A Multi-Dimensional Specialized Equipment Simulation Training Resource Scheduling Method

Abstract

This invention discloses a multi-dimensional resource scheduling method for professional equipment simulation training, specifically relating to the field of professional equipment simulation training management and control. It addresses issues such as inconsistent resource states, execution time conflicts, and on-site occupancy insertions leading to plan mismatches that are difficult to review and correct during concurrent training tasks. The method unifies and merges scheduling console operation logs, node heartbeats, and task registration forms to construct a schedulable resource base map. It then generates a task package matching table and training execution sequence by combining the resource demand and execution dependencies of injected task packages. During execution, it continuously verifies and records abnormal occupancy. Finally, it corrects the matching table and scheduling rules by comparing the baseline image and reset image with the standard playback results, forming an iteratively updatable next round of training execution sequence. Simultaneously, it accumulates samples to support subsequent stable scheduling decisions.

Description

Technical Field

[0001] This invention relates to the field of professional equipment simulation training management and control, and more specifically, to a multi-dimensional professional equipment simulation training resource scheduling method. Background Technology

[0002] In specialized equipment simulation training scenarios, training organizations typically need to integrate simulation execution, task injection, state reset, process replay, and result evaluation within the same training workflow. For this type of business, the scheduling platform must not only allocate computing power and nodes but also uniformly orchestrate training units, injected task packages, state snapshots, and change trajectories to ensure a coherent training process, complete records, and verifiable conclusions. Existing solutions generally possess resource management and task allocation capabilities and can support the initiation and execution of training tasks. However, in real-world training scenarios involving the collaborative participation of multiple resources, the scheduling object is no longer merely whether resources are available, but rather whether the training process is continuous, comparable, and traceable.

[0003] However, existing technologies still suffer from the following problems when scheduling resources for multi-dimensional professional equipment simulation training: Within the same training process, the binding relationship between injected task packages and training units is difficult to maintain consistently throughout the entire process; isolated write domains and state tracking are difficult to implement stably; post-injection state snapshots and change trajectories are difficult to preserve completely; and standard-compliant playback of baseline and post-reset images is difficult to achieve stable comparison. This leads to inconsistent conclusions, difficulty in closing process evidence, and inaccurate evaluation results reflecting the true training level in different rounds of the same training operation. This problem directly affects the reliability and practicality of training scheduling and has become the core technical problem that this invention needs to solve.

[0004] To address the aforementioned problems, a technical solution is provided. Summary of the Invention

[0005] To overcome the aforementioned deficiencies in the prior art, embodiments of the present invention provide a multi-dimensional professional equipment simulation training resource scheduling method. This method unifies and merges the scheduling console operation logs, node heartbeats, and task registration forms to construct a schedulable resource base map. It then generates a task package matching table and a training execution sequence by combining the resource demand relationships and execution dependencies of injected task packages. During execution, it continuously verifies and records abnormal occupancy. Finally, it corrects the matching table and scheduling rules by comparing the baseline image and reset image with the standard playback results, forming an iteratively updatable next round of training execution sequence. Simultaneously, it accumulates samples to support subsequent stable scheduling decisions, thereby solving the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] S1: Read the scheduling console operation log, node heartbeat, and task registration form according to the training unit, complete the source alignment and record merging, map the original state to the available state, occupied state, and switched state, and generate a resource state set after resolving concurrent conflicts;

[0008] S2: Read the resource requirement relationship and execution dependency relationship of the injected task package, generate candidate resource combinations and candidate execution time periods based on the resource state set, calculate the availability continuity of the same window and the neighborhood conflict density, obtain the steady-state matching probability and form a task package matching table;

[0009] S3: Output the training execution sequence and issue scheduling instructions according to the task package matching table. Before execution, verify the consistency between the candidate resource combination and the current resource status. During execution, collect status snapshots and change trajectories. After identifying abnormal occupancy, backfill the task package matching table.

[0010] S4: Based on the training execution sequence, drive the baseline image and reset image to perform standard playback, compare the playback trajectory with the execution record, output the verification result, correct the task package matching table and scheduling rules according to the verification result, and generate the next round of training execution sequence.

[0011] Furthermore, based on the training unit identifier and resource instance identifier, the scheduling console operation log, node heartbeat, and task registration form are collected, and the unified source status observation is mapped to available status, occupied status, and switched status. Based on the unified event sequence, the aligned status observation and conflict resolution are completed, and the current status and status start record are determined.

[0012] Furthermore, a resource state set is generated based on the current state, state start record, training unit identifier, and resource instance identifier. The resource state set is synchronously written to the most recent state change record and forms a resource base map. The resource base map maintains a unique mapping from the training unit identifier and resource instance identifier to the current state, which is used by the task package matching table.

[0013] Furthermore, based on the resource state set and resource base map, the resource demand relationship and execution dependency relationship of the injected task package are analyzed. Candidate resource combinations are formed according to the consistency of resource roles and the occupancy boundary. Candidate execution periods are extracted from the unified state sequence and written into the complete period marker.

[0014] Furthermore, the availability continuity and neighborhood conflict density of candidate resource combinations and candidate execution periods are calculated. The availability continuity of the same window represents the level of continuous execution readiness within the period, and the neighborhood conflict density represents the level of aggregation of abnormal occupancy markers in historical execution records. The two-parameter input Gaussian Naive Bayes outputs the steady-state matching probability.

[0015] Furthermore, the matching status in the task package matching table is determined according to the steady-state matching probability and the execution dependency relationship. The task package matching table uniformly records the injected task package identifier, candidate resource combination identifier, candidate execution time period identifier and two-parameter results, and generates training execution sequences according to resource conflict resolution rules.

[0016] Furthermore, scheduling instructions are triggered one by one according to the training execution sequence. Before scheduling, the consistency verification between the candidate resource combination in the task package matching table and the current resource status is performed. Only after the consistency verification is passed can execution begin. During the execution phase, status snapshots are continuously collected and archived in the order of events. Based on adjacent status snapshots, status change segments are extracted and change trajectories are generated and written into the execution record.

[0017] Furthermore, based on the two types of situations identified in the execution records—unregistered task occupancy and state switching intrusion into candidate execution periods—abnormal occupancy markers are generated and merged to form abnormal occurrence intervals. The task package matching table is backfilled according to the injected task package identifier, candidate resource combination identifier, and candidate execution period identifier. The matching status is updated synchronously, and a complete execution record is output for verification result comparison.

[0018] Furthermore, based on the same training execution sequence, the baseline image and reset image are driven to perform playback according to the same standard. The playback trajectory of the baseline image, the playback trajectory of the reset image, and the execution record are compared segment by segment according to a unified event sequence to form a verification result and locate the inconsistency position. The verification result includes consistent and inconsistent states.

[0019] Furthermore, based on the verification results, the task package matching table and scheduling rules are updated. In the consistent state, the candidate resource combination and candidate execution time period are maintained. In the inconsistent state, the candidate resource combination and candidate execution time period are reselected according to the inconsistent position. At the same time, the samples of this round are written into the model sample library, and the next round of training execution sequence is generated according to the updated task package matching table and scheduling rules.

[0020] The technical effects and advantages of the multi-dimensional specialized equipment simulation training resource scheduling method of the present invention are as follows:

[0021] This invention transforms training resource scheduling from a "queue and try your luck" approach to a closed-loop process of "verifying status first, matching time slots, verifying during execution, and correcting based on results." The injected task package undergoes dual screening of resource status and dependencies before execution, continuously verifies and records abnormal usage during execution, and corrects errors through dual-mirror playback after execution. Therefore, training arrangements no longer rely on temporary adjustments based on human experience. On-site problems such as resource conflicts, time slot congestion, and execution interruptions are more easily avoided and accurately located in advance, resulting in a more coherent training process and more controllable training organization costs.

[0022] This invention goes beyond a single successful scheduling attempt; rather, it ensures that the results of each round of execution are distilled into reusable criteria, driving the generation of the next sequence to better align with the actual training order. The resulting scheduling mechanism retains rule interpretability while possessing continuous correction capabilities, enabling the same batch of specialized equipment to maintain a stable resource supply rhythm under different training task densities, reducing ineffective waiting and redundant switching, and improving the efficiency of training resource utilization and the fulfillment of training plans. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating a multi-dimensional professional equipment simulation training resource scheduling method according to the present 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 embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Example 1: Figure 1 This invention presents a multi-dimensional specialized equipment simulation training resource scheduling method, comprising:

[0026] S1: Read the scheduling console operation log, node heartbeat, and task registration form according to the training unit, complete the source alignment and record merging, map the original state to the available state, occupied state, and switched state, and generate a resource state set after resolving concurrent conflicts.

[0027] S2: Read the resource requirement relationship and execution dependency relationship of the injected task package, generate candidate resource combinations and candidate execution time periods based on the resource state set, calculate the availability continuity of the same window and the neighborhood conflict density, obtain the steady-state matching probability and form a task package matching table.

[0028] S3: Output the training execution sequence and issue scheduling instructions according to the task package matching table. Before execution, verify the consistency between the candidate resource combination and the current resource status. During execution, collect status snapshots and change trajectories. After identifying abnormal occupancy, backfill the task package matching table.

[0029] S4: Based on the training execution sequence, drive the baseline image and reset image to perform standard playback, compare the playback trajectory with the execution record, output the verification result, correct the task package matching table and scheduling rules according to the verification result, and generate the next round of training execution sequence.

[0030] The present invention aims to transform the training scheduling of specialized equipment from a one-off process into a continuous scheduling process that is verifiable, replayable, and correctable. First, it aggregates the scheduling console's operation logs, node heartbeats, and task registration forms to create a unified resource status set and resource base map, thus resolving the inconsistency in resource status definitions. Then, by combining the resource demand relationships and execution dependencies of the injected task packages, it generates candidate resource combinations and candidate execution time periods. Finally, it uses steady-state matching probabilities to form a task package matching table and outputs the training execution sequence, freeing the scheduling results from reliance on manual experience.

[0031] During the execution phase, the consistency between the candidate resource combination and the current resource status is first verified. Then, a status snapshot and change trajectory are collected to form an execution record, and abnormal occupancy markers are backfilled into the task package matching table. After execution is completed, the baseline image and reset image are replayed according to the same standard, and the verification results are obtained by comparing them with the execution record. If they are consistent, the current rules are retained; if they are inconsistent, the task package matching table and scheduling rules are corrected according to the inconsistency. At the same time, the samples are written into the model sample library to generate the next round of training execution sequence. Through this path, training scheduling realizes a complete solution from availability to controllability, from execution to review, and from review to continuous improvement.

[0032] When training tasks enter the scheduling phase concurrently, the visible state of resources is often recorded separately by the scheduling console operation log, node heartbeat, and task registration form. The recording criteria and update time are not consistent. Directly using these for scheduling can easily lead to conflicts where multiple states of the same resource instance coexist. Therefore, it is necessary to first align and merge the source records into a unified state expression, and then establish a resource view that can be directly read, in order to ensure that the scheduling entry information has consistency and decidability.

