Intelligent production monitoring management method and system for aviation ring forgings

Abstract

The invention relates to the technical field of manufacturing operation management, in particular to an intelligent production monitoring management method and system for aviation ring forgings. The method comprises the following steps: acquiring a micro-area grid and probe layout scheme and an equipment list, and carrying out coordinate and time anchor point binding and channel self-inspection and cleaning to form an available sequence and confidence label set and an influence factor candidate set; carrying out quantile adaptive threshold learning, probabilistic switching and rollback configuration, and carrying out causal impact reinforcement to obtain an adaptive threshold set and a version mirror image; realizing robust fusion and trigger item explanation of the environment, the equipment and the production line based on the time-varying weight and the consistency measurement, and outputting a disposal suggestion; multi-target scheduling, amplitude limiting and receipt registration are executed in combination with a production resource mirror image, and a closed loop from learning to execution to re-learning is constructed. The method is beneficial to improving the judgment accuracy, reducing false alarm and missing alarm, and supporting traceable auditing and stable scheduling.

Description

Technical Field

[0001] This invention relates to the field of manufacturing operations management technology, and in particular to an intelligent production monitoring and management method and system for aerospace ring forgings. Background Technology

[0002] In the field of manufacturing operations management, existing solutions for aerospace ring forging production typically involve a combined system of environmental monitoring, equipment operation monitoring, and process event recording around production line workstations, heating, and forming processes. This system employs a fixed-point deployment based on workstations or the entire line, rule-based threshold settings, and offline report analysis. However, this approach suffers from limitations such as restricted data collection range and spatial resolution, a lack of adaptive threshold settings, and a lack of closed-loop evidence for assessment and handling. Existing methods often rely on aggregating raw data from a single or limited source, focusing on static threshold determination and fixed evaluation formulas. In scenarios with frequent changes in operating condition labels, tight coupling of production line resources, and constraints of both maintenance windows and quality stability, difficulties arise in data alignment and channel health management, as well as arbitration challenges related to threshold drift and lack of evidence for probability switching. These issues make it difficult to meet the stable requirements for the formation and implementation of handling recommendations. To address the need for joint processing based on available sequences, adaptive threshold sets, and evaluation records combined with operational condition labels, existing technologies generally suffer from fragmented processes and insufficient evidence transmission in various stages, including: extraction of calibration segments and historical baselines; drift verification and homogeneous deduplication; robust denoising and minimum exposure screening of outlier segments; quantile adaptive threshold learning and causal impact detection; data consistency measurement and online estimation of time-varying weights; trigger identification and factor-level localization; causal graph construction and disposal suggestion assembly; multi-objective constraint assembly and conflict resolution; priority sequence solving and amplitude limiting configuration; execution list compilation and implementation; and effect feedback registration and feedback writing. This makes it difficult to establish a consistent process of collection-alignment-judgment-control-recording in application scenarios oriented towards production resource mirroring, resulting in inconsistencies in the transformation of disposal suggestions into execution lists and the accumulation of effect feedback. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides an intelligent production monitoring and management method for aerospace ring forgings, comprising: Obtain the micro-area grid, probe deployment scheme and equipment list, and perform coordinate binding, time anchor point registration, channel health self-check, drift verification, homogeneity deduplication, working condition label mapping, abnormal segment robust denoising and minimum exposure screening to generate the candidate set structure of impact factors; The structure of the candidate set of impact factors is obtained, and quantile adaptive threshold learning, threshold learning input set construction, probability switching, and backoff strategy configuration are performed. Item-level threshold candidates containing upper and lower boundaries, candidate positions, morphological labels and applicable working conditions are generated and aggregated into learning group-level threshold candidate units. Causal impact detection and boundary reinforcement processing are performed to generate an adaptive threshold set. Based on an adaptive threshold set, data consistency measurement, online estimation of time-varying weights, robust fusion, trigger item identification, factor-level localization and cause graph construction are performed to generate a disposal suggestion structure. Based on the proposed disposal structure, the system performs multi-objective constraint assembly, conflict resolution, priority sequence solving, amplitude limit configuration, execution list compilation, implementation, effect feedback registration and feedback writing, and generates an effect feedback structure.

[0004] Furthermore, the micro-area grid and probe deployment scheme and equipment list include: The micro-grid refers to dividing the production line area into stable regions with fixed boundaries and numbers according to heating stations, forming stations, heat treatment stations, bulging stations, straightening stations, grinding stations, and inspection stations; the probe deployment scheme refers to the written configuration of the installation positions, installation methods, numbering rules, and channel mapping relationships of environmental micro-area probes, equipment status probes, and process cycle probes; the equipment list refers to the list of all monitoring-related equipment names, equipment numbers, station affiliations, unit levels, maintenance record entry points, and available sensor channels within the production line.

[0005] Furthermore, the process of performing causal impact detection and boundary reinforcement also includes: Intervention markers are read from the execution records of the previous or adjacent cycles, and the corresponding sample intervals of the candidates are aligned with the intervention markers on the time axis. Parallel control groups with impact windows and non-impact windows are constructed based on the alignment relationship.

[0006] Furthermore, the boundary reinforcement process also includes: For items affected by the intervention, a guard band is applied to the upper and lower boundaries of the item-level candidates. The width of the guard band is related to the dispersion description and sample coverage within the learning group, and a constraint condition for preferentially triggering bypass backoff is set within the guard band.

[0007] Furthermore, the process of generating the adaptive threshold set also includes: The item-level hardening results are aggregated within the learning group to generate a learning group-level threshold item set. The references to the probability switching registration domain and rollback registration domain used in the current period are recorded at the set layer. For all learning groups and all impact factor items, a hardened threshold form set is formed. The source label, version label, impact registration domain, temporary output registration domain, and protection band parameter description are recorded at the set layer to construct an adaptive threshold set. At the same time, a threshold version mirror structure is generated. The snapshot of the threshold form, item source, form label, protection band parameter, rollback registration reference, and probability switching registration reference of all items in the current period is recorded by period as the organizational unit. The pointer of the previous period's mirror is written into the mirror to form a continuous spectrum.

[0008] Furthermore, the process of performing data consistency measurement, online estimation of time-varying weights, robust fusion, trigger identification, factor-level localization, and causal graph construction also includes: The correspondence between the aggregated sequence and the confidence interval is read from the evaluation record. When the aggregated sequence crosses the boundary between any window and the upper or lower boundary of the confidence interval, a trigger candidate is registered. When the boundary crossing relationship and the boundary hit record of the adaptive threshold entry appear in the same window and are in the same direction, the trigger candidate is upgraded to a trigger item, and a source label and window index are attached to the trigger item. Within the trigger window, the intra-domain fusion sequence and window-level weight snapshot are backtracked to retrieve the entry that contributes the most to cross-domain convergence within the window and has experienced a replacement event or weight jump in intra-domain fusion. Possible causal factors are registered at the item level.

[0009] Furthermore, the process of constructing the causal map also includes: A multi-layered node graph covering workstations, micro-areas, equipment, items, working condition labels, and threshold patterns is constructed. The edge types include trigger association, replacement association, fallback association, and boundary hit association. Each edge is accompanied by a trigger window, direction description, and source label. When multiple trigger items appear consecutively in adjacent windows within the same learning group, the cause graph establishes a time sequence among these trigger items and records the cluster identifier.

