a timely manner in traditional computing power allocation. It realizes the dynamic allocation of computing power resources among different parallel mapping combinations, making full use of redundant computing power and accurately replenishing scarce computing power. It improves the overall utilization rate of computing power resources and the processing efficiency of training tasks, ensuring that all kinds of training tasks can obtain appropriate computing power support. It reduces the waste of some task processing delays and idle computing power resources caused by uneven computing power allocation, and optimizes the overall configuration efficiency of computing power resources.

[0122] In one embodiment of the present invention, the step of performing computing power allocation analysis on the computing power shortage coefficient based on the computing power redundancy coefficient to obtain computing power allocation analysis data includes:

[0123] Screen parallel mapping combinations that have publicly disclosed data levels and exclude parallel mapping combinations that have sensitive or classified data levels.

[0124] The computing power redundancy coefficients of multiple parallel mapping combinations are sorted from largest to smallest to obtain a computing power redundancy sequence.

[0125] Obtain the difference between the computing power redundancy coefficient and the computing power shortage coefficient in each computing power redundancy sequence to obtain the computing power allocation difference;

[0126] Obtain the parallel mapping combination corresponding to the computing power redundancy coefficient with the smallest (positive) computing power allocation difference, and determine the computing power allocation source;

[0127] By allocating computing resources based on the computing redundancy coefficient of the computing power allocation source and the computing power shortage coefficient, computing power resource allocation data is obtained.

[0128] The working principle and technical effect of the above technical solution are as follows: Parallel mapping combinations with adjustable computing power are selected based on data security levels. Only redundant computing power combinations with publicly available data levels are retained, while combinations with sensitive / confidential data levels are excluded to prevent the misappropriation of computing power for high-security data. The selected redundant combinations are then sorted from largest to smallest computing power redundancy coefficient to form a computing power redundancy sequence, clarifying the priority of the adjustable computing power for each combination. The difference between the computing power redundancy coefficient of each redundant combination in the sequence and the computing power shortage coefficient of the scarce combination is calculated. The redundant combination with the smallest positive difference is found, as its idle computing power best matches the needs of the scarce combination, and it is identified as the computing power allocation source. The redundant computing power of this allocation source is used to supplement the computing power of the scarce combination, achieving precise allocation of computing power resources. This process, through the logic of security screening, sorting, difference calculation, allocation source determination, and allocation execution, achieves efficient computing power allocation under the premise of security compliance.

[0129] This method addresses the security risks of misappropriation of computing power for sensitive / classified data due to the lack of consideration for data security levels during computing power allocation, as well as the technical issues of mismatch between allocated computing power and demand. It achieves precise matching and efficient allocation of computing resources under the premise of security and compliance, ensuring both data security and adequate computing power support for scarce tasks. It improves the security, accuracy, and efficiency of computing power allocation, balancing data security and resource utilization efficiency. It reduces data security risks caused by computing power allocation and minimizes problems such as poor allocation results and resource waste due to low computing power matching, further optimizing the quality of computing power resource configuration.

[0130] According to one embodiment of the present invention, the system includes:

[0131] The quad-camera data processing module is used to build a quad-camera matrix sensor group, collect training quad-camera data on the training area, obtain training quad-camera collected data, perform functional analysis and optimization and data feature processing based on the training quad-camera collected data, and obtain training feature processing data.

[0132] The computing power status analysis module is used to perform parallel mapping of the data collected by the four training cameras and the training feature processing data to analyze the computing power processing status and obtain computing power processing status determination information.

[0133] The computing power allocation module is used to perform computing power resource allocation analysis based on the computing power processing status determination information, allocate computing power resources based on the computing power resource allocation analysis information, and obtain computing power resource allocation data.

[0134] The working principle and technical effects of the above technical solution are as follows: Based on the core requirements of the AI ​​training computing power workstation, this system constructs a closed-loop process encompassing data acquisition, status analysis, and resource allocation. Multi-dimensional data acquisition of the training area is completed through a four-camera matrix sensor group. After functional optimization and feature security processing, the raw acquired data is transformed into usable data that meets the training requirements. The acquired data and processed data are associated through a parallel mapping method to accurately analyze the real-time status of computing power processing (scarcity / redundancy / reasonable idleness). Based on the status judgment results, targeted computing power resource allocation is triggered, realizing the dynamic distribution of computing power among different training tasks.

[0135] This invention solves the technical problems of single data collection, static computing power allocation, and inefficient resource utilization in traditional AI training; it realizes the fully automated processing from data collection to computing power scheduling, which can adapt to the training needs of multiple scenarios without human intervention; it improves the overall operating efficiency of AI training, the comprehensiveness of data processing and the utilization rate of computing power resources; and it reduces the operational errors, waste of computing power resources and data security risks caused by human intervention.

[0136] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A data processing method for an AI training computing workstation, characterized in that, The method includes: S1. Build a four-camera matrix sensor group, collect training data from the training area using the four cameras, obtain training data collected by the four cameras, perform functional analysis and optimization and data feature processing based on the training data collected by the four cameras, and obtain training feature processing data. S2. Analyze the computing power processing status by performing parallel mapping between the training four-camera acquisition data and the training feature processing data, and obtain computing power processing status determination information. S3. Perform computing power resource allocation analysis based on the computing power processing status determination information, and allocate computing power resources based on the computing power resource allocation analysis information to obtain computing power resource allocation data.

