Backup policy intelligent tuning method and system based on ai inference

Abstract

The application belongs to the technical field of data backup, and particularly relates to a backup strategy intelligent optimization method and system based on AI inference, which comprises the following steps: collecting business attribute features and resource load features to obtain a target feature sequence, and generating an importance label according to a directory level depth; inputting feature discrete degree into a negative normalization algorithm to obtain data analysis weight; weighting the target feature sequence according to the data analysis weight to obtain comprehensive variable resistance, and inputting the comprehensive variable resistance and the target feature sequence into an intelligent classification network to obtain an activated confidence scalar; routing micro-change data to a high-priority buffer pool or a throughput buffer pool according to the activated confidence scalar and the importance label; extracting the current backlog volume and the residence waiting time length of the high-priority buffer pool, and triggering a sequential merge write instruction in combination with the activated confidence scalar. The application executes a dynamic scheduling process according to the calculated activated confidence scalar, thereby improving the safety and scheduling efficiency of data backup.

Description

Technical Field

[0001] This invention relates to the field of data backup technology. More specifically, this invention relates to a method and system for intelligent optimization of backup strategies based on AI inference. Background Technology

[0002] In the daily operation and continuous data protection of cross-data center disaster recovery systems, real-time security assessment of minor changes to underlying data is a fundamental step in ensuring the tamper-proofing of core data. This process requires continuously collecting node read / write change signals to assess the security urgency of underlying configuration modifications, and then outputting instructions to drive dynamic routing and disk merging to offset the impact of data loss caused by high-risk abnormal changes.

[0003] For example, Chinese patent document CN117149527B discloses a server data backup and recovery method. This method includes: collecting and analyzing server backup data through a collection module; obtaining the update frequency of the backup data through an update module, performing high-frequency incremental backups on frequently changing data, and performing full backups or low-frequency incremental backups on less frequently changing data; and using a removal module to perform sparsification processing on the local model using model pruning techniques, and identifying and removing redundant data based on the global model generated by the cloud server. Furthermore, this related technology achieves real-time data synchronization by setting up a first server and a second server, switching to the second server to ensure availability when the first server fails.

[0004] However, related technologies primarily rely on static backup scheduling based on update frequency and preset time periods. In extreme high-frequency concurrency scenarios such as massive data intensive reception or high-frequency instantaneous command interaction in combat command systems, changes to core configuration files often exhibit discrete characteristics with extremely small data volumes and high concealment. These minute changes are easily overwhelmed by massive business flows, causing the system to fail to accurately perceive the urgency of high-risk data remaining in memory. This lack of an identification mechanism leads to a delay in disaster recovery and disk write-down actions against covert tampering, leaving core data in an unprotected exposed state for an extended period before being written to disk, increasing the risk of data corruption or loss. Summary of the Invention

[0005] To address the technical problem that existing static backup scheduling struggles to identify subtle, hidden changes, leading to delays in high-risk data disaster recovery and increased risks of core data exposure, this invention provides solutions in the following aspects.

[0006] In a first aspect, the present invention provides an AI-based intelligent optimization method for backup strategies, comprising: S1 collecting the business attribute features of the underlying target file and the resource load features of the current device, generating importance labels accordingly, and constructing a target feature sequence after preprocessing the various features; S2 calculates the temporal dispersion of the target feature sequence within the feature sampling window to allocate data analysis weights; calculates the comprehensive change resistance based on the data analysis weights, and inputs the comprehensive change resistance and the target feature sequence into a preset intelligent classification network, outputting an activation confidence scalar; S3 if the activation confidence scalar reaches a preset upgrade threshold, then the minor change data of the corresponding target file is forcibly routed to the high-priority buffer pool; otherwise, the minor change data is routed to the high-priority buffer pool or the throughput buffer pool respectively according to the importance label; S4 obtains the current backlog volume of the high-priority buffer pool in a non-empty state, as well as the waiting time for minor change data; when the current backlog volume reaches a preset level, then a sequential merge write command is forcibly triggered; otherwise, the risk judgment status is determined based on the activation confidence scalar, and the capacity limit and time limit are allocated to the high-priority buffer pool accordingly, and a sequential merge write command is triggered when either limit is met.

[0007] This invention constructs a target feature sequence by extracting business attribute features and resource load features, measuring the overall operating load of the underlying architecture and the activity level of core configurations, thus optimizing the omission defects of static path monitoring. By calculating the feature dispersion, it allocates data analysis weights and calculates the overall change resistance, measuring the true fluctuation baseline within historical stable operating cycles and suppressing extraction interference caused by conventional network jitter and local business mutations. Based on the activation confidence scalar, it assesses the underlying security conditions and dynamically adjusts capacity and time limits, measuring the true carrying capacity of the high-priority buffer pool and the security risk level of changing data. When the backlog volume or waiting time reaches the limit, it triggers a sequential merge write instruction, measuring the memory overflow and crash boundary under extreme concurrency peaks, ensuring the rapid truncation and secure persistence of core changed data in emergency situations, and improving the anti-tampering response agility and dynamic disk scheduling accuracy of the storage system under complex operating conditions.