[0010] Furthermore, the process of generating the disposal recommendation structure includes: Based on the trigger items and the node and edge information in the graph, abnormal evidence elements are compiled, including the correspondence between trigger window snapshot, trigger item source label, item-level contribution description, replacement event summary, rollback registration reference and adaptive threshold form, and the above elements are organized into abnormal evidence packages; each package corresponds to a trigger item or a group of consecutive trigger items, and contains indexes and references for review.

[0011] Furthermore, the process of generating the abnormal evidence package and the proposed handling structure also includes: The candidate strategy modules include reducing workstation load, switching equipment operating status, adjusting cycle time parameters, requesting short-term maintenance, using backup machines, and limiting operating boundaries. The assembly logic of the candidate strategy modules reads the weight parameters, the source and version labels of the adaptive threshold entry protection band and the threshold version mirror, and forms a structured suggestion that can be read by the executable system.

[0012] Furthermore, an intelligent production monitoring and management system for aerospace ring forgings, applied to any of the methods described above, includes: The micro-area grid and probe deployment and calibration module is used to install environmental micro-area probes, equipment status probes and process cycle probes and complete coordinate binding, time anchor point registration and channel health self-check, and provide multi-source raw data with location confidence labels to the data acquisition and access module; The data acquisition and admission module is used to receive raw data from multiple sources and perform calibration fragment extraction, historical baseline association, drift verification, homogeneous deduplication, robust denoising of outlier segments and minimum exposure screening to generate usable sequences, confidence label sets and impact factor candidate sets. The threshold learning and version management module is used to perform quantile adaptive threshold learning, probability switching, and rollback strategy configuration based on available sequences and operating condition labels. The evaluation fusion module is used to perform data consistency measurement and online estimation of time-varying weights on the candidate set of impact factors and the adaptive threshold set under confidence label constraints. The anomaly interpretation and handling generation module is used to identify triggers, locate factors, and construct cause maps based on assessment records and confidence intervals, forming an anomaly evidence package and handling recommendations. The multi-objective constraint and conflict resolution module is used to receive disposal suggestions, threshold version images and production resource images, assembly output scale constraints, power consumption constraints, quality stability constraints, maintenance window constraints and safety barrier constraints, perform conflict resolution and generate priority sequences and equipment scheduling schemes, which are then provided to the safety barrier and execution programming module. The safety barrier and execution compilation module is used to extract execution unit and safety barrier parameters and priority sequences from the equipment scheduling scheme, complete priority sequence solving, limit configuration and execution list compilation, and output safety barrier configuration and execution list; The execution and receipt registration module is used to issue instructions, perform shadow execution and rollback calls based on the execution list, and register and write effect receipts to form effect receipts.

[0013] The key innovations of this invention include: (1) Around the threshold learning input set, through quantile adaptive threshold learning, probability switching and backoff strategy configuration, and causal impact detection and boundary reinforcement of threshold candidates, an adaptive threshold set and threshold version mirror are formed under the constraint of working condition label, realizing the integrated management of threshold form and version spectrum, involving the linkage processing of the influence factor candidate set, the available sequence and the threshold candidate set.

[0014] (2) Based on the evaluation input, under the constraints of the candidate set of impact factors, the adaptive threshold set, and the confidence label set, perform data consistency measurement and online estimation and robust fusion of time-varying weights, and link trigger item identification, factor-level localization and cause map construction to generate weight parameters, evaluation records, confidence intervals, abnormal evidence packages and disposal suggestions, and complete the link organization from item to intra-domain fusion to cross-domain interpretation.

[0015] (3) Based on the scheduling input, implement multi-objective constraint assembly, conflict resolution, priority sequence solution, limit configuration and execution list compilation based on the disposal suggestions, threshold version mirror and production resource mirror, and carry out implementation and effect feedback registration and feedback writing, output safety barrier configuration and execution list and effect feedback, and construct a closed-loop control link from judgment to control to recording.

[0016] The following are its main beneficial effects: (1) It is applied to the threshold governance and boundary management process. The adaptive threshold set and the threshold version mirror image make the threshold learning and rollback consistent with the source labeling and form switching record under the intervention context and working condition label constraints, forming a verifiable threshold boundary output, and providing a stable judgment basis for the evaluation record and multi-objective constraint assembly. It is applicable to scenarios such as strategy adjustment, maintenance operation and parameter switching.

[0017] (2) In the multi-source assessment and interpretation process, the data consistency measurement, the online estimation of time-varying weights, and the robust fusion complete the organization of entries within and across domains under the confidence label constraint. The abnormal evidence package and handling suggestions output by the trigger item identification, the factor-level localization, and the cause map construction provide structured input for subsequent multi-objective constraint assembly, and maintain the continuity and traceability of the assessment link under the conditions of working condition label changes and batch differences.

[0018] (3) It is applied to the scheduling and execution process. The multi-objective constraint assembly, conflict resolution, priority sequence solution, amplitude limit configuration, execution list compilation, and integrated handling suggestions and production resource mirroring, the landing, effect receipt registration, and receipt writing form a unified record of safety barrier configuration, execution list, and effect receipt. It provides a historical baseline that can be called for the subsequent update of the adaptive threshold set and weight parameters. It is suitable for the operating conditions where production line resource coupling and maintenance window coexist. Attached Figure Description

[0019] Figure 1 A flowchart illustrating an intelligent production monitoring and management method for aerospace ring forgings provided in this application embodiment. Figure 2 This is a structural block diagram of an intelligent production monitoring and management system for aerospace ring forgings provided in an embodiment of this application. Detailed Implementation

[0020] Example 1: Refer to Figure 1 This is a flowchart illustrating an intelligent production monitoring and management method for aerospace ring forgings provided in an embodiment of the present invention. The process may include at least steps S100-S400: S100: Obtain the micro-area grid, probe deployment scheme and equipment list, perform coordinate binding, time anchor point registration, channel health self-check, drift verification, homogeneous deduplication, working condition label mapping, abnormal segment robust denoising and minimum exposure screening, and generate the impact factor candidate set structure. S200: Obtain the candidate set of impact factors, perform quantile adaptive threshold learning, threshold learning input set construction, probability switching, and backoff strategy configuration, generate item-level threshold candidates containing upper and lower boundaries, candidate positions, morphological labels, and applicable working conditions, and aggregate them into learning group-level threshold candidate units, perform causal impact detection and boundary reinforcement processing, and generate adaptive threshold set and threshold version mirror structure. S300, based on an adaptive threshold set, performs data consistency measurement, online estimation of time-varying weights, robust fusion, trigger item identification, factor-level localization and cause graph construction, and generates anomaly evidence packages and disposal suggestion structures. S400: Based on the disposal suggestion structure, perform multi-objective constraint assembly, conflict resolution, priority sequence solving, amplitude limit configuration, execution list compilation, implementation, effect feedback registration and feedback writing processing, and generate effect feedback structure.