2. The data processing method for an AI training computing workstation according to claim 1, characterized in that, S1 includes: Acquire training area information and build a four-camera matrix sensor group based on the training area information; Acquire the preset training target data, and perform four-camera data acquisition on the preset training target data of the training area information based on the four-camera matrix sensor group to obtain the training four-camera acquisition data; Feature acquisition data analysis was performed on the data collected by the four cameras in the training exercise to obtain feature acquisition analysis data; Based on the feature acquisition and analysis data, the quad-camera acquisition function is adjusted to obtain quad-camera acquisition function adjustment data; Based on the adjustment data of the four-camera acquisition function, adjust the four-camera data acquisition to obtain the updated data of the four-camera training. The training data feature analysis and processing were performed on the updated data from the four cameras to obtain the training feature processing data.

3. The data processing method for an AI training computing workstation according to claim 2, characterized in that, The process of performing feature acquisition data analysis on the data collected by the four cameras in the training exercise to obtain feature acquisition analysis data includes: The training quad-camera data was divided into target quad-camera function data, obtaining main camera function data, ultra-wide-angle function data, telephoto function data and auxiliary function data; The main camera function data, ultra-wide-angle function data, telephoto function data, and auxiliary function data are compared and analyzed with the preset training target data to obtain the four-camera function comparison data. Based on the comparison data of the four camera functions, determine the unqualified function data; The non-compliant functional data refers to the feature acquisition and analysis data.

4. The data processing method for an AI training computing workstation according to claim 2, characterized in that, The process of performing training data feature analysis on the updated data from the four training cameras to obtain training feature-processed data includes: The training feature data is extracted from the updated data of the four training cameras to obtain the training feature extraction data. The extracted training features are classified to obtain training feature category data; Perform security feature analysis on the training feature category data to obtain security feature analysis data; Based on the security feature analysis data, feature security extraction is performed on the training feature category data to obtain training feature category security extraction data. The data extracted from the training feature categories is the training feature processing data.

5. The data processing method for an AI training computing workstation according to claim 4, characterized in that, The process of performing security feature analysis and processing on the training feature category data to obtain security feature analysis data includes: Feature extraction of key information of preset encryption is performed on the training feature category data to obtain training encrypted feature extraction data; Obtain the proportion of the extracted encryption features in the preset encryption key information to obtain the sensitivity coefficient of the training features; The sensitivity coefficients of the training features are compared with preset sensitivity thresholds and preset confidentiality thresholds to obtain sensitivity level comparison results and confidentiality level comparison results. The results of the sensitivity level comparison and the classification level comparison are the security feature analysis data.

6. The data processing method for an AI training computing workstation according to claim 1, characterized in that, S2 includes: The data collected by the four training cameras and the training feature processing data are mapped in parallel to obtain parallel mapped data. Obtain computing power processing information for parallel mapping data; Based on the computing power processing information, the computing power processing status of the parallel mapping data is determined to obtain computing power processing status determination information. The computing power processing status determination information triggers the computing power processing allocation command.

7. The data processing method for an AI training computing workstation according to claim 6, characterized in that, The step of determining the computing power processing status of the parallel mapping data based on the computing power processing information to obtain computing power processing status determination information includes: Group the parallel mapping data to obtain multiple combinations of parallel mappings; Obtain CPU core utilization data for parallel mapping data of each barracks mapping combination; The ratio of the CPU core utilization rate data to the preset CPU core utilization rate data is obtained to obtain the computing power utilization coefficient. The computing power utilization coefficient is compared with the preset computing power utilization threshold to obtain the computing power utilization comparison result; When the computing power utilization comparison result is that the computing power utilization coefficient is greater than the preset computing power utilization threshold, a computing power processing status shortage judgment is made. When the computing power utilization comparison result is that the computing power utilization coefficient is less than or equal to the preset computing power utilization threshold, a computing power processing state redundancy determination is performed.

8. The data processing method for an AI training computing workstation according to claim 1, characterized in that, S3 includes: When a computing power allocation command is triggered, the ratio of the computing power utilization coefficient to the preset computing power utilization threshold is obtained based on the computing power processing status shortage judgment information, and the computing power shortage coefficient is obtained. The ratio of the computing power utilization coefficient to the preset computing power utilization threshold is obtained based on the computing power processing status redundancy determination information, thus obtaining the computing power redundancy coefficient. Based on the computing power redundancy coefficient, the computing power shortage coefficient is analyzed to obtain computing power allocation analysis data. Based on the computing power allocation analysis data, computing power resources are allocated to obtain computing power resource allocation data.

9. The data processing method for an AI training computing workstation according to claim 8, characterized in that, The step of performing computing power allocation analysis on the computing power shortage coefficient based on the computing power redundancy coefficient to obtain computing power allocation analysis data includes: The computing power redundancy coefficients of multiple parallel mapping combinations are sorted from largest to smallest to obtain a computing power redundancy sequence. Obtain the difference between the computing power redundancy coefficient and the computing power shortage coefficient in each computing power redundancy sequence to obtain the computing power allocation difference; Obtain the parallel mapping combination corresponding to the computing power redundancy coefficient with the smallest computing power allocation difference, and determine the computing power allocation source; By allocating computing resources based on the computing redundancy coefficient of the computing power allocation source and the computing power shortage coefficient, computing power resource allocation data is obtained.

10. A data processing system for an AI training computing workstation, characterized in that, The system includes: The quad-camera data processing module is used to build a quad-camera matrix sensor group, collect training quad-camera data on the training area, obtain training quad-camera collected data, perform functional analysis and optimization and data feature processing based on the training quad-camera collected data, and obtain training feature processing data. The computing power status analysis module is used to perform parallel mapping of the data collected by the four training cameras and the training feature processing data to analyze the computing power processing status and obtain computing power processing status determination information. The computing power allocation module is used to perform computing power resource allocation analysis based on the computing power processing status determination information, allocate computing power resources based on the computing power resource allocation analysis information, and obtain computing power resource allocation data.

Images