An artificial intelligence-based computing resource scheduling method, device and medium

Abstract

This invention discloses an artificial intelligence-based method, device, and medium for scheduling computing resources, relating to the field of computing resource scheduling technology. The method includes: inputting a vector alignment table into a resource evaluation deep neural network to calculate resource suitability scores and conflict risk scores; performing candidate resource clustering and lower limit constraint verification through a scheduling optimization strategy to generate a candidate resource list; setting the resource suitability scores and conflict risk scores in the candidate resource list as sorting keys, performing fine-grained ranking within clusters, selecting head resources for distribution, and collecting acceptance receipts, initiation receipts, and operation telemetry receipts to generate a feedback packet; inputting the feedback packet into a feedback learning deep neural network to perform behavior replay comparison, locating key conflict points, dynamically updating the scheduling optimization strategy, and registering weight migration trajectories to generate a computing resource scheduling record set. This invention achieves continuous optimization of the scheduling strategy, improving the decision-making accuracy and operational stability when facing dynamic loads.

Description

Technical Field

[0001] This invention relates to the field of computing resource scheduling technology, and in particular to a computing resource scheduling method, device and medium based on artificial intelligence. Background Technology

[0002] With the continuous evolution of artificial intelligence technology, the demand for computing resource scheduling has expanded from single scenarios to complex application systems integrating multiple modalities. In emerging fields such as generative model training, edge intelligence deployment, and cross-domain collaborative computing, computing resources exhibit highly heterogeneous characteristics, encompassing various forms such as general-purpose processors, dedicated acceleration chips, and hybrid architectures. Currently, the industry is improving computing efficiency by constructing dynamic resource allocation frameworks. For example, strategy optimization mechanisms based on multi-objective optimization models achieve a balance between energy consumption and performance, driving the intelligent development of data center resource management and providing fundamental support for large-scale parallel computing tasks.

[0003] However, existing technologies have two key limitations in resource scheduling: First, traditional scheduling strategies are difficult to effectively handle the fragmentation of heterogeneous resources. Static allocation methods are prone to resource gaps and lack the ability to aggregate and integrate resources, resulting in low utilization of fragmented resources. Second, existing scheduling systems are difficult to achieve dynamic evolution of strategies. Most systems use preset rule bases or single-trained neural network models, which cannot continuously optimize decision logic through real-time feedback data and lack a strategy optimization mechanism based on operational feedback. This leads to a decrease in scheduling efficiency when sudden load fluctuations or new task modes occur. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an artificial intelligence-based computing resource scheduling method to solve the problem of heterogeneous resource clustering and dynamic optimization of scheduling strategies.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a computing resource scheduling method based on artificial intelligence, comprising, Collect computing resource data and task requirement data, perform field integrity checks, register missing field markers, and generate scheduling input packages; Based on the scheduling input packet, the task behavior range is located for the task requirement field, the resource trend is segmented for the computing resource field, and behavior matching and dimension completion are performed to generate a vector alignment table. The vector alignment table is input into the resource evaluation deep neural network to calculate the resource fit score and conflict risk score. Then, the candidate resource clustering and candidate resource lower limit constraint verification are performed through the scheduling optimization strategy to generate a candidate resource list. Set the resource adaptation score and conflict risk score in the candidate resource list as the sorting key, perform fine sorting within the cluster, select the head resource for distribution and collect acceptance receipts, start receipts and running telemetry receipts, and generate a feedback package; The feedback packet is input into the feedback learning deep neural network, the behavior is replayed and compared, key conflict points are located, the scheduling optimization strategy is dynamically updated and the weight migration trajectory is registered, and a set of computing resource scheduling records is generated.

[0007] As a preferred embodiment of the AI-based computing resource scheduling method of the present invention, the specific steps of collecting computing resource data and task requirement data, performing field integrity checks, registering missing field markers, and generating a scheduling input package are as follows: The collected computing resource data and task requirement data are written into time anchors, and segmented according to the collection sequence to generate a time anchor collection record set. Based on the time anchor point collection record set, the system performs location locking, field integrity verification, and missing field marking trajectory registration, and creates field cross-mapping entries, which are then integrated into a field locking verification set. Extract the task requirement field and computing resource field from the field lock check set, perform field bundle reorganization, write missing field markers and field cross-mapping entries, and generate a scheduling input package.

[0008] As a preferred embodiment of the AI-based computing resource scheduling method of the present invention, the following steps are taken: based on the scheduling input packet, task behavior interval positioning is performed on the task requirement field, resource trend segmentation is performed on the computing resource field, and behavior matching and dimension completion are performed to generate a vector alignment table. For the task requirement field and computing resource field in the scheduling input package, the alignment benchmark is solidified based on the time anchor point and sequence number, and an alignment sequence number is added to generate a field directory; Based on the field directory, track the switching position of task requirement fields, write the start time anchor and end time anchor of the task behavior interval, and generate a set of behavior intervals; Based on the behavior interval set, the scanning boundary is locked, a segmented scanning window is constructed, monotonic direction determination is performed and inflection point time anchors are registered to generate a trend segment set; The behavior interval set and trend segment set are paired according to the overlap relationship of time anchor points, and a dimension list is established. Missing dimensions are filled in and units are consistent to generate a vector alignment table.

[0009] As a preferred embodiment of the AI-based computing resource scheduling method of the present invention, the specific steps for inputting the vector alignment table into the resource evaluation deep neural network and calculating the resource suitability score and conflict risk score are as follows: The vector alignment table is grouped, its position is checked and cleaned, and then it is divided according to a fixed neural input batch size to generate neural input batches. The neural input batches are fed into the resource evaluation deep neural network to calculate the resource adaptation score and conflict risk score, and bind them with the alignment number. At the same time, range checking and labeling are performed to generate a score record set.