[0008] Preferably, the process of collecting the business attribute features of the underlying target files and the resource load features of the current device, generating importance tags based on these features, and constructing a target feature sequence after preprocessing the various features includes: intercepting target files whose content has changed at the underlying level within a set time window; extracting the file extension, directory level depth, and read / write call frequency of the target files as business attribute features; simultaneously extracting the directory level depth and comparing it with a preset level threshold; if the directory level depth is less than or equal to the preset level threshold, the target file is determined to be core configuration and a high importance tag is generated; if the directory level depth is greater than the preset level threshold, the target file is determined to be regular business and a low importance tag is generated; simultaneously collecting the queuing backlog depth and available network bandwidth of the current device as resource load features; and constructing a target feature sequence after preprocessing the various features.

[0009] Preferably, the step of constructing the target feature sequence after preprocessing the various features includes: performing linear normalization calculations on directory level depth, read / write call frequency, queue backlog depth, and available network bandwidth to obtain continuous feature vectors; performing one-hot encoding conversion on file extensions to obtain discrete feature vectors; and concatenating the continuous feature vectors and discrete feature vectors dimensionally to output the target feature sequence.

[0010] This invention constructs a target feature sequence by collecting directory hierarchy depth, read / write call frequency, and resource load characteristics, performing normalization and concatenation operations. The constructed target feature sequence measures the overall operating status and data activity of the underlying storage architecture, providing a standardized data foundation for subsequent multi-dimensional risk analysis, optimizing the blind spots of single static path monitoring, and preventing the omission of key configuration status characteristics under high-frequency interaction.

[0011] Preferably, the step of calculating the temporal dispersion of the target feature sequence within the feature sampling window to allocate data analysis weights includes: setting a feature sampling window containing a preset number of sampling points; extracting the target feature sequence corresponding to each historical sampling time within the feature sampling window to form a sequence set containing multiple target feature sequences; for each feature dimension in the sequence set, extracting the feature values ​​of that dimension from all target feature sequences to form a feature value set for that dimension; calculating the variance of the feature value set for each dimension; extracting the preset benchmark variance of the corresponding dimension within the historical stable operating period of the device, dividing the variance of each dimension by the corresponding preset benchmark variance to obtain the dispersion of each dimension; and allocating data analysis weights according to the dispersion of each dimension.

[0012] Preferably, the step of allocating data analysis weights according to the degree of dispersion of each dimension includes: extracting the degree of dispersion corresponding to each dimension in the target feature sequence to construct a dispersion input sequence, and inputting it into the negative Softmax normalization algorithm to perform weight reduction and normalization calculations to generate a feature weight sequence, thereby allocating data analysis weights to each feature dimension.

[0013] This invention obtains the degree of dispersion by calculating the variance of the feature value set within the feature sampling window and the preset benchmark variance, and combines it with a negative normalization algorithm to allocate data analysis weights. The allocation of data analysis weights measures the real fluctuation state of different feature dimensions within the historical stable operation period, suppresses the interference of local feature mutations caused by routine business jitter, and reduces the error risk of extracting data features with minor changes.

[0014] Preferably, the step of calculating the comprehensive resistance to change based on data analysis weights includes: using data analysis weights to perform a linear weighted summation calculation on the target feature sequence at the latest moment to obtain the comprehensive resistance to change.

[0015] This invention utilizes data analysis weights to perform a linear weighted summation operation on the target feature sequence at the latest time to obtain the comprehensive change resistance. The comprehensive change resistance measures the global change anti-interference baseline after weight reduction and smoothing, and absorbs the local redundancy features caused by network jitter in high-frequency concurrent scenarios, providing a robust reference dimension for subsequent intelligent network calculation of activation confidence scalar.

[0016] Preferably, obtaining the current backlog volume of the high-priority buffer pool in a non-empty state, and the dwell time of minor change data, includes: counting the total amount of pending data that has been received but not yet executed on disk in the high-priority buffer pool, defining it as the current backlog volume; when the high-priority buffer pool is not empty, extracting the current backlog volume, marking the minor change data with the longest dwell time in the pool as the first change data, and then calculating the dwell time of the data from the time it entered the buffer pool until the current time.

