Adaptive communication resource arrangement and efficiency optimization system based on synesthesia intelligent algorithm storage integration

Abstract

The application relates to the technical field of wireless communication and algorithm network cross, and discloses an adaptive communication resource arrangement and efficiency optimization system based on a sense and intelligence storage integration, a sense and intelligence storage full-dimension resource unified abstract model is constructed, normalized quantitative representation of four types of heterogeneous resources of communication time-frequency-space, sensing detection, algorithm and storage is realized, technical obstacles that different dimension resources are not unified and cannot be jointly scheduled are solved; a global resource state benchmark library is constructed through air load step loading and filtering denoising, hardware inherent overhead and noise interference are stripped, and the accuracy of effective load calculation is guaranteed; an asymmetric resource weight system is established, hardware rated capacity and service load level quantitative weight coefficients are combined, and a multi-dimensional efficiency benchmark self-learning matrix is constructed by fusing resource coupling constraint correction, coupling relationships such as sense mutual exclusion, algorithm transmission positive correlation and storage algorithm mapping are accurately quantified, and the problem that resource coupling effects are difficult to quantify is solved.

Description

Technical Field

[0001] This invention relates to the technical field of the intersection of wireless communication and computing networks, specifically to an adaptive communication resource orchestration and performance optimization system based on the integration of sensing, intelligent computing, and storage. Background Technology

[0002] As 6G mobile communication technology evolves towards integrated air-space-ground-sea communication and deep convergence of sensing, computing, and communication, the deep coupling of four types of resources—communication, sensing, computing, and storage—has become the core architecture of future edge networks. In typical scenarios such as the Industrial Internet, smart transportation, low-altitude security, and immersive XR, the network not only needs to provide high-speed and reliable wireless communication transmission, but also needs to simultaneously support diverse services such as environmental radar perception, edge AI inference, and hierarchical storage of massive amounts of data. Sensing data acquisition, communication backhaul processing, real-time computing power analysis, and hierarchical caching of storage constitute a complete end-to-end service link, and the collaborative scheduling capability of these four types of resources directly determines the network's service quality and overall operational efficiency. Current integrated sensing, intelligent computing, and storage networks exhibit typical characteristics such as strong resource heterogeneity, complex coupling relationships, and large dynamic fluctuations in services. Traditional resource management models based on independent scheduling and local optimization are no longer suitable for the operational requirements of integrated systems. Existing technologies suffer from several technical defects: First, resource scheduling is fragmented, generally only covering the joint optimization of two-dimensional wireless resources (communication and sensing), without incorporating computing and storage resources into a unified scheduling system. This leads to a mismatch between supply and demand of resources across the entire link from sensing and detection to data transmission, computing processing, and data storage. The benefits of communication resource scheduling are constrained by the computing and storage bottleneck, failing to unleash the synergistic gains of integrated sensing, intelligent computing, and storage. Second, scheduling strategies are static and fixed. Resource reuse, power allocation, and time slot avoidance in sensing and communication often rely on preset rules or local optimization, without dynamically adjusting strategies based on the real-time load status of computing and storage resources. This can easily lead to increased inter-sensor interference and low resource utilization in dynamic service scenarios. Third, it lacks a comprehensive linkage and orchestration mechanism, making it impossible to automatically complete joint decisions on sensing time slot planning, communication time-frequency space allocation, computing node matching, and storage resource deployment based on global resource status and business QoS requirements. In dynamic scenarios, resource reconfiguration latency is high, and end-to-end business performance is difficult to guarantee. Fourth, the performance evaluation dimension is singular, and the optimization target only focuses on local indicators such as communication spectrum efficiency and energy efficiency. A multi-dimensional system-level performance evaluation model of communication-sensing-computing-storage has not been established, making it impossible to achieve the overall optimal performance of the integrated system. Fifth, it lacks forward prediction for sudden scenarios. For typical events such as sudden access of sensing tasks and migration of cold and hot data, relying solely on feedback adjustment has significant lag and is prone to causing instantaneous fluctuations in business performance. Sixth, it has not deeply adapted to the characteristics of the storage-computing integrated architecture, making it unable to match the deployment mode of near-memory computing and cold and hot data stratification. The efficiency of sensor data interaction between storage and computing nodes is low, which restricts the overall processing performance of the system.

[0003] Chinese invention patent CN106385709B, published on December 20, 2022, discloses a resource scheduling method and device. This method involves informing the terminal of the subcarrier interval used during communication, the number of symbols in each time slot, and the length of each symbol. When communication resources need to be allocated to the terminal, the method indicates the position of the first time slot in a plurality of aggregated time slots and the number of such time slots, thus indicating the allocated communication resources to the terminal. This invention effectively reduces signaling overhead during resource allocation when multiple terminals use different subcarrier intervals, and ensures that users accurately know the time resources allocated to them, improving system throughput. However, the above technical solutions lack a clear resource scheduling dimension, generally only covering the joint optimization of two-dimensional wireless resources (communication and sensing), without incorporating computing power and storage resources into a unified scheduling system. This leads to a mismatch between supply and demand of resources across the entire link of sensing detection, data transmission, computing power processing, and data storage. The benefits of communication resource scheduling are constrained by the computing and storage bottleneck, failing to unleash the synergistic gains of integrated sensing, intelligent computing, and storage. Summary of the Invention

[0004] (a) Technical problems to be solved To address the several technical shortcomings of existing technologies: First, resource scheduling is fragmented, generally only covering the joint optimization of two-dimensional wireless resources (communication and sensing), without incorporating computing and storage resources into a unified scheduling system. This leads to a mismatch between supply and demand of resources across the entire link of sensing detection, data transmission, computing processing, and data storage. The benefits of communication resource scheduling are constrained by computing and storage bottlenecks, failing to unleash the synergistic gains of integrated sensing, intelligent computing, and storage. Second, scheduling strategies are statically fixed. Resource reuse, power allocation, and time slot avoidance in sensing and communication often rely on preset rules or local optimizations, without dynamically adjusting strategies based on the real-time load status of computing and storage resources. This easily leads to increased mutual interference between sensing and communication and low resource utilization in dynamic service scenarios. Third, there is a lack of a full-dimensional linkage orchestration mechanism, making it impossible to automatically complete joint decisions on sensing time slot planning, communication time-frequency space allocation, computing node matching, and storage resource deployment based on global resource status and service QoS requirements. This results in high latency for resource reconfiguration in dynamic scenarios, making it difficult to guarantee end-to-end service performance. Fourth, efficiency... The evaluation dimensions are singular, and the optimization targets only focus on local indicators such as communication spectrum efficiency and energy efficiency. A multi-dimensional system-level performance evaluation model encompassing communication, sensing, computing power, and storage has not been established, making it impossible to achieve globally optimal overall system performance. Fifth, there is a lack of forward-looking prediction for sudden scenarios. For typical events such as sudden access to sensing tasks and migration of hot and cold data, relying solely on feedback-based adjustments has significant lags and can easily cause instantaneous fluctuations in business performance. Sixth, it is not deeply adapted to the characteristics of the integrated storage-computing architecture, and cannot match the deployment mode of near-memory computing and hot-cold data tiering. The low efficiency of sensor data interaction between storage and computing nodes restricts the overall processing performance of the system. To address these issues, the system should construct a unified orchestration system for all-dimensional resources of sensory intelligent computing and storage. Through asymmetric weighted load transfer and multi-dimensional self-learning benchmarks, it should solve the problem of supply and demand mismatch of heterogeneous resources. Combined with an event-driven forward-looking compensation mechanism, it should significantly improve the response speed in sudden scenarios, achieve globally optimal overall system performance, deeply adapt to the integrated storage-computing architecture, and significantly improve network operation stability and business service quality.