[0010] As a preferred embodiment of the AI-based computing resource scheduling method of the present invention, the step of generating a candidate resource list by performing candidate resource clustering and lower limit constraint verification through scheduling optimization strategies is as follows: Based on the score record set, resource adaptation scores and conflict risk scores are converted into cluster keys through scheduling optimization strategies and merged to form candidate resource clusters. Based on the candidate resource cluster, the lower limit constraint of the candidate resources is checked, the check mark is registered, the scoring record items are dynamically summarized, and a candidate resource list is generated.

[0011] As a preferred embodiment of the AI-based computing resource scheduling method of the present invention, the following steps are taken: setting the resource matching score and conflict risk score in the candidate resource list as the sorting key, performing intra-cluster fine sorting, selecting the head resource for distribution and collecting acceptance receipts, start receipts, and running telemetry receipts, and generating a feedback packet. Based on the candidate resource list, the resource adaptation score and conflict risk score are set as sorting keys, and gradient conversion and jitter detection are performed to construct a sorting vector for fine sorting within the cluster and generate a sorting table for distribution. According to the sorting table, the head resources are selected in sequence and distributed to the nodes in the cluster. At the same time, the acceptance receipt, the start receipt and the running telemetry receipt are collected to generate a receipt chain record set. The receipt chain record set is chained together to build a receipt chain index, and status verification and association are performed to generate a feedback package.

[0012] As a preferred embodiment of the artificial intelligence-based computing resource scheduling method of the present invention, the specific steps of inputting the feedback packet into the feedback learning deep neural network and performing behavior playback comparison are as follows: Based on the feedback package, expand the restoration receipt chain index, and split and interleave the acceptance receipt, start receipt and running telemetry receipt into three tracks, while binding them with the alignment sequence number to generate a multi-track playback arrangement set. The multitrack playback orchestration set is input into a feedback learning deep neural network, and then aligned and cross-linked with the running telemetry receipts to generate a playback deviation scale.

[0013] As a preferred embodiment of the artificial intelligence-based computing resource scheduling method of the present invention, the specific steps for generating the computing resource scheduling record set are as follows: Based on the playback deviation scale, a candidate pool is constructed through classification and screening, and key conflict points are located in a series and cross-location manner to generate a list of key conflict points. The list of key conflict points is mapped to the scheduling optimization strategy and dynamically updated. At the same time, the weight migration trajectory is registered to generate a set of computing resource scheduling records.

[0014] In a second aspect, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the computer program, when executed by the processor, implements any step of the artificial intelligence-based computing resource scheduling method described in the first aspect of the present invention.

[0015] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the artificial intelligence-based computing resource scheduling method described in the first aspect of the present invention.

[0016] The beneficial effects of this invention are as follows: by clustering candidate resources, the structured integration of heterogeneous computing resources is realized, overcoming the resource gap problem and improving the utilization rate of fragmented resources; by using feedback learning deep neural networks for behavior playback comparison, the scheduling strategy is continuously optimized, improving the decision-making accuracy and operational stability when facing dynamic loads. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a computing resource scheduling method based on artificial intelligence.

[0019] Figure 2 A flowchart for generating a vector alignment table.

[0020] Figure 3 A flowchart for resource evaluation and candidate resource generation.

[0021] Figure 4 A flowchart for feedback learning and scheduling strategy optimization. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides a computing resource scheduling method based on artificial intelligence, including the following steps: S1: Collect computing resource data and task requirement data, perform field integrity checks, register missing field markers, and generate scheduling input packages; S1.1: Write the collected computing resource data and task requirement data into time anchors, and perform segmentation according to the collection sequence number to generate a time anchor collection record set; Specifically, computing resource data and task requirement data are collected, and time anchor values ​​are attached to them respectively. These values ​​are then fixed to the record header field. The time anchor values ​​are truncated or padded with a uniform time precision to maintain consistency (e.g., millisecond-level time precision). Based on the arrival order of the collected data, a sequence number is generated for each of the collected computing resource data and task requirement data and written into the number field. Concurrent arrival records are processed to resolve concurrency issues, and a unique sequence number is determined according to a preset sorting rule to avoid number conflicts. Segment boundaries are constructed based on the time anchor values ​​and the collection sequence number, and segment markers are written into them. The segment boundary determination uses a segment threshold (the segment threshold is defined as the judgment boundary for triggering the registration of a new segment based on the time anchor value interval or the collection sequence number interval of adjacent records; the example range is 1 millisecond to 10 seconds) to judge the interval between adjacent records and register a new segment (e.g., a time interval threshold). Records with time anchor values, collection sequence numbers, and segment markers are summarized and integrated into a time anchor collection record set.

[0026] It should be noted that the time anchor value is derived from the collection access time or timestamp field of the collected computing resource data and task requirement data. When the timestamp field exists, the timestamp field is normalized according to a unified time zone and a unified time precision and then used as the time anchor value. When the timestamp field is missing or the timestamp field is inconsistent with the collection access time, the collection access time is used as the time anchor value after being substituted or corrected.

[0027] The preset sorting rule refers to setting a fixed sorting key priority and judgment order for concurrent arrival records, and comparing concurrent arrival records according to the sorting key priority to output a unique sequence number. For example, the records are judged in the order of time anchor value, data source identifier and access port identifier until a unique sorting result is obtained.