[0017] Preferably, when the current backlog volume reaches a preset level, a sequential merge write instruction is forcibly triggered, including: when the current backlog volume is greater than or equal to the preset level, it is determined that the underlying memory resources are facing the risk of overflow and crash, the subsequent hierarchical logic is stopped, a sequential merge write instruction for remote hardware is directly triggered, and the high-priority buffer pool is cleared after the data is successfully persisted, thereby performing the underlying forced intervention.

[0018] Preferably, the step of determining the risk assessment status based on the activation confidence scalar, and allocating capacity and time limits for the high-priority buffer pool accordingly, and triggering a sequential merge write instruction when either limit is met, includes: when the current backlog volume is less than a preset water level, assessing the safety status of the underlying hardware by comparing the activation confidence scalar with a preset high-risk threshold, and then dynamically initiating an emergency truncation branch or a regular merge branch; extracting the corresponding first capacity limit and first time limit, or second capacity limit and second time limit, according to the selected branch, and comparing the current backlog volume and waiting time with the extracted limit parameters in a dual-dimensional numerical comparison; as long as the current backlog volume or waiting time reaches the corresponding limit condition, a sequential merge write instruction for the remote hardware is issued and the high-priority buffer pool is cleared.

[0019] This invention assesses the underlying security status by statistically analyzing the current backlog volume and waiting time of the high-priority buffer pool, combined with the activation confidence scalar. Then, it dynamically extracts capacity and time limits to execute branch disk write scheduling. The execution of the branch disk write scheduling measures the actual load and security risk level of the underlying hardware. While intercepting the risk of memory overflow and crash during extreme concurrency peaks, it ensures that core changed data can be quickly truncated and triggered sequential merge writing in emergency situations.

[0020] Secondly, the present invention provides an AI-based intelligent optimization system for backup strategies, including a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned AI-based intelligent optimization method for backup strategies is implemented.

[0021] By adopting the above technical solution, the AI-based inference-based backup strategy intelligent tuning method is generated into a computer program and stored in the memory for loading and execution by the processor. This allows for the creation of terminal devices based on the memory and processor, making them convenient to use.

[0022] The beneficial effects of this invention are as follows: By extracting underlying state data and business attribute features to construct a target feature sequence, this invention measures the overall operating status of the underlying architecture and the activity level of core configurations, thus optimizing the omission defects of single static path monitoring; by combining data dispersion analysis with intelligent classification network to calculate activation confidence scalars, it measures the security risk status and tampering probability of minor data changes, smoothing out sudden interference caused by regular network jitter; by performing dynamic routing isolation based on importance labels and activation confidence scalars, it measures the high urgency characteristics of discrete writes, intercepts regular business data with sudden high-risk anomalies, and improves the matching degree of the environment in which minor data changes reside; furthermore, by allocating backlog capacity limits and retention time limits based on risk diversion results to perform branch disk scheduling, it measures the risk diversion level and memory backlog of the underlying hardware, suppresses the risk of disk fragmentation or memory overflow crashes caused by single fixed judgments, and improves the accuracy of disk response and data security of the system under concurrent peaks. Attached Figure Description

[0023] Figure 1 This is a flowchart of an AI-based intelligent optimization method for backup strategies. Figure 2 This is a comparison chart of the performance of existing technologies and the protection mechanism of this invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0026] This invention discloses an intelligent optimization method for backup strategies based on AI inference, referring to... Figure 1 This includes steps S1-S4: S1. Collect underlying state data and construct the target feature sequence.

[0027] It should be noted that in extreme high-frequency concurrent scenarios such as massive data intensive reception or high-frequency interaction of instantaneous commands in combat command systems, core production data is distributed in different logical directories. Traditional fixed monitoring rules based on a single static path or time period cannot accurately intercept the core configuration change signals that are actively modified. Therefore, this invention measures the comprehensive operating status of the underlying architecture through target feature sequences, providing a standardized data foundation for subsequent multi-dimensional data analysis and backup strategy optimization.

[0028] Specifically, within a set time window, target files whose underlying content has changed are intercepted; the file extension, directory level depth, and read / write call frequency of the target files are extracted as business attribute features; the directory level depth is extracted simultaneously and compared with a preset level threshold. If the directory level depth is less than or equal to the preset level threshold, the target file is determined to be core configuration and a high importance label is generated; if the directory level depth is greater than the preset level threshold, the target file is determined to be regular business and a low importance label is generated; the current device's queue backlog depth and network available bandwidth are collected simultaneously as resource load features; the directory level depth, read / write call frequency, queue backlog depth, and network available bandwidth are linearly normalized to obtain continuous feature vectors; the file extension is one-hot encoded to obtain discrete feature vectors; the continuous feature vectors and discrete feature vectors are concatenated dimensionally to output the target feature sequence.