[0033] S101 Data Extraction and State Semantic Mapping

[0034] To ensure consistency in source data, the dispatch console operation logs, node heartbeats, and task registration forms utilize the publicly available data acquisition links, without repeating details of the acquisition hardware. The current segment only supplements the state semantic processing procedure. Each source record is first bound to a training unit identifier and a resource instance identifier, then a source record sequence is formed according to the event order, followed by state semantic mapping. State semantic mapping only allows output of four types of results: available state, occupied state, switching state, and empty observation. Available state corresponds to idle completion and running ready actions; occupied state corresponds to registered occupation and executed occupation actions; switching state corresponds to start switching and stop switching actions; and empty observation indicates that the current source did not effectively report in the current event. The range of source state observation values ​​is fixed to four discrete values, and all subsequent judgments are based on this fixed set of values.

[0035] S102 Event Alignment and Record Merging

[0036] To ensure comparisons can be made at the same event location from different sources, the runtime log event sequence, heartbeat event sequence, and registration form event sequence are first merged into a unified event sequence, and then numbered in natural order to form an event sequence chain. For each event sequence number, forward piling is performed. The processing order is as follows: first, check if the source has a valid state in the current event; if a valid state exists, it is directly written into the aligned state observation; if no valid state exists, it inherits the most recent valid state from the previous event of the same source; if the previous event has no valid state, an empty observation is retained. The merged result adopts a three-source parallel record structure, including runtime log aligned state observations, node heartbeat aligned state observations, and task registration form aligned state observations. The value range of the aligned state observations follows the four discrete value categories, and the event sequence number range is the natural order number in the unified event sequence. The merged record output here already meets the conditions for in-place comparison, and step S103 can directly perform conflict resolution.

[0037] S103 State conflict resolution and state start record generation.

[0038] To ensure the current state conforms to the business action chain and is resistant to momentary fluctuations, conflict resolution employs a combined state transition legality verification algorithm and a lexicographical order priority comparison algorithm. For each event sequence number, a state support count is first calculated for each of the three candidate states. This count represents the number of sources that match the candidate state, ranging from none to all sources. Then, transition legality verification is performed between the confirmed state of the previous event and the current candidate state. The transition legality flag is either legal or illegal, with the legality relationship determined by the registration rules and the action semantic table. The comparison order is fixed: first, compare the transition legality flag; then, compare the state support count; then, compare the source confirmation order, which is fixed as: task registration form first, then dispatch console operation log, then node heartbeat. If the above comparisons still do not yield a result, the confirmed state of the previous event remains unchanged. A state start record is generated synchronously, and the processing order is: compare the current confirmed state with the confirmed state of the previous event. If the state changes, the current event sequence number is written as the state start record; otherwise, the state start record of the previous event is inherited.

[0039] Example: When an injection task is initiated, the task registration form initially shows an occupied state, while the runtime log briefly reports an available state, and the node heartbeat reports a switched state. Conflict resolution first checks the legality flags of the transition from the previously confirmed state to the current candidate state, then compares the state support counts, and finally determines the current confirmed state as occupied based on the source confirmation order. The state start record is written to the current event sequence number. In practice, after a resource instance enters the occupied state, it is no longer rewritten by instantaneous heartbeat jitter, and the scheduling action remains consistent with the registration action.

[0040] S104 Resource Status Set and Resource Base Map Output.

[0041] To ensure that step S2 can directly read stable input, each resource instance under each training unit extracts its current state from the end of a unified event sequence and combines it with the previous event to form a recent state change record. First, the sequence number of the end event is located, then the confirmed state of the end event is read as the current state, and then the confirmed state of the previous event is compared with the confirmed state of the end event to generate a recent state change record. If a resource instance has only a single event, the recent state change record is written as the initial entry into the current state. The resource state set record fields are fixed as resource instance identifier, current state, training unit identifier, state start record, and recent state change record. The resource base map uses a mapping structure from training unit identifier to resource instance identifier to current state. With step S1 completed, step S2 can generate candidate resource combinations and candidate execution periods by reading the current state and state start record from the resource state set. The cross-step call relationship is clear and there is no terminology drift.

[0042] Step S1 completes the normalization transformation from the original records to the underlying scheduling data, outputting a resource state set and a resource base map. The fields are fixed as resource instance identifier, current state, training unit identifier, state start record, and most recent state change record. Resource conflicts have been resolved before entering the scheduling process, and the resource availability boundary and state start point are clearly marked.

[0043] Once the resource status is unified, the scheduling problem changes from "seeing resources" to "placing resources in the right tasks and at the right time periods". The injection of task packages is constrained by both resource demand and execution dependency relationships. Simply using the first-come-first-served approach will introduce time period overlap and insufficient continuity. Therefore, it is necessary to establish a common decision link between candidate resource combinations, candidate execution time periods and stability assessment.

[0044] Step S2 selects the continuity of available resources and the density of neighborhood conflicts for comprehensive analysis because these two parameters characterize the two most critical and complementary dimensions of interference judgment: whether the resources are continuously executable within the current candidate execution period and whether abnormal occupancy in adjacent training intervals is clustered. The former reflects the stability of on-site executability, while the latter reflects the risk of external interference propagation. The combination of the two can fully cover the dominant path of interference formation, so that the interference confidence coefficient maintains both the sensitivity to the identification of sudden conflicts and the stability of the judgment of short-term fluctuations, while reducing misjudgment and model drift caused by parameter coupling, and improving the interpretability, consistency of verification and scheduling implementation of the results.

[0045] S201 candidate resource portfolio construction.

[0046] Step S1 has already provided the resource status set and resource base map. Step S2 will not repeat the source collection action. Instead, it will directly read the resource instance identifier, current status, training unit identifier, and recent status change record from the resource status set, and read the resource requirement relationship and execution dependency relationship from the injected task package.

[0047] First, a resource role list is established based on resource demand relationships. Then, resource instances with consistent roles are filtered out from the resource base map. The filtering results are judged based on the occupancy boundary. Subsequently, a candidate resource combination list is assembled. The list fields are fixed as follows: injection task package identifier, candidate resource combination identifier, resource instance sequence, occupancy boundary marker, and role consistency marker. The occupancy boundary marker has a value range of pass / fail, and the role consistency marker has a value range of consistent / inconsistent.

[0048] Subsequently, a unified status metric is executed. Resource instances entering the candidate resource combination list retain only the available state. Status transitions are only permitted as boundary observations and do not participate in the determination of the execution time period start point. After standardization, candidate resource combinations are output. The output object name remains the candidate resource combination for direct use by S202.

[0049] S202 Candidate Execution Period Extraction.

[0050] Candidate resource combinations only reflect resource composition. Training execution sequences also require time period information. Time period extraction must be placed simultaneously with execution dependencies to avoid records where the order is satisfied but the time period conflicts occur.

[0051] For each candidate resource combination, scan point by point along a unified event sequence. The scanning action synchronously checks the availability status according to the resource instance sequence. Once all resource instances are simultaneously in an available state, a time period start is established. The time period end is established when any resource instance leaves the available state, thus obtaining the candidate execution time period. Each candidate execution time period is written with a start event position and an end event position, and a time period integrity flag is written. The time period integrity flag has a value range of complete and interrupted.

[0052] Next, filter candidate execution periods by execution dependency. The filtering rule is fixed: the end of the preceding injected task package's period must precede the start of the following injected task package's period. Candidate execution periods that do not satisfy the relationship are directly eliminated, and the results are output as a list of candidate execution periods. The output object name remains the candidate execution period, which is used by S203 to calculate two parameters.

[0053] S203 Two parameters are calculated.

[0054] The continuity of available segments is calculated using normalized continuous proportion logic. First, all event positions within the candidate execution time period are read to form the total time period length. Then, continuous available segments are extracted within the time period, and the longest continuous available segment is selected. Finally, the longest continuous available segment is divided by the total time period length to obtain the continuity of available segments. The value range is the normalized proportion interval.

[0055] Using the start and end event positions of the candidate execution period as the central boundary, records with resource instances intersecting with the candidate resource combination are selected from the historical execution records of the same training unit. The immediately preceding and immediately following training intervals are selected in chronological order, and intervals overlapping with or touching the candidate execution period are merged to form the neighborhood training interval. Then, abnormal occupancy markers are extracted from the historical execution records corresponding to the neighborhood training intervals. These markers are triggered by non-registered task occupancy events or state switching intrusion events. The marker fields are fixed as resource instance identifier, abnormal occurrence interval, affected injected task package identifier, trigger type, and evidence source identifier. After deduplicating the abnormal occupancy markers by resource instance identifier, abnormal occurrence interval, and affected injected task package identifier, the number of deduplicated markers is counted against the total number of records in the neighborhood training interval. When the total number of records in the neighborhood training interval is not empty, the ratio of the two is recorded as the neighborhood conflict density, maintaining a normalized ratio value.

[0056] When the total number of records in the neighborhood training interval is empty, the median value of the conflict ratio is first extracted from the historical sequence of the same training unit and the same resource role as the neighborhood conflict density filling value, and then written into the neighborhood missing sample marker; when there are still no historical samples of the same training unit and the same resource role, the neighborhood conflict density is recorded as zero and the neighborhood missing sample marker is retained. The output object is kept as the available continuity and neighborhood conflict density of the same window, for S204 to call.

[0057] Example: In the same training unit, two candidate resource combinations emerge. The first candidate resource combination remains available throughout the candidate execution period, while the second candidate resource combination experiences a brief interruption in the middle of the period. The availability continuity of the first candidate resource combination within the same window is higher than that of the second candidate resource combination. Within the neighborhood training interval, the first candidate resource combination has fewer associated abnormal occupancy markers, while the second candidate resource combination has more associated abnormal occupancy markers. The neighborhood conflict density of the first candidate resource combination is lower than that of the second candidate resource combination.

[0058] S204 steady-state matching probability calculation.

[0059] The Gaussian Naive Bayes model is constructed using candidate execution time periods as sample units. Sample sources are limited to historical execution records and historical verification results from the same training unit and resource role. For each sample, continuity and neighborhood conflict density are used as input features, and consistent and inconsistent states from historical verification results are used as labels. Sample preprocessing first removes duplicates based on the injected task package identifier, candidate resource combination identifier, and candidate execution time period identifier, then removes samples with missing fields, reversed time sequences, and conflicting labels. Subsequently, the training and validation sets are divided chronologically, and stratified sampling is performed within the training set to maintain coverage of both label classes. In the parameter estimation stage, the prior probabilities of the two labels are calculated and a prior smoothing parameter is added. Then, the mean and dispersion parameters of the two-dimensional features under the two labels are estimated, with a lower limit set for the dispersion parameter to avoid numerical instability. In the parameter optimization stage, grid search is performed on the prior smoothing parameter, the lower limit parameter of dispersion, and the decision threshold parameter. The decision stability and probability calibration consistency on the validation set are used as joint selection criteria to determine and solidify the model version. During the inference phase, the available continuity and neighborhood conflict density of each candidate execution period are read. The conditional probability density of the consistent and inconsistent states is calculated separately, and then multiplied by the corresponding prior probabilities to form two types of posterior values. After normalization, the steady-state matching probability is obtained. The steady-state matching probabilities are written into the sorting list from high to low. When probabilities are tied, they are resolved and tied in order of execution dependency priority and the order of the most recent state change record. In the neighborhood missing sample scenario, the neighborhood conflict density is generated according to the defined missing sample caliber and then enters the same inference process. The final output object is maintained as the steady-state matching probability sorting list and is called by S205.

[0060] Example: The injection task packages are sequentially dependent. The preceding injection task package and the following injection task package each have multiple candidate execution time periods. After inference, the preceding injection task package obtains a higher steady-state matching probability in one time period, and the following injection task package obtains a higher steady-state matching probability in another time period. The sorting list forms a clear priority order accordingly.

[0061] S205 Task Package Matching Table Generation and Training Execution Sequence Output.

[0062] The sorting results only reflect local priority relationships. The training execution sequence also needs to handle resource overlap caused by the concurrency of multiple injected task packages. The handling actions need to be clear and fixed.

[0063] First, within a single injection task package, select the candidate resource combination and candidate execution time period with the highest steady-state matching probability and satisfying execution dependencies. Then, perform resource conflict resolution across multiple injection task packages. The conflict resolution order is fixed as follows: first, execute dependencies; second, sort by steady-state matching probability; and third, the order of most recent state change records. After conflict resolution is completed, write the results to the task package matching table. The table fields are fixed as follows: injection task package identifier, candidate resource combination identifier, candidate execution time period identifier, continuity of availability within the same window, neighborhood conflict density, steady-state matching probability, and matching status.

[0064] Finally, the task package matching table is topologically sorted according to execution dependencies to generate a training execution sequence. When a loop occurs in the execution dependency, it is broken according to the priority of the execution dependency and the order of registration of the injected task packages before sorting. The fields of the training execution sequence are fixed as the injected task package identifier, execution order, candidate resource combination identifier, and candidate execution time period identifier. The output object name remains the training execution sequence, which is directly read and executed in step S3.

[0065] Example: If the pre-injection task package and the post-injection task package overlap on the same resource instance within the training unit, the resource conflict resolution first retains the pre-injection task package, and then selects a non-overlapping time period from the candidate time period of the post-injection task package. The task package matching table forms two conflict-free records, and the training execution sequence is output in the order of dependency.

[0066] Step S2 takes the resource state set as input, generates candidate resource combinations and candidate execution time periods, calculates the availability continuity and neighborhood conflict density of the same window, and obtains the steady-state matching probability through Gaussian Naive Bayes. Finally, a task package matching table and training execution sequence are formed, and task orchestration is transformed from empirical selection to interpretable and comparable probabilistic matching.

[0067] Completing the schedule does not guarantee reliable execution. In training sessions, situations such as unregistered tasks occupying the space, state switching intruding into the execution window, and temporary state jitter often occur. Without process traceability and anomaly localization, the correctness of the schedule cannot be verified, and the verification results cannot be fed back into the matching strategy. Therefore, verification, recording, and anomaly marking must be incorporated into the same action chain during the execution phase.

[0068] S301 performs gate verification and position activation.

[0069] The training execution sequence has given the execution position, injection task package identifier, candidate resource combination identifier, and candidate execution time period identifier. This sub-step first converts the static orchestration into a deployable action. Only after the gating verification is passed will the action be issued. If the gating verification fails, the execution position will enter the waiting queue.

[0070] The gating verification process employs a rule-based gating algorithm, first performing resource consistency checks, then time-period consistency checks, and finally a merge check. Resource consistency checks read the current resource status item by item for each candidate resource combination; a pass is recorded when all resource instances are available, and a fail is recorded when any resource instance is unavailable. Time-period consistency checks compare the current event position with the candidate execution time period boundary; a pass is recorded when the event position falls within the candidate execution time period, and a fail is recorded when the event position does not fall within the candidate execution time period. The merge check uses an AND logic; a consistency verification flag is marked as passed if both resource and time-period consistency checks pass simultaneously, and marked as failed otherwise. The value range for resource consistency checks is pass or fail, the value range for time-period consistency checks is pass or fail, and the value range for consistency verification flags is pass or fail.

[0071] S302 Status Snapshot Acquisition and Execution Record Archiving.

[0072] After the gate verification is passed, the execution action enters the operation phase. The operation phase needs to form a verifiable evidence chain. Under the premise of not repeating the details of the source collection in step S1, this sub-step writes the operation log, node heartbeat, and task execution receipt into the status snapshot. Among them, the task execution receipt comes from the scheduling instruction execution link receipt queue and is associated with the execution position according to the injected task package identifier.

[0073] State snapshot acquisition employs an event-driven sampling algorithm, writing a state snapshot at each event location. The state snapshot fields are fixed as follows: injection task package identifier, candidate resource combination identifier, event location, and the current resource status of each resource instance within the candidate resource combination. The current resource status of each resource instance remains within the range of available, occupied, and switched states. Execution record archiving uses a structured archiving algorithm, with the archive primary key fixed as the training unit identifier, injection task package identifier, and execution position. State snapshots are written to the execution records in the order of event location.

[0074] After archiving is completed, the execution records and training execution sequences form a one-to-one correspondence. Step S303 can directly calculate the change trajectory based on the execution records, with the field names remaining unchanged and the reading path clearly defined.

[0075] S303 change trajectory generation.

[0076] The state snapshots have been sorted by event location. This sub-step writes the discrete states into a continuous range of change, forming a locatable change trajectory.

[0077] The change trajectory generation employs an adjacent difference algorithm and a continuous segment merging algorithm. The adjacent difference algorithm first determines the initialization rule for the first event position, which is defaulted to "unchanged." Then, it compares the current state with the previous event state for each resource instance; different states are recorded as changed, and the same states are recorded as "unchanged." The continuous segment merging algorithm combines consecutive changed event positions into change intervals, recording the start and end event positions for each interval. Finally, the resource instance-level change trajectory is output and written to the execution record.

[0078] Example: Within a certain execution position, the candidate resource combination corresponding to the injected task package contains two resource instances. The status of the first resource instance changes from available to occupied and then back to available. The second resource instance always remains available. The change trajectory results show that the first resource instance forms two change intervals, while the second resource instance has no change interval. The on-site actions are consistent with the recorded results.

[0079] S304 Abnormal Occupation Detection and Abnormal Occupation Marker Generation.

[0080] The change trajectory has been formed. This sub-step identifies abnormal occupancy on the same event sequence and generates abnormal occupancy markers. The detection process maintains dual-condition gating logic to avoid misjudgment based on a single condition.

[0081] Anomaly occupancy detection employs a dual-condition gating algorithm. Condition 1 is a non-registered task occupancy determination, based on the condition that the actual occupant task identifier exists and is not equal to the currently injected task package identifier. Condition 2 is a state switching intrusion determination, based on the condition that the resource instance's current resource state is in a switched state and the event location falls within the candidate execution time period. Anomaly occupancy markers are generated using an OR logic: if either condition 1 or condition 2 is true, the anomaly occupancy marker is recorded as existing; if neither condition is true, the anomaly occupancy marker is recorded as non-existent. The anomaly occupancy marker value ranges from existing to non-existent.

[0082] The actual task occupancy identifier is read from the occupancy record item in the task execution receipt, without repeatedly expanding the receipt collection link. After the abnormal occupancy marker is generated, continuous segments are merged according to the event location, the abnormal occurrence interval is output and written to the execution record.

[0083] Example: During execution, a resource instance is temporarily occupied by another injected task package, and the non-registered task occupation determination is established. At the same time, another resource instance enters the switching state and is within the candidate execution period, and the switching state intrusion determination is established. The abnormal occupation mark is recorded as existing on both resource instances, and the abnormal occurrence interval is fully recorded.

[0084] S305 Task Package Matching Table Backfilling and Step Output.

[0085] An abnormal occupancy marker has been formed. This sub-step completes the task package matching table backfilling and outputs the directly readable object from step S4, ensuring that the same key can be used for comparison during the verification phase.

[0086] The backfilling action uses a row key positioning rule, with the row key fixed as the injection task package identifier, candidate resource combination identifier, and candidate execution time period identifier. The backfilling fields are fixed as the abnormal occupancy flag set, the abnormal occurrence interval set, and the matching status. The matching status value range is fixed as "reserved" and "pending correction." If an abnormal occupancy flag exists, the matching status is recorded as "pending correction"; if the abnormal occupancy flag does not exist, the matching status is recorded as "reserved."

[0087] The execution record output fields are fixed as archive primary key, status snapshot sequence, change trajectory set, abnormal occupation mark set, and abnormal occurrence interval set. After the task package matching table and execution record are output synchronously, step S4 directly reads the two types of objects to complete the baseline image and reset image standard playback verification. The steps are completely connected and the fields are consistent.

[0088] Step S3 involves completing consistency verification and instruction issuance around the training execution sequence, continuously collecting status snapshots and generating change trajectories to form structured execution records. At the same time, it identifies and backfills abnormal occupancy markers into the task package matching table, establishing a one-to-one correspondence between execution facts and scheduling entries, and enabling the abnormal locations and affected task packages to be clearly tracked.

[0089] The execution log already contains event-level evidence, but whether the scheduling strategy is valid still needs to be determined through repeatable comparison. Single trajectory review is easily affected by occasional factors on site. Therefore, by using baseline mirror and reset mirror to play back under the same standard conditions and comparing them segment by segment with the execution log, a stable and consistent conclusion and an operable correction rule can be obtained.

[0090] S401 standard playback trajectory construction.

[0091] The execution log already provides the sequence of events and the facts of the state. For mirror playback to be comparable, the same input sequence must be used. First, the inputs must be standardized, then the output structure must be standardized, so that the on-site records and mirror results can be compared under the same coordinate system.

[0092] First, the injection task package identifier, candidate resource combination identifier, and candidate execution time period identifier are read from the training execution sequence. Then, the event sequence is read from the execution record to form a replay event sequence. The baseline image and reset image generate image replay status event by event along the same replay event sequence. The image replay status value range is fixed as available, occupied, and switched. When there is a gap in the replay event sequence, the gap position is filled along the previous event state and a replay complete flag is written. When the gap occurs at the first event position, it is filled along the first valid state of the execution record. If there is still no valid state, a switched state is written and marked as incomplete. The replay complete flag value range is fixed as complete and incomplete. The output objects are fixed as the baseline image replay trajectory, the reset image replay trajectory, and the replay complete flag.

[0093] S402 Three-Track Alignment and Consistency Indicator Generation.

[0094] The consistency conclusion is only meaningful for verification when the mirror playback trajectory and the execution record trajectory are compared item by item at the same event location. First, align the three trajectories, and then add point-by-point consistency indicators to make the logical chain tighter.