[0005] (II) Technical Solution This invention is achieved through the following technical solution: an adaptive communication resource orchestration and performance optimization system based on integrated sensing, intelligent computing, and storage, the system comprising: The module for abstracting and constructing communication benchmarks for synesthetic computing and storage resources builds a unified abstract model of synesthetic intelligent computing and storage resources across all dimensions. It normalizes and quantifies communication time-frequency-space resources, sensing and detection resources, computing power resources, and storage resources, collects various resource operation parameters under no-load benchmark conditions, and constructs a global resource status benchmark library. The multi-dimensional performance self-learning and communication weight configuration module allocates asymmetric resource weight coefficients based on the hardware capabilities and service carrying priorities of each resource node during the benchmark calibration phase, records benchmark performance indicators under stable business scenarios, and constructs a multi-dimensional performance benchmark self-learning matrix by combining time-frequency resource allocation, perception accuracy requirements, computing power load rate, storage occupancy rate and service QoS parameters. The real-time load monitoring and communication performance calculation module collects the current actual resource usage data and service performance data of each node in real time during the dynamic operation phase. It calls the global resource status benchmark library to deduct the benchmark idle overhead to obtain the current actual effective load. The current actual effective load is compared with the benchmark value in the multi-dimensional performance benchmark self-learning matrix to calculate the real-time load deviation of each dimension of resources and the overall system performance deviation. The adaptive communication orchestration and dynamic load transfer module calculates the absolute value of the real-time load deviation difference between associated resource dimensions and the absolute value of the overall system performance deviation. When the absolute value of the real-time load deviation difference is greater than the dynamic balance adaptive threshold, or the absolute value of the overall system performance deviation is greater than the dynamic balance adaptive threshold, the module locks the configuration parameters of the high-weight core resource dimension according to the asymmetric resource weight coefficient, calculates the resource fine-tuning compensation amount of the low-weight associated dimension to perform load smooth transfer and performance convergence, and updates the multi-dimensional performance benchmark self-learning matrix. When the absolute value of both the real-time load deviation difference and the absolute value of the overall system performance deviation are not greater than the dynamic balance adaptive threshold, the module smoothly updates the current actual effective load to the multi-dimensional performance benchmark self-learning matrix. The event-driven communication feedforward and performance compensation module, when a preset triggering event is detected, calls the multi-dimensional performance benchmark self-learning matrix, matches the feature parameters of the preset triggering event, calculates the expected performance change caused by the event, and converts the expected performance change into a feedforward resource orchestration compensation signal and superimposes it into the corresponding communication resource scheduling link.

[0006] Preferably, a unified abstract model of all-dimensional resources for intelligent computing and storage is constructed, which normalizes and quantifies communication time-frequency-space resources, sensing and detection resources, computing power resources, and storage resources, collects various resource operation parameters under idle benchmark conditions, and constructs a global resource status benchmark library, including the following steps: S11. In the system's no-load initialization state, control the four types of resource nodes—communication, sensing, computing power, and storage—to load in a stepped manner within a preset load range, and collect the original operating parameters of each node. S12. Perform spectrum analysis and noise filtering on the original operating parameters to separate and shield the inherent periodic fluctuation components of the hardware, and obtain the noise-reduced reference no-load parameters. S13. Use a fitting algorithm to perform curve fitting on the denoised baseline idle parameters to generate continuous load-idle overhead baseline curves for resources in each dimension, and integrate them to form a global resource status baseline library.

[0007] Preferably, during the benchmark calibration phase, asymmetric resource weight coefficients are allocated based on the hardware capabilities and service carrying priorities of each resource node. Benchmark performance indicators under stable service scenarios are recorded. A multi-dimensional performance benchmark self-learning matrix is ​​constructed by combining time-frequency resource allocation, perception accuracy requirements, computing load rate, storage occupancy rate, and service QoS parameters. The specific steps for allocating asymmetric resource weight coefficients include: S21. Obtain the rated capacity set of four types of resource nodes: communication, sensing, computing power, and storage, and extract the maximum rated capacity value among them; S22. Divide the rated capacity of each resource node by the maximum rated capacity value to obtain the basic weight factor of each node; S23. Assign a level correction factor based on the bearing level in the end-to-end business link according to the resource dimension, and multiply the basic weight factor by the level correction factor to obtain the final asymmetric resource weight coefficient.

[0008] Preferably, during the benchmark calibration phase, asymmetric resource weight coefficients are allocated based on the hardware capabilities and service carrying priorities of each resource node. Benchmark performance indicators under stable service scenarios are recorded. Combining time-frequency resource allocation, perception accuracy requirements, computing load rate, storage occupancy rate, and service QoS parameters, a multi-dimensional performance benchmark self-learning matrix is ​​constructed. The construction of the multi-dimensional performance benchmark self-learning matrix also incorporates a resource coupling constraint correction step. Based on the mutual exclusion constraints between communication bandwidth and sensing time slots, the positive correlation constraints between computing load and data transmission volume, and the mapping constraints between storage capacity and cache hit rate, the coupling gain coefficient and loss coefficient under different resource combinations are calculated; and the benchmark performance index in the multidimensional performance benchmark self-learning matrix is ​​scaled and corrected according to the coupling gain coefficient and loss coefficient.

[0009] Preferably, during the dynamic operation phase, the current actual resource occupancy data and business performance data of each node are collected in real time. The global resource status benchmark library is called to deduct the benchmark idle overhead to obtain the current actual effective load. The current actual effective load is compared with the benchmark value in the multi-dimensional performance benchmark self-learning matrix to calculate the real-time load deviation and the overall system performance deviation of each dimension of resources. The specific steps for calculating the real-time load deviation and the overall system performance deviation include: S31. Within the current control cycle, read the current actual total load and current business operation parameters of the target resource node; S32. Based on the current service operation parameters, match the corresponding baseline idle overhead in the global resource status baseline library; S33. Subtract the baseline no-load overhead from the current actual total load to obtain the current actual effective load; S34. Extract the benchmark load value and benchmark performance value under the corresponding working condition from the multi-dimensional performance benchmark self-learning matrix, calculate the relative deviation between the current actual effective load and the benchmark load value respectively, and obtain the single-dimensional real-time load deviation; calculate the relative deviation between the current comprehensive performance value and the benchmark performance value, and obtain the system comprehensive performance deviation.

[0010] Preferably, the absolute values ​​of the real-time load deviation difference between related resource dimensions and the absolute value of the system overall performance deviation are calculated. When the absolute value of the real-time load deviation difference is greater than the dynamic balance adaptive threshold, or the absolute value of the system overall performance deviation is greater than the dynamic balance adaptive threshold, the configuration parameters of the high-weight core resource dimension are locked according to the asymmetric resource weight coefficient, the resource fine-tuning compensation amount of the low-weight related dimension is calculated to perform load smoothing transfer and performance convergence, and the multi-dimensional performance benchmark self-learning matrix is ​​updated. When the absolute values ​​of the real-time load deviation difference and the absolute value of the system overall performance deviation are both not greater than the dynamic balance adaptive threshold, the current actual effective load is smoothly updated to the multi-dimensional performance benchmark self-learning matrix. Specific steps include: S41. Calculate the absolute value of the real-time load deviation difference between related resource dimensions, and the absolute value of the overall system performance deviation. S42. Determine whether the absolute value of the real-time load deviation difference is greater than the dynamic balance adaptive threshold, or whether the absolute value of the system comprehensive performance deviation is greater than the dynamic balance adaptive threshold; if yes, then execute S43; if neither is greater than, then execute S44. S43. Compare the asymmetric resource weight coefficients of the associated resource dimensions, keep the configuration of the core resource dimensions with larger weight coefficients unchanged, calculate the resource fine-tuning compensation amount of the low-weight dimensions based on the ratio of the deviation difference to the weight coefficients; send the resource fine-tuning compensation amount to the corresponding scheduling node, and transfer the excess load to the high-weight resource dimension to be carried by adjusting the resource allocation ratio. S44. Keep the current configuration of all resource dimensions unchanged, and use the exponential smoothing formula to update the current actual effective load and performance value to the multi-dimensional performance benchmark self-learning matrix.