[0029] It's important to clarify that directory depth refers to the position level of a target file within the file system. Starting from the root directory, the depth increases with each additional subdirectory in the file path. In storage security logic, this metric is often used to measure the importance of data. Core system configuration files are typically stored in shallower system paths, while regular business files are often nested in deeper directory structures. For example, the core network configuration file ` / etc / nginx / nginx.conf` in a Linux system has a depth of only 3. Because such files directly control the global network traffic distribution and the start / stop rules of core services, tampering with them can cause widespread business paralysis or traffic hijacking, making them high-risk core targets vulnerable to malicious attacks. Conversely, the business log file ` / data / app / user / logs / 2026 / 05 / click.log` for ordinary users has a depth of 7. Since such files only serve as append records of historical interaction flows, their single-point damage or tampering will neither disrupt core business operations nor cause the underlying architecture to go out of control. Therefore, the system will classify it as regular business data with a lower priority for tamper prevention.

[0030] Read / write call frequency refers to the total number of read and write commands received by the underlying storage interface for specific data within a preset time window. It is a dynamic indicator of data activity; high-frequency read / write indicates that data is in a state of intensive interaction. In anti-tampering scenarios, abnormal spikes in this indicator are often important precursors to malicious encryption or illegal scanning. For example, the system permission file of a database typically experiences only a few read / write operations per day during normal operation. If the system detects a sudden surge of thousands of rewrite commands for this file within one second, this precipitous increase in call frequency can be detected and identified as a high-risk mutation characteristic of ransomware performing high-speed malicious encryption.

[0031] Queue backlog depth refers to the number of pending requests in the current storage controller or scheduling queue that are in a state of being issued but not completed. It reflects the real-time load of the underlying hardware. The deeper the backlog depth, the greater the system processing pressure, and the higher the risk of data remaining in memory. It is an important limiting parameter for determining whether to skip the regular merging logic and force disk write. For example, when encountering a sudden surge in business or a concurrent attack, the storage controller's I / O queue may instantly accumulate tens of thousands of discrete requests waiting to be written to disk. At this time, an extremely deep queue backlog depth means that the system memory is at a very high risk of overflow (OOM) and crash. Based on this parameter, the system will decisively abandon the regular strategy of waiting for the data to be fully merged and forcibly trigger an emergency disk write command to salvage the memory data.

[0032] In this embodiment, the preset level threshold is set to 3. Implementers can set the preset level threshold according to the file directory deployment specifications of the underlying operating system. For example, when the core configuration files are concentrated in a very shallow system root directory, the preset level threshold can be appropriately reduced to strictly tighten the distribution scope of core tags; when the file structure of business applications is deeply nested, the preset level threshold can be appropriately increased to prevent the omission of critical configurations.

[0033] S2. Activation confidence scalar is obtained based on data analysis and intelligent networks.

[0034] It should be noted that in extreme scenarios such as massive data reception or high-frequency interaction of instantaneous commands in combat command systems, minor data changes occur frequently. The original feature dimensions contain data redundancy and lack contextual correlation, making it impossible to assess the security risks of these minor data changes, thus leading to misjudgments in backup buffer cycles. Therefore, this invention uses data dispersion analysis combined with intelligent classification networks for AI reasoning to assess the aforementioned security risk status, providing a benchmark for determining the timing of subsequent backup buffer pool disk writes.

[0035] Specifically, a feature sampling window with a preset number of sampling points is set; the target feature sequence corresponding to each historical sampling time within the feature sampling window is extracted to form a sequence set containing multiple target feature sequences; for each feature dimension in the sequence set, the feature values ​​of that dimension in all target feature sequences are extracted to form a feature value set for the corresponding dimension; the variance of the feature value set for each dimension is calculated; the preset benchmark variance of the corresponding dimension is extracted during the historical stable operation period of the device, and the variance of each dimension is divided by the corresponding preset benchmark variance to obtain the degree of dispersion of each dimension.

[0036] Extract the degree of dispersion of each dimension in the target feature sequence to construct the discreteness input sequence, and input it into the negative Softmax normalization algorithm to perform weight reduction and normalization calculation to generate the feature weight sequence, thereby assigning data analysis weights to each feature dimension.

[0037] It should be added that, based on the calculation logic of exponential mapping and normalization, the larger the discreteness value of a certain dimension, the more severe the abnormal mutation in that dimension, and the smaller the corresponding decay mapping value, resulting in a smaller final data analysis weight. Thus, the smoothing and suppression of abnormal mutation feature interference are achieved through a determined algorithm model.