[0095] First, obtain the common event sequence of the baseline mirror playback trajectory, reset mirror playback trajectory, and execution record trajectory. If the common event sequence is not empty, proceed directly to point-by-point comparison. If the common event sequence is empty, use the execution record event sequence as the benchmark, project the two mirror trajectories onto the benchmark sequence according to the nearest neighbor event positions, and fill in the positions where projection fails along the previous valid state and write an alignment padding mark; if the first event position fails to project, fill in along the first valid state of the execution record. Perform consistency indicator generation for each resource instance and each event position. If the baseline mirror playback state is consistent with the execution record state and the reset mirror playback state is consistent with the execution record state, it is recorded as consistent; otherwise, it is recorded as inconsistent. The value range of the consistency indicator is fixed as consistent and inconsistent. The output objects are fixed as the consistency indicator sequence and the inconsistency candidate position sequence.

[0096] S403 segment-by-segment comparison and verification results generation.

[0097] The sequence of candidate inconsistencies has been obtained. The next step is to convert these discrete inconsistencies into segment-level conclusions, facilitating positional adjustments to the task packet matching table and scheduling rules. The segmentation rules must be fixed for the conclusions to be stable.

[0098] First, read the change trajectory and anomaly occurrence interval from the execution record, merge them to obtain the segment boundary, and then divide the sequence into segments according to the segment boundary. When segments conflict, they are handled according to a fixed priority order, with anomaly occurrence intervals taking precedence over change trajectories. Within each segment, first count the number of comparisons indicating consistency, then count the total number of comparisons within the segment, and finally use the number of consistent comparisons and the total number of comparisons to form the segment consistency degree. The segment consistency degree value range is fixed as a percentage range. Next, read the training unit-level consistency threshold from the scheduling rules; if the consistency threshold is not configured, read the default threshold of the training unit. Compare the consistency degree of each segment with the consistency threshold. If all segments meet the threshold, the verification result is recorded as consistent; if any segment does not meet the threshold, the verification result is recorded as inconsistent. The verification result value range is fixed as consistent and inconsistent. Inconsistent positions are merged and output from the inconsistent candidate position sequence by segment.

[0099] Example: In a replay, a state switching intrusion occurs within the candidate execution period. The interval where the anomaly occurs is first used as the segment boundary cutting trajectory. After segment comparison, the consistency of the relevant segments is lower than the consistency threshold. The verification result is recorded as inconsistent, and the inconsistent position is located to the corresponding event interval of the affected resource instance.

[0100] S404 Task Package Matching Table and Scheduling Rules Correction.

[0101] The verification results are clear. The correction action needs to directly map the inconsistent positions to the task package matching table rows, and then synchronously correct the scheduling rule sorting priority relationship to avoid hitting the same type of conflict in the next round.

[0102] When the verification results are consistent, the task package matching table retains the original candidate resource combinations and candidate execution time periods, and the matching status is marked as "reserved". When the verification results are inconsistent, the task package matching table is jointly corrected according to the inconsistent position and the abnormal occupancy flag. The correction content is fixed as candidate resource combination replacement, candidate execution time period reselection, and matching status rewriting. The matching status value range is fixed as "reserved" and "pending correction". When multiple task packages conflict concurrently, the conflict resolution order is fixed as execution dependency priority, resource exclusivity relationship second, and most recent status change record third. Subsequently, the sorting priority relationship in the scheduling rules is updated according to the corrected task package matching table, and the updated fields are fixed as task package position priority and candidate execution time period priority.

[0103] Example: Two injected task package records simultaneously contend for the same resource instance, and one of the records has an inconsistent position in the replay segment. The conflict resolution order first ensures the execution dependency, then rewrites the inconsistent record as pending correction and reselects a candidate execution time period. The scheduling rules adjust the priority accordingly, and the conflicting records no longer overlap.

[0104] S405 model sample library write-back and next round of training execution sequence generation.

[0105] If the correction results are not included in the model sample library, the steady-state matching probability will not absorb the latest verification facts. Sample write-back and sequence regeneration must be completed in the same closed loop for the correction effect to be reflected in the next scheduling round.

[0106] First, the available continuity and neighborhood conflict density of the same window are read from the task package matching table. Then, the consistent labels are read from the verification results, and these are combined to form the write-back samples. The fields of the write-back samples are fixed as available continuity of the same window, neighborhood conflict density, and consistent labels. The sample deduplication key is fixed as the injected task package identifier, candidate resource combination identifier, candidate execution time period identifier, and verification round identifier. Duplicate samples are overwritten and updated. After the model sample library is written back, the dependency constraint sorting is performed according to the revised task package matching table and the revised scheduling rules, and the next round of training execution sequence is output. The output object name remains the same as the next round of training execution sequence.

[0107] Step S4 outputs the verification results and inconsistency locations, and corrects the task package matching table and scheduling rules accordingly. At the same time, the samples of this round are written into the model sample library to form the next round of training execution sequence. The scheduling mechanism is changed from one-time scheduling to a closed operation mode of continuous verification, continuous correction and continuous convergence.

[0108] Specifically, the above are merely preferred embodiments of this application and are not intended to limit this application.

[0109] The thresholds and preset parameters can be pre-calibrated through offline simulation testing or set to fixed values ​​according to on-site operating procedures.

