A data center storage space dynamic monitoring method based on twin modeling

Abstract

This invention discloses a dynamic monitoring method for data center storage space based on twin modeling, comprising the following steps: deploying a reversible release boundary projection DeepONet to form a release boundary and twin synchronization result; performing boundary folding with a release boundary encoder to generate a Branch output; performing a joint query with a boundary position query operator to generate a Trunk output; coupling the Branch output and the Trunk output to project the boundary projection result; performing a difference-aligned reverse search for the minimum boundary change to determine the source of the anomaly; re-updating the release conditions with candidate handling actions and recalculating the boundary projection; and outputting the dynamic monitoring result, anomaly source, and handling priority according to the three-state change. This invention relates to the field of data center storage resource monitoring technology, realizing accurate monitoring of storage space and improving the efficiency of anomaly diagnosis and the reliability of handling.

Description

Technical Field

[0001] This invention relates to the field of data center storage resource monitoring technology, and in particular to a method for dynamic monitoring of data center storage space based on twin modeling. Background Technology

[0002] Data center storage resource monitoring technology has been widely applied to storage pool capacity management, logical volume occupancy monitoring, snapshot space statistics, backup copy management, and hot / cold data migration tracking. Existing solutions mostly rely on capacity thresholds, growth trends, access load, and task logs for alerts, which can detect issues such as near-exhaustion of space, increased write pressure, abnormal interruption of migration tasks, and increasing snapshot occupancy, providing a basis for operations and maintenance personnel to carry out expansion, cleanup, and migration.

[0003] Existing digital twin methods can map storage resource status to a virtual model, forming a visual monitoring object. Existing deep learning prediction methods can predict future occupancy changes based on historical capacity sequences. However, these methods mainly focus on "whether capacity is increasing" and "whether space is insufficient," making it difficult to determine whether the space released on paper can be converted into actual releaseable space. Snapshot lineage, reference occupancy, migration balancing, and retention limits can cause release distortion in deletion, migration, and redundancy reduction actions, leading to inconsistencies between alarm results and actual releaseable capacity.

[0004] Existing models often use action scoring or rule ranking when processing candidate actions, lacking a process of backfeeding actions to the storage release boundary before extrapolation. Anomaly identification mostly uses threshold comparison or classification models, making it difficult to search for anomaly sources in reverse along the locked, transitional, and releaseable states. Storage space monitoring still lacks a dynamic monitoring method that can combine DeepONet operator learning, twin synchronization, release boundary projection, and counterfactual backfeeding.

[0005] Therefore, how to provide a dynamic monitoring method for data center storage space based on twin modeling is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a dynamic monitoring method for data center storage space based on twin modeling. This invention identifies the true releaseable capacity and the source of anomalies by releasing boundary projection, reverse search of minimum boundary change, and counterfactual backfeeding. It has the advantages of accurate release determination, clear anomaly location, and reliable handling priority.

[0007] A method for dynamic monitoring of data center storage space based on twin modeling according to an embodiment of the present invention includes the following steps: Deploy DeepONet, a reversible release boundary projection, in the data center storage space twin to divide the storage release process into locked state, transitional state, and releaseable state, forming a release boundary, and update the twin to obtain the twin synchronization result; Replace DeepONet's Branch with a release boundary encoder to perform boundary folding on release conditions formed by snapshot lineage, reference holding, migration balancing, and retention constraints, so that the Branch output carries the release conditions; Replace DeepONet's Trunk with a boundary location query operator, perform a joint query on the monitored object, event stage, and candidate action, and make the Trunk output point to the release boundary location to be deduced; Operator coupling is performed on the Branch output and Trunk output. The coupling result is then subjected to boundary projection according to the release condition. Capacities that do not meet the release condition are mapped to locked states, capacities in the disposal simulation are mapped to transition states, and capacities that meet the release condition are mapped to releaseable states, thus obtaining the boundary projection result. The boundary projection results are aligned with the twin synchronization results by difference. The minimum boundary change that makes the difference converge is searched in reverse along the locked state, transition state, and release state. The source of the anomaly is determined according to the boundary state to which the minimum boundary change belongs. The candidate disposal action is converted into a candidate boundary change, fed back to the release boundary encoder, the release condition carried in the Branch output is updated, and the operator coupling and boundary projection are re-executed. Based on the reduction in the locked state, the transfer in the transition state, and the increase in the release state before and after reinjection, the system outputs dynamic monitoring results, sources of anomalies, and handling priorities.

[0008] Optionally, the step of updating the twin to obtain the twin synchronization result includes: A three-state release channel is established at the output of the reversible release boundary projection DeepONet, and a boundary residual correction gate is embedded in the three-state release channel; Write the storage release process into the locked state, and use the boundary residual correction gate to compare the boundary residual between the current release state and the current state of the three-state release channel. When the release condition corresponding to the boundary residual is not met, the gated path from the locked state to the transition state is closed; When the deduction of the corresponding handling of boundary residuals is not completed, the gating path from the transition state to the release state is closed; When the release condition corresponding to the boundary residual is met and the disposal simulation is completed, the gated path from the locked state to the transition state and the gated path from the transition state to the releaseable state are opened. Update the locked state, transitional state, and releaseable state according to the gating path after opening, and write the three-state release channel back to the data center storage space twin to obtain the twin synchronization result.

[0009] Optionally, the step of performing boundary folding includes: Set a release condition folding core within the release boundary encoder and input the release condition into the release condition folding core; The release condition folding kernel compresses the release condition in a unidirectional state sequence of locked state, transition state, and releaseable state, forming the locked state output component, the transition state output component, and the releaseable state output component; When the release condition is compressed to the locked state, the transition state output component and the releaseable state output component are turned off, while the locked state output component is retained. When the release condition is compressed to the transition state, the locked state output component and the releaseable state output component are turned off, while the transition state output component is retained. When the release condition is compressed to the releaseable state, the locked state output component and the transition state output component are turned off, while the releaseable state output component is retained. The retained output components are combined into a Branch output, which carries the release conditions consistent with the release boundary.

[0010] Optionally, the step of performing the join query includes: On the Trunk side, a counterfactual boundary coordinate latch kernel is established, which maps the monitored object, event stage, and candidate disposal action to the release boundary to form the release boundary coordinates. Forward coordinate locking is performed along the direction from the locked state to the transition state and from the transition state to the releaseable state. Read back the release boundary coordinates after locking the forward coordinates in the direction from the release state to the transition state and from the transition state to the lock state to obtain the readback object, readback stage, and readback action; When the readback object is consistent with the monitored object, the readback phase is consistent with the event phase, and the readback action is consistent with the candidate disposal action, latch and release the boundary coordinates. If any of the readback object, readback phase, or readback action is inconsistent, the boundary coordinates of the occlusion release will be interrupted. Write the latched release boundary coordinates into the Trunk output, so that the Trunk output points to the release boundary position to be calculated.

[0011] Optionally, the step of performing boundary projection includes: Set a release boundary projection kernel at the coupling end of the reversible release boundary projection DeepONet, and use the Branch output as the release condition constraint and the Trunk output as the release boundary position to be deduced. The release boundary projection kernel performs state-by-state coupling on the Branch output and Trunk output according to the locked state, transition state and releaseable state, so that the coupling result is simultaneously constrained by the release condition and the release boundary position to be deduced. The release boundary projection kernel performs three-state adsorption on the coupling result according to the release conditions. When the release conditions are not met and the candidate disposal action is not entered, the coupling result is adsorbed to the locked state. When the candidate disposal action is entered and the disposal action is not completed, the coupling result is adsorbed to the transition state. When the release conditions are met, the coupling result is adsorbed to the releasable state. The coupling result after the three-state adsorption is used as the boundary projection result.

[0012] Optionally, the step of performing difference alignment includes: Inverse boundary perturbation search kernel is set at the causal end of DeepONet with reversible release boundary projection, and the boundary projection result is used as the search starting point and the twin synchronization result is used as the convergence target. The reverse boundary perturbation search kernel backtracks the boundary projection results in the order of release state to transition state and transition state to locked state, generating a boundary change amount each time it backtracks; After applying each boundary change to the boundary projection result, perform difference alignment, retain the boundary changes that reduce the difference, and discard the boundary changes that do not reduce the difference. Among the retained boundary changes, the boundary change that minimizes the difference between the boundary projection result and the twin synchronization result and minimizes the backtracking steps is selected as the minimum boundary change. The source of the anomaly is determined based on the boundary state to which the minimum boundary change belongs.