[0028] S1.2: Based on the time anchor point collection record set, perform location locking, field integrity verification and missing field marking trajectory registration, and create field cross-mapping entries, integrating them into a field locking verification set; Specifically, a location pointer is generated for the time anchor collection record set using the time anchor value and collection sequence number as index keys, and the start and end boundaries are fixed. The location pointer is segmented to limit the coverage area, forming a location locking record. Based on the location locking record, the collected computing resource data and task requirement data are checked for required fields, types, and value ranges according to field integrity rules. For fields that do not meet the field integrity rules, missing field markers are generated and bound to the location pointer. The location pointer is used as the primary key to register the missing field markers as trajectory entries in the order of occurrence, and the time anchor value and collection sequence number are attached to form a missing field marker trajectory record set. Using the collected computing resource data and task requirement data as input, a field comparison relationship is constructed and the mapping direction, mapping conditions, and mapping confidence markers are registered. Under the constraints of the missing field marker trajectory record set, the field comparison relationship generates field cross-mapping entries and retains them according to the consistency threshold (e.g., 0.8 to 0.99), generating a field locking verification set.

[0029] It should be noted that the field integrity rules are defined as the judgment rules for field verification and missing field marking based on the preset list of required fields, field type constraint list, and field value range constraint list of the collected computing resource data and task requirement data.

[0030] The consistency threshold is defined as the decision limit for retaining field cross-mapping entries based on the consistency of the mapping confidence mark and the pass rate of the check, with an example range of 0.8 to 0.99.

[0031] S1.3: Extract the task requirement field and computing resource field from the field lock check set, perform field bundle reorganization, write missing field markers and field cross-mapping entries, and generate scheduling input package.

[0032] Specifically, the corresponding records are located based on the position pointer and time anchor point. The task requirement field and computing resource field are extracted separately, and the collection sequence number and segmentation mark are retained as association keys. The missing field mark is copied to the extraction result along with the field and the source position pointer is registered. In the field bundle reorganization process, the task requirement field and computing resource field are merged and split using a preset field bundle template. Fields under the same association key are sorted according to the field bundle template and the field bundle completeness is obtained. When the field bundle completeness is lower than the completeness threshold, the missing field mark is kept unchanged and a field bundle gap mark is added (for example, the field bundle completeness judgment boundary is defined based on the proportion of required fields in the field bundle entry and the number of missing field marks, with an example range of 0.8 to 0.99). The fields after the field bundle reorganization are filled or replaced with values ​​according to the mapping direction and mapping conditions in the field cross-mapping entry. When the mapping confidence mark does not meet the consistency threshold, value filling is prohibited and the missing field mark is retained. The field bundle, missing field mark and field cross-mapping entry are merged and packaged according to the association key to generate the scheduling input package.

[0033] It should be noted that the preset field bundle template pre-fixes the merging order, splitting boundary, field bundle entry structure and field bundle entry sorting of task requirement fields and computing resource fields based on the encapsulation requirements of the scheduling input package, which is used to rearrange fields under the same association key into field bundle entries.

[0034] S2: Based on the scheduling input packet, perform task behavior range positioning on the task requirement field, perform resource trend segmentation on the computing resource field, and perform behavior matching calibration and dimension completion to generate a vector alignment table; S2.1: For the task requirement field and computing resource field in the scheduling input package, perform alignment benchmark solidification based on time anchor point and sequence number, and add alignment sequence number to generate field directory; Specifically, the task requirement field and computing resource field in the scheduling input package are extracted and registered according to the time anchor value and sequential number. The time anchor value is truncated or padded according to a uniform time precision to ensure that the time anchor value is comparable (e.g., millisecond level). The sequential number maintains the collection arrival order and serves as the basis for resolving concurrent conflicts. The alignment benchmark solidification process combines the time anchor value and sequential number of each record into an alignment benchmark entry and solidifies the alignment benchmark entry into an immutable index key. At the same time, a reverse jump threshold judgment is performed on adjacent records whose time anchor values ​​do not meet the monotonic order. The alignment sequence number is generated by sorting according to the time anchor value and sequential number, and the alignment sequence number is bound to the task requirement field position pointer and the computing resource field position pointer respectively to form a one-to-one correspondence, generating a field directory.

[0035] It should be noted that the reverse jump threshold is defined as the allowable rollback limit based on the amount of backslid in the alignment sorting direction of the time anchor values ​​of adjacent records, with an example range of 1 millisecond to 100 milliseconds.

[0036] S2.2: Based on the field directory, track the switching position of task requirement fields, write the start time anchor and end time anchor of the task behavior interval, and generate a set of behavior intervals; Specifically, the task requirement fields in the field directory are read in the order of alignment number and associated with the corresponding time anchor values. After the task requirement field enters the switching position tracking, the status of the task requirement field with adjacent alignment number is compared according to the preset switching judgment condition of the task requirement field. When the status of the task requirement field meets the switching judgment condition, the switching position is registered and the corresponding alignment number of the switching position is locked as the interval boundary. For each interval boundary, the most recent switching position or the starting alignment number is traced backward to determine the starting time anchor value of the task behavior interval, and the next switching position or the last alignment number is traced backward to determine the ending time anchor value of the task behavior interval. The starting time anchor value and the ending time anchor value of the task behavior interval are written into the behavior interval entry and the interval number is bound to the corresponding alignment number range. At the same time, the missing field marker and the field cross-mapping entry reference identifier in the field directory are attached to the behavior interval entry for subsequent verification, generating a behavior interval set.

[0037] It should be noted that the preset switching judgment condition sets the number of consecutive confirmations as the switching judgment parameter and binds it to the alignment number window. When the value of the task requirement field remains at the new value within the number of consecutive confirmations and adjacent alignment numbers, the switching position is determined (for example, the number of consecutive confirmations is three to five).