[0021] Step S100 includes at least steps S110-S130: S110. Obtain the micro-area grid and probe deployment scheme and equipment list, perform coordinate binding, time anchor point registration and channel health self-check processing, and obtain multi-source raw data packets; The input sources for this section are the micro-area grid, probe deployment scheme, and equipment list that have been pre-established in this invention. The micro-area grid refers to the division of the production line site into stable areas with fixed boundaries and numbers according to spatial regions such as heating stations, forming stations, heat treatment stations, bulging stations, straightening stations, grinding stations, and testing stations. The probe deployment scheme refers to the written configuration of the installation positions, installation methods, numbering rules, and channel mapping relationships of environmental micro-area probes, equipment status probes, and process cycle probes. The equipment list refers to the list of all monitoring-related equipment names, equipment numbers, station affiliations, unit levels, maintenance record entry points, and available sensor channels within the production line. Specifically, the micro-area grid and probe deployment scheme and equipment list are used as parallel inputs. First, the coordinate binding process establishes a one-to-one correspondence between the physical location of the probe and the spatial unit of the micro-area grid. Then, the equipment number in the equipment list is cross-registered with the micro-area unit number and workstation number, so that different channels of the same equipment can be traced back to a unique micro-area in space. Furthermore, in order to avoid ambiguity in subsequent data alignment, the time anchor registration process uniformly records the clock reference, shift start and end, work day boundary and process instruction trigger time of various acquisition terminals. The output of the time anchor registration is a unified time table that can be called by the whole domain. The unified time table is not output as an independent field in this section, but as the time index source when packaging multi-source raw data. Understandably, the coordinate binding process has built-in position tolerance judgment rules to handle the case where there is a very small deviation between the probe installation and the micro-area boundary, and to classify the probe into the nearest micro-area according to the preset tolerance; when the probe number is duplicated or a channel is missing from the equipment list, the system will generate a conflict registration record and write the record into the abnormal remarks field of the current batch. The abnormal remarks field is stored together with the multi-source original data packets.

[0022] Furthermore, channel health self-checking is integrated throughout the data access phase of this section. Channel health self-checking refers to the rapid detection of the availability of a single channel. The judgment indicators include online status, stable sampling interval, amplitude exceeding limits, presence of silent segments, and short pulse jitter. The detection logic is automatically triggered at the beginning of each acquisition window. When a channel is determined to be offline, this section does not perform difference estimation; instead, it writes a missing marker in the multi-source raw data and retains the time anchor information for subsequent steps of missing measurement only labeling. When a channel has short pulse jitter but does not continuously exceed the sampling window threshold, the channel health self-check writes this situation into the quality label field, which is used to prompt subsequent steps for robust processing. When the same physical quantity is provided by two parallel channels, the channel health self-check first compares the time coverage and basic statistics of the two channels and gives a same-source determination result. The same-source determination result is only used for labeling in this section and does not involve channel merging. After completing the aforementioned process, this section will package the raw sampling points within the window according to a three-level index: micro-area, device, and channel. The packaged structure includes a spatial index, time anchor point, channel number, quality label field, and anomaly note field, and will be uniformly written into the multi-source raw data packet. During the packaging process, if the probe deployment scheme includes process cycle probes, the process event trigger time and associated device number will be written into the event index field. The event index field will be used for subsequent steps to map operating condition labels. The output of this section is a multi-source raw data packet, which is explicitly recorded as an output field name and is called as the sole input by the multi-source raw data packet in subsequent step S120. At the same time, the spatial index, time anchor point, and quality label field in the multi-source raw data packet are referenced by the main processes S200 and S300 in cross-main step scenarios and will not be separately split into new fields.

[0023] S120. Extract calibration segments, historical baselines and operating condition labels from the multi-source raw data packets, perform drift verification and same-source deduplication and operating condition label mapping registration, and generate usable sequences and confidence label sets. The input source in this section is the multi-source raw data packet output by S110. The calibration segment refers to a continuous sampling segment within a time period where the equipment status is stable, environmental fluctuations are controlled, and process events are not triggered. The historical baseline refers to a set of reference data within a recent historical range for the same equipment, micro-area, shift, or similar shifts in the current batch. The operating condition label refers to the process status identifier and operation stage identifier recorded by the process cycle probe and parsed through the event index field. Specifically, the event index field is first parsed from the multi-source raw data packet to obtain the operating condition label. Then, based on the operating condition label, the statically stable segment and the loading segment that can be used as references within the current window are located. The statically stable segment is used to extract calibration segments from the equipment status channel and the environmental micro-area channel, while the loading segment is used to provide dynamic comparison in subsequent evaluation steps. The extraction strategy for the calibration segment follows constraints of temporal continuity, good quality labels, and consistent spatial attribution, and is automatically truncated when there is a change in operating condition at the boundary. The truncation rule is written into the truncation record field and stored together with the calibration segment. Subsequently, historical baselines consistent with the current equipment, micro-area, and operating condition labels are retrieved from the historical data archive. Historical baselines are selected with priority given to the same shift, followed by adjacent shifts, then adjacent dates, and sorted by time of recentity. The time range, sampling coverage, and quality label distribution of historical baseline entries are registered in the baseline overview field. The baseline overview field, together with the calibration segments, is used for drift verification.

[0024] Furthermore, the drift verification process is responsible for determining whether the statistical distribution of the current calibration segment has shifted relative to the historical baseline. The triggering conditions for drift verification come from the coverage requirements and quality label thresholds in the baseline overview field. When the triggering conditions are met, the system will perform segmented comparisons between the calibration segment and the historical baseline and output a drift evidence description. The drift evidence description includes a description of the offset direction, an amplitude segmentation description, and a micro-area coverage description. When the drift evidence description shows that the offset exceeds a preset threshold, this section will record a probability switching suggestion and write it into the strategy indication field. The strategy indication field is used as an intermediate registration item in this section and is not used as an output field. It will be referenced in the subsequent configuration of probability switching and rollback strategies in S220. For cases where multiple channels operate in parallel for the same physical quantity, this section performs deduplication based on source. Deduplication based on source means determining whether two channels are from the same source by considering time coverage overlap, event response consistency, and statistical similarity of stable segments. When determined to be from the same source, the channel with the better quality label is retained as the primary channel, while the discarded channel is retained as a bypass channel and marked with a bypass flag. The bypass flag is used for subsequent rollback. When two channels cannot be determined to be from the same source, the system retains both and assigns a confidence level to each channel in the confidence label set. The condition label mapping registration is established based on the cross-relationship between the event index domain and the equipment list. Each sampling point is associated with the most recent process event segment and written into the condition label mapping table. The condition label mapping table will be written into the output and used as a matching input with the available sequences in subsequent steps. After completing the aforementioned processing, this section organizes the channel data after drift verification and deduplication to generate a coherent, fully labeled time series set with quality labels, defined as usable sequences. Simultaneously, each usable sequence is scored and graded according to channel health self-check results, drift evidence descriptions, deduplication judgments, and operational condition label stability, generating a confidence label set. Understandably, usable sequences are the direct data input for subsequent evaluation and threshold learning, while the confidence label set provides a reliable basis for subsequent weight estimation and threshold boundary reinforcement. The outputs of this section are usable sequences and confidence label sets, explicitly recorded as output field names and directly called by the usable sequences in subsequent step S130. Furthermore, the mapping table between usable sequences, confidence label sets, and operational condition labels will be used by the main process S300 for data consistency measurement and weight estimation, and will not be repeatedly generated in cross-main-step scenarios.