[0038] It should be added that the intelligent classification network adopts a multilayer perceptron network, which includes an input layer, a feature mapping layer, and an output layer. The output layer is configured with an activation function that maps the comprehensive features to continuous values ​​between 0 and 1.

[0039] The intelligent classification network is trained through the following steps: historical feature sequences and historical resistance to change are extracted from historical logs, concatenated to construct historical training vectors, and the index memory ratio and total recovery time for the corresponding time period are extracted simultaneously. If the index memory ratio reaches a preset threshold or the total recovery time reaches a preset limit, the device is determined to be abnormal, and the corresponding historical training vector is marked as a high-risk training sample and assigned a label value of 1; conversely, if neither of these thresholds is reached, the device is determined to be operating stably, and it is marked as a normal training sample and assigned a label value of 0. Based on the comprehensive training set constructed above, the samples are input into the initial intelligent classification network in batches for forward computation to obtain predicted output values. The predicted output values ​​and actual label values ​​are substituted into a preset loss function to quantify the network error, and the weights and biases of the network nodes are iteratively updated based on this network error using a backpropagation mechanism; until the output value of the loss function converges to a stable range, the intelligent classification network finally establishes a precise mapping relationship from feature sequences with resistance to change to activation confidence scalars.

[0040] The comprehensive resistance to change is obtained by performing a linear weighted summation on the target feature sequence at the latest moment using data analysis weights; the comprehensive resistance to change is then concatenated with the target feature sequence and input into a pre-set intelligent classification network; the intelligent classification network is used to calculate and output the activation confidence scalar.

[0041] The larger the activation confidence scalar, the higher the matching degree between the current underlying feature sequence and historical high-risk samples, and the greater the possibility that the data change belongs to core configuration tampering. This guides the system to shorten the residence isolation period, thereby achieving the technical benefit of ensuring the real-time disk storage of core data. The smaller the activation confidence scalar, the closer the current feature sequence is to the regular business change characteristics in historical logs, and the lower the urgency of the data change. This guides the system to extend the residence period to carry out sufficient version merging, thereby achieving the technical benefit of reducing the pressure on metadata indexing of remote nodes.

[0042] For example, the feature sampling window is set to include sampling points from the most recent 10 time points. When the read / write call frequency dimension of the target file undergoes a sudden change at the current time, the corresponding temporal variance and temporal dispersion increase. The system assigns a data analysis weight of 0.05 to this dimension based on negative correlation logic, and obtains a stable comprehensive change resistance through linear weighted summation. This effectively absorbs and smooths the local sudden change interference that may be caused by conventional network jitter. Subsequently, this stable comprehensive change resistance is used as a reference baseline and concatenated with the target feature sequence that retains the original change value. Both are then input into the intelligent classification network. This allows the network to reference both the global resistance base against interference and directly capture the original feature of the sudden surge in read / write frequency. Finally, the system determines that the sequence feature matches the historical high-risk behavior feature and outputs an activation confidence scalar of 0.92.

[0043] In this embodiment, the preset water level is set to 80%. Implementers can dynamically adjust this water level based on the hardware configuration and disaster recovery requirements of the backup nodes: for example, when node memory resources are scarce, the preset water level can be appropriately lowered to improve the intelligent classification network's sensitivity to labeling high-risk samples, prompting earlier triggering of backup optimization and disk dumping defenses; conversely, when node memory is ample and larger fluctuations are permissible, the preset water level can be appropriately increased to prevent data mislabeling and frequent backup oscillations caused by fluctuations in regular business operations.

[0044] S3. Perform dynamic route isolation based on importance labels and confidence scalars.

[0045] It should be noted that in extreme scenarios such as massive data reception or high-frequency instantaneous command interactions in combat command systems, if cross-node backup data streams are isolated solely based on static tags, it is difficult to cope with extreme conditions of sudden high-risk changes in routine business data. This can lead to potentially risky minor changes being incorrectly intercepted in low-priority throughput buffers, resulting in the negative consequence of data protection lag. Therefore, this invention corrects importance tags by activating confidence scalars, intercepting discrete writes with high urgency characteristics, and providing a more accurate residence environment for minor data changes.

[0046] Specifically, if the activation confidence scalar is greater than or equal to the preset upgrade threshold, the current minor data change is determined to be of high urgency. Ignoring the original indication of the importance label, the minor data change is forcibly routed to the high-priority buffer pool. If the activation confidence scalar is less than the preset upgrade threshold, the static routing logic is restored. When the importance label indicates core configuration, the data is routed to the high-priority buffer pool for merged storage. When the importance label indicates regular business, the data is routed to the throughput buffer pool for independent temporary storage.

[0047] It should be further explained that the high-priority buffer pool refers to a high-priority memory area specifically used to carry core configuration data or data facing high-risk abnormal mutations. Its disk write scheduling strategy focuses on ultra-fast response and low-latency truncation, aiming to ensure the safety of critical data with minimal data retention time and prevent core data loss due to memory overflow or sudden crashes. The throughput backup buffer pool, on the other hand, refers to a low-priority memory area used to carry daily routine business change data. Its disk write scheduling strategy focuses on sufficient merging of large data blocks, aiming to reduce the physical read and write wear of the underlying storage disk and improve the overall I / O throughput performance by aggregating massive small discrete writes.

[0048] For example, the preset upgrade threshold determines the system's sensitivity to forced bypass routing in response to sudden abnormal changes in routine business data. An empirically chosen value range is [0.75, 0.95]. In this embodiment, the preset upgrade threshold is set to 0.85. Implementers can set the preset upgrade threshold based on the production system's security tolerance for core data tampering. For instance, when the system is in a high-risk ransomware attack environment and data security requirements are extremely high, the preset upgrade threshold can be appropriately lowered to accelerate the isolation response to abnormal writes. When the system is in a closed intranet and prioritizes the throughput performance of routine business operations, the preset upgrade threshold can be appropriately increased to avoid misjudgments that excessively consume high-priority memory resources.

[0049] S4. Execute branch disk scheduling based on multi-dimensional parameters.

[0050] It should be noted that, taking the sudden issuance of massive tactical commands by the combat command system during actual combat exercises, or the exponential surge of discrete telemetry data into the astronomical monitoring / satellite receiving system in a short period of time as an example, the small changes in the data residing in the device memory face business concurrency and large-scale changes. A single fixed time judgment or fixed capacity judgment cannot adaptively match the high-frequency and drastically fluctuating discrete write traffic, resulting in the high-priority buffer pool experiencing fragmentation due to premature disk write or memory overflow and crash due to late disk write. Therefore, this invention activates the confidence scalar to perform underlying risk diversion, allocates backlog capacity limits and retention time limits at different levels according to the risk diversion results, and uses overflow boundaries to perform forced intervention. It uses procedural branch judgment logic to replace numerical limits, providing multi-dimensional trigger judgment for the execution order and merge write commands.

[0051] Specifically, the total amount of pending data that has been received but not yet executed on disk in the high-priority buffer pool is defined as the current backlog volume. If the high-priority buffer pool is not empty, the current backlog volume is extracted, and the smallest change data that has stayed in the pool for the longest time is marked as the first change data. Then, the waiting time of the data from the time it entered the buffer pool until the current time is calculated.

[0052] If the current backlog volume is greater than or equal to the preset water level, it is determined that the underlying memory resources are at risk of overflow and crash. The subsequent hierarchical logic is stopped, and the sequential merge write instruction for remote hardware is directly triggered. The high-priority buffer pool is cleared after the data is successfully persisted and saved, thereby performing the underlying forced intervention. If the current backlog volume is less than the preset water level, read the pre-acquired activation confidence scalar; compare the activation confidence scalar with the preset high-risk threshold to execute the graded triggering process.

[0053] If the activated confidence scalar is greater than or equal to the preset high-risk threshold, the underlying hardware is determined to be in a high-risk abnormal operating condition. An emergency truncation branch is initiated and the first capacity limit and the first time limit are extracted. The current backlog volume is compared with the first capacity limit, and the waiting time is compared with the first time limit. When the current backlog volume is greater than or equal to the first capacity limit, or the waiting time is greater than or equal to the first time limit, a sequential merge write instruction for remote hardware is triggered and the high-priority buffer pool is cleared. If the activation confidence scalar is less than the preset high-risk threshold, the underlying hardware is determined to be in a stable operating condition. A regular merge branch is initiated, and the second capacity limit and the second time limit are extracted. The current backlog volume is compared with the second capacity limit, and the waiting time is compared with the second time limit. When the current backlog volume is greater than or equal to the second capacity limit, or the waiting time is greater than or equal to the second time limit, a sequential merge write instruction for remote hardware is triggered, and the high-priority buffer pool is cleared.

[0054] Among them, the value of the first capacity limit is strictly less than the value of the second capacity limit, and the value of the first time limit is less than the value of the second time limit; the throughput buffer pool is monitored synchronously, and when the normal backlog volume of the throughput buffer pool reaches the preset throughput limit, an asynchronous disk flushing instruction is triggered to prevent memory overflow and crash.

[0055] Among them, the first capacity limit and the first time limit are used for agile truncation under abnormal operating conditions. The smaller the value of the first capacity limit or the first time limit, the lower the tolerance of the equipment to the data retention under risk conditions. This makes the high-quality buffer pool more likely to trigger the sequential merge write instruction under low load conditions, thereby more likely to achieve the technical benefits of cutting off the malicious tampering link and accelerating the landing of core data. Conversely, the larger the value of the first capacity limit or the first time limit, the more buffer space the device retains under abnormal operating conditions, which raises the threshold for triggering sequential merge write instructions and thus achieves the technical benefit of preventing occasional misjudgments from causing high-frequency read and write jitter in the system. The second capacity limit and the second time limit are used for full merging under normal operating conditions. The larger the value of the second capacity limit or the second time limit, the longer the device allows data to reside in the memory isolation layer under stable operating conditions, which delays the triggering of sequential merge write instructions and thus achieves the technical benefit of fully integrating small changes in data into continuous large blocks for disk and reducing disk read and write wear. Conversely, the smaller the value of the second capacity limit or the second time limit, the faster the memory resources of the high-priority buffer pool are cleared and released under stable operating conditions, thus more likely to achieve the technical benefit of preventing memory overflow crashes caused by concurrent floods.

[0056] For example, the second capacity limit determines the upper limit of absorption of the high-efficiency buffer pool under stable operating conditions, and the empirical value range is [200, 800] megabytes. In this embodiment, the second capacity limit is set to 512 megabytes.

[0057] For example, the first capacity limit determines the responsiveness of the device when it is attacked, with an empirical value range of [10, 50] megabytes. In this embodiment, the first capacity limit is set to 30 megabytes.

[0058] In this embodiment, the preset water level is set to 95%; the preset high-risk threshold determines the risk judgment boundary for the underlying hardware to start the emergency truncation branch, and the empirical value range is [0.75, 0.95]. In this embodiment, the preset high-risk threshold is set to 0.85.

[0059] In this embodiment, the second time limit is set to 60 seconds; the first time limit determines the response period boundary of forced truncation under abnormal operating conditions, and the empirical value range is [1,5] seconds. In this embodiment, the first time limit is set to 2 seconds.

[0060] For example, the preset throughput limit determines the capacity threshold for asynchronous disk write-to-database operations performed by the throughput buffer pool. The empirical range is [1024, 4096] megabytes. In this embodiment, the preset throughput limit is set to 2048 megabytes. Implementers can adjust the configuration of the time limit and capacity limit according to the security tolerance and hardware network status. For example, when the underlying environment has strict anti-tampering requirements and low tolerance for data loss, the first time limit and the first capacity limit can be appropriately reduced to achieve rapid disk write-to-database interception of minor data changes. When a node encounters frequent network jitter, the second time limit can be appropriately increased to prevent excessive concurrent disk write-to-database operations caused by network fluctuations.

[0061] For example, Figure 2 The diagram compares the performance of existing technologies with the protection mechanism of this invention. Existing technologies, limited by a fixed capacity threshold, passively wait for data to accumulate as the runtime step progresses. Furthermore, they lack preventative mechanisms when entering high-risk anomaly zones, causing the current backlog volume to continuously increase and exceed the disaster recovery high-risk exposure warning line. In contrast, this invention, during normal operation, uses coordinated control of conventional capacity and time limits to ensure regular disk write operations, reducing the read / write burden on the underlying system. When entering high-risk anomaly zones, it determines the high-risk state based on the activation confidence scalar output by the intelligent classification network, and then dynamically allocates more stringent capacity and time limits, thereby triggering early truncation at a lower current backlog volume percentage, reducing the risk of core data loss due to data retention.

[0062] This invention also discloses an AI-based intelligent optimization system for backup strategies, comprising a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement an AI-based intelligent optimization method for backup strategies according to the present invention.