[0038] S2.3: Based on the behavior interval set, lock the scanning boundary, construct a segmented scanning window, perform monotonic direction determination and register the inflection point time anchor point, and generate a trend segment set; Specifically, the starting and ending time anchor values ​​of the task behavior interval entries are read according to the interval sequence number and fixed as the scanning boundary. Within the scanning boundary, the scanning windows are continuously divided into segments according to the window step size (1–10 sampling intervals), and the segmented scanning windows are bound to the interval sequence number and alignment number range. In each segmented scanning window, the value sequence of the computing power resource field is extracted and adjacent values ​​are compared. If the comparison result is continuously rising, an upward direction mark is registered; if the comparison result is continuously falling, a downward direction mark is registered; if the comparison result is flat, a flat direction mark is registered. When the direction mark reverses in an adjacent window and the reversal meets the confirmation number, the time anchor value corresponding to the reversal position is registered as the inflection point time anchor value (e.g., the confirmation number is 2 to 5, and the confirmation number is inversely linked to the confirmation step size). The adjacent inflection point time anchor values ​​or the scanning boundary endpoints are used as segment boundaries. The segment starting time anchor value, segment ending time anchor value, direction mark and segment sequence number are written and bound to the interval sequence number to form trend segment entries. All trend segment entries are summarized to generate a trend segment set.

[0039] S2.4: Pair the behavior interval set and trend segment set according to the overlap relationship of time anchor points, establish a dimension list, perform missing dimension completion and unit consistency, and generate a vector alignment table.

[0040] Furthermore, the behavior interval set and trend segment set are paired according to the overlap relationship of time anchor points. The existence of an intersection between two time anchor point intervals is used as the determination result of the time anchor point overlap relationship, and this is used as the pairing condition. When the pairing is successful, the time anchor point values ​​of the intersection start point and intersection end point are locked, and the interval number, segment number, and alignment number range are recorded to form a pairing entry. After the pairing entry enters the dimension list creation process, the dimension list is summarized from the task requirement field and computing power resource field in the scheduling input package, and the dimension position number and unit identifier are written into the dimension list. After the dimension list enters the missing dimension completion process, the pairing entries are checked according to the dimension position number. The system checks the existence of values ​​and performs value filling by referencing field cross-mapping entries when values ​​are missing. If a field cross-mapping entry does not meet the consistency threshold, a missing dimension flag is written, and a placeholder value is generated according to a preset filling value rule (which refers to the rule of specifying a fixed placeholder value for the missing dimension and registering the missing dimension flag when value filling is unavailable, such as filling zero values ​​for numeric dimensions or filling null values ​​for non-numeric dimensions). After the filled values ​​enter the unit consistency process, the unit of the values ​​is checked and converted using the dimension list unit identifier as the target unit, and a conversion flag is registered (e.g., milliseconds to seconds). The paired entries and the filled values ​​are encapsulated in a fixed column order to generate a vector alignment table.

[0041] S3: Input the vector alignment table into the resource evaluation deep neural network, calculate the resource adaptation score and conflict risk score, and perform candidate resource clustering and candidate resource lower limit constraint verification through scheduling optimization strategy to generate a candidate resource list; S3.1: Perform resource grouping, positional verification and cleaning on the vector alignment table, and divide it according to the fixed neural input batch size to generate neural input batches; Specifically, the vector alignment table is grouped into resource groups, and a grouping index is generated for each resource group. Records after resource grouping are sorted by alignment sequence number and dimension position number, and the positional continuity is checked. If the positional continuity is not satisfied, a positional gap marker is registered, and placeholder values ​​corresponding to the missing dimension markers are filled at the gap positions to maintain positional alignment. Records with completed positional alignment are cleaned to remove duplicate records, and abnormal values ​​are marked as abnormal. Abnormal values ​​are truncated or replaced with placeholder values ​​according to a preset processing method (a fixed constraint specifying a unified handling action and processing order for abnormal values, such as truncating abnormal values ​​that exceed the value range and replacing them with placeholder values ​​when truncation is unavailable). The alignment sequence is continuously sliced ​​according to a fixed neural input batch size, while maintaining the same grouping index at the slice boundaries to prevent splitting across batches. After slicing, a batch number, grouping index range, and alignment sequence number range are written to each slice and encapsulated as a batch entry. The batch entries are then aggregated to generate a neural input batch.

[0042] S3.2: Input the neural input batch into the resource evaluation deep neural network, calculate the resource adaptation score and conflict risk score, and bind them with the alignment number. At the same time, perform range verification and labeling to generate a score record set. Furthermore, batch entries of the neural input batches are extracted according to batch numbers, and the vector sequences carried by the batch entries are read. The vector sequences are fed into the resource assessment deep neural network according to their sequence positions in a fixed input order to complete forward propagation. The resource assessment deep neural network generates resource adaptation scores and conflict risk scores at the output end and binds them to the alignment sequence numbers one by one and writes them into the score entries. The score entries are simultaneously written with the batch number and grouping index to locate the source of the vector sequence. When generating resource adaptation scores, the sequence length is normalized and the input feature vector is located using the sequence position index. After the input feature vector is linearly transformed by the weight matrix and bias vector, it enters the hyperbolic tangent activation function to obtain the intermediate... The intermediate representation is aggregated using the L2 norm and summarized across sequence positions, then mapped by an exponential function to obtain the resource adaptation score. When generating the conflict risk score, the sequence length is normalized and the input feature vector is located using the sequence position index. The input feature vector undergoes a linear transformation of the weight matrix and bias vector before entering a smoothing positive activation function to obtain the intermediate representation. This intermediate representation is aggregated using the L1 norm and summarized across sequence positions, then mapped by an exponential function and its complement is taken to obtain the conflict risk score. Range verification and marking are performed on both types of scores, checking their value ranges and writing range markers. Out-of-bounds values ​​are registered with out-of-bounds markers, then boundary truncation or placeholder value replacement is performed and written into the score entries. All these steps are then combined to generate a score record set.