[0013] Optionally, the step of re-executing operator coupling and boundary projection includes: Set a counterfactual boundary rewrite kernel at the feedback end of the reversible release boundary projection DeepONet; Write the candidate disposal action into the counterfactual boundary rewriting kernel, and rewrite the gating path according to the one-way release boundary transfer path from the locked state to the transition state and from the transition state to the releaseable state, thus forming the candidate boundary change quantity; The candidate boundary change is injected into the release condition folding kernel set in the release boundary encoder, and the release condition is controlled to be recompressed to one of the locked state, transition state, and releaseable state, and the Branch output is updated. The updated Branch output and Trunk output are fed into the release boundary projection kernel. The state-by-state coupling and three-state adsorption are re-executed according to the release boundary position to be simulated, and the boundary projection result after re-feedback is obtained.

[0014] Optionally, the steps of outputting dynamic monitoring results, anomaly sources, and handling priorities include: Set a boundary migration sorting kernel at the output of the reversible release boundary projection DeepONet, and input the boundary projection results before and after reflow into the boundary migration sorting kernel. The boundary migration sorting kernel performs boundary migration comparisons on candidate actions based on the decrease in locked state, the transition in transition state, and the increase in released state. The boundary migration comparison results are obtained, and candidate actions that have not decreased in locked state and have not increased in released state are eliminated, while candidate actions that have decreased in locked state and increased in released state are retained. The remaining candidate actions are sorted in the following order: first, the increase in the release state; second, the decrease in the lock state; and third, the transfer in the transition state. This yields the action priority. The boundary migration comparison results, the boundary status to which the minimum boundary change belongs, and the handling priority are combined and output to form dynamic monitoring results, anomaly sources, and handling priorities.

[0015] The beneficial effects of this invention are: (1) This invention divides the storage release process into locked state, transition state and releaseable state by DeepONet, which is a reversible release boundary projection. By using the release boundary projection kernel to constrain and couple the Branch output and Trunk output, the changes in book capacity are mapped to the real release boundary, reducing misjudgments of release caused by snapshot lineage, reference occupation, migration balancing and retention restrictions, and improving the accuracy of data center storage space monitoring.

[0016] (2) The present invention uses a reverse boundary perturbation search kernel to perform difference alignment between the boundary projection result and the twin synchronization result, and searches for the minimum boundary change in reverse from the release state to the transition state and from the transition state to the locked state. Then, it determines the source of the anomaly based on the boundary state to which the minimum boundary change belongs, avoiding reliance on threshold alarms and improving the accuracy of anomaly location such as snapshot residue, unremoved references, and incomplete migration.

[0017] (3) This invention converts candidate disposal actions into candidate boundary changes by rewriting the counterfactual boundary kernel, feeds them back to the release boundary encoder, re-executes operator coupling and boundary projection, and generates disposal priorities based on the decrease in locked state, the transfer in transitional state, and the increase in releaseable state, so that disposal actions such as cleanup, migration, expansion, and quota can be sorted according to the actual release contribution, thereby improving the efficiency of storage operation and maintenance. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a dynamic monitoring method for data center storage space based on twin modeling proposed in this invention; Figure 2 This is a schematic diagram of the reversible release boundary projection DeepONet structure, which is a dynamic monitoring method for data center storage space based on twin modeling proposed in this invention. Figure 3 This is a schematic diagram of the counterfactual backfilling and priority generation structure of a data center storage space dynamic monitoring method based on twin modeling proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figures 1-3 A method for dynamic monitoring of data center storage space based on twin modeling includes the following steps: Deploy DeepONet, a reversible release boundary projection, in the data center storage space twin to divide the storage release process into locked state, transitional state, and releaseable state, forming a release boundary, and update the twin to obtain the twin synchronization result; In a specific embodiment of the present invention, the network structure and training method of the reversible release boundary projection DeepONet include: The reversible release boundary projection DeepONet includes a release boundary encoder, a boundary position query operator, a release boundary projection kernel, a reverse boundary perturbation search kernel, a counterfactual boundary rewriting kernel, and a boundary migration sorting kernel. The release boundary encoder, as a branch-side structure, receives release conditions and compresses them into a state expression of locked, transitional, or releaseable states, forming the branch output. The boundary position query operator, as a trunk-side structure, receives monitored objects, event stages, and candidate disposal actions, mapping the three types of inputs to the same release boundary position, forming the trunk output. The release boundary projection kernel is located after the branch output and trunk output, performing operator coupling and boundary projection on both. The Branch output uses a three-state representation, corresponding to the locked state output component, the transition state output component, and the releaseable state output component, respectively. After the release condition folding kernel completes the boundary folding, it retains only one output component corresponding to the current release condition and closes the other two output components, so that the Branch output carries only one release boundary state. This process is used to prevent the same memory release process from falling into the locked state, the transition state, and the releaseable state at the same time. Trunk output uses a three-state coordinate representation, corresponding to the locked state position, transition state position, and releaseable state position in the release boundary, respectively; the counterfactual boundary coordinate latch performs forward coordinate locking on the monitored object, event stage, and candidate disposal action, and then verifies whether the three types of inputs are consistent through reverse readback; if the verification is consistent, the release boundary coordinates are written to Trunk output; if the verification is inconsistent, the release boundary coordinates are blocked and do not enter subsequent operator coupling; The release boundary projection kernel performs state-by-state coupling on the Branch and Trunk outputs. State-by-state coupling proceeds sequentially through the locked state, transitional state, and releaseable state. The Branch output defines the release conditions, and the Trunk output defines the release boundary position to be derived. Together, they determine whether the coupling result can enter the corresponding state. When the Branch output corresponds to the locked state, the coupling result must not enter the transitional or releaseable state. When the Branch output corresponds to the transitional state, the coupling result must not enter the releaseable state. When the Branch output corresponds to the releaseable state, the coupling result enters the releaseable state. After state-by-state coupling, the release boundary projection kernel performs three-state adsorption; if the release condition is not met and the candidate disposal action is not entered, the coupling result is adsorbed to the locked state; if the candidate disposal action is entered but the disposal action is not completed, the coupling result is adsorbed to the transition state; if the release condition is met, the coupling result is adsorbed to the releasable state; the coupling result after the three-state adsorption is completed is used as the boundary projection result. The training process of DeepONet for reversible release boundary projection includes release condition folding training, boundary position query training, boundary projection training, inverse perturbation search training, and counterfactual backfeeding sorting training. Release condition folding training is used to ensure that the Branch output is consistent with the known release state; boundary position query training is used to ensure that the Trunk output is consistent with the monitored object, event stage, and candidate disposal action; boundary projection training is used to ensure that the boundary projection result is consistent with the adsorption rules of the locked state, transition state, and releaseable state; inverse perturbation search training is used to ensure that the minimum boundary change amount can make the boundary projection result converge to the twin synchronization result; and counterfactual backfeeding sorting training is used to ensure that the disposal priority is consistent with the release boundary migration effect before and after backfeeding. During training, historical samples with known release states, known candidate actions, and known action results are input into the DeepONet reversible release boundary projection. The release boundary encoder generates branch output, the boundary position query operator generates trunk output, the release boundary projection kernel generates boundary projection results, the inverse boundary perturbation search kernel generates minimum boundary change, the counterfactual boundary rewriting kernel generates candidate boundary change, and the boundary migration sorting kernel generates action priority. Based on the differences between each output and the corresponding labeled results in the historical samples, the parameters of the release boundary encoder, boundary position query operator, release boundary projection kernel, inverse boundary perturbation search kernel, counterfactual boundary rewriting kernel, and boundary migration sorting kernel are updated until the training results meet the preset training stopping conditions. During prediction, the network parameters remain unchanged after training. The current release conditions, monitoring objects, event stages, and candidate actions are input, and the Branch output, Trunk output, boundary projection results, minimum boundary change, candidate boundary change, and action priority are generated in sequence. The above results are used to perform dynamic monitoring of data center storage space, anomaly source determination, and action priority output. Replace DeepONet's Branch with a release boundary encoder to perform boundary folding on release conditions formed by snapshot lineage, reference holding, migration balancing, and retention constraints, so that the Branch output carries the release conditions; Replace DeepONet's Trunk with a boundary location query operator, perform a joint query on the monitored object, event stage, and candidate action, and make the Trunk output point to the release boundary location to be deduced; Operator coupling is performed on the Branch output and Trunk output. The coupling result is then subjected to boundary projection according to the release condition. Capacities that do not meet the release condition are mapped to locked states, capacities in the disposal simulation are mapped to transition states, and capacities that meet the release condition are mapped to releaseable states, thus obtaining the boundary projection result. The boundary projection results are aligned with the twin synchronization results by difference. The minimum boundary change that makes the difference converge is searched in reverse along the locked state, transition state, and release state. The source of the anomaly is determined according to the boundary state to which the minimum boundary change belongs. The candidate disposal action is converted into a candidate boundary change, fed back to the release boundary encoder, the release condition carried in the Branch output is updated, and the operator coupling and boundary projection are re-executed. Based on the reduction in the locked state, the transfer in the transition state, and the increase in the release state before and after reinjection, the system outputs dynamic monitoring results, sources of anomalies, and handling priorities.