[0063] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

Claims

1. A backup strategy intelligent optimization method based on AI inference, characterized in that, include: S1: Collect the business attribute characteristics of the underlying target files and the resource load characteristics of the current device, generate importance labels accordingly, and construct a target feature sequence after preprocessing the various features. This includes: intercepting target files whose content has changed within a set time window; extracting the file extension, directory level depth, and read / write call frequency of the target files as business attribute characteristics; synchronously extracting the directory level depth and comparing it with a preset level threshold. If the directory level depth is less than or equal to the preset level threshold, the target file is determined to be a core configuration and a high importance label is generated; if the directory level depth is greater than the preset level threshold, the target file is determined to be a regular business and a low importance label is generated; synchronously collecting the queue backlog depth and available network bandwidth of the current device as resource load characteristics; and constructing a target feature sequence after preprocessing the various features. The queue backlog depth refers to the number of pending requests in the current storage controller or scheduling queue that are in the issued but not completed state. S2: Calculate the temporal dispersion of the target feature sequence within the feature sampling window to allocate data analysis weights; calculate the comprehensive change resistance based on the data analysis weights, including: performing a linear weighted summation calculation on the target feature sequence at the latest moment using the data analysis weights. S3: Obtain the comprehensive change resistance and input the comprehensive change resistance and target feature sequence into a preset intelligent classification network, outputting an activation confidence scalar; S4: If the activation confidence scalar reaches the preset upgrade threshold, the minor change data of the corresponding target file is forcibly routed to the high-priority buffer pool; otherwise, the minor change data is routed to the high-priority buffer pool or the throughput buffer pool according to the importance label; S5: Obtain the current backlog volume of the high-priority buffer pool in a non-empty state, as well as the waiting time of the minor change data. Obtaining the current backlog volume includes: the total amount of pending data received but not yet executed by the disk write command in the high-priority buffer pool. The quantity is defined as the current backlog volume. When the current backlog volume reaches a preset level, a sequential merge write instruction is forcibly triggered, including: if the current backlog volume is greater than or equal to the preset level, it is determined that the underlying memory resources are facing an overflow and crash risk, the subsequent hierarchical logic is stopped, a sequential merge write instruction for remote hardware is directly triggered, and the high-priority buffer pool is cleared after the data is successfully persisted, thereby performing the underlying forced intervention; otherwise, the risk judgment status is determined based on the activation confidence scalar, and the capacity limit and time limit are allocated to the high-priority buffer pool accordingly, and a sequential merge write instruction is triggered when either limit is met.

2. The intelligent optimization method for backup strategies based on AI inference according to claim 1, characterized in that, The process of constructing the target feature sequence after preprocessing various features includes: performing linear normalization calculations on directory level depth, read / write call frequency, queue backlog depth, and available network bandwidth to obtain continuous feature vectors; performing one-hot encoding conversion on file extensions to obtain discrete feature vectors; and concatenating the continuous feature vectors and discrete feature vectors dimensionally to output the target feature sequence.

3. The intelligent optimization method for backup strategies based on AI inference according to claim 1, characterized in that, The calculation of the temporal dispersion of the target feature sequence within the feature sampling window to allocate data analysis weights includes: setting a feature sampling window containing a preset number of sampling points; extracting the target feature sequence corresponding to each historical sampling time within the feature sampling window to form a sequence set containing multiple target feature sequences; for each feature dimension in the sequence set, extracting the feature values ​​of that dimension from all target feature sequences to form a feature value set for the corresponding dimension; calculating the variance of the feature value set for each dimension; extracting the preset benchmark variance of the corresponding dimension within the historical stable operating period of the device, dividing the variance of each dimension by the corresponding preset benchmark variance to obtain the dispersion of each dimension; and allocating data analysis weights according to the dispersion of each dimension.

4. The intelligent optimization method for backup strategies based on AI inference according to claim 3, characterized in that, The step of allocating data analysis weights based on the degree of dispersion of each dimension includes: extracting the degree of dispersion corresponding to each dimension in the target feature sequence to construct a dispersion input sequence, and inputting it into the negative Softmax normalization algorithm to perform weight reduction and normalization calculations to generate a feature weight sequence, thereby allocating data analysis weights to each feature dimension.

5. The intelligent optimization method for backup strategies based on AI inference according to claim 1, characterized in that, The process of obtaining the retention time of minor change data includes: extracting the current backlog volume when the high-priority buffer pool is not empty, marking the minor change data with the longest residence time in the pool as the first change data, and then calculating the retention time of the data from the time it entered the buffer pool until the current time.

6. The intelligent optimization method for backup strategies based on AI inference according to claim 1, characterized in that, The process of determining the risk assessment status based on the activation confidence scalar, and allocating capacity and time limits for the high-priority buffer pool accordingly, and triggering a sequential merge write instruction when either limit is met, includes: when the current backlog volume is less than a preset water level, assessing the safety status of the underlying hardware by comparing the activation confidence scalar with a preset high-risk threshold, and then dynamically initiating an emergency truncation branch or a regular merge branch; extracting the corresponding first capacity limit and first time limit, or second capacity limit and second time limit, according to the selected branch, and comparing the current backlog volume and waiting time with the extracted limit parameters in a dual-dimensional numerical comparison; when the current backlog volume or waiting time reaches the corresponding limit condition, issuing a sequential merge write instruction and clearing the high-priority buffer pool; the limit parameters are the first capacity limit and first time limit, or the second capacity limit and second time limit.

7. An AI-based intelligent optimization system for backup strategies, characterized in that: include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement the AI-based inference-based intelligent optimization method for backup strategies according to any one of claims 1-6.

Images