[0043] The formula for calculating the resource adaptation score is: ; in, Indicates resource compatibility score. Represents an exponential function. This represents the length of the vector sequence within the alignment index range and is a positive integer. Represents the L2 norm, This represents the linear transformation weight matrix corresponding to the resource adaptation score output header. The input feature vector represents the sequence position t. This represents the bias vector corresponding to the resource adaptation score output header. This represents the hyperbolic tangent activation function.

[0044] The formula for calculating the conflict risk score is: ; in, Conflict risk scoring Describes the norm, Represents a smooth positive activation function. This represents the linear transformation weight matrix corresponding to the conflict risk score output header. This represents the bias vector corresponding to the conflict risk score output header.

[0045] It should be noted that the resource assessment deep neural network is constructed using the dimensional position index order of the vector alignment table as the input channel order and the alignment index range as the temporal expansion order. At the input end, the placeholder values ​​corresponding to missing dimension markers and valid values ​​are normalized on the same scale and then aggregated into batch-entry-level vector sequences according to the grouping index. The main body of the resource assessment deep neural network adopts a four-layer nonlinear mapping backbone, with a normalization layer and a dropout layer connected sequentially after each nonlinear mapping layer to stabilize the training process and suppress overfitting. The number of hidden nodes in the nonlinear mapping backbone decreases layer by layer or remains constant (e.g., from 1024 to 256 or 512 to 512). The backbone output enters the temporal convergence layer, where the representation expanded by the alignment index is converged to produce a resource adaptation score output head and a conflict risk score output head. Each output head contains two fully connected layers and outputs at the end... Output resource fit scores and conflict risk scores (e.g., 256 to 64 output head nodes); the training of the resource assessment deep neural network takes neural input batches as input and aligns supervision labels according to batch number and alignment sequence number. The supervision labels convert acceptance receipts and initiation receipts into resource fit labels (acceptance receipts and initiation receipts are both recorded as high fit labels, acceptance receipts are recorded as medium fit labels, and acceptance receipts are recorded as low fit labels) and convert running telemetry receipts into conflict risk labels. The generation of conflict risk labels uses an error code density threshold to classify continuous indicators (e.g., 0.01 to 0.2). The training process updates the network weights in reverse with dual output head loss and stops when the discrimination of resource fit scores and conflict risk scores on validation samples improves, thus obtaining a resource assessment deep neural network for inference.

[0046] It should be noted that the error code density threshold (example range: 0.01–0.20) is set with the lower limit based on the upper bound of the error code noise floor density in the historical stable operation window, and the upper limit based on the tolerable error density threshold before startup failure or degradation is triggered.

[0047] S3.3: Based on the score record set, the resource adaptation score and conflict risk score are converted into cluster keys through scheduling optimization strategies and merged to form candidate resource clusters; Specifically, the resource adaptation score and conflict risk score in the scoring entries are read according to the alignment sequence number, and the range markers are retained. For each scoring entry, the conflict risk score is placed into the corresponding risk level according to the risk level boundary set by the scheduling optimization strategy (referring to the list of boundary value points used to divide the conflict risk score into multiple risk level intervals) and written into the risk level marker. At the same time, the resource adaptation score is placed into the corresponding adaptation level according to the adaptation level boundary set by the scheduling optimization strategy (referring to the list of boundary value points used to divide the resource adaptation score into multiple adaptation level intervals) and written into the adaptation level marker. The risk level marker and the matching level marker are concatenated in a fixed order to form a cluster key, which is written into the cluster record along with the alignment sequence number, batch number, and grouping index (for example, if the boundary is 0.2, 0.5, 0.8, the risk level marker is R0-R3, and the matching level marker is A0-A3, then the cluster key is R2|A3). Cluster records with the same cluster key are merged into the same cluster entry and a cluster index is generated. An alignment sequence number list and a scoring entry reference are appended to the cluster entry, and the number within the cluster is recorded for subsequent filtering. All cluster entries are summarized to generate a candidate resource cluster.

[0048] It should be noted that the scheduling optimization strategy refers to solidifying the risk level boundary, the adaptation level boundary, the concatenation order of the risk level marker and the adaptation level marker, and the cluster merging rules based on the value distribution of the resource adaptation score and the conflict risk score. This strategy is used to convert the score record set into candidate resource clusters and support subsequent fine-tuning and distribution.

[0049] S3.4: Based on the candidate resource cluster, perform candidate resource lower limit constraint verification, register verification marks, dynamically summarize scoring record items, and generate a candidate resource list.

[0050] Specifically, the alignment sequence list and score entry references in each cluster entry are read according to the cluster index, and the corresponding score entry is located by back-linking the score record set. The candidate resource lower limit constraint verification is based on the candidate resource constraint list. The resource adaptation score, conflict risk score and range mark in the score entry are compared with the candidate resource constraint list item by item, and the verification conclusion is output. When the verification conclusion meets the candidate resource constraint list, a pass mark is registered and the score entry reference is retained. When the verification conclusion does not meet the candidate resource constraint list, an interception mark is registered and an interception reason mark is written (e.g., the resource adaptation score is lower than the lower limit or the conflict risk score is higher than the upper limit). After the verified score entries enter the dynamic summary process, the minimum value and median value of the resource adaptation score within the cluster are taken according to the cluster index, and the maximum value and median value of the conflict risk score within the cluster are taken. At the same time, the alignment sequence list is retained for subsequent fine-tuning and positioning, and a candidate resource list is generated.