[0110] In the description of this specification, references to terms such as "an embodiment," "example," and "specific example" indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0111] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for scheduling resources in multi-dimensional professional equipment simulation training, characterized in that, Including the following steps: S1: Read the scheduling console operation log, node heartbeat, and task registration form according to the training unit, complete the source alignment and record merging, map the original state to the available state, occupied state, and switched state, and generate a resource state set after resolving concurrent conflicts; S2: Read the resource requirement relationships and execution dependencies of the injected task package, and generate candidate resource combinations and candidate execution periods based on the resource state set; Read all event locations within the candidate execution period to form the total length of the period. Extract continuous available segments within the candidate execution period. Select the maximum length among the continuous available segments and divide the maximum length by the total length of the period to obtain the available continuity of the window, which represents the level of continuous execution readiness within the period. Using the start and end event positions of the candidate execution time period as the central boundary, records that intersect with the candidate resource combination in the historical execution records of the same training unit are selected. The adjacent previous training interval and the adjacent subsequent training interval are selected in chronological order, and the intervals that overlap with or touch the beginning and end of the candidate execution time period are merged into the neighboring training interval. Subsequently, abnormal occupancy markers are extracted from the historical execution records corresponding to the neighboring training interval. After deduplication by resource instance identifier, abnormal occurrence interval, and affected injection task package identifier, the number of deduplicated markers is counted and compared with the total number of records in the neighboring training interval. When the total number of records in the neighboring training interval is not empty, the ratio of the two is used as the neighborhood conflict density to characterize the aggregation level of abnormal occupancy markers in the historical execution records. When the total number of records in the neighborhood training interval is empty, the median value of the conflict ratio is first extracted from the historical sequence of the same training unit and the same resource role as the neighborhood conflict density filling value. When there are still no historical samples of the same training unit and the same resource role, the neighborhood conflict density is recorded as zero. The available continuity of the same window and the neighborhood conflict density are used as two parameters to input Gaussian Naive Bayes, and the conditional probability density of the consistent state and the inconsistent state is calculated respectively. Then, it is multiplied with the corresponding prior probability to form two types of posterior values. After normalization, the steady-state matching probability is obtained. The matching status in the task package matching table is determined according to the steady-state matching probability and execution dependency. The task package matching table uniformly records the injected task package identifier, candidate resource combination identifier, candidate execution time period identifier, window availability continuity, neighborhood conflict density, steady-state matching probability and matching status, and generates training execution sequences according to resource conflict resolution rules. S3: Output the training execution sequence and issue scheduling instructions according to the task package matching table. Before execution, verify the consistency between the candidate resource combination and the current resource status. During execution, collect status snapshots and change trajectories. After identifying abnormal occupancy, backfill the task package matching table. S4: Based on the training execution sequence, drive the baseline image and reset image to perform standard playback, compare the playback trajectory with the execution record, output the verification result, correct the task package matching table and scheduling rules according to the verification result, and generate the next round of training execution sequence.

2. The multi-dimensional specialized equipment simulation training resource scheduling method according to claim 1, characterized in that, Step S1 includes: Based on the training unit identifier and resource instance identifier, the scheduling console operation log, node heartbeat, and task registration form are collected, and the unified source status observation is mapped to available status, occupied status, and switched status. Based on the unified event sequence, the aligned status observation and conflict resolution are completed, and the current status and the status start record are determined.

3. The multi-dimensional specialized equipment simulation training resource scheduling method according to claim 2, characterized in that, Step S1 also includes: A resource state set is generated based on the current state, the state start record, the training unit identifier, and the resource instance identifier. The resource state set is synchronously written to the most recent state change record and forms a resource base map. The resource base map maintains a unique mapping from the training unit identifier and the resource instance identifier to the current state, which is used by the task package matching table.

4. The multi-dimensional specialized equipment simulation training resource scheduling method according to claim 3, characterized in that, Step S2 includes: Based on the resource state set and resource base map, the resource demand relationship and execution dependency relationship of the injected task package are analyzed. Candidate resource combinations are formed according to the consistency of resource roles and the occupancy boundary. Candidate execution periods are extracted from the unified state sequence and written into the complete period marker.

5. The multi-dimensional specialized equipment simulation training resource scheduling method according to claim 4, characterized in that, Step S3 also includes: The scheduling instructions are triggered one by one according to the training execution sequence. Before scheduling, the consistency verification between the candidate resource combination in the task package matching table and the current resource status is performed. Only after the consistency verification is passed can the execution begin. During the execution phase, status snapshots are continuously collected and archived in the order of events. Based on adjacent status snapshots, the status change segments are extracted and the change trajectory is generated and written into the execution record.

6. The multi-dimensional specialized equipment simulation training resource scheduling method according to claim 5, characterized in that, Step S3 also includes: The system identifies two scenarios based on execution records: unregistered tasks occupying the execution space and switching states intruding into candidate execution periods. It generates abnormal occupation markers and merges them to form abnormal occurrence intervals. The system then backfills the task package matching table according to the injected task package identifier, candidate resource combination identifier, and candidate execution period identifier. The system also updates the matching status synchronously and outputs complete execution records for verification results comparison.

7. The multi-dimensional specialized equipment simulation training resource scheduling method according to claim 6, characterized in that, Step S4 includes: Based on the same training execution sequence, the baseline image and reset image are driven to perform playback according to the same standard. The playback trajectory of the baseline image, the playback trajectory of the reset image, and the execution record are compared segment by segment according to a unified event sequence to form a verification result and locate the inconsistency position. The verification result includes consistent and inconsistent states.

8. The multi-dimensional specialized equipment simulation training resource scheduling method according to claim 7, characterized in that, Step S4 also includes: Based on the verification results, update the task package matching table and scheduling rules. In the consistent state, maintain the candidate resource combination and candidate execution time period. In the inconsistent state, reselect the candidate resource combination and candidate execution time period according to the inconsistent position. At the same time, write the samples of this round into the model sample library, and generate the next round of training execution sequence according to the updated task package matching table and scheduling rules.

Images