[0021] 2. The method for dynamic monitoring of data center storage space based on twin modeling according to claim 1, characterized in that the step of updating the twin to obtain the twin synchronization result includes: A three-state release channel is established at the output of the reversible release boundary projection DeepONet, and a boundary residual correction gate is embedded in the three-state release channel; In this invention, the three-state release channel is set at the output end of the DeepONet reversible release boundary projection to carry the state of the storage release process in the release boundary; the three-state release channel consists of a locked state channel, a transition state channel, and a releaseable state channel; the locked state channel receives storage release processes that do not meet the release conditions, the transition state channel receives storage release processes that have entered the candidate disposal action deduction but have not reached the releaseable state, and the releaseable state channel receives storage release processes that meet the release conditions and can participate in the capacity release calculation; In this invention, the boundary residual correction gate is embedded in the three-state release channel to control the opening and closing of the gated path between the locked state, the transition state, and the releaseable state. The boundary residual correction gate receives the current release state and the current state of the three-state release channel, compares the state difference between the two at the release boundary, and adjusts the gated path from the locked state to the transition state and the gated path from the transition state to the releaseable state according to the state difference. Write the storage release process into the locked state, and use the boundary residual correction gate to compare the boundary residual between the current release state and the current state of the three-state release channel. In this invention, the current release state is determined by the data center storage space twin during synchronous updates, and is used to represent the release position corresponding to the current moment in the storage release process; the current release state is determined based on the release conditions, which are formed by snapshot lineage, reference occupancy, migration balancing, and retention restrictions; when the release conditions are not met, the current release state corresponds to the locked state; when candidate disposal actions enter the deduction and the release conditions are not all met, the current release state corresponds to the transitional state; when the release conditions are met and the disposal deduction is completed, the current release state corresponds to the releaseable state. In this invention, the current state of the three-state release channel is the current distribution position of the storage release process in the locked state channel, the transition state channel, and the releaseable state channel; the current state of the three-state release channel is determined by the channel position after the last synchronization update and is used to compare with the current release state; when the two are inconsistent, the boundary residual correction gate controls the opening and closing of the gate path according to the difference direction; In this invention, the boundary residual is the offset between the current release state and the current state of the three-state release channel on the release boundary. When the boundary residual indicates that the release condition is not met, the boundary residual correction gate blocks the gated path from the locked state to the transition state. When the boundary residual indicates that the disposal deduction is not completed, the boundary residual correction gate blocks the gated path from the transition state to the releaseable state. When the boundary residual indicates that the release condition is met and the disposal deduction is completed, the boundary residual correction gate opens the gated path from the locked state to the transition state and the gated path from the transition state to the releaseable state. In this invention, the determination of boundary residuals is performed according to the state position in the release boundary. The locked state, transitional state, and releasable state are sequentially used as the three state positions on the release boundary. When the current release state is at the same state position as the current state of the three-state release channel, the boundary residual is determined to be zero. When the current release state is after the current state of the three-state release channel, the boundary residual indicates insufficient release progress. When the current release state is before the current state of the three-state release channel, the boundary residual indicates a rollback of the release condition. When the boundary residual indicates insufficient release progress, the boundary residual correction gate opens the corresponding gated path according to the degree to which the release condition is satisfied. When the boundary residual indicates a rollback of the release condition, the boundary residual correction gate closes the gated path from the transitional state to the releasable state and rewrites the corresponding stored release process into the locked state or transitional state. In this invention, the boundary residual can also be determined according to the three-state coding distance; the locked state, transition state, and releaseable state correspond to three mutually exclusive state codes respectively. When the current release state code is consistent with the current state code of the three-state release channel, the boundary residual is zero; when the two codes are inconsistent, the boundary residual direction is determined according to the state distance between the two in the release boundary; the boundary residual direction is used to control the gating path from the locked state to the transition state and the gating path from the transition state to the releaseable state. When the release condition corresponding to the boundary residual is not met, the gated path from the locked state to the transition state is closed; When the deduction of the corresponding handling of boundary residuals is not completed, the gating path from the transition state to the release state is closed; When the release condition corresponding to the boundary residual is met and the disposal simulation is completed, the gated path from the locked state to the transition state and the gated path from the transition state to the releaseable state are opened. In this invention, the gated path from the locked state to the transition state is used to transfer the memory release process from the locked state channel to the transition state channel; the gated path from the transition state to the releaseable state is used to transfer the memory release process from the transition state channel to the releaseable state channel; both gated paths are unidirectional, prohibiting rollback from the releaseable state to the transition state and prohibiting rollback from the transition state to the locked state. When a new release condition is not met, the memory release process is rewritten into the locked state channel by the boundary residual correction gate. In this invention, the release condition being met means that snapshot lineage is released, reference occupancy is released, migration balancing is completed, and retention restrictions are released; the release condition not being met means that there is an unreleased state among snapshot lineage, reference occupancy, migration balancing, and retention restrictions; the release condition is used to control the opening and closing of the gated path of the boundary residual correction gate, and also to determine whether the storage release process allows the transition from the locked state to the transition state, and from the transition state to the releaseable state; In this invention, "disposal simulation complete" means that after a candidate disposal action enters the reversible release boundary projection DeepONet, the storage release process has completed the boundary projection, and the corresponding release condition is met; "disposal simulation incomplete" means that a candidate disposal action has entered the reversible release boundary projection DeepONet, but the boundary projection is not completed, or the release condition is still not met after the boundary projection is completed; the disposal simulation state is used to control the gating path from the transition state to the releaseable state. Update the locked state, transitional state, and releaseable state according to the gating path after opening, and write the three-state release channel back to the data center storage space twin to obtain the twin synchronization result.

[0022] In this invention, after the three-state release channel completes the gated path update, it writes the current states of the locked, transitional, and releaseable channels back to the data center storage space twin. The written-back data center storage space twin forms a twin synchronization result. The twin synchronization result is used for subsequent boundary projection result difference alignment, minimum boundary change reverse search, and anomaly source determination.