[0051] S4: Set the resource adaptation score and conflict risk score in the candidate resource list as the sorting key, perform fine sorting within the cluster, select the head resource for distribution and collect the acceptance receipt, start receipt and running telemetry receipt, and generate a feedback package; S4.1: Based on the candidate resource list, set the resource adaptation score and conflict risk score as the sorting key, and perform gradient conversion and jitter detection to construct a sorting vector for fine sorting within the cluster and generate a sorting table for distribution. Specifically, the summarized values ​​of resource adaptation scores and conflict risk scores are written into the sort key field. Gradient reduction is based on the difference in sort keys between adjacent candidate resource items. The sort key field is processed and the reduction mark is retained to weaken the abrupt change at the boundary point. Jitter detection combines the alignment sequence list in the candidate resource items with the backlink score record set to extract the corresponding score items. The direction of change of resource adaptation scores and conflict risk scores on consecutive alignment sequences is checked. When the direction of change reverses multiple times within a short window (the short window length in jitter detection is W consecutive alignment sequences, where W is a positive integer and W ranges from 3 to 10), a jitter mark is registered and... The sort key field is restricted to a range allowed by the jitter threshold (defined by the frequency and magnitude of the reversal of resource adaptation scores and conflict risk scores on consecutively aligned ordinal numbers within a short window, with an example range of 0.02 to 0.10). A sort vector is constructed using the cluster index, cluster key, converted resource adaptation score sort key, converted conflict risk score sort key, jitter flag, and score entry reference list. Within the range corresponding to the cluster index, intra-cluster sort numbers are generated in order of priority for resource adaptation score sort key and second-best for conflict risk score sort key. The intra-cluster sort numbers and sort vector are then aggregated and integrated into a sort table for distribution.

[0052] S4.2: Based on the sorting table, select the head resources in sequence and distribute them to the nodes in the cluster. At the same time, collect the acceptance receipt, start receipt and operation telemetry receipt, and generate a receipt chain record set. Specifically, sorted entries are read sequentially according to the cluster index and intra-cluster sorting number. Candidate resource entries are located based on the cluster index and alignment number list in the sorted entries. Candidate resource entries are selected based on the distribution quota, with the entries with the highest intra-cluster sorting number being registered as header resources. After the header resources enter the distribution process to intra-cluster nodes, the computing power resource field, task requirement field, and alignment number corresponding to the header resources are encapsulated into a distribution payload and distributed to the intra-cluster nodes. To facilitate the association of receipts, time anchor values ​​and collection sequence numbers are added to the distribution payload. When an intra-cluster node returns an acceptance receipt, the acceptance receipt is bound to the alignment number and the acceptance time anchor value is registered. When an intra-cluster node returns a start receipt, the start receipt and acceptance receipt binding entries are concatenated and the start time anchor value is registered. When an intra-cluster node returns a running telemetry receipt, the running telemetry receipt is attached to the start receipt binding entry according to the time anchor value order and the running telemetry receipt sequence number is registered. The acceptance receipt, start receipt, and running telemetry receipt are summarized and integrated into a receipt chain record set according to the alignment number.

[0053] S4.3: Chain the receipt chain record set, build the receipt chain index, perform status verification and association, and generate a feedback package.

[0054] Furthermore, the sequence of acceptance receipts, initiation receipts, and operational telemetry receipts is read according to the alignment sequence number and chained together according to the time anchor value order. The acceptance receipt is used as the head node of the chain, and the initiation receipt is attached to the acceptance receipt node. The operational telemetry receipts are then attached to the initiation receipt node in sequence according to the operational telemetry receipt sequence number. A receipt chain index is generated using the alignment sequence number, time anchor value, and acquisition sequence number, and written into the index field for locating nodes within the chain. A status consistency check is performed on the nodes within the chain corresponding to the receipt chain index, covering both the acceptance status and the initiation status. The timing consistency and the continuity of status codes in the telemetry receipts (meaning that no unauthorized transition sets appear in the status codes within the same alignment sequence number range) are checked. When the status consistency check fails, an abnormal status flag is registered and the time anchor value of the abnormal position is retained. When the status consistency check passes, a status pass flag is registered. The association process establishes an association entry with the receipt chain index, the cluster index, the intra-cluster sorting sequence number, and the alignment sequence number in the scoring record set in the issued sorting table and writes it into the association field. The receipt chain index, intra-chain nodes, status verification flags, and association entries are summarized and encapsulated to generate a feedback package.

[0055] It should be noted that the timing consistency between the acceptance state and the startup state means that within the same alignment sequence number range, an acceptance receipt must appear first and be accepted successfully before the corresponding startup receipt can appear and be bound to that acceptance receipt. If startup precedes acceptance or startup still appears after acceptance failure, it is considered inconsistent. S5: Input the feedback packet into the feedback learning deep neural network, perform behavior playback comparison, locate key conflict points, dynamically update the scheduling optimization strategy and register the weight migration trajectory, and generate a computing resource scheduling record set.

[0056] S5.1: Based on the feedback package, expand the restoration receipt chain index, and split and interleave the acceptance receipt, start receipt and running telemetry receipt into three tracks, while binding them with the alignment sequence number to generate a multi-track playback arrangement set; Specifically, the system expands and restores the acceptance receipt, initiation receipt, and operation telemetry receipt according to the receipt chain index, retaining the alignment number, time anchor value, and acquisition sequence number in the receipt chain index as positioning information. The three tracks are split: the acceptance receipt is extracted to form an acceptance receipt track and sorted by time anchor value; the initiation receipt is extracted to form an initiation receipt track and sorted by time anchor value; and the operation telemetry receipt is sorted by operation telemetry receipt number and time anchor value to form an operation telemetry receipt track. The system is then staggered and aligned to the same alignment number, and the order of the time anchor value is arranged within the same alignment number. If a track is missing, a missing mark is written, generating a multi-track playback arrangement set.