[0025] S130. Perform robust denoising and minimum exposure screening on the available sequences to generate a candidate set structure of impact factors. The input source for this section is the available sequences and confidence label set output by S120. The available sequences have unified spatial indexes, unified time anchors and operating condition label mapping information, and the confidence label set provides sequence-level confidence level and bypass markers. This section first performs anomaly segment identification and robust denoising. Anomaly segment identification employs two rules: event proximity judgment and abrupt change proximity judgment. Event proximity judgment marks a sampling point near the boundary of a process event as a sensitive segment, while abrupt change proximity judgment marks a continuous segment as an abrupt change segment when the sampling change in the continuous segment shows an irregular increase or decrease within a short time window. For segments marked as sensitive or abrupt changes, this section does not perform difference estimation but instead uses a robust denoising strategy to process their influence range. The robust denoising strategy performs mild smoothing on high-confidence sequences and segment replacement on low-confidence sequences. Segment replacement prioritizes stable segments in the same channel as replacement sources. When stable segment coverage is insufficient and bypass markers exist, the corresponding segment under the same operating condition label in the bypass channel is called for replacement. When a replacement occurs, the replacement source, replacement interval, and replacement reason are written into the replacement registration field. The replacement registration field is stored in the output structure as evidence for subsequent anomaly interpretation.

[0026] Furthermore, the minimum exposure screening process, based on the principle of outputting only the minimum fields necessary for subsequent threshold learning and evaluation fusion, performs field pruning on the available sequences channel by channel and micro-region by micro-region. Field pruning retains the continuous samples after robust denoising, the corresponding time anchors, spatial indices, operating condition labels, and quality labels, while removing redundant annotation information that is not directly related to subsequent calculations. When multiple channels are substitutable under the same micro-region and the same operating condition label, the minimum exposure screening prioritizes outputting the main channel with high confidence, and retains a bypass reference pointer in the output structure. The bypass reference pointer records the substitution order and applicable boundaries, which are used in subsequent steps to retrieve the bypass source when the fallback strategy is triggered. This section then proceeds to extract and construct candidate impact factors. The impact factor candidate set refers to a collection of basic indicator entries that can be used to describe states and changes across three dimensions: environment, equipment, and production line cycle time. Environment domain entries are derived from the statistics and change descriptions of stable segments of environmental micro-area probes under corresponding operating condition labels. Equipment domain entries are derived from the statistics and response characteristic descriptions of stable segments of equipment status probes. Production line cycle time domain entries are derived from the execution duration, interval, and synchronization relationship descriptions corresponding to process cycle time probes. To ensure the feasibility of subsequent weight estimation and threshold boundary reinforcement, each entry in the impact factor candidate set includes associated information such as the micro-area, equipment, operating condition label, time coverage, and quality label, and is written into the candidate registration domain. Understandably, the impact factor candidate set does not contain calculation models or weights; it only provides standardized candidate entries for subsequent learning and fusion. After the above processing, this section outputs an impact factor candidate set as a structured set. The structure includes robustly processed minimum exposure sequence references, bypass reference pointers, candidate registration fields, and replacement registration fields. The impact factor candidate set is explicitly recorded as an output field name and is directly used as input in subsequent step S210. It also cross-steps with the adaptive threshold set generated in the threshold candidate generation in the main process S200, without requiring additional intermediate fields. In summary, this step achieves the following technical effect: by performing robust denoising and minimum exposure screening on available sequences and constructing an impact factor candidate set structure with spatial, temporal, and operational condition mapping information, subsequent threshold learning and evaluation are integrated on the input side with a unified caliber and traceable evidence.

[0027] Step S200 includes at least steps S210-S230: S210. Obtain the candidate set of impact factors, available sequences and working condition labels, perform quantile adaptive threshold learning and threshold learning input set construction to obtain the threshold learning input set. The input sources for this section are the candidate set of impact factors, available sequences, and operating condition labels output in the previous steps. The candidate set of impact factors refers to a set of indicator entries that have undergone robust processing in the environmental, equipment, and production line cycle time domains and have spatial indices, time anchors, and quality labels. Available sequences refer to a set of continuous data sequences formed under a unified spatial and temporal caliber. Operating condition labels refer to the mapping information of the operation stage, event start and end, and equipment association registered by the process cycle time probe. Specifically, the above three types of inputs are aligned in both time and spatial dimensions. The alignment process unfolds according to the secondary index of micro-area and workstation, and time slices are established according to shift and workday boundaries. By merging event trigger markers within the same time period, sample slices that can be used for learning are formed. Cases of alignment anomalies are truncated at the slice boundaries, and the truncation information is recorded in the input registration field of the current batch for subsequent traceability and sample rollback. Furthermore, to ensure that subsequent learning processes focus on representative segments, this section performs stable segment screening on the aligned available sequences. Stable segment screening eliminates high-perturbation regions based on quality labels and event proximity markers. The retained segments undergo sequence integrity and time coverage verification. Segments that pass the verification are labeled as learning candidate segments and associated one-to-one with corresponding impact factor candidate entries. These associations are written into the candidate segment mapping table. Understandably, the candidate segment mapping table is one of the core input structures for threshold learning, used to describe the index and binding relationship between sample segments and impact factor entries, while also storing the corresponding operating condition labels, equipment numbers, and micro-region numbers. Segments that fail the verification are not included in the subsequent processing of this step but are registered as bypass entries.

[0028] After organizing the samples, this section divides the candidate fragments into several learning groups according to the work condition labels. Each learning group contains only samples with consistent work condition labels and spatial index compatibility. Within each learning group, quantile statistics and dispersion descriptions are performed on the same influencing factor candidate items. Quantile statistics are used to characterize the basic distribution characteristics of candidate items under the same work condition, while dispersion descriptions are used to reflect the fluctuation range of items under the same work condition label. To avoid the bias caused by sparse samples in the statistical description, this section sets a minimum sample coverage requirement for each learning group. When the sample coverage is insufficient, a neighbor merging strategy is triggered. Neighbor merging follows the rule of prioritizing the same work position and then adjacent shifts, and the merging source is written into the merging registration field. Subsequently, this section performs quantile adaptive thresholding, which involves calculating a set of threshold candidate positions with upper and lower boundaries for each impact factor candidate item within the learning group, and combining this with the dispersion description to form a threshold candidate baseline. When generating the threshold candidate baseline, operating condition label stability information is simultaneously introduced. When the stability is low, this section robustly shrinks and widens the boundaries of the threshold candidate positions, and records the triggering reasons for shrinkage and widening in the threshold labeling domain. To form a unified and usable learning input, this section integrates quantile statistics, dispersion description, operating condition label stability, sample coverage, merged registration domain, and threshold labeling domain to generate a threshold learning input set. The threshold learning input set is a structured set containing a learning group index, an impact factor item index, and a threshold candidate baseline, where each record corresponds to the statistical induction result of an impact factor candidate item under a specific operating condition. After completing the above processing, the output field name of this section is Threshold Learning Input Set. It is clearly stated that the Threshold Learning Input Set is directly called as the only input for the subsequent step S220. At the same time, the learning group index and entry index in the Threshold Learning Input Set are referenced by the evaluation fusion process without being changed in the cross-main step scenario.