[0023] In this embodiment, the step of performing boundary folding includes: Set a release condition folding core within the release boundary encoder and input the release condition into the release condition folding core; In this invention, the release condition folding kernel is set inside the release boundary encoder, located between the release condition input and the Branch output. The release condition folding kernel receives the release condition formed by snapshot lineage, reference occupancy, migration balancing, and retention constraints, performs three-state compression on the release condition, and outputs locked state output components, transition state output components, and releasable state output components. The release condition folding kernel is used to compress discrete release constraints into a single state expression corresponding to the release boundary. The release condition folding kernel compresses the release condition in a unidirectional state sequence of locked state, transition state, and releaseable state, forming the locked state output component, the transition state output component, and the releaseable state output component; In this invention, the unidirectional state sequence is locked state, transition state, and releasable state. The release condition folding core processes the release conditions in this order. The locked state is used to carry the state where the release conditions are not met, the transition state is used to carry the state where some release conditions have been released but the releasable state has not yet been reached, and the releasable state is used to carry the state where all release conditions are met. The unidirectional state sequence restricts the release conditions to advance only from the locked state to the transition state and from the transition state to the releasable state during the folding process. In this invention, the compressed release condition refers to the release condition folding kernel performing a unified state merging of snapshot lineage, reference holding, migration balancing, and retention restrictions; when there are unresolved snapshot lineages, unresolved reference holdings, incomplete migration balancing, or unresolved retention restrictions, the release condition is compressed to a locked state; when some of the above release conditions are resolved and the remaining release restrictions still need to be addressed, the release condition is compressed to a transitional state; when all of the above release conditions are resolved, the release condition is compressed to a releaseable state. In this invention, the locked-state output component, the transition-state output component, and the releaseable-state output component are the output parts of the three-state release boundary corresponding to the release condition folding kernel; the locked-state output component is used to write the non-releaseable state to the Branch output, the transition-state output component is used to write the disposal deduction state to the Branch output, and the releaseable-state output component is used to write the releaseable state to the Branch output; the three types of output components are mutually exclusive in the same boundary folding. When the release condition is compressed to the locked state, the transition state output component and the releaseable state output component are turned off, while the locked state output component is retained. When the release condition is compressed to the transition state, the locked state output component and the releaseable state output component are turned off, while the transition state output component is retained. When the release condition is compressed to the releaseable state, the locked state output component and the transition state output component are turned off, while the releaseable state output component is retained. In this invention, when the release condition is compressed to the locked state, it indicates that the storage release process is still constrained by at least one type of release constraint among snapshot lineage, reference occupancy, migration balancing, and retention constraint; when the release condition is compressed to the transition state, it indicates that the release constraint has begun to be released but has not yet formed a releaseable state; when the release condition is compressed to the releaseable state, it indicates that snapshot lineage, reference occupancy, migration balancing, and retention constraint have all met the release requirements. In this invention, closing the output component means that the release condition folding check performs invalidation processing on the output component that does not belong to the current compression state; when the release condition is compressed to the locked state, the transition state output component and the releasable state output component are invalidated; when the release condition is compressed to the transition state, the locked state output component and the releasable state output component are invalidated; when the release condition is compressed to the releasable state, the locked state output component and the transition state output component are invalidated. In this invention, retaining the output component means that the release condition folding kernel only retains the output component corresponding to the current compression state; the retained output component continues to participate in Branch output synthesis, while the closed output component does not participate in Branch output synthesis; the release condition folding kernel, through this process, makes the Branch output correspond to only one of the locked state, transition state, and releaseable state; In this invention, the mutual exclusion retention mechanism is used to prevent the same memory release process from falling into multiple release states simultaneously; the release condition folding kernel allows only one type of output component to remain in a valid state in each boundary fold, while the other two types of output components remain in a closed state; the mutual exclusion retention mechanism ensures that the release condition output by the release boundary encoder is consistent with a single state in the release boundary. The retained output components are combined into a Branch output, which carries the release conditions consistent with the release boundary.

[0024] In this invention, after the release condition folding kernel completes mutual exclusion retention, the output component that remains in a valid state is written to the Branch output. The Branch output no longer retains independent constraint expressions for snapshot lineage, reference occupancy, migration balancing, and retention restrictions, but instead retains a single release boundary expression among the locked state, transition state, and releaseable state.

[0025] In this invention, the release condition carried by the Branch output is consistent with the release boundary, meaning that the state expression in the Branch output corresponds to one of the locked, transitional, or releaseable states. After the Branch output enters the subsequent operator coupling, DeepONet participates in the boundary projection according to this state expression, so that the coupling result is constrained by the release boundary.

[0026] In this embodiment, the step of performing a joint query includes: On the Trunk side, a counterfactual boundary coordinate latch kernel is established, which maps the monitored object, event stage, and candidate disposal action to the release boundary to form the release boundary coordinates. Forward coordinate locking is performed along the direction from the locked state to the transition state and from the transition state to the releaseable state. In this invention, the counterfactual boundary coordinate latch kernel is set inside the boundary position query operator, located between the Trunk input and the Trunk output. The counterfactual boundary coordinate latch kernel receives the monitored object, event stage, and candidate action, and constrains the three types of inputs to the same release boundary position to avoid inconsistent Trunk outputs after the monitored object, event stage, and candidate action are queried separately. In this invention, the release boundary coordinates are the joint position of the monitored object, event stage, and candidate disposal action within the release boundary; the release boundary coordinates fall into one of the state positions of locked state, transition state, and releaseable state, and correspond to one of the release boundary transition path positions from locked state to transition state and from transition state to releaseable state; the release boundary coordinates are written as Trunk output into the subsequent operator coupling process; In this invention, when the coordinates are mapped to the release boundary, the counterfactual boundary coordinate latch core binds the monitored object to the object position in the release boundary, the event phase to the state position in the release boundary, and the candidate disposal action to the action position in the release boundary. After the three types of positions are bound, they are combined into the same release boundary coordinates. If the same release boundary coordinates are not formed, they are not written to the Trunk output. In this invention, forward coordinate locking is performed along the direction from the locked state to the transition state and from the transition state to the releaseable state; the counterfactual boundary coordinate latch kernel restricts the release boundary coordinates to the unidirectional release boundary transfer path, so that the Trunk output can only point to the release boundary position to be deduced that conforms to the release advancement direction, and prohibits the release boundary coordinates from crossing the unidirectional sequence of the locked state, transition state and releaseable state; Read back the release boundary coordinates after locking the forward coordinates in the direction from the release state to the transition state and from the transition state to the lock state to obtain the readback object, readback stage, and readback action; In this invention, reverse readback is performed in the direction from the release state to the transition state and from the transition state to the locked state; the counterfactual boundary coordinate latch kernel performs reverse parsing on the release boundary coordinates after forward coordinate locking, and reads the object information, stage information and action information corresponding to the release boundary coordinates, which is used to verify whether the forward coordinate locking maintains input consistency; In this invention, the readback object, readback stage, and readback action are obtained by reverse readback of the release boundary coordinates; the readback object corresponds to the monitored object locked in the release boundary coordinates, the readback stage corresponds to the event stage locked in the release boundary coordinates, and the readback action corresponds to the candidate disposal action locked in the release boundary coordinates; the three types of readback results serve as the basis for determining whether to latch or block the release boundary coordinates. When the readback object is consistent with the monitored object, the readback phase is consistent with the event phase, and the readback action is consistent with the candidate disposal action, latch and release the boundary coordinates. In this invention, when the readback object is consistent with the monitoring object, the readback stage is consistent with the event stage, and the readback action is consistent with the candidate disposal action, the counterfactual boundary coordinate latch kernel keeps the release boundary coordinates valid; after the valid release boundary coordinates are latched, they are written to the Trunk output and used as the release boundary position to be deduced to participate in the operator coupling between the Branch output and the Trunk output. If any of the readback object, readback phase, or readback action is inconsistent, the boundary coordinates of the occlusion release will be interrupted. In this invention, when any of the readback object, readback stage, or readback action is inconsistent with the input, the counterfactual boundary coordinate latch kernel will set the release boundary coordinate to an invalid state; the invalid release boundary coordinate is not written to the Trunk output, and the boundary position query operator stops passing the corresponding query result to the subsequent operator coupling process; In this invention, consistency verification is used to confirm whether the release boundary coordinates simultaneously correspond to the monitored object, the event stage, and the candidate action; the counterfactual boundary coordinate latch kernel compares the readback object with the monitored object, the readback stage with the event stage, and the readback action with the candidate action item by item; latching is performed when all three comparisons are consistent; blocking is performed when any one comparison is inconsistent. Write the latched release boundary coordinates into the Trunk output, so that the Trunk output points to the release boundary position to be calculated.

[0027] In this invention, the latched release boundary coordinates are written as Trunk outputs into the reversible release boundary projection DeepONet. The Trunk output points to the release boundary position to be deduced and is coupled with the Branch output formed by the release boundary encoder into the operator. This writing process enables subsequent boundary projections to be executed under the same monitored object, the same event stage, and the same candidate action.

[0028] In this embodiment, the step of performing boundary projection includes: Set a release boundary projection kernel at the coupling end of the reversible release boundary projection DeepONet, and use the Branch output as the release condition constraint and the Trunk output as the release boundary position to be deduced. In this invention, the release boundary projection kernel is set at the coupling end of the reversible release boundary projection DeepONet to receive Branch output and Trunk output. The release boundary projection kernel performs operator coupling between the release condition carried by the Branch output and the release boundary position to be deduced pointed to by the Trunk output, and restricts the coupling result within the release boundary formed by the locked state, transition state and releaseable state. In this invention, the coupling end is located after the release boundary encoder and the boundary position query operator, and receives the Branch output formed by the release boundary encoder and the Trunk output formed by the boundary position query operator. The coupling end does not directly output the capacity prediction value, but sends the Branch output and Trunk output into the release boundary projection kernel to form a coupling result constrained by the release conditions and the release boundary position to be deduced. In this invention, the Branch output is obtained by performing boundary folding on the release condition using a release boundary encoder. After the Branch output enters the release boundary projection kernel, it is used to limit whether the coupling result can enter the transition state or the releaseable state. When the Branch output corresponds to the locked state, the release boundary projection kernel blocks the coupling result from entering the transition state and the releaseable state. When the Branch output corresponds to the transition state, the release boundary projection kernel blocks the coupling result from entering the releaseable state. When the Branch output corresponds to the releaseable state, the release boundary projection kernel allows the coupling result to enter the releaseable state. In this invention, the Trunk output is obtained by performing a joint query on the monitored object, event stage, and candidate disposal action by the boundary position query operator; after the Trunk output enters the release boundary projection kernel, it is used to limit the release boundary position corresponding to the coupling result, so that the boundary projection is performed under the same monitored object, the same event stage, and the same candidate disposal action. The release boundary projection kernel performs state-by-state coupling on the Branch output and Trunk output according to the locked state, transition state and releaseable state, so that the coupling result is simultaneously constrained by the release condition and the release boundary position to be deduced. In this invention, state-by-state coupling refers to the release boundary projection kernel performing coupling on the Branch output and Trunk output in sequence according to the locked state, transition state, and releaseable state. During the coupling process, the locked state corresponds to the boundary position that cannot be released, the transition state corresponds to the boundary position in the candidate disposal action deduction, and the releaseable state corresponds to the boundary position that meets the release condition. The release boundary projection kernel filters the coupling results according to the state order to avoid the coupling result falling into more than two states at the same time. In this invention, the coupling result is jointly determined by the release condition carried by the Branch output and the release boundary position to be deduced pointed to by the Trunk output; the release condition determines the range of states that the coupling result is allowed to enter, and the release boundary position to be deduced determines the monitoring object, event stage and candidate disposal action corresponding to the coupling result; when the two are inconsistent, the release boundary projection kernel will keep the coupling result in a locked state or a transition state and will not output a releaseable state result. The release boundary projection kernel performs three-state adsorption on the coupling result according to the release conditions. When the release conditions are not met and the candidate disposal action is not entered, the coupling result is adsorbed to the locked state. When the candidate disposal action is entered and the disposal action is not completed, the coupling result is adsorbed to the transition state. When the release conditions are met, the coupling result is adsorbed to the releasable state. In this invention, three-state adsorption refers to the release boundary projection kernel classifying the coupling result into one of the locked state, transition state, or releasable state. Three-state adsorption is performed after state-by-state coupling and is used to adsorb coupling results that do not meet the release conditions into the locked state, to adsorb coupling results that have entered the candidate disposal action deduction but the disposal deduction is not completed into the transition state, and to adsorb coupling results that meet the release conditions into the releasable state. In this invention, the coupling result is adsorbed into a locked state when the release condition is not met and the candidate disposal action deduction has not yet begun; the coupling result is adsorbed into a transitional state when the candidate disposal action deduction has begun but the deduction is not yet complete. The coupling result is adsorbed into a releasable state when the release condition is met; the release boundary projection kernel retains only one adsorbed state in the same boundary projection. In this invention, "not entering the candidate action deduction" means that the candidate action has not yet been written into the boundary position query operator, or the candidate action has not been locked by the coordinate latching kernel of the counterfactual boundary coordinate latching kernel; the coupling result in this state does not participate in the release boundary migration, and the release boundary projection kernel keeps it in the locked state; In this invention, "disposal simulation incomplete" means that the candidate disposal action has been written into the boundary position query operator and the Trunk output has pointed to the release boundary position to be simulated, but the release condition has not yet been met or the three-state adsorption has not yet entered the releaseable state; when the disposal simulation is incomplete, the release boundary projection kernel will keep the coupling result in the transition state. The coupling result after the three-state adsorption is used as the boundary projection result.

[0029] In this invention, after the release boundary projection kernel completes the three-state adsorption, the coupling result of the adsorbed kernel to the locked state, transition state, or releasable state is used as the boundary projection result. The boundary projection result is used for subsequent difference alignment with the twin synchronization result, serving as input for the inverse search of the minimum boundary change.

[0030] In this embodiment, the step of performing difference alignment includes: Inverse boundary perturbation search kernel is set at the causal end of DeepONet with reversible release boundary projection, and the boundary projection result is used as the search starting point and the twin synchronization result is used as the convergence target. In this invention, the reverse boundary perturbation search kernel is set at the causal end of the reversible release boundary projection DeepONet, and receives the boundary projection result and the twin synchronization result. The reverse boundary perturbation search kernel takes the boundary projection result as the search starting point and the twin synchronization result as the convergence target, and searches along the reverse state sequence of the release boundary to reduce the boundary change amount that can reduce the difference between the two. In this invention, the cause-determining end is located after the release boundary projection kernel and is used to receive the boundary projection results; the cause-determining end does not perform capacity prediction, but performs reverse perturbation search between the boundary projection results and the twin synchronization results, and outputs the minimum boundary change and the source of the anomaly; In this invention, the search starting point is the boundary projection result output by the release boundary projection kernel; the reverse boundary perturbation search kernel starts from the search starting point and retreats back to the release boundary state, gradually changing the boundary state in the boundary projection result, so that the boundary projection result moves closer to the twin synchronization result; In this invention, the convergence objective is the twin synchronization result. After each backtracking, the reverse boundary perturbation search kernel aligns the adjusted boundary projection result with the twin synchronization result based on the difference, and uses the reduction of the difference as the criterion for whether to retain the boundary change. The reverse boundary perturbation search kernel backtracks the boundary projection results in the order of release state to transition state and transition state to locked state, generating a boundary change amount each time it backtracks; In this invention, the reverse rollback sequence is from the releasable state to the transition state, and from the transition state to the locked state. This sequence is opposite to the forward transfer direction of the release boundary and is used to trace the boundary position that caused the deviation from the already projected release state, so that the source of the anomaly can be located to the specific state in the release boundary. In this invention, the boundary change is generated by a reverse backtracking; the boundary change records the state change that occurs when the boundary projection result backtracks from the release state to the transition state, or from the transition state to the locked state; each boundary change participates in a difference alignment. After applying each boundary change to the boundary projection result, perform difference alignment, retain the boundary changes that reduce the difference, and discard the boundary changes that do not reduce the difference. In this invention, when the boundary change is applied to the boundary projection result, the inverse boundary perturbation search kernel modifies the state position in the boundary projection result according to the boundary change. The modified boundary projection result is used to recalculate the difference between the result and the twin synchronization result. In this invention, difference alignment refers to comparing the modified boundary projection result with the twin synchronization result in the same release boundary state sequence; the comparison object is the difference in state position in the locked state, transition state, and releaseable state; when the difference in state position decreases, the boundary change amount enters the retention set; when the difference in state position does not decrease, the boundary change amount is removed. In this invention, the reduction of difference is determined based on the state position difference. The boundary projection result and the twin synchronization result are mapped to the state positions in the locked state, transition state, and release state, respectively, and the state position difference before and after the boundary change is applied is calculated. If the state position difference between the boundary projection result and the twin synchronization result is less than the state position difference before the boundary change is applied, the difference is determined to be reduced. If the state position difference is equal to or greater than the state position difference before the boundary change is applied, the difference is determined not to be reduced. In this invention, the reduction of difference can also be judged according to the difference in state sequence; the boundary projection results of the same monitored object in a continuous synchronization cycle are arranged into a projection state sequence, and the twin synchronization results are arranged into a synchronization state sequence. The number of inconsistent positions between the projection state sequence and the synchronization state sequence before and after the boundary change is applied is compared; when the number of inconsistent positions decreases, the difference is judged to be reduced; when the number of inconsistent positions does not decrease, the difference is judged to be not reduced. In this invention, the reverse boundary perturbation search kernel filters each boundary change; boundary changes that reduce the difference are retained, while boundary changes that do not reduce the difference are removed; the filtered boundary changes are used to determine the minimum boundary change in the subsequent process. Among the retained boundary changes, the boundary change that minimizes the difference between the boundary projection result and the twin synchronization result and minimizes the backtracking steps is selected as the minimum boundary change. In this invention, difference convergence means that after the boundary projection result is adjusted by the boundary change amount, the difference in state position between it and the twin synchronization result reaches the preset convergence condition; after difference convergence, the reverse boundary perturbation search kernel stops to continue backtracking and selects the minimum boundary change amount from the retained boundary change amounts. In this invention, difference convergence is determined by at least one of state position difference, three-state coding distance, and state sequence difference. When using state position difference determination, difference convergence is determined when the boundary projection result and the twin synchronization result are in the same state position, or the state position difference between the two reaches a preset allowable range. When using three-state coding distance determination, difference convergence is determined when the three-state code corresponding to the boundary projection result is consistent with the three-state code corresponding to the twin synchronization result, or the coding distance between the two reaches a preset allowable range. When using state sequence difference determination, difference convergence is determined when the number of inconsistencies between the projected state sequence and the synchronized state sequence reaches a preset allowable number. The reverse boundary perturbation search kernel stops backtracking after the difference converges; if there are several boundary changes that all cause the difference to converge, the boundary change with the fewest backtracking steps is selected as the minimum boundary change; if the backtracking steps are the same, the boundary change that minimizes the deviation of the release state is selected as the minimum boundary change; the boundary state to which the minimum boundary change belongs is used to determine the source of the anomaly. In this invention, the backtracking step count is used to represent the number of backward backtracking steps required to reach the difference convergence state from the boundary projection result; when multiple boundary changes can all make the difference converge, the reverse boundary perturbation search kernel selects the boundary change with the fewest backtracking steps as the minimum boundary change. The source of the anomaly is determined based on the boundary state to which the minimum boundary change belongs.

[0031] In this invention, the boundary state to which the minimum boundary change belongs is the release boundary state that the minimum boundary change falls into after it has completed its rollback. The boundary states include locked, transitional, and releasable states, used to determine the release phase where the anomaly originates.

[0032] In this invention, when the boundary state to which the minimum boundary change belongs is locked, the release condition corresponding to the anomaly source is not met; when the boundary state to which the minimum boundary change belongs is transitional, the deduction of the handling corresponding to the anomaly source is not completed; when the boundary state to which the minimum boundary change belongs is releasable, there is a releasable state deviation between the boundary projection result corresponding to the anomaly source and the twin synchronization result.

[0033] In this embodiment, the step of re-executing operator coupling and boundary projection includes: Set a counterfactual boundary rewrite kernel at the feedback end of the reversible release boundary projection DeepONet; In this invention, the counterfactual boundary rewriting kernel is set at the feedback end of the DeepONet projected by the reversible release boundary. It receives candidate disposal actions and converts the candidate disposal actions into gated path rewriting quantities that act on the release boundary. The counterfactual boundary rewriting kernel does not directly output disposal scores, but changes the passage conditions from the locked state to the transition state and from the transition state to the release state, forming candidate boundary change quantities. In this invention, the feedback end is located after the boundary projection and before the release boundary encoder, and is used to send the candidate boundary change amount generated by the candidate disposal action back to the release boundary encoder; the feedback end is connected to the counterfactual boundary rewriting kernel, the release condition folding kernel and the release boundary projection kernel, so that the candidate disposal action can re-influence the Branch output and subsequent boundary projection. Write the candidate disposal action into the counterfactual boundary rewriting kernel, and rewrite the gating path according to the one-way release boundary transfer path from the locked state to the transition state and from the transition state to the releaseable state, thus forming the candidate boundary change quantity; In this invention, the counterfactual boundary rewriting kernel rewrites the gating path according to the unidirectional release boundary transfer path from the locked state to the transition state and from the transition state to the releaseable state; when the release condition corresponding to the candidate disposal action is released, the gating path from the locked state to the transition state is opened; when the disposal deduction corresponding to the candidate disposal action is completed, the gating path from the transition state to the releaseable state is opened; the gating path that does not meet the release advancement condition remains closed; In this invention, the candidate boundary change amount is the state rewriting amount formed by the candidate disposal action in the release boundary; the candidate boundary change amount records the change result of the candidate disposal action on the transition conditions between the locked state, the transition state, and the releaseable state, and is used to trigger the release condition folding kernel to recompress the release condition; The candidate boundary change is injected into the release condition folding kernel set in the release boundary encoder, and the release condition is controlled to be recompressed to one of the locked state, transition state, and releaseable state, and the Branch output is updated. In this invention, after the candidate boundary change enters the release condition folding kernel, the release condition folding kernel adjusts the mutual exclusion retention relationship of the locked state output component, the transition state output component, and the releaseable state output component according to the candidate boundary change; when the candidate boundary change does not change the release propulsion condition, the release condition folding kernel maintains the original compression state; when the candidate boundary change changes the release propulsion condition, the release condition folding kernel reselects the corresponding output component. In this invention, when the release condition is recompressed, the release condition folding kernel determines the state corresponding to the release condition based on the change in the candidate boundary; when the change in the candidate boundary has not released the locking condition, the release condition is compressed to the locked state; when the change in the candidate boundary triggers the processing deduction but has not completed the release, the release condition is compressed to the transition state; when the change in the candidate boundary completes the release advancement, the release condition is compressed to the releaseable state. In this invention, after the release condition folding kernel completes recompression, the re-preserved output components are written into the Branch output to replace the Branch output before the refeedback; the updated Branch output carries the release conditions after the candidate boundary change, which is used by the release boundary projection kernel to re-execute operator coupling. The updated Branch output and Trunk output are fed into the release boundary projection kernel. The state-by-state coupling and three-state adsorption are re-executed according to the release boundary position to be simulated, and the boundary projection result after re-feedback is obtained.

[0034] In this invention, the updated Branch output is used as a new release condition constraint in the release boundary projection kernel. The Trunk output still points to the same release boundary position to be simulated, which is used to keep the monitored object, event stage, and candidate disposal action unchanged, so that the boundary projections before and after the recharge are comparable.

[0035] In this invention, after receiving the updated Branch and Trunk outputs, the release boundary projection kernel re-executes state-by-state coupling according to the locked state, transitional state, and releasable state, and performs three-state adsorption based on the updated release conditions. The re-execution result is used to determine whether the candidate disposal action changes the state position of the stored release process within the release boundary.

[0036] In this invention, the boundary projection result after reinjection is the boundary projection result obtained again after the candidate boundary change. The boundary projection result after reinjection is used to compare with the boundary projection result before reinjection to determine the impact of the candidate treatment action on the reduction of the locked state, the transfer of the transition state, and the increase of the releaseable state.

[0037] In this embodiment, the steps of outputting dynamic monitoring results, anomaly sources, and handling priorities include: Set a boundary migration sorting kernel at the output of the reversible release boundary projection DeepONet, and input the boundary projection results before and after reflow into the boundary migration sorting kernel. In this invention, the boundary migration sorting kernel is set at the output end of the reversible release boundary projection DeepONet to receive the boundary projection results before and after the re-feedback. The boundary migration sorting kernel filters and sorts the candidate actions according to the state changes of locked state, transition state and release state, and outputs the action priority. The boundary migration sorting kernel performs boundary migration comparisons on candidate actions based on the decrease in locked state, the transition in transition state, and the increase in released state. The boundary migration comparison results are obtained, and candidate actions that have not decreased in locked state and have not increased in released state are eliminated, while candidate actions that have decreased in locked state and increased in released state are retained. In this invention, the boundary migration comparison is performed between the boundary projection results before and after recharge; the boundary migration sorting kernel compares the changes in the locked state, transition state, and releaseable state before and after the same candidate treatment action to obtain the decrease in locked state, the transfer amount of transition state, and the increase in releaseable state. In this invention, the reduction in locked states is obtained by subtracting the number of locked states before rework from the number of locked states after rework, and is used to represent the scale of the state released from the locked state by the candidate disposal action; when the reduction in locked states is zero, it means that the candidate disposal action has not reduced the locked storage release process. In this invention, the transition state transfer quantity is obtained from the change in the transition state before and after refilling, and is used to represent the degree to which the candidate disposal action drives the storage release process into the disposal simulation stage; the transition state transfer quantity is used to distinguish the contribution of the candidate disposal action in the middle stage of release promotion; In this invention, the increase in releasable states is obtained by subtracting the number of releasable states before refilling from the number of releasable states after refilling. It is used to represent the increment by which the candidate disposal action causes the storage release process to enter the releasable state. The increase in releasable states serves as the primary basis for prioritizing disposal actions. In this invention, the boundary migration sorting kernel removes candidate actions where the locked state has not decreased and the release state has not increased; such candidate actions do not drive the release boundary from the locked state to the release state and are not included in the processing priority sorting; the boundary migration sorting kernel retains candidate actions where the locked state has decreased and the release state has increased, so that subsequent sorting is only performed on candidate actions that generate effective release boundary migration; The remaining candidate actions are sorted in the following order: first, the increase in the release state; second, the decrease in the lock state; and third, the transfer in the transition state. This yields the action priority. The boundary migration comparison results, the boundary status to which the minimum boundary change belongs, and the handling priority are combined and output to form dynamic monitoring results, anomaly sources, and handling priorities.

[0038] Example 1: To verify the feasibility of this invention in practice, it was applied to a data center storage resource monitoring environment. This environment includes several storage pools, virtual volumes, snapshot chains, backup copies, and hot / cold data migration tasks. A typical problem arose during operation: the storage management platform showed that some files had been deleted, and the business side had submitted cleanup tasks, but the available space in the storage pools did not increase synchronously. Manual investigation revealed that some space was locked by snapshot lineage, shared references, retention restrictions, and incomplete migration status. Conventional capacity monitoring methods can only display the capacity occupancy percentage, remaining capacity, and growth trend, making it difficult to determine whether the apparent release actions resulted in actual releaseable capacity, and also difficult to provide a ranking of disposal actions.

[0039] In this embodiment, DeepONet, a reversible release boundary projection, is first deployed in the data center storage space twin. The storage release process is divided into a locked state, a transitional state, and a releaseable state. Storage release processes with unresolved snapshot lineage, still referenced shared data blocks, incomplete migration balancing, or unresolved retention restrictions enter the locked state. Storage release processes where candidate disposal actions have been simulated but the boundary projection has not yet entered the releaseable state enter the transitional state. Once the release conditions are met, the storage release process enters the releaseable state. The twin is updated according to the three-state release channel to obtain the twin synchronization result, which is used for subsequent difference alignment.

[0040] During operation, the release boundary encoder replaces DeepONet's Branch. The release condition folding kernel compresses snapshot lineage, reference holding, migration balancing, and retention constraints into one of the following states: locked, transitional, or releaseable. The boundary position query operator replaces DeepONet's Trunk. The counterfactual boundary coordinate latch kernel maps the monitored object, event phase, and candidate disposal actions to the same release boundary position, verifying the query position through forward coordinate locking and backward readback. Release boundary coordinates with inconsistent verification are occluded and do not enter subsequent operator coupling.

[0041] During the boundary projection phase, the release boundary projection kernel receives Branch and Trunk outputs. The Branch output defines the release conditions, and the Trunk output defines the location of the release boundary to be deduced. The release boundary projection kernel performs state-by-state coupling according to the locked state, transition state, and releaseable state, and then forms the boundary projection result through three-state adsorption. When the release conditions are not met and the candidate action has not entered the deduction, the coupling result is adsorbed to the locked state; when the candidate action enters the deduction but the deduction is not completed, the coupling result is adsorbed to the transition state; when the release conditions are met, the coupling result is adsorbed to the releaseable state.

[0042] In the anomaly determination phase, the reverse boundary perturbation search kernel uses the boundary projection result as the search starting point and the twin synchronization result as the convergence target. The search process backtracks along the direction from the release state to the transition state and from the transition state to the locked state, with each backtracking forming a boundary change. After the boundary change is applied to the boundary projection result, if the state position difference between the boundary projection result and the twin synchronization result decreases, the boundary change is retained; if the state position difference does not decrease, the boundary change is discarded. After the search is completed, the boundary change that minimizes the difference and requires the fewest backtracking steps is selected as the minimum boundary change. The boundary state to which the minimum boundary change belongs is used to determine the source of the anomaly.

[0043] During the disposal simulation phase, the counterfactual boundary rewriting kernel receives candidate disposal actions, rewrites them into candidate boundary changes, and feeds them back into the release boundary encoder. The release condition folding kernel recompresses the release conditions and updates the Branch output. The updated Branch output and Trunk output re-enter the release boundary projection kernel, performing state-by-state coupling and three-state adsorption to obtain the fed-back boundary projection result. The boundary migration sorting kernel compares the reduction in locked states, the transition state shift, and the increase in releasable states before and after the feed-back, eliminates candidate disposal actions that do not lead to a reduction in locked states or an increase in releasable states, and assigns disposal priorities to the retained candidate disposal actions.

[0044] Table 1: Comparison of the implementation effects of the present invention and conventional capacity monitoring methods

[0045] The table above shows a comparison of conventional capacity monitoring methods and the method of this invention in the same type of storage resource monitoring environment. Conventional capacity monitoring methods are sensitive to capacity thresholds, growth rates, and alarm triggering, but lack the ability to determine release boundaries such as snapshot lineage, reference occupancy, and migration balancing, resulting in low release fulfillment rates and low priority hit rates. This invention, through a three-state release boundary, boundary projection results, and a counterfactual backfeeding process, links candidate disposal actions with actual releaseable capacity, thereby reducing invalid disposal actions.

[0046] In continuous verification, conventional capacity monitoring methods can provide alerts on "space tightness," but are insufficient in determining "whether space can actually be released." This invention maps book release actions to release boundaries, enabling the differentiation of snapshot remnants, unremoved references, incomplete migrations, and unremoved retention restrictions. In the verification log, the accuracy rate of identifying truly releaseable capacity increased from 82.6% to 95.8%, the anomaly identification rate of snapshot remnants increased from 78.4% to 94.6%, and the time for anomaly source location decreased from 18.5 minutes to 6.2 minutes. After prioritizing candidate disposal actions, maintenance personnel no longer need to try and fail item by item according to fixed rules, but prioritize actions that can increase the number of releaseable states, reducing the number of manual reviews from 31 to 12.

[0047] The table shows a significant reduction in the time required to locate anomalies. This is because the present invention does not involve checking each log field individually, but rather searches backwards from the difference between the boundary projection result and the twin synchronization result for the minimum boundary change. When the minimum boundary change falls into a locked state, the focus shifts to the failure to meet the release condition; when it falls into a transitional state, the focus shifts to the incomplete handling simulation. This process makes anomaly source location more focused, thereby reducing the number of manual reviews.

[0048] The improved priority handling hit rate is mainly due to the comparison of state changes before and after the backfilling during boundary migration sorting. Only candidate handling actions that result in a decrease in locked states and an increase in releaseable states are retained for sorting. The increase in releaseable states serves as the primary sorting criterion, ensuring that cleanup, migration, expansion, and quota-limiting actions are arranged according to their actual release contribution. The proportion of invalid candidate handling actions in the table decreased from 28.9% to 8.7%, indicating that counterfactual boundary backfilling has a screening effect on operational handling.

[0049] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for dynamic monitoring of data center storage space based on twin modeling, characterized in that, Includes the following steps: Deploy DeepONet, a reversible release boundary projection, in the data center storage space twin to divide the storage release process into locked state, transitional state, and releaseable state, forming a release boundary, and update the twin to obtain the twin synchronization result; Replace DeepONet's Branch with a release boundary encoder to perform boundary folding on release conditions formed by snapshot lineage, reference holding, migration balancing, and retention constraints, so that the Branch output carries the release conditions; Replace DeepONet's Trunk with a boundary location query operator, perform a joint query on the monitored object, event stage, and candidate action, and make the Trunk output point to the release boundary location to be deduced; Operator coupling is performed on the Branch output and Trunk output. The coupling result is then subjected to boundary projection according to the release condition. Capacities that do not meet the release condition are mapped to locked states, capacities in the disposal simulation are mapped to transition states, and capacities that meet the release condition are mapped to releaseable states, thus obtaining the boundary projection result. The boundary projection results are aligned with the twin synchronization results by difference. The minimum boundary change that makes the difference converge is searched in reverse along the locked state, transition state, and release state. The source of the anomaly is determined according to the boundary state to which the minimum boundary change belongs. The candidate disposal action is converted into a candidate boundary change, fed back to the release boundary encoder, the release condition carried in the Branch output is updated, and the operator coupling and boundary projection are re-executed. Based on the reduction in the locked state, the transfer in the transition state, and the increase in the release state before and after reinjection, the system outputs dynamic monitoring results, sources of anomalies, and handling priorities.

2. The method for dynamic monitoring of data center storage space based on twin modeling according to claim 1, characterized in that, The steps for updating the twin to obtain the twin synchronization result include: A three-state release channel is established at the output of the reversible release boundary projection DeepONet, and a boundary residual correction gate is embedded in the three-state release channel; Write the storage release process into the locked state, and use the boundary residual correction gate to compare the boundary residual between the current release state and the current state of the three-state release channel. When the release condition corresponding to the boundary residual is not met, the gated path from the locked state to the transition state is closed; When the deduction of the corresponding handling of boundary residuals is not completed, the gating path from the transition state to the release state is closed; When the release condition corresponding to the boundary residual is met and the disposal simulation is completed, the gated path from the locked state to the transition state and the gated path from the transition state to the releaseable state are opened. Update the locked state, transitional state, and releaseable state according to the gating path after opening, and write the three-state release channel back to the data center storage space twin to obtain the twin synchronization result.

3. The method for dynamic monitoring of data center storage space based on twin modeling according to claim 1, characterized in that, The step of performing boundary folding includes: Set a release condition folding core within the release boundary encoder and input the release condition into the release condition folding core; The release condition folding kernel compresses the release condition in a unidirectional state sequence of locked state, transition state, and releaseable state, forming the locked state output component, the transition state output component, and the releaseable state output component; When the release condition is compressed to the locked state, the transition state output component and the releaseable state output component are turned off, while the locked state output component is retained. When the release condition is compressed to the transition state, the locked state output component and the releaseable state output component are turned off, while the transition state output component is retained. When the release condition is compressed to the releaseable state, the locked state output component and the transition state output component are turned off, while the releaseable state output component is retained. The retained output components are combined into a Branch output, which carries the release conditions consistent with the release boundary.

4. The method for dynamic monitoring of data center storage space based on twin modeling according to claim 1, characterized in that, The steps for performing the join query include: On the Trunk side, a counterfactual boundary coordinate latch kernel is established, which maps the monitored object, event stage, and candidate disposal action to the release boundary to form the release boundary coordinates. Forward coordinate locking is performed along the direction from the locked state to the transition state and from the transition state to the releaseable state. Read back the release boundary coordinates after locking the forward coordinates in the direction from the release state to the transition state and from the transition state to the lock state to obtain the readback object, readback stage, and readback action; When the readback object is consistent with the monitored object, the readback phase is consistent with the event phase, and the readback action is consistent with the candidate disposal action, latch and release the boundary coordinates. If any of the readback object, readback phase, or readback action is inconsistent, the boundary coordinates of the occlusion release will be interrupted. Write the latched release boundary coordinates into the Trunk output, so that the Trunk output points to the release boundary position to be calculated.

5. The method for dynamic monitoring of data center storage space based on twin modeling according to claim 1, characterized in that, The steps for performing boundary projection include: Set a release boundary projection kernel at the coupling end of the reversible release boundary projection DeepONet, and use the Branch output as the release condition constraint and the Trunk output as the release boundary position to be deduced. The release boundary projection kernel performs state-by-state coupling on the Branch output and Trunk output according to the locked state, transition state and releaseable state, so that the coupling result is simultaneously constrained by the release condition and the release boundary position to be deduced. The release boundary projection kernel performs three-state adsorption on the coupling result according to the release conditions. When the release conditions are not met and the candidate disposal action is not entered, the coupling result is adsorbed to the locked state. When the candidate disposal action is entered and the disposal action is not completed, the coupling result is adsorbed to the transition state. When the release conditions are met, the coupling result is adsorbed to the releasable state. The coupling result after the three-state adsorption is used as the boundary projection result.

6. The method for dynamic monitoring of data center storage space based on twin modeling according to claim 1, characterized in that, The steps for performing difference alignment include: Inverse boundary perturbation search kernel is set at the causal end of DeepONet with reversible release boundary projection, and the boundary projection result is used as the search starting point and the twin synchronization result is used as the convergence target. The reverse boundary perturbation search kernel backtracks the boundary projection results in the order of release state to transition state and transition state to locked state, generating a boundary change amount each time it backtracks; After applying each boundary change to the boundary projection result, perform difference alignment, retain the boundary changes that reduce the difference, and discard the boundary changes that do not reduce the difference. Among the retained boundary changes, the boundary change that minimizes the difference between the boundary projection result and the twin synchronization result and minimizes the backtracking steps is selected as the minimum boundary change. The source of the anomaly is determined based on the boundary state to which the minimum boundary change belongs.

7. The method for dynamic monitoring of data center storage space based on twin modeling according to claim 5, characterized in that, The step of re-executing operator coupling and boundary projection includes: Set a counterfactual boundary rewrite kernel at the feedback end of the reversible release boundary projection DeepONet; Write the candidate disposal action into the counterfactual boundary rewriting kernel, and rewrite the gating path according to the one-way release boundary transfer path from the locked state to the transition state and from the transition state to the releaseable state, thus forming the candidate boundary change quantity; The candidate boundary change is injected into the release condition folding kernel set in the release boundary encoder, and the release condition is controlled to be recompressed to one of the locked state, transition state, and releaseable state, and the Branch output is updated. The updated Branch output and Trunk output are fed into the release boundary projection kernel. The state-by-state coupling and three-state adsorption are re-executed according to the release boundary position to be simulated, and the boundary projection result after re-feedback is obtained.

8. The method for dynamic monitoring of data center storage space based on twin modeling according to claim 7, characterized in that, The steps for outputting dynamic monitoring results, anomaly sources, and handling priorities include: Set a boundary migration sorting kernel at the output of the reversible release boundary projection DeepONet, and input the boundary projection results before and after reflow into the boundary migration sorting kernel. The boundary migration sorting kernel performs boundary migration comparisons on candidate actions based on the decrease in locked state, the transition in transition state, and the increase in released state. The boundary migration comparison results are obtained, and candidate actions that have not decreased in locked state and have not increased in released state are eliminated, while candidate actions that have decreased in locked state and increased in released state are retained. The remaining candidate actions are sorted in the following order: first, the increase in the release state; second, the decrease in the lock state; and third, the transfer in the transition state. This yields the action priority. The boundary migration comparison results, the boundary status to which the minimum boundary change belongs, and the handling priority are combined and output to form dynamic monitoring results, anomaly sources, and handling priorities.

Images