[0057] S5.2: Input the multi-track playback orchestration set into the feedback learning deep neural network, align it with the running telemetry receipts and cross-link them to generate a playback deviation scale. Specifically, the alternating order of the acceptance receipt track, start receipt track, and running telemetry receipt track is extracted according to the alignment sequence number and encapsulated into a playback input segment. This segment is then input into a feedback learning deep neural network. The feedback learning deep neural network performs forward propagation on the playback input segment and outputs the playback comparison result bound to the alignment sequence number, along with a status comparison marker and a timing comparison marker. Running telemetry receipts are aligned and cross-linked, using the alignment sequence number and time anchor value as the alignment key to match the running telemetry receipts. The status code, error code, and time anchor value are linked to the playback comparison result to form a cross-linked record. The status comparison marker is compared with the status code to form a status deviation marker. The timing reference mark is checked against the time anchor value to form a timing deviation mark. The state deviation mark and the timing deviation mark are mapped to a deviation scale according to the scale threshold (defined based on the deviation intensity value range of the state deviation mark and the timing deviation mark under the same alignment number, the example range is 0.2 to 0.8) and the deviation scale level is recorded (for example, the deviation scale falling into 0.2 to 0.4 is recorded as mild, the deviation scale falling into 0.4 to 0.6 is recorded as moderate, and the deviation scale falling into 0.6 to 0.8 is recorded as severe). The alignment number, receipt chain index, deviation scale level and cross-linked record reference are summarized and encapsulated to generate a playback deviation scale table.

[0058] It should be noted that the construction of the feedback learning deep neural network uses replay input segments as input carriers. First, the alternating order of the acceptance receipt track, the initiation receipt track, and the running telemetry receipt track is encoded into an event sequence according to the alignment sequence number. The event type, time anchor value, and state-related fields are mapped into event embedding vectors of the same dimension. At the same time, the alignment sequence number is written into the position encoding to fix the unfolding order at time. The event embedding vectors are input into the alternating sequence encoding backbone of gated temporal convolutional layers and multi-head self-attention layers to simultaneously capture intra-window mutations and cross-window dependencies. Finally, the backbone output is compared with the number of running telemetry receipts. The input value is represented by a cross-attention alignment layer to perform bidirectional alignment to obtain a playback reference representation. A state discrimination output structure and a temporal offset output structure are set on the playback reference representation, and the output is a state reference label and a temporal reference label bound to the alignment sequence number. In the training phase, the event sequence is first pre-trained by mask reconstruction to obtain a robust representation of the missing dimension. Then, the state classification loss and the temporal offset regression loss are jointly optimized in a supervised manner. The consistency rate of the deviation scale level in the verification sample is improved and the offset error converges as a solidification condition, so as to obtain a feedback learning deep neural network for generating the playback deviation scale table.

[0059] S5.3: Based on the playback deviation scale, a candidate pool is constructed through classification and filtering, key conflict points are located in a series and cross-location manner, and a list of key conflict points is generated. Specifically, records that meet the screening criteria for deviation scale level are retained through classification and screening (e.g., records with a deviation scale level of moderate or severe are directly retained, and records with a deviation scale level of mild are retained only when both the status deviation mark and the timing deviation mark exist). Records with both the status deviation mark and the timing deviation mark exist are prioritized for retention. The records retained through classification and screening are summarized to form a candidate pool, and candidate numbers are written to the records in the candidate pool. After the candidate pool enters the serial processing, the candidate numbers are serialized into continuous segments according to the alignment sequence number in the receipt chain index, and the alignment sequence number of the start and end points of the continuous segments is recorded. After the candidate pool enters the cross-location, the cross-attached records are used to reference the back-connected telemetry receipts and extract the status code, error code, and time anchor value corresponding to the same alignment sequence number. At the same time, the time anchor value of the acceptance receipt and the start receipt is used as a reference to check the abnormal positions of the running telemetry receipts. The positions in the continuous segments where the status code jumps or the error codes are dense are registered as key conflict points, and a list of key conflict points is generated.

[0060] S5.4: Map the list of key conflict points to the scheduling optimization strategy and perform dynamic updates. At the same time, register the weight migration trajectory and generate a set of computing resource scheduling records.

[0061] Specifically, the key conflict points are read according to the receipt chain index and alignment sequence number, and the sorting table and scoring record set are back-linked to locate the corresponding cluster key, cluster index and intra-cluster sorting sequence number. The mapping process establishes a correspondence between the key conflict points and the risk level boundary, adaptation level boundary, cluster key concatenation order and merging method in the scheduling optimization strategy, and writes it into the mapping record. The dynamic update performs increases or decreases on the risk level boundary and adaptation level boundary and adjusts the cluster key concatenation order according to the mapping record. At the same time, the updated version number and effective time anchor value are written. The weight migration trajectory registration writes the changes in the risk level boundary and adaptation level boundary values ​​before and after each update, the number of key conflict points triggered by the update and the corresponding alignment sequence number range into the migration record. The migration record and the updated scheduling optimization strategy are encapsulated together to generate a computing resource scheduling record set.

[0062] This embodiment also provides a computer device applicable to the artificial intelligence-based computing resource scheduling method, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the artificial intelligence-based computing resource scheduling method proposed in the above embodiment.

[0063] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0064] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the artificial intelligence-based computing resource scheduling method proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0065] In summary, this invention achieves structured integration of heterogeneous computing resources through the candidate resource clustering step, overcoming the resource gap problem and improving the utilization rate of fragmented resources; and achieves continuous optimization of scheduling strategies through feedback learning deep neural networks for behavior playback comparison, thereby improving the decision-making accuracy and operational stability when facing dynamic loads.

[0066] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for scheduling computing resources based on artificial intelligence, characterized in that: include, Collect computing resource data and task requirement data, perform field integrity checks, register missing field markers, and generate scheduling input packages; Based on the scheduling input packet, the task behavior range is located for the task requirement field, the resource trend is segmented for the computing resource field, and behavior matching and dimension completion are performed to generate a vector alignment table. The vector alignment table is input into the resource evaluation deep neural network to calculate the resource fit score and conflict risk score. Then, the candidate resource clustering and candidate resource lower limit constraint verification are performed through the scheduling optimization strategy to generate a candidate resource list. Set the resource adaptation score and conflict risk score in the candidate resource list as the sorting key, perform fine sorting within the cluster, select the head resource for distribution and collect acceptance receipts, start receipts and running telemetry receipts, and generate a feedback package; The feedback packet is input into the feedback learning deep neural network, the behavior is replayed and compared, key conflict points are located, the scheduling optimization strategy is dynamically updated and the weight migration trajectory is registered, and a set of computing resource scheduling records is generated.

2. The artificial intelligence-based computing resource scheduling method as described in claim 1, characterized in that: The process of collecting computing resource data and task requirement data, performing field integrity checks, registering missing field markers, and generating a scheduling input package involves the following steps: The collected computing resource data and task requirement data are written into time anchors, and segmented according to the collection sequence to generate a time anchor collection record set. Based on the time anchor point collection record set, the system performs location locking, field integrity verification, and missing field marking trajectory registration, and creates field cross-mapping entries, which are then integrated into a field locking verification set. Extract the task requirement field and computing resource field from the field lock check set, perform field bundle reorganization, write missing field markers and field cross-mapping entries, and generate a scheduling input package.

3. The artificial intelligence-based computing resource scheduling method as described in claim 2, characterized in that: Based on the scheduling input packet, the process involves locating task behavior intervals for the task requirement field, segmenting resource trends for the computing resource field, performing behavior matching and dimension completion, and generating a vector alignment table. The specific steps are as follows: For the task requirement field and computing resource field in the scheduling input package, the alignment benchmark is solidified based on the time anchor point and sequence number, and an alignment sequence number is added to generate a field directory; Based on the field directory, track the switching position of task requirement fields, write the start time anchor and end time anchor of the task behavior interval, and generate a set of behavior intervals; Based on the behavior interval set, the scanning boundary is locked, a segmented scanning window is constructed, monotonic direction determination is performed and inflection point time anchors are registered to generate a trend segment set; The behavior interval set and trend segment set are paired according to the overlap relationship of time anchor points, and a dimension list is established. Missing dimensions are filled in and units are consistent to generate a vector alignment table.

4. The artificial intelligence-based computing resource scheduling method as described in claim 3, characterized in that: The specific steps for inputting the vector alignment table into the resource evaluation deep neural network to calculate the resource fit score and conflict risk score are as follows. The vector alignment table is grouped, its position is checked and cleaned, and then it is divided according to a fixed neural input batch size to generate neural input batches. The neural input batches are fed into the resource evaluation deep neural network to calculate the resource adaptation score and conflict risk score, and bind them with the alignment number. At the same time, range checking and labeling are performed to generate a score record set.

5. The artificial intelligence-based computing resource scheduling method as described in claim 4, characterized in that: The process of generating a candidate resource list involves performing candidate resource clustering and lower limit constraint verification through scheduling optimization strategies. The specific steps are as follows: Based on the score record set, resource adaptation scores and conflict risk scores are converted into cluster keys through scheduling optimization strategies and merged to form candidate resource clusters. Based on the candidate resource cluster, the lower limit constraint of the candidate resources is checked, the check mark is registered, the scoring record items are dynamically summarized, and a candidate resource list is generated.

6. The artificial intelligence-based computing resource scheduling method as described in claim 5, characterized in that: The steps involve setting the resource compatibility score and conflict risk score in the candidate resource list as the sorting key, performing intra-cluster fine sorting, selecting the head resource for distribution, and collecting acceptance receipts, start receipts, and operation telemetry receipts to generate a feedback packet. The specific steps are as follows: Based on the candidate resource list, the resource adaptation score and conflict risk score are set as sorting keys, and gradient conversion and jitter detection are performed to construct a sorting vector for fine sorting within the cluster and generate a sorting table for distribution. According to the sorting table, the head resources are selected in sequence and distributed to the nodes in the cluster. At the same time, the acceptance receipt, the start receipt and the running telemetry receipt are collected to generate a receipt chain record set. The receipt chain record set is chained together to build a receipt chain index, and status verification and association are performed to generate a feedback package.

7. The artificial intelligence-based computing resource scheduling method as described in claim 6, characterized in that: The specific steps for inputting the feedback packet into the feedback learning deep neural network and performing behavior playback comparison are as follows. Based on the feedback package, expand the restoration receipt chain index, and split and interleave the acceptance receipt, start receipt and running telemetry receipt into three tracks, while binding them with the alignment sequence number to generate a multi-track playback arrangement set. The multitrack playback orchestration set is input into a feedback learning deep neural network, and then aligned and cross-linked with the running telemetry receipts to generate a playback deviation scale.

8. The artificial intelligence-based computing resource scheduling method as described in claim 7, characterized in that: The specific steps for generating the computing resource scheduling record set are as follows: Based on the playback deviation scale, a candidate pool is constructed through classification and screening, and key conflict points are located in a series and cross-location manner to generate a list of key conflict points. The list of key conflict points is mapped to the scheduling optimization strategy and dynamically updated. At the same time, the weight migration trajectory is registered to generate a set of computing resource scheduling records.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the artificial intelligence-based computing resource scheduling method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the artificial intelligence-based computing resource scheduling method according to any one of claims 1 to 8.

Images