[0029] S220. Extract quantile statistics and drift evidence and policy indications from the threshold learning input set, perform probability switching and backoff policy configuration and threshold candidate generation, and generate a threshold candidate set. This section builds upon the aforementioned threshold learning input set to generate and manage threshold morphologies. Quantile statistics provide the basic position and upper and lower boundaries, drift evidence provides a description of the relative difference from historical baselines, and strategy indications provide suggestions for type switching and bypass usage for registered probabilistics from the previous round or historical batches. Specifically, the threshold learning input set is expanded at the learning group granularity, and for each influencing factor entry's threshold candidate baseline, corresponding drift evidence and strategy indications are read. The type switching suggestions in the strategy indications are used to determine whether the entry should switch from its current morphology to another morphology. Morphology switching can manifest as asymmetrical adjustments to the upper and lower boundaries of the threshold, reordering of stable segment priorities, or a stepwise migration of candidate positions. When the strategy indication is empty or the switching trigger condition is not met, the entry maintains its current morphology and enters the candidate generation branch. When the strategy indication meets the switching trigger condition, the entry enters the type switching branch, and the learning group index and triggering reason for the switching are recorded in the type switching registration field. Furthermore, the backoff strategy configuration addresses situations where the threshold form undergoes abrupt changes within a short time window. For each influencing factor entry, a bypass source and bypass application boundary are configured. The bypass source originates from the historical stable form of the same entry within the learning group or from a substitute entry with a higher quality label under the same conditions. The bypass application boundary is jointly determined by the sample coverage and dispersion description within the learning group. When the threshold generation branch encounters insufficient sample coverage, unstable drift evidence, or conflicting strategy indications, this section activates the backoff strategy and adopts the form from the bypass source as a temporary threshold form. Simultaneously, the source and effective time period of the temporary form are recorded in the backoff registration domain.

[0030] During the threshold candidate generation process, this section generates entry-level threshold candidates based on the quantile statistical position and dispersion of each entry within the learning group. These candidates include upper and lower boundaries, candidate positions, morphological labels, and applicable operating conditions. After generation, the entry-level threshold candidates are aggregated into learning group-level threshold candidate units. Each threshold candidate unit contains the threshold candidates for all entries within the learning group, along with their morphological labels and backtracking registration information, and includes a reference pointer to the probability switching registration field. To ensure compatibility with subsequent boundary reinforcement and causal impact detection, this section outputs both source and version labels for the threshold candidates during candidate generation. The source label distinguishes between three sources: stable segment statistics, merging neighboring entries, and bypass backtracking. The version label records the candidate generation batch and time interval. Understandably, after threshold candidate generation is complete, this section defines the learning group-level threshold candidate unit set as the threshold candidate set. The threshold candidate set structure includes entry-level candidates, morphological labels, source labels, version labels, probability switching registration fields, and backtracking registration fields, maintaining index consistency with the impact factor candidate set at the entry level. After completing this section, the output field name of this section is Threshold Candidate Set. The Threshold Candidate Set is called as the sole input of the Threshold Candidate Set in the subsequent step S230. At the same time, the source label and version label in the Threshold Candidate Set are used to form audit records by multi-target device scheduling and online retraining in cross-main step scenarios, without setting up additional intermediate fields.

[0031] S230. Perform causal impact detection and boundary reinforcement processing on the threshold candidate set to generate an adaptive threshold set and a threshold version mirror structure. This section, building upon the existing threshold candidate set, constructs a causal impact detection chain and implements boundary reinforcement based on scenarios involving strategy changes, equipment status changes, and environmental disturbances encountered by the candidates in real production processes. Specifically, the threshold candidate set is expanded along both the operating condition label and learning group dimensions. Intervention markers are read from the execution records of the previous or adjacent cycles. Intervention markers refer to event descriptions with intervention attributes, such as strategy adjustments, maintenance operations, or process parameter switching, written by the production control system. The sample intervals corresponding to the candidates are aligned with the intervention markers on the timeline. Parallel control groups with impact windows and non-impact windows are constructed through the alignment relationship. The impact window is used to evaluate the stability of the threshold candidates before and after the intervention, while the non-impact window is used to describe the natural fluctuations under no-intervention conditions. Furthermore, causal impact detection is performed at the item level. For each impact factor item, the difference in candidate performance between the impact window and the non-impact window is determined. The input for the difference determination comes from the source label and version label of the threshold candidate and the dispersion description within the learning group. When the difference determination shows that the candidate has a significant shift in the impact window and that the shift does not appear in the non-impact window, this section marks the item as an item affected by the intervention and writes the intervention type, impact direction and impact period into the impact level field. When the difference determination shows that the shift appears in both the impact window and the non-impact window or is indistinguishable, the item is marked as uncertain and is handled by a robust strategy when strengthening the boundary.

[0032] After completing the causal impact detection, this section proceeds to boundary hardening. Different hardening methods are used for entries marked as affected by intervention and for uncertain entries. For affected entries, a protective band is applied to the upper and lower boundaries of the entry-level candidates. The width of the protective band is related to the dispersion description and sample coverage within the learning group. Priority triggering of bypass backoff is set within the protective band to ensure the threshold maintains more robust upper and lower boundaries during the intervention period. For uncertain entries, the candidate position is slightly narrowed and compared with the bypass source recorded in the backoff registration field. If the morphology of the bypass source is more stable in adjacent periods, the bypass morphology is used as the temporary threshold output in this period, and this decision is written to the temporary output registration field. For entries that have passed the causal impact detection and do not require narrowing, the upper and lower boundaries and positions of the candidates remain unchanged, and the batch number of this period is incremented in the version label to form a new candidate version. Subsequently, this section aggregates the item-level hardening results within the learning group, generating a learning group-level threshold item set, and records references to the probability switching registration domain and rollback registration domain used in the current period at the set layer. Understandably, to support traceability and relevance to threshold changes in subsequent production stages, this section constructs an adaptive threshold set based on the learning group-level threshold item set. The adaptive threshold set refers to the set of hardened threshold forms formed for all learning groups and all impact factor items, and records source labels, version labels, impact registration domains, temporary output registration domains, and protection band parameter descriptions at the set layer. To make version management more systematic, this section also generates a threshold version mirror structure. The threshold version mirror structure is a snapshot, organized by period, recording the threshold form, item source, form label, protection band parameters, rollback registration references, and probability switching registration references for all items in the current period, and writing pointers to the previous period's mirror into the mirror to form a continuous lineage. After completing the above processing, the output fields of this section are "Adaptive Threshold Set" and "Threshold Version Mirror". The adaptive threshold set is used as input to the adaptive threshold set in subsequent step S310 for data consistency measurement and online estimation of time-varying weights. The threshold version mirror is used as input to the threshold version mirror in subsequent step S410 for multi-objective constraint assembly and execution strategy auditing. Furthermore, the output of this section, in cross-main-step scenarios, serves as a trigger reference for online retraining and is no longer split into other fields. In summary, the technical effect of this step is: through causal impact detection and boundary reinforcement, this step transforms the threshold candidates into an adaptive threshold form with anti-interference stability and forms a traceable threshold version mirror, supporting parameter consistency and version management for subsequent evaluation fusion and scheduling execution.