[0011] Preferably, before sending the resource fine-tuning compensation amount to the corresponding scheduling node in step S43, a service QoS boundary verification step is also included: S431. Read the target configuration parameters after overlaying the resource fine-tuning compensation amount on the low-weight dimension; S432. Calculate the service end-to-end latency, reliability, and perception accuracy indicators corresponding to the target configuration parameters; S433. Determine whether the indicator falls within the preset service QoS allowable range; if it exceeds the upper limit of the range, force the resource fine-tuning compensation amount to be truncated to the configuration value corresponding to the upper limit of the range; if it is lower than the lower limit of the range, force the resource fine-tuning compensation amount to be truncated to the configuration value corresponding to the lower limit of the range.

[0012] Preferably, when a preset triggering event is detected, the multi-dimensional performance benchmark self-learning matrix is ​​invoked to match the feature parameters of the preset triggering event, the expected performance change caused by the event is calculated, and the expected performance change is converted into a feedforward resource orchestration compensation signal and superimposed on the corresponding communication resource scheduling link. The preset triggering event includes a sudden access event for a sensing task. Specific steps include: S51. The signaling monitoring module captures the access request of the sensing task in real time and analyzes it to obtain the detection accuracy requirements, single-cycle detection time slot ratio and beam configuration parameters of the sensing task. S52. When a sensing task is detected to be accessing the network, the detection accuracy requirement and the single-cycle detection time slot ratio of the sensing task are extracted to determine the occupancy type. If the detection accuracy requirement is lower than a preset accuracy threshold and the single-cycle detection time slot ratio is lower than a preset ratio threshold, the occupancy type of the task on the communication time and frequency resources is determined to be shared occupancy. If the detection accuracy requirement is not lower than the preset accuracy threshold, or the single-cycle detection time slot ratio is not lower than the preset ratio threshold, the occupancy type of the task on the communication time and frequency resources is determined to be exclusive occupancy. S53. Extract the expected communication resource consumption and computing power increment requirements corresponding to the detection parameters of the perception task from the multi-dimensional performance benchmark self-learning matrix. S54. If it is determined to be a shared occupation, a feedforward resource orchestration compensation signal for communication power allocation optimization and computing node pre-scheduling is generated. Sensing coexistence is achieved through power orthogonality and beam isolation within the same frequency and time slot. If it is determined to be an exclusive occupation, a feedforward resource orchestration compensation signal for communication time slot avoidance and storage cache pre-deployment is generated. Communication services in the corresponding time slot are migrated to the backup time slot and sensing data cache resources are deployed in advance. S55. The feedforward resource orchestration compensation signal is superimposed onto the scheduling and control channel of the corresponding communication resource.

[0013] Preferably, when a preset triggering event is detected, the multi-dimensional performance benchmark self-learning matrix is ​​invoked to match the feature parameters of the preset triggering event, the expected performance change caused by the event is calculated, and the expected performance change is converted into a feedforward resource orchestration compensation signal and superimposed on the corresponding communication resource scheduling link. The preset triggering event also includes the cold and hot data migration event of the in-memory computing node. When a cold / hot data migration command is detected, the changes in storage resource usage and computing power processing requirements are calculated based on the data migration amount and access popularity. The changes are multiplied by the current actual effective load to obtain the expected performance change, and a corresponding feedforward communication bandwidth compensation signal is generated and sent to the associated communication resource scheduling node.

[0014] (III) Beneficial Effects This invention provides an adaptive communication resource orchestration and performance optimization system based on integrated intelligent computing and storage. It has the following beneficial effects: I. By unifying the abstraction of resources across all dimensions and constructing a multi-dimensional self-learning benchmark, the core problem of the difficulty in coordinating and scheduling heterogeneous resources is solved at the architectural level. A unified abstraction model of all-dimensional resources for communication, intelligent computing, and storage is constructed to achieve normalized and quantitative representation of four types of heterogeneous resources: communication time-frequency-space, sensing and detection, computing power, and storage. This solves the technical obstacle of inconsistent resource dimensions and the inability to jointly schedule resources. A global resource state benchmark library is constructed through idle-load step-by-step loading and filtering and denoising to remove inherent hardware overhead and noise interference, ensuring the accuracy of effective load calculation. An asymmetric resource weight system is established, which combines the rated hardware capacity and the business carrying level to quantify weight coefficients and integrates resource coupling constraint correction to construct a multi-dimensional performance benchmark self-learning matrix. This accurately quantifies coupling relationships such as mutual exclusivity of communication, positive correlation between computing and transmission, and storage-computing mapping, solving the problem of difficulty in quantifying resource coupling effects and large benchmark deviations. This lays a high-precision reference benchmark for subsequent adaptive orchestration and fundamentally breaks down the dimensional barriers of traditional resource scheduling.

[0015] Second, through dynamic deviation perception and asymmetric load transfer mechanisms, the system achieves both adaptive resource optimization and service quality assurance in dynamic scenarios. By collecting resource operation data in real time and deducting baseline idle overhead to obtain the effective load, it calculates the single-dimensional load deviation and the overall system performance deviation, achieving highly sensitive monitoring of service load fluctuations. Based on an asymmetric weight system, dynamic load transfer is executed, locking high-weight core resource dimensions to ensure stable operation. By fine-tuning the configuration of low-weight dimensions, it achieves smooth cross-dimensional transfer of excess load, solving the problems of local resource overload and supply-demand mismatch. This avoids the failure of weak resource nodes due to overload and ensures the stable operation of core services. At the same time, a service QoS boundary verification step is added to ensure that resource adjustments are always within the allowable range of the process, solving the problem of balancing resource optimization and service quality. Combined with threshold judgment and exponential smoothing update mechanisms, the system responds quickly to abnormal changes and iterates the baseline during normal fluctuations, giving it both shock resistance and self-evolution capabilities, achieving continuous convergence and optimization of system performance.

[0016] Third, through an event-driven feedforward compensation and closed-loop self-learning system, the response speed in sudden scenarios and the long-term robustness of the system are significantly improved. For typical triggering events such as sudden access to sensing tasks and migration of hot and cold data, a feedforward resource orchestration compensation mechanism is established. This mechanism calls a multi-dimensional performance benchmark self-learning matrix to match event feature parameters, accurately calculates expected performance changes, and generates corresponding compensation signals. For sensing tasks, the occupancy type is determined by both detection accuracy and time slot percentage, and differentiated compensation strategies for shared coexistence or time slot avoidance are generated, solving the problem of large communication performance jitter during sudden sensing task access. For cold and hot data migration... Data migration reserves communication bandwidth in advance to avoid migration traffic crowding out business bandwidth and deeply adapts to the characteristics of the storage-computing architecture. The entire mechanism upgrades the traditional passive feedback to proactive prediction and adaptation, significantly reducing resource reconfiguration latency and solving the technical problem of scheduling lag in sudden scenarios. At the same time, a closed-loop self-learning system is formed throughout the process. The multi-dimensional performance benchmark self-learning matrix can be continuously iterated and updated as the system runs, automatically adapting to slow changes such as business characteristic drift and hardware performance degradation. It can maintain scheduling accuracy for a long time without manual recalibration, comprehensively improving the overall operating efficiency and long-term operating stability of the integrated intelligent computing and storage system. Attached Figure Description

[0017] Figure 1 A flowchart of the operating method of the adaptive communication resource orchestration and performance optimization system based on the integration of sensing, intelligent computing and storage provided by the present invention; Figure 2 This is a schematic diagram of the module of the adaptive communication resource orchestration and performance optimization system based on the integration of sensing, intelligent computing and storage provided by the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] An example of the adaptive communication resource orchestration and performance optimization system based on integrated intelligent computing and storage is as follows: Example: Please see Figures 1-2 An adaptive communication resource orchestration and performance optimization system based on integrated sensing, intelligent computing, and storage, comprising: The module for abstracting and constructing communication benchmarks for synesthetic computing and storage resources builds a unified abstract model of synesthetic intelligent computing and storage resources across all dimensions. It normalizes and quantifies communication time-frequency-space resources, sensing and detection resources, computing power resources, and storage resources, collects various resource operation parameters under no-load benchmark conditions, and constructs a global resource status benchmark library. The multi-dimensional performance self-learning and communication weight configuration module allocates asymmetric resource weight coefficients based on the hardware capabilities and service carrying priorities of each resource node during the benchmark calibration phase, records benchmark performance indicators under stable business scenarios, and constructs a multi-dimensional performance benchmark self-learning matrix by combining time-frequency resource allocation, perception accuracy requirements, computing power load rate, storage occupancy rate and service QoS parameters. The real-time load monitoring and communication performance calculation module collects the current actual resource usage data and business performance data of each node in real time during the dynamic operation phase. It calls the global resource status benchmark library to deduct the benchmark idle overhead and obtain the current actual effective load. It compares the current actual effective load with the benchmark value in the multi-dimensional performance benchmark self-learning matrix and calculates the real-time load deviation of each dimension of resources and the overall system performance deviation. The adaptive communication orchestration and dynamic load transfer module calculates the absolute value of the real-time load deviation difference between related resource dimensions and the absolute value of the overall system performance deviation. When the absolute value of the real-time load deviation difference is greater than the dynamic balancing adaptive threshold, or the absolute value of the overall system performance deviation is greater than the dynamic balancing adaptive threshold, the module locks the configuration parameters of the high-weight core resource dimension based on the asymmetric resource weight coefficient, calculates the resource fine-tuning compensation amount of the low-weight related dimension to perform load smooth transfer and performance convergence, and updates the multi-dimensional performance benchmark self-learning matrix. When the absolute value of both the real-time load deviation difference and the absolute value of the overall system performance deviation are not greater than the dynamic balancing adaptive threshold, the module smoothly updates the current actual effective load to the multi-dimensional performance benchmark self-learning matrix. The event-driven communication feedforward and performance compensation module, when a preset trigger event is detected, calls the multi-dimensional performance benchmark self-learning matrix, matches the feature parameters of the preset trigger event, calculates the expected performance change caused by the event, and converts the expected performance change into a feedforward resource orchestration compensation signal and superimposes it into the corresponding communication resource scheduling link.

[0020] This embodiment provides a detailed description based on a 6G edge-sensing intelligent computing and storage integrated smart industrial park network scenario. The network covers an industrial park of 1 square kilometer and includes one macro base station with an integrated sensing antenna array, three edge computing nodes, and two distributed storage and computing nodes. It simultaneously supports four core services: industrial high-definition video backhaul, environmental millimeter-wave radar sensing, machine vision AI quality inspection inference, and production data cold and hot tiered storage. The network resource configuration has significant heterogeneity and coupling: the total bandwidth on the communication side is 200MHz, supporting time-division multiplexing and space-division multiplexing. The sensing side uses millimeter-wave radar with a detection range of 1km and a distance resolution of 0.1m; the computing power side has a total computing power of 128TOPS, including heterogeneous computing power of GPU and CPU; the storage side has a total capacity of 2PB, divided into two layers: hot storage SSD and cold storage HDD; the QoS requirements of the services vary significantly: industrial video transmission requires end-to-end latency <20ms and reliability of 99.999%; environmental perception requires detection accuracy <0.2m and update cycle of 100ms; AI inference requires single-frame processing latency <50ms; and data storage requires read / write throughput >1GB / s.

[0021] A unified abstract model of all-dimensional resources for intelligent computing and storage is constructed, which normalizes and quantifies communication time-frequency-space resources, sensing and detection resources, computing power resources, and storage resources. Various resource operation parameters under idle baseline conditions are collected, and a global resource status baseline library is built. This includes the following steps: S11. In the system's no-load initialization state, control the four types of resource nodes—communication, sensing, computing power, and storage—to load in a stepped manner within a preset load range, and collect the original operating parameters of each node. Specifically, before any service load is connected, the system idle calibration process is initiated; communication resources are controlled to occupy bandwidth in 10% increments from 0% to 100%; sensing resources are increased in 10% increments from 0% to full power; computing nodes are loaded with computing tasks in 10% increments from 0% to full load; and storage nodes are increased in 10% increments from 0% to full load. Each load step is allowed to remain stable for 30 seconds to ensure the system's operating status is stable before data is collected. The raw operating parameters collected include: transmission power and baseband processing overhead on the communication side; detection power consumption and signal processing overhead on the sensing side; CPU utilization, memory utilization, and idle power consumption on the computing side; and disk I / O overhead and idle read / write latency on the storage side.

[0022] S12. Perform spectrum analysis and noise filtering on the original operating parameters, separate out the inherent periodic fluctuation components of the hardware and shield them to obtain the noise-reduced reference no-load parameters. Specifically, due to the inherent characteristics of hardware such as the local oscillator of communication radio frequency devices, the cooling fan of computing power, and the rotation of storage disks, the original operating parameters contain periodic noise components corresponding to the hardware operating frequency. The system performs a fast Fourier transform on the collected original parameters to separate the periodic fluctuation components of the corresponding hardware operating fundamental frequency and its second and third harmonics. These components are shielded by a digital band-stop filter, and the filtered steady-state mean is extracted as the noise reduction reference idle parameter to avoid subsequent scheduling misjudging the inherent hardware fluctuations as changes in service load.

[0023] S13. Use a fitting algorithm to perform curve fitting on the denoised baseline idle parameters, generate continuous load-idle overhead baseline curves for resources in each dimension, and integrate them to form a global resource status baseline library.

[0024] Specifically, the system collects denoising baseline no-load parameters for 10 load levels across various dimensions; taking communication bandwidth as an example, it uses the least squares method to perform quadratic polynomial fitting to generate a bandwidth occupancy-power consumption baseline curve, the curve expression of which is: ,in This represents the noise reduction idle power overhead under the corresponding bandwidth usage, in watts (W). This indicates the amount of communication bandwidth used, in MHz. The coefficients represent the quadratic fitting terms, characterizing the nonlinear device overhead that increases with the square of the bandwidth, and are expressed in W / MHz. 2 ; The coefficient represents the linear fitting coefficient, which characterizes the linear power consumption proportional to the bandwidth, and is expressed in W / MHz. The constant term fitting coefficient represents the basic static power consumption of the device, and its unit is W.

[0025] In this embodiment, the parameters of the communication dimension baseline curve are obtained through fitting: W / MHz 2 , W / MHz, W.

[0026] Similarly, benchmark curves are generated for sensing power-processing overhead, computing load-energy consumption overhead, and storage throughput-latency overhead. All benchmark curves are integrated to form a global resource status benchmark library, which is used to deduct idle overhead and extract the actual business effective load in the future.

[0027] During the benchmark calibration phase, asymmetric resource weight coefficients are allocated based on the hardware capabilities and service carrying priorities of each resource node. Benchmark performance indicators under stable service scenarios are recorded. Combining time-frequency resource allocation, perception accuracy requirements, computing load rate, storage utilization rate, and service QoS parameters, a multi-dimensional performance benchmark self-learning matrix is ​​constructed. The specific steps for allocating asymmetric resource weight coefficients include: S21. Obtain the rated capacity set of four types of resource nodes: communication, sensing, computing power, and storage, and extract the maximum rated capacity value among them; Specifically, the system reads hardware configuration parameters to obtain the rated physical capabilities of four types of resources: total communication bandwidth of 200MHz, maximum sensing power of 43dBm, total computing power of 128TOPS, and total storage capacity of 2PB.

[0028] Since the four types of resources have different dimensions, they are first mapped to a unified resource capacity equivalent through a resource equivalent conversion model: the conversion factor for communication bandwidth is 0.544 TOPS / MHz, the conversion factor for sensing power is 1.97 TOPS / dBm, and the conversion factor for storage capacity is 48 TOPS / PB.

[0029] The converted rated capacity equivalents for each dimension are as follows: Communication dimension Equivalent; Perceptual dimension Equivalent; Computing power dimension Equivalent; Storage Dimension Equivalent; maximum rated capacity value extracted Equivalent, corresponding to the computing power dimension.

[0030] S22. Divide the rated capacity of each resource node by the maximum rated capacity value to obtain the basic weight factor of each node. Specifically, based on the maximum rated capacity value of 128 TOPS equivalent, the basic weighting factors for each dimension are calculated: Communication dimension basic weight factor ; Perception Dimension Basic Weighting Factor ; Basic weighting factor for computing power ; Storage dimension basic weight factor .

[0031] S23. Assign a level correction factor based on the bearing level in the end-to-end business link according to the resource dimension, and multiply the basic weight factor by the level correction factor to obtain the final asymmetric resource weight coefficient.

[0032] Specifically, based on the end-to-end carrying logic of smart park business, computing power processing is the core carrying layer of business, and is given a layer correction factor of 1.0; communication transmission is the backbone layer of business carrying, and is given a layer correction factor of 1.05; perception and detection is the data input layer, and is given a layer correction factor of 0.95; storage is the data persistence layer, and is given a layer correction factor of 0.9.

[0033] Multiply the base weight factor by the corresponding hierarchical correction factor to obtain the final asymmetric resource weight coefficient: Communication Dimension Weight ; Perception Dimension Weights ; Computing power dimension weight ; Storage dimension weight .

[0034] The system defines the computing power dimension with the highest weight as the core load bearing dimension.

[0035] The hierarchical correction factor is obtained through multiple regression analysis of historical business data. The system extracts the Pearson correlation coefficients of resource fluctuations and QoS degradation in each dimension from the historical business degradation samples, calculates the variance contribution rate of each dimension of resources to the final business performance, and obtains the hierarchical correction factor for each dimension after normalization.

[0036] During the benchmark calibration phase, asymmetric resource weight coefficients are allocated based on the hardware capabilities and service carrying priorities of each resource node. Benchmark performance indicators under stable service scenarios are recorded. Combining time-frequency resource allocation, perception accuracy requirements, computing load rate, storage occupancy rate, and service QoS parameters, a multi-dimensional performance benchmark self-learning matrix is ​​constructed. The construction of this matrix also incorporates resource coupling constraint correction steps. Based on the mutual exclusion constraints of communication bandwidth and sensing time slots, the positive correlation constraints of computing load and data transmission volume, and the mapping constraints of storage capacity and cache hit rate, the coupling gain coefficient and loss coefficient under different resource combinations are calculated; and the benchmark performance index in the multidimensional performance benchmark self-learning matrix is ​​scaled and corrected according to the coupling gain coefficient and loss coefficient.

[0037] Specifically, the system incorporates three types of core coupling constraint models: L1, the mutual exclusion model of sensing time and frequency resources: when sensing and communication share the same frequency band, the interference loss coefficient is 0.15, that is, for every 10% increase in the shared bandwidth ratio, the communication transmission efficiency decreases by 1.5%; there is no additional loss when time-division multiplexing; L2, a positive correlation model between computing power and data transmission: for every 10% increase in computing power load, the corresponding data transmission volume increases by 8%, and the transmission link multiplexing gain brings a 0.8% improvement in overall system efficiency; L3, in-memory mapping model: For every 10% increase in the proportion of hot storage, the cache hit rate increases, the corresponding computing power processing latency decreases by 5%, and the overall system efficiency improves by 0.5%.

[0038] Under the baseline operating conditions of this embodiment: the shared bandwidth of the sensing is 40MHz, accounting for 20% of the total bandwidth; the baseline computing load rate is 65%; and the hot storage accounts for 40%.

[0039] Calculate the impact of each coupling term on the overall performance: Synesthesia mutual exclusion loss: That is, the efficiency decreased by 3%; Calculate the positive correlation gain: This translates to a 5.2% improvement in efficiency; In-store mapping gain: This translates to a 2% increase in efficiency. Overall Coupling Gain Coefficient This means a net increase of 4.2% in overall efficiency.

[0040] The system scales up the original baseline performance value by 4.2% based on this coefficient, eliminating the baseline bias caused by coupling effects and improving the accuracy of the self-learning matrix.

[0041] When all four types of services are running stably and all QoS standards are met, the system records all parameters under the current state and constructs a structured multidimensional performance benchmark self-learning matrix. The matrix contains the following dimension fields and their corresponding benchmark values: Resource allocation dimensions: communication transmission bandwidth 120MHz, sensing shared bandwidth 40MHz, computing load rate 65%, hot storage utilization rate 40%, sensing detection accuracy 0.1m; Business QoS dimensions: end-to-end latency 15ms, reliability 99.999%, perception update cycle 100ms, AI inference latency 40ms; Performance metrics: spectral efficiency 12bit / s / Hz, energy efficiency 0.8Mbit / J, computing efficiency 1.2TOPS / W, storage efficiency 0.5GB / s / TB; Coupling correction dimension: Overall coupling gain coefficient 1.042, corrected overall performance value 1.042 times the original value.

[0042] The system encapsulates the above multidimensional parameters in a structured manner, constructs and saves them as a multidimensional performance benchmark self-learning matrix.

[0043] During the dynamic operation phase, real-time data on the actual resource usage and business performance of each node are collected. The global resource status benchmark library is called to deduct the benchmark idle overhead, yielding the current actual effective load. The current actual effective load is then compared with the benchmark value in the multi-dimensional performance benchmark self-learning matrix to calculate the real-time load deviation and overall system performance deviation for each dimension of resources. The specific steps for calculating the real-time load deviation and overall system performance deviation include: S31. Within the current control cycle, read the current actual total load and current business operation parameters of the target resource node; Specifically, the system uses a 100ms control cycle to read the parameters within the current cycle: total transmission power on the communication side is 29.5W, corresponding to an actual bandwidth of 150MHz; total power consumption of the computing node is 210W, corresponding to an actual load rate of 82%; sensing and detection power is 40dBm; and hot storage occupancy rate is 45%.

[0044] S32. Based on the current business operation parameters, match the corresponding baseline idle overhead in the global resource status baseline library; Specifically, using linear interpolation, the current communication bandwidth of 150MHz is substituted into the communication bandwidth-power overhead baseline curve generated by S13 to calculate the corresponding baseline no-load power: W; Similarly, the baseline no-load power consumption of the computing node at 82% load is found to be 45W.

[0045] S33. Subtract the baseline no-load cost from the current actual total load to obtain the current actual effective load. Specifically, the power corresponding to the effective load on the communication side. W, which translates to an effective load bandwidth of 135MHz; Power consumption corresponding to effective load on the computing side W, which translates to an effective load factor of 76%; Similarly, the effective power of the sensing load is 38dBm, and the effective throughput of storage is 420MB / s; This effective load figure purely represents the amount of resources actually consumed by the business, excluding the idle load loss of the hardware itself.

[0046] S34. Extract the benchmark load value and benchmark performance value under the corresponding working condition from the multi-dimensional performance benchmark self-learning matrix, calculate the relative deviation between the current actual effective load and the benchmark load value respectively, and obtain the single-dimensional real-time load deviation; calculate the relative deviation between the current comprehensive performance value and the benchmark performance value, and obtain the system comprehensive performance deviation.

[0047] The formula for calculating the real-time load deviation in a single dimension is: ,in, Indicates the first Real-time load deviation across resource dimensions; Indicates the first The current actual effective load in each dimension; Indicates the first The baseline load value for each dimension.

[0048] Specific calculations: Extracting baseline load values: Communication baseline effective bandwidth 105MHz, computing power baseline effective load rate 60%; Communication load deviation ; Computing load deviation ; The formula for calculating the deviation of the overall system performance is: ,in, Indicates the deviation of the system's overall performance; This represents the current overall system performance value; This represents the benchmark overall performance value.

[0049] In this embodiment, the calculated current overall performance value is 6.2% lower than the benchmark, which is the deviation of the system's overall performance. .

[0050] Calculate the absolute value of the real-time load deviation difference between related resource dimensions and the absolute value of the overall system performance deviation. When the absolute value of the real-time load deviation difference is greater than the dynamic balancing adaptive threshold, or the absolute value of the overall system performance deviation is greater than the dynamic balancing adaptive threshold, lock the configuration parameters of the high-weight core resource dimension based on the asymmetric resource weight coefficient, calculate the resource fine-tuning compensation amount of the low-weight related dimensions to smoothly transfer load and achieve performance convergence, and update the multi-dimensional performance benchmark self-learning matrix. When the absolute values ​​of the real-time load deviation difference and the overall system performance deviation are both not greater than the dynamic balancing adaptive threshold, smoothly update the current actual effective load to the multi-dimensional performance benchmark self-learning matrix. Specific steps include: S41. Calculate the absolute value of the real-time load deviation difference between related resource dimensions, and the absolute value of the overall system performance deviation. Specifically, in this scenario, the demand for computing power surged due to the addition of AI quality inspection tasks in the park; the calculated computing power load deviation was 28.57%, and the communication load deviation was 10.2%, with the absolute value of the difference between the two deviations being... The absolute value of the deviation of the overall system performance is .

[0051] S42. Determine whether the absolute value of the real-time load deviation difference is greater than the dynamic balancing adaptive threshold, or whether the absolute value of the system comprehensive performance deviation is greater than the dynamic balancing adaptive threshold; if yes, proceed to S43; if neither is greater than, proceed to S44. Specifically, the system's preset dynamic balance adaptive threshold is 5%; if the deviation difference of 18.37% and the performance deviation of 6.2% both exceed the threshold, it is determined that a sudden change in business load has occurred, and the dynamic resource reordering procedure of S43 is triggered.

[0052] Dynamic balancing adaptive thresholding involves collecting several sets of deviation samples under stable business scenarios during the system trial operation phase and calculating the standard deviation of the sample set. ,in accordance with The criteria set a threshold, which was calculated in this embodiment. This is used to distinguish between normal business fluctuations and abnormal load changes.

[0053] S43. Compare the asymmetric resource weight coefficients of related resource dimensions, keep the configuration of core resource dimensions with larger weight coefficients unchanged, calculate the resource fine-tuning compensation amount of low weight dimensions based on the ratio of deviation difference to weight coefficients; distribute the resource fine-tuning compensation amount to the corresponding scheduling nodes, and transfer the excess load to high weight resource dimensions through resource allocation adjustment. Specifically, the system compares the weights of the computing power dimension ( ) and communication dimension weights ( The configuration of the computing power core dimension with higher weight is kept unchanged, while the communication dimension is treated as an adjustable dimension. The formula for calculating resource fine-tuning compensation is: ,in This indicates the amount of compensation for fine-tuning the communication bandwidth. This represents the preset load transfer coefficient, in this embodiment... ; This represents the absolute value of the deviation difference in the related dimensions; These are the weighting coefficients for high-weight dimensions; This represents the current configuration value for the low-weight dimension.

[0054] Substitute the numerical values ​​into the calculation: MHz, rounded up to an additional 15MHz of transmission bandwidth (constrained by the total bandwidth margin).

[0055] Because the current link already occupies 135MHz of bandwidth, and the system hardware's maximum total available bandwidth is 150MHz, the remaining bandwidth margin is only 15MHz. The algorithm theoretically calculates a bandwidth increment of 24.8MHz, which is greater than the system's remaining bandwidth margin. Therefore, due to hardware bandwidth resource constraints, the actual bandwidth increment is taken as the maximum value of the remaining margin, 15MHz. After adjustment, the total working bandwidth of the link is 150MHz.

[0056] By increasing communication bandwidth to improve data transmission rate, the data input and output of computing nodes can be accelerated, the queuing load of computing nodes can be alleviated, and the excess load on the computing side can be smoothly transferred through improved transmission efficiency.

[0057] Before the resource fine-tuning compensation is sent to the corresponding scheduling node in S43, a service QoS boundary verification step is also included: S431. Read the target configuration parameters after overlaying the resource fine-tuning compensation amount on the low-weight dimension; Specifically, after adding a 15MHz increment to the communication bandwidth, the target configuration is 150MHz.

[0058] S432. Calculate the end-to-end latency, reliability, and perception accuracy indicators of the service corresponding to the target configuration parameters. Specifically, the system calculates through a link-level simulation model that after the bandwidth is increased to 150MHz, the end-to-end latency of industrial video is 18ms, the reliability is 99.999%, and the sensing and detection accuracy is 0.18m.

[0059] S433. Determine whether the indicator falls within the preset service QoS allowable range; if it exceeds the upper limit of the range, force the resource fine-tuning compensation amount to be truncated to the configuration value corresponding to the upper limit of the range; if it is lower than the lower limit of the range, force the resource fine-tuning compensation amount to be truncated to the configuration value corresponding to the lower limit of the range. Specifically, the preset QoS allowable range is: latency <20ms, reliability ≥99.999%, and perception accuracy <0.2m; if all calculation results fall within the allowable range, the verification is passed, and the 15MHz bandwidth increment is sent to the resource scheduling node.

[0060] If the calculated sensing accuracy drops to 0.22m, exceeding the allowable limit, then the bandwidth increment will be forcibly truncated to 10MHz to ensure that the sensing accuracy meets the standard.

[0061] S44. Keep the current configuration of all resource dimensions unchanged, and use the exponential smoothing formula to update the current actual effective load and performance value to the multi-dimensional performance benchmark self-learning matrix.

[0062] Specifically, if the deviation difference is 3.2% (not greater than the 5% threshold), the system determines it as normal business fluctuation; without adjusting any resource configuration, it uses the exponential smoothing formula to slowly absorb and update the current effective load and performance value into the multi-dimensional performance benchmark self-learning matrix, thereby achieving adaptive drift correction of the benchmark.

[0063] The formula for exponential smoothing is: ;in This is the updated baseline load value; This represents the current actual effective load. The baseline load value before the update; This is a smoothing coefficient, with a value range of 0 < λ < 1. The smaller the value, the smoother the baseline update.

[0064] In this embodiment This ensures slow iteration of the benchmark, balancing stability and adaptability; the smoothing coefficient is calculated by performing autocorrelation analysis on low-frequency drift signals from historical normal business fluctuations, thus determining the correlation time constant. The result was derived later.

[0065] When a preset trigger event is detected, the multi-dimensional performance benchmark self-learning matrix is ​​invoked to match the feature parameters of the preset trigger event, calculate the expected performance change caused by the event, and convert the expected performance change into a feedforward resource orchestration compensation signal, which is then superimposed on the corresponding communication resource scheduling link. The preset trigger events include sudden access events for sensing tasks. Specific steps include: S51. The signaling monitoring module captures the access requests of the sensing task in real time and analyzes them to obtain the detection accuracy requirements, single-cycle detection time slot ratio and beam configuration parameters of the sensing task. Specifically, the park's security system suddenly initiated a high-precision area detection request. The signaling monitoring module captured the request in real time and analyzed the parameters: detection accuracy requirement of 0.05m, single-cycle detection time slot ratio of 40%, and beam azimuth angle covering the main road area of ​​the park.

[0066] S52. When a sensing task is detected to be accessing the network, the detection accuracy requirement and the proportion of single-cycle detection time slots of the sensing task are extracted to determine the occupancy type. If the detection accuracy requirement is lower than a preset accuracy threshold and the proportion of single-cycle detection time slots is lower than a preset proportion threshold, the occupancy type of the communication time and frequency resources of the task is determined to be shared occupancy. If the detection accuracy requirement is not lower than a preset accuracy threshold, or the proportion of single-cycle detection time slots is not lower than a preset proportion threshold, the occupancy type of the communication time and frequency resources of the task is determined to be exclusive occupancy. Specifically, the system has a preset accuracy threshold of 0.1m and a preset percentage threshold of 30%. The detection accuracy of this mission is 0.05m, which is not lower than the preset accuracy threshold. It meets the judgment condition of "accuracy not lower than the threshold". Therefore, it is judged as an exclusive occupation type. This mission will occupy the communication transmission time slot.

[0067] S53. Extract the expected communication resource consumption and computing power increment requirements corresponding to the detection parameters of the perception task from the multi-dimensional performance benchmark self-learning matrix. Specifically, the system calls the multi-dimensional performance benchmark self-learning matrix and, based on the feature parameters of a perception resolution of 0.05m and a time slot ratio of 40%, matches and extracts the following: This task will result in a 25MHz reduction in available communication bandwidth and an increase of 18TOPS in computing power processing requirements; Based on this resource loss, the expected overall performance decrease is calculated to be 9.2%, that is, the expected performance change is -9.2%.

[0068] S54. If it is determined to be a shared occupation, a feedforward resource orchestration compensation signal for communication power allocation optimization and computing node pre-scheduling is generated. Sensing coexistence is achieved through power orthogonality and beam isolation within the same frequency and time slot. If it is determined to be an exclusive occupation, a feedforward resource orchestration compensation signal for communication time slot avoidance and storage cache pre-deployment is generated. Communication services in the corresponding time slot are migrated to the backup time slot and sensing data cache resources are deployed in advance. Specifically, since the system is determined to be exclusively occupied, it generates three feedforward compensation signals to offset the expected performance loss: H1, Communication Time Slot Avoidance: Adjust the communication time slot allocation, postpone the transmission time slots of non-critical services to spare time slots, and reserve 40% of time slots for sensing tasks; H2, Storage Cache Pre-deployment: Pre-schedule hot storage nodes to reserve 20GB of cache space for temporary storage of raw sensing data, reducing subsequent burst transmission pressure; H3, pre-scheduled computing nodes: pre-allocate 18 TOPSGPU computing power to computing nodes and pre-load the perception processing algorithm model.

[0069] If it is determined to be a shared occupation, the generated compensation signal is as follows: optimize the power allocation ratio of communication and sensing, and adopt power orthogonal design; achieve spatial domain isolation through beamforming; pre-schedule the corresponding computing resources to achieve coexistence of sensing under the same frequency and time slot.

[0070] S55. Superimpose the feedforward resource orchestration compensation signal onto the scheduling and control channel of the corresponding communication resource.

[0071] Specifically, within the same control cycle when a sensing task is accessed, the system sends a feedforward compensation signal to the communication scheduler, computing scheduler, and storage scheduler; and completes resource pre-configuration before the sensing task is officially started, avoiding resource contention and performance jitter at the moment of task access, and ensuring the smoothness of service switching.

[0072] When a preset triggering event is detected, the multi-dimensional performance benchmark self-learning matrix is ​​invoked to match the feature parameters of the preset triggering event, calculate the expected performance change caused by the event, and convert the expected performance change into a feedforward resource orchestration compensation signal and superimpose it into the corresponding communication resource scheduling link. The preset triggering event also includes the cold and hot data migration event of the in-memory computing node. When a cold / hot data migration command is detected, the changes in storage resource usage and computing power processing requirements are calculated based on the data migration amount and access frequency. The changes are multiplied by the current actual effective load to obtain the expected performance change, and a corresponding feedforward communication bandwidth compensation signal is generated and sent to the associated communication resource scheduling node.

[0073] Specifically, the system triggers a cold backup migration command for historical production data, planning to migrate 500GB of hot data to cold storage within a 10-minute time window. The first step is to calculate the changes in storage resources: hot storage utilization decreased by 10%, cold storage utilization increased by 5%, and the demand for computing power indexing increased by 3%. The second step is to calculate the incremental communication bandwidth requirement: 500GB of data is migrated within 10 minutes, with an average transmission rate of about 85MB / s, which corresponds to a wireless communication bandwidth requirement of about 10MHz. The third step is to calculate the expected change in performance: the proportion of migration bandwidth to the current effective communication bandwidth is... If bandwidth is not reserved, it will consume business bandwidth, and the overall efficiency is expected to decrease by 7.4%. The fourth step is to generate a feedforward communication bandwidth compensation signal: reserve 10MHz of dedicated migration bandwidth from the system's spare bandwidth pool, which does not occupy the service transmission bandwidth and offsets the performance loss. The fifth step is to send the bandwidth compensation signal to the communication resource scheduling node and complete the bandwidth reservation configuration before the migration starts to avoid migration traffic crowding out the service transmission bandwidth and ensure the stable operation of production services.

[0074] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An adaptive communication resource orchestration and performance optimization system based on integrated sensing, intelligent computing, and storage, characterized in that: The system includes: The module for abstracting and constructing communication benchmarks for synesthetic computing and storage resources builds a unified abstract model of synesthetic intelligent computing and storage resources across all dimensions. It normalizes and quantifies communication time-frequency-space resources, sensing and detection resources, computing power resources, and storage resources, collects various resource operation parameters under no-load benchmark conditions, and constructs a global resource status benchmark library. The multi-dimensional performance self-learning and communication weight configuration module allocates asymmetric resource weight coefficients based on the hardware capabilities and service carrying priorities of each resource node during the benchmark calibration phase, records benchmark performance indicators under stable business scenarios, and constructs a multi-dimensional performance benchmark self-learning matrix by combining time-frequency resource allocation, perception accuracy requirements, computing power load rate, storage occupancy rate and service QoS parameters. The real-time load monitoring and communication performance calculation module collects the current actual resource usage data and service performance data of each node in real time during the dynamic operation phase. It calls the global resource status benchmark library to deduct the benchmark idle overhead to obtain the current actual effective load. The current actual effective load is compared with the benchmark value in the multi-dimensional performance benchmark self-learning matrix to calculate the real-time load deviation of each dimension of resources and the overall system performance deviation. The adaptive communication orchestration and dynamic load transfer module calculates the absolute value of the real-time load deviation difference between associated resource dimensions and the absolute value of the overall system performance deviation. When the absolute value of the real-time load deviation difference is greater than the dynamic balance adaptive threshold, or the absolute value of the overall system performance deviation is greater than the dynamic balance adaptive threshold, the module locks the configuration parameters of the high-weight core resource dimension according to the asymmetric resource weight coefficient, calculates the resource fine-tuning compensation amount of the low-weight associated dimension to perform load smooth transfer and performance convergence, and updates the multi-dimensional performance benchmark self-learning matrix. When the absolute value of both the real-time load deviation difference and the absolute value of the overall system performance deviation are not greater than the dynamic balance adaptive threshold, the module smoothly updates the current actual effective load to the multi-dimensional performance benchmark self-learning matrix. The event-driven communication feedforward and performance compensation module, when a preset triggering event is detected, calls the multi-dimensional performance benchmark self-learning matrix, matches the feature parameters of the preset triggering event, calculates the expected performance change caused by the event, and converts the expected performance change into a feedforward resource orchestration compensation signal superimposed on the corresponding communication resource scheduling link.

2. The adaptive communication resource orchestration and performance optimization system based on integrated intelligent computing and storage according to claim 1, characterized in that, A unified abstract model of all-dimensional resources for intelligent computing and storage is constructed, which normalizes and quantifies communication time-frequency-space resources, sensing and detection resources, computing power resources, and storage resources. Various resource operation parameters under idle baseline conditions are collected, and a global resource status baseline library is built. This includes the following steps: S11. In the system's no-load initialization state, control the four types of resource nodes—communication, sensing, computing power, and storage—to load in a stepped manner within a preset load range, and collect the original operating parameters of each node. S12. Perform spectrum analysis and noise filtering on the original operating parameters to separate and shield the inherent periodic fluctuation components of the hardware, and obtain the noise-reduced reference no-load parameters. S13. Use a fitting algorithm to perform curve fitting on the denoised baseline idle parameters to generate continuous load-idle overhead baseline curves for resources in each dimension, and integrate them to form a global resource status baseline library.

3. The adaptive communication resource orchestration and performance optimization system based on integrated intelligent computing and storage according to claim 1, characterized in that, During the benchmark calibration phase, asymmetric resource weight coefficients are allocated based on the hardware capabilities and service carrying priorities of each resource node. Benchmark performance indicators under stable service scenarios are recorded. Combining time-frequency resource allocation, perception accuracy requirements, computing load rate, storage utilization rate, and service QoS parameters, a multi-dimensional performance benchmark self-learning matrix is ​​constructed. The specific steps for allocating asymmetric resource weight coefficients include: S21. Obtain the rated capacity set of four types of resource nodes: communication, sensing, computing power, and storage, and extract the maximum rated capacity value among them; S22. Divide the rated capacity of each resource node by the maximum rated capacity value to obtain the basic weight factor of each node; S23. Assign a level correction factor based on the bearing level in the end-to-end business link according to the resource dimension, and multiply the basic weight factor by the level correction factor to obtain the final asymmetric resource weight coefficient.

4. The adaptive communication resource orchestration and performance optimization system based on integrated intelligent computing and storage according to claim 1, characterized in that, During the benchmark calibration phase, asymmetric resource weight coefficients are allocated based on the hardware capabilities and service carrying priorities of each resource node. Benchmark performance indicators under stable service scenarios are recorded. Combining time-frequency resource allocation, perception accuracy requirements, computing load rate, storage occupancy rate, and service QoS parameters, a multi-dimensional performance benchmark self-learning matrix is ​​constructed. The construction of this matrix also incorporates resource coupling constraint correction steps. Based on the mutual exclusion constraints between communication bandwidth and sensing time slots, the positive correlation constraints between computing load and data transmission volume, and the mapping constraints between storage capacity and cache hit rate, the coupling gain coefficient and loss coefficient under different resource combinations are calculated; and the benchmark performance index in the multidimensional performance benchmark self-learning matrix is ​​scaled and corrected according to the coupling gain coefficient and loss coefficient.

5. The adaptive communication resource orchestration and performance optimization system based on integrated intelligent computing and storage according to claim 1, characterized in that, During the dynamic operation phase, real-time data on the actual resource usage and business performance of each node are collected. The global resource status benchmark library is called to deduct benchmark idle overhead, yielding the current actual effective load. This current actual effective load is compared with the benchmark value in the multi-dimensional performance benchmark self-learning matrix to calculate the real-time load deviation and overall system performance deviation for each dimension of resources. The specific steps for calculating the real-time load deviation and overall system performance deviation include: S31. Within the current control cycle, read the current actual total load and current business operation parameters of the target resource node; S32. Based on the current service operation parameters, match the corresponding baseline idle overhead in the global resource status baseline library; S33. Subtract the baseline no-load overhead from the current actual total load to obtain the current actual effective load; S34. Extract the benchmark load value and benchmark performance value under the corresponding working condition from the multi-dimensional performance benchmark self-learning matrix, calculate the relative deviation between the current actual effective load and the benchmark load value respectively, and obtain the single-dimensional real-time load deviation; calculate the relative deviation between the current comprehensive performance value and the benchmark performance value, and obtain the system comprehensive performance deviation.

6. The adaptive communication resource orchestration and performance optimization system based on integrated intelligent computing and storage according to claim 1, characterized in that, Calculate the absolute value of the real-time load deviation difference between related resource dimensions and the absolute value of the system overall performance deviation; when the absolute value of the real-time load deviation difference is greater than the dynamic balance adaptive threshold, or the absolute value of the system overall performance deviation is greater than the dynamic balance adaptive threshold, lock the configuration parameters of the high-weight core resource dimension according to the asymmetric resource weight coefficient, calculate the resource fine-tuning compensation amount of the low-weight related dimension to perform load smoothing transfer and performance convergence, and update the multi-dimensional performance benchmark self-learning matrix; when the absolute value of the real-time load deviation difference and the absolute value of the system overall performance deviation are both not greater than the dynamic balance adaptive threshold, smoothly update the current actual effective load to the multi-dimensional performance benchmark self-learning matrix. Specific steps include: S41. Calculate the absolute value of the real-time load deviation difference between related resource dimensions, and the absolute value of the overall system performance deviation. S42. Determine whether the absolute value of the real-time load deviation difference is greater than the dynamic balance adaptive threshold, or whether the absolute value of the system comprehensive performance deviation is greater than the dynamic balance adaptive threshold; if yes, then execute S43; if neither is greater than, then execute S44. S43. Compare the asymmetric resource weight coefficients of the associated resource dimensions, keep the configuration of the core resource dimensions with larger weight coefficients unchanged, calculate the resource fine-tuning compensation amount of the low-weight dimensions based on the ratio of the deviation difference to the weight coefficients; send the resource fine-tuning compensation amount to the corresponding scheduling node, and transfer the excess load to the high-weight resource dimension to be carried by adjusting the resource allocation ratio. S44. Keep the current configuration of all resource dimensions unchanged, and use the exponential smoothing formula to update the current actual effective load and performance value to the multi-dimensional performance benchmark self-learning matrix.

7. The adaptive communication resource orchestration and performance optimization system based on integrated intelligent computing and storage according to claim 6, characterized in that, Before the resource fine-tuning compensation amount is sent to the corresponding scheduling node in step S43, a service QoS boundary verification step is also included: S431. Read the target configuration parameters after overlaying the resource fine-tuning compensation amount on the low-weight dimension; S432. Calculate the service end-to-end latency, reliability, and perception accuracy indicators corresponding to the target configuration parameters; S433. Determine whether the indicator falls within the preset service QoS allowable range; If the range is exceeded, the resource fine-tuning compensation will be forcibly truncated to the configuration value corresponding to the range limit. If the value is below the lower limit of the range, the resource fine-tuning compensation amount will be forcibly truncated to the configuration value corresponding to the lower limit of the range.

8. The adaptive communication resource orchestration and performance optimization system based on integrated intelligent computing and storage according to claim 1, characterized in that, When a preset triggering event is detected, the multi-dimensional performance benchmark self-learning matrix is ​​invoked to match the feature parameters of the preset triggering event, calculate the expected performance change caused by the event, and convert the expected performance change into a feedforward resource orchestration compensation signal superimposed on the corresponding communication resource scheduling link. The preset triggering event includes a sudden access event for a sensing task. Specific steps include: S51. The signaling monitoring module captures the access request of the sensing task in real time and analyzes it to obtain the detection accuracy requirements, single-cycle detection time slot ratio and beam configuration parameters of the sensing task. S52. When a sensing task is detected to be accessing the network, the detection accuracy requirement and the single-cycle detection time slot ratio of the sensing task are extracted to determine the occupancy type. If the detection accuracy requirement is lower than a preset accuracy threshold and the single-cycle detection time slot ratio is lower than a preset ratio threshold, the occupancy type of the task on the communication time and frequency resources is determined to be shared occupancy. If the detection accuracy requirement is not lower than the preset accuracy threshold, or the single-cycle detection time slot ratio is not lower than the preset ratio threshold, the occupancy type of the task on the communication time and frequency resources is determined to be exclusive occupancy. S53. Extract the expected communication resource consumption and computing power increment requirements corresponding to the detection parameters of the perception task from the multi-dimensional performance benchmark self-learning matrix. S54. If it is determined to be a shared occupation, a feedforward resource orchestration compensation signal for communication power allocation optimization and computing node pre-scheduling is generated. Sensing coexistence is achieved through power orthogonality and beam isolation within the same frequency and time slot. If it is determined to be an exclusive occupation, a feedforward resource orchestration compensation signal for communication time slot avoidance and storage cache pre-deployment is generated. Communication services in the corresponding time slot are migrated to the backup time slot and sensing data cache resources are deployed in advance. S55. The feedforward resource orchestration compensation signal is superimposed onto the scheduling and control channel of the corresponding communication resource.

9. The adaptive communication resource orchestration and performance optimization system based on integrated intelligent computing and storage according to claim 1, characterized in that, When a preset triggering event is detected, the multi-dimensional performance benchmark self-learning matrix is ​​invoked to match the feature parameters of the preset triggering event, calculate the expected performance change caused by the event, and convert the expected performance change into a feedforward resource orchestration compensation signal and superimpose it into the corresponding communication resource scheduling link. The preset triggering event also includes the cold and hot data migration event of the in-memory computing node. When a cold / hot data migration command is detected, the changes in storage resource usage and computing power processing requirements are calculated based on the data migration amount and access popularity. The changes are multiplied by the current actual effective load to obtain the expected performance change, and a corresponding feedforward communication bandwidth compensation signal is generated and sent to the associated communication resource scheduling node.

Images