[0033] Step S300 includes at least steps S310-S330: S310. Obtain the candidate set of impact factors, the adaptive threshold set, and the confidence label set; perform data consistency measurement and online estimation of time-varying weights to obtain weight parameters and evaluation input set. The input sources for this section are the candidate set of impact factors, the adaptive threshold set, and the confidence label set output in the previous steps. The candidate set of impact factors is a set of entries formed after robust processing in the environmental domain, equipment domain, and production line cycle time domain. The entries include structured fields such as the micro-region to which they belong, the equipment to which they belong, the operating condition label to which they belong, time coverage, spatial index, and quality label. The adaptive threshold set is a set of threshold forms formed after causal impact detection and boundary reinforcement. It includes information such as the upper and lower boundaries of the entries, candidate positions, form labels, source annotations, and version annotations. The confidence label set is a set of labels formed after classifying the available sequences and their derived entries according to dimensions such as collection completeness, quality label stability, and bypass reference substitutability. Specifically, the candidate set of impact factors is divided into continuous windows on the time axis according to the micro-area grid and work station code, and aligned one-to-one with the adaptive threshold set in the dimensions of work condition label and time coverage. The alignment step adopts a matching method with work condition label as primary key and micro-area and equipment as joint index. For entries that cross windows, the back-end of the segmentation is spliced ​​and boundary markers are registered at the segmentation boundary. When the time coverage of the candidate impact factor entries and the time coverage of the adaptive threshold entries only partially overlap, the data consistency measurement process registers a consistency benchmark in the overlapping area and a reference benchmark in the non-overlapping area. Subsequently, it only participates in estimation and fusion in the consistency benchmark area. The reference benchmark area is reserved in read-only mode for traceability. Furthermore, data consistency is verified across three dimensions: sequence integrity, consistency stability, and condition label consistency. Sequence integrity verification uses quality labels and missing data markers to confirm whether consecutive samples meet the minimum coverage requirement. Consistency stability verification identifies fluctuating and stable segments in the trajectory of the same entry across adjacent windows and registers stability markers in stable segments. Condition label consistency verification uses cross-registration of event indexes and equipment lists to confirm whether the condition label of an entry is consistent with the applicable condition of the threshold form. When any verification fails to meet the minimum condition, the entry is marked as a candidate restricted entry within the current window and is only allowed to participate as a bypass reference in subsequent estimation processes. After data consistency measurement is completed, the entries that pass verification within the consistency benchmark area are extracted as estimation candidate units. Each estimation candidate unit contains an impact factor candidate entry and its corresponding threshold form, confidence level, stability marker, and boundary marker. Understandably, the above estimation candidate units are organized in windows, which preserves the contextual information generated by alignment while avoiding the introduction of segments that do not meet the verification conditions into subsequent estimations.

[0034] After obtaining the candidate units for estimation, the time-varying weight online estimation process is triggered. This process operates with a dual granularity of learning groups and windows. The learning group is defined by both working condition labels and spatial indices to ensure that entries under the same working condition are aggregated and estimated under similar production states. The window granularity is used to track the contribution changes of entries in adjacent time slices. The inputs to the time-varying weight online estimation are the candidate units for estimation, a set of confidence labels, and an adaptive threshold entry shape. The confidence labels are used to provide the initial weight starting point, and the adaptive threshold entry shape is used to provide boundary constraints on the entry behavior. Specifically, a candidate list is constructed for each learning group, and the candidate list is sorted according to the confidence level and stability label of the entries. During the sorting process, when two entries have the same confidence level but different stability, the one with higher stability is given priority; when the stability is the same but the source labels of the entries are different, the order is: priority is given to the statistics of stable segments in the source labels, followed by merging neighboring entries, and then bypassing and backtracking. Subsequently, the online estimation iterates through the candidate list item by item within the window, reading the performance of each item within the consistency benchmark region. Performance descriptions include the item's proximity to the threshold pattern within the window, boundary hit frequency, and whether the backoff pointer is triggered after a boundary hit. Based on this, the contribution weight of the item in the current window is adjusted. When an item is hit by a boundary within the window and triggers a backoff pointer, its contribution weight is downgraded within the current window and gradually restored to its pre-stabilization level in the next window using a recovery factor. When an item shows a continuous increase in stability within the window, its contribution weight is upgraded within the current window, and the reason for the upgrade is recorded. After the online estimation is completed, a window-level weight snapshot is generated. This snapshot records the item's contribution weight in the current window, the reasons for upgrading or downgrading, and boundary hit records, while also retaining its association with the learning group. To ensure seamless integration with subsequent evaluations, this section associates and integrates window-level weight snapshots with estimated candidate units, generating a structured set containing entries, weights, threshold patterns, time coverage, and operating condition labels. This set is named the evaluation input set. Simultaneously, the contribution records of entries in the learning group and window dimensions are extracted as weight parameters, stored using the entry index and learning group index as keys. After the above processing, the output fields of this section are named weight parameters and the evaluation input set. The evaluation input set is the sole input source for subsequent step S320. The weight parameters are referenced in multi-target device scheduling and online retraining across main steps, used for result write-back and version spectrum system calculation, without additional intermediate transformations.

[0035] S320. Extract environmental domain entries, equipment domain entries, and production line domain entries from the evaluation input set, perform robust fusion processing, and generate evaluation records and confidence intervals. After the evaluation input set is available, this section first extracts environmental, equipment, and production line domain entries from the input set according to domain division. Domain division is based on the entry attribution from the candidate set of influencing factors. Environmental domain entries are derived from the stable segment statistics and change descriptions corresponding to environmental micro-area probes; equipment domain entries are derived from the stable segment statistics and response descriptions corresponding to equipment status probes; and production line domain entries are derived from the execution duration, interval, and synchronization relationship descriptions corresponding to process cycle time probes. Specifically, three domain subsets are established within each learning group, each containing one of the three types of entries and their weight parameters and boundary hit records in the window-level weighted snapshot, organized into a time series according to window order. In cases where entries overlap across domains, the domain where the entry is first registered is the primary domain, and other domains are used as referencing domains with reference pointers. Robust fusion processing is performed in two stages within the learning group: the first stage is intra-domain fusion, and the second stage is cross-domain convergence. Intra-domain fusion is constrained by the proximity of entries to adaptive threshold patterns and boundary hit records, and weighted aggregation is performed according to the weight parameters of window-level weighted snapshots. During intra-domain fusion, when an entry's boundary hit occurs within a window and triggers a backtracking pointer, the entry's immediate contribution is replaced by the immediate contribution of the bypass source entry, and the replacement event is registered in the fusion result. When an entry within a window remains in a low-contribution state for an extended period, accompanied by a decrease in confidence, the entry is added to the intra-domain observation list, retaining its impact on the fusion result with only a lower weight. The first stage outputs an intra-domain fusion sequence, which includes the aggregation results of the domain in each window, replacement event records, and the status of the observation list.

[0036] The second stage of cross-domain convergence uses the intra-domain fusion sequence as input to establish a cross-domain weighting framework for the aggregation results of the environment, equipment, and production line domains. The cross-domain weighting framework is based on learning groups, and the weight starting point is initialized at the cross-domain level according to the weight parameters provided in the online estimation stage. When any of the three domains experiences frequent replacement events or an inflated observation list in multiple adjacent windows, the cross-domain weight of that domain is reduced in the current window and increased in subsequent windows according to a recovery strategy. When a domain exhibits stable continuity and no boundary hit records within adjacent windows, the cross-domain weight of that domain is increased in the current window. Cross-domain convergence generates learning group-level aggregation sequences, which represent the synthetic state of the learning group from a global perspective. To support subsequent factor-level localization and causal graph construction, this section calculates confidence boundary descriptions on the aggregation sequences. Confidence boundary descriptions define intervals based on the window discreteness of the aggregation sequences, the density of replacement events, and the proportion of the observation list. These interval definitions are written into the confidence intervals. Understandably, intra-domain fusion and cross-domain aggregation are organized using the same data structure and indexing method, ensuring a clear traceability path between the item level, domain level, and learning group level. During implementation, when the window span crosses shift or workday boundaries, a marker is inserted into the aggregation sequence using a time anchor. This marker is used only for visualization and auditing records and does not participate in weight calculation. After completing the robust fusion, this section generates learning group-level evaluation items, each carrying an aggregation value, cross-domain weight status, replacement event record, and confidence boundary description. Subsequently, all learning group evaluation items are concatenated along the production line dimension according to workstation order to form evaluation records. To ensure continuity with subsequent steps, the output fields of this section are named Evaluation Record and Confidence Interval. The evaluation record serves as the sole input source for subsequent step S330, while the confidence interval is read as an auxiliary boundary for trigger item identification in subsequent steps. Simultaneously, the evaluation record and confidence interval are read by multi-target device scheduling and online retraining in cross-main-step scenarios for policy auditing and trigger criterion registration, and are no longer split into other fields.

[0037] S330. Perform trigger item identification, factor-level localization, and cause map construction on the evaluation records to generate an abnormal evidence package and disposal suggestion structure. This section takes the aforementioned evaluation records and confidence intervals as input and performs identification and localization in two dimensions: learning group and workstation. Trigger identification first reads the correspondence between the aggregated sequence and the confidence interval in the evaluation records. When the aggregated sequence crosses the boundary between any window and the upper or lower boundary of the confidence interval, a trigger candidate is registered. When the boundary crossing relationship and the boundary hit record of the adaptive threshold item appear in the same window and are in the same direction, the trigger candidate is upgraded to a trigger item, and a source label and window index are attached to the trigger item. For the registered trigger items, this section further performs factor-level localization. Factor-level localization backtracks the domain-wide fusion sequence and window-level weight snapshot within the trigger window, retrieves the item that contributes the most to cross-domain convergence within the window and has experienced a replacement event or weight jump in domain-wide fusion, and registers possible causal factors at the item level. In the case of bypass replacement of an item, factor-level localization simultaneously registers the association between the main item and the bypass source item, and indicates the associated work condition label and micro-area location. Understandably, factor-level localization does not change the original values ​​of the evaluation record, but rather overlays interpretive meta-information on them, providing a candidate list of nodes and edges for subsequent graph construction.

[0038] After factor-level localization is completed, this section constructs the cause graph. The cause graph uses a multi-layered node representation of workstation—micro-area—equipment—item—condition label—threshold. Edge types include trigger association, replacement association, fallback association, and boundary hit association. Each edge includes a trigger window, direction description, and source label. When multiple triggers appear consecutively in adjacent windows within the same learning group, the cause graph establishes a temporal order among these triggers and records cluster identifiers. These cluster identifiers indicate the temporal continuity of the triggers in the group. During graph construction, if an entry corresponding to a trigger has duplicate nodes in different domains, they are merged into the same entry node using reference pointers to avoid redundancy in interpretation. After the cause graph is constructed, this section compiles abnormal evidence elements based on the trigger items and the node and edge information in the graph. These elements include trigger window snapshots, trigger item source annotations, item-level contribution descriptions, replacement event summaries, rollback registration references, and the correspondence between these elements and adaptive threshold forms. These elements are then organized into abnormal evidence packages. An abnormal evidence package is a structured collection generated for auditing and handling strategies. Each package corresponds to one trigger item or a group of consecutive trigger items, containing indexes and references for review. Based on the abnormal evidence packages, this section generates handling suggestions without modifying the assessment records. These suggestions use strategy modules such as reducing workstation load, switching equipment operating status, adjusting cycle time parameters, requesting short-term maintenance, using backup equipment, and limiting operating boundaries as candidates. The assembly logic of the candidate strategy modules reads weight parameters, adaptive threshold item protection bands, and the source and version annotations of the threshold version mirror, forming structured suggestions that can be read by the executable system. After completing the above processing, the output fields of this section are named "Abnormal Evidence Package and Handling Suggestions." The handling suggestions serve as the sole input source for subsequent step S410. The abnormal evidence package is referenced in multi-target device scheduling and online retraining across main step scenarios, used for trigger registration of execution strategy auditing and online parameter updates. In summary, the technical effect of this step is: by implementing trigger identification, factor-level localization, and cause graph construction on the evaluation records, this step outputs a verifiable abnormal evidence package and structured handling suggestions, enabling subsequent scheduling and retraining to be executed and written back on the same evidence chain.

[0039] Step S400 includes at least steps S410-S430: S410: Obtain disposal suggestions, threshold version images, and production resource images; perform multi-objective constraint assembly and conflict resolution processing to obtain priority sequences and equipment scheduling schemes. The input sources for this section are the handling suggestions, threshold version mirrors, and production resource mirrors already output by the higher-level steps. The handling suggestions refer to a set of structured strategy entries formed by combining the assessment records and abnormal evidence packages. Each entry specifies the trigger window, the workstation involved, the associated equipment, the suggested action, and the limiting conditions. The threshold version mirror refers to a set of version snapshots organized and recorded according to the cycle, including the threshold form, source label, protection band parameters, and rollback registration references. The production resource mirror refers to a unified mapping structure around the availability status, switching costs, maintenance windows, standby machine list, and tooling resource occupancy of production line resources in the time and space dimensions. Specifically, the proposed handling measures are used as strategy inputs, the threshold version image is used as boundary inputs, and the production resource image is used as resource inputs to the multi-objective constraint assembly process. The multi-objective constraint assembly process supplements the constraint metadata for each proposed handling measure. The constraint metadata includes at least capacity constraints, energy consumption constraints, yield constraints, maintenance window constraints, and safety barrier constraints. The safety barrier constraints are determined by the protection band parameters and rollback registration references in the threshold version image. When a proposed handling measure involves cross-workstation linkage, the multi-objective constraint assembly process marks the linkage relationship and registers the linkage boundary in the same time slice to facilitate consistency processing in the subsequent conflict resolution stage. Furthermore, during constrained assembly, for equipment and tooling marked as temporarily occupied in the production resource image, the system marks them as unschedulable during the corresponding time period and writes the source of occupation and the expected release time into the constraint metadata. For standby machine switching involved in the disposal suggestion, the assembly process reads the compatibility field and switching cost field in the standby machine list and forms a switching candidate. The switching candidate forms a one-to-many mapping with the original equipment and comes with a switching order. When a boundary conflict occurs between the disposal suggestion and the threshold version image (for example, the disposal suggestion attempts to relax the operating boundary within the protection band), the assembly process prioritizes retaining the protection band restrictions in the threshold version image and downgrades the disposal suggestion action to shadow execution or limited execution. The downgrade record is written into the assembly registration field.

[0040] After completing the multi-objective constraint assembly, this section proceeds to conflict resolution. The input for conflict resolution is the assembled strategy-constraint-resource triplet. Conflict sources include multiple actions competing for the same equipment within the same time period, concurrent occupation of the same resource across workstations, and operational actions proposed within the maintenance window versus amplitude adjustments proposed within the protection zone. To handle these conflicts, conflict resolution first generates a conflict candidate list. Each candidate in the list is associated with its trigger window, involved resources, and conflict level. The conflict level is determined first based on safety barrier constraints and maintenance window constraints, followed by yield constraints and energy consumption constraints. Subsequently, conflict resolution invokes the action relationships in the disposal suggestions, storing... Actions with strong dependencies are executed sequentially under the same ID. Actions with weak or no dependencies are marked with equivalence levels within the commutative set. For concurrent use of the same resource, conflict resolution calculates the insertion order based on the switching cost and release time in the production resource image, prioritizing the action combination with the lowest switching cost and earliest release time, and adding the delayed actions to the candidate sequence. When a candidate action conflicts with a rollback registration reference in the threshold version image, conflict resolution calls the rollback registration reference for boundary adjustment and writes the rollback effective time slice as a strong constraint into the execution context of the current batch. Understandably, after conflict resolution, an execution-oriented sequence expression is formed based on the resolved actions and the order of retention. This sequence expression is defined as a priority sequence. While generating the priority sequence, the system reorganizes action blocks by workstation and equipment dimensions and fills in necessary pre-checks and subsequent confirmation steps to form an equipment scheduling scheme. The equipment scheduling scheme includes fields such as equipment number, action time slice, action type, linkage reference, shadow execution flag, and limiting parameters. At this point, the output field name of this section is priority sequence and device scheduling scheme. The priority sequence and device scheduling scheme are directly called as inputs to the subsequent step S420. At the same time, the shadow execution flag and limiting parameter in the priority sequence are read online during cross-main step scenarios to generate trigger criterion registration.

[0041] S420. Extract the execution unit and safety barrier parameters and priority sequence from the equipment scheduling scheme, perform priority sequence solving, limit configuration and execution list compilation, and generate safety barrier configuration and execution list; This section receives the device scheduling scheme and the priority sequence as input, and parses the execution units, safety barrier parameters, and reads the shadow execution flags within the scheme. An execution unit refers to an atomic action implemented on a single device within a single time slice. An atomic action includes the action target, action amplitude, action duration, linkage reference, and confirmation method. Safety barrier parameters refer to the upper and lower limits, acceleration and deceleration constraints, bypass triggering conditions, and alternative sources for rollback that are allowed under the current threshold condition, all related to the action target. The priority sequence refers to the order of actions and their sibling sets obtained after conflict calculation. Specifically, the device scheduling scheme is aggregated by workstation and expanded by device number within each workstation, generating a hierarchical index of workstation—device—execution unit. Under this index, the safety barrier parameters of each execution unit are read. The safety barrier parameters come from the protection band parameters and rollback registration references in the threshold version image, supplemented as necessary by switching cost constraints from the production resource image. Subsequently, the priority sequence solver expands the execution units step by step according to the hierarchical index and sorts the exchangeable actions within the same set. The sorting is based on the shadow execution flag and the switching cost field. Actions with the shadow execution flag "yes" are executed first in the shadow channel, while actions with higher switching costs are postponed to the feasible time period. When there are linked references within the same set, the solver generates a synchronization bundle according to the linked reference relationship, and the execution units within the synchronization bundle are issued simultaneously in the same time slice.

[0042] In the amplitude limiting configuration stage, this section configures the upper and lower limits of the action amplitude for each execution unit under the current threshold mode, and configures the maximum continuous duration and minimum interval for the action duration. When the execution unit is marked as a shadow execution, the amplitude limiting configuration only takes effect within the shadow channel and writes a read-only observation mark to the main channel. When the execution unit is marked as a source that needs to be rolled back, the amplitude limiting configuration reads the rollback registration reference and writes the rollback candidate and trigger window into the execution context. For execution units involving standby machine switching, the amplitude limiting configuration reads the compatibility field of the standby machine list and generates a blocking entry when incompatible. The blocking entry is marked as pending review when the execution list is compiled subsequently. After completing the priority sequence solution and amplitude limiting configuration, this section proceeds to the execution list compilation. The execution list is compiled based on the workstation-equipment-time slice order, displaying the execution units on the timeline. Each time slice includes four nodes: pre-check, action issuance, subsequent confirmation, and exception rollback. The pre-check reads the occupancy status and maintenance window from the production resource image; if a conflict occurs, the current time slice is marked as delayed, and the reason for the delay is written into the execution annotation. The action issuance node reads the action target and amplitude range and sends it to the execution system (in this specification, the execution system refers to the control platform that manages equipment action commands). The subsequent confirmation node reads the event index and current threshold form of the multi-source raw data packets, registering the actual effective time of the action and boundary hit status. The exception rollback node calls rollback candidates within the trigger window and registers the rollback result. After compilation, the system outputs a safety barrier configuration and an execution list. The safety barrier configuration includes the amplitude range, duration, bypass trigger conditions, and rollback candidates for each execution unit. The execution list includes the action arrangement of the workstation-equipment-time slice, pre-check and subsequent confirmation nodes, shadow execution markers, and amplitude limiting parameters. The safety barrier configuration and the execution list are explicitly recorded as output field names. The execution list is called as the only input in the subsequent step S430. The safety barrier configuration is read by threshold learning and online retraining in cross-main step scenarios and is used to audit boundaries and record the implementation parameters of the policy.

[0043] S430. Perform landing and effect feedback registration and feedback writing processing on the execution list to generate an effect feedback structure; This section receives the aforementioned execution list and safety barrier configuration as input. The implementation process is initiated by the execution system and issued item by item at the workstation-equipment-time slice granularity. Specifically, at the beginning of each time slice, the execution system calls the pre-check nodes in the execution list, reads the occupancy status and maintenance window of the production resource image, and verifies whether the execution channel is available. If the pre-check shows unavailable, a rejection record is generated for this time slice, and the rejection reason, standby duration, and alternative time slices are written into the receipt draft. The receipt draft will be written into the effect receipt at the end of this time slice. If the pre-check shows available, the execution system issues an execution command to the device side according to the action target and amplitude range in the execution list, and registers the command number, issuance time, and target amplitude in the execution log. During action execution, the system reads the channel data in the multi-source raw data packets and aligns it on the time anchor point, identifies the actual effective time of the action, and records the boundary hit and bypass trigger status according to the safety barrier configuration after the action takes effect. When a boundary hit is detected and the bypass trigger condition is met, the execution system starts a rollback candidate in this time slice and records the device switching, channel switching, and amplitude contraction during the rollback process in the rollback log. The rollback log will be written into the effect receipt as a receipt element at the end of this time slice.

[0044] Effect feedback registration is triggered at the end of each time slice, and the registration content covers four aspects: First, action implementation information, including action target, target range, actual range, execution duration and command number; second, boundary and rollback information, including boundary hit time, whether bypass is triggered, whether rollback candidate is adopted, adoption source and effective period; third, resource and maintenance information, including changes in the occupancy status of production resource mirrors before and after execution and whether maintenance windows are occupied; fourth, linkage consistency information, including the action consistency and timing deviation of each execution unit within the synchronization bundle. After the above registration elements are registered, a time slice feedback record is formed. The time slice feedback record is accumulated into a batch feedback record at the workstation-equipment dimension, and the batch feedback record is summarized into an effect feedback structure at the production line dimension. The receipt writing process is executed immediately after the effect receipt structure is generated. The write destinations fall into two categories: one is the threshold version mirror audit channel, where the written content includes boundary hit status, bypass trigger status, rollback adoption status, and corresponding time slice pointers, used to form a cross-cycle boundary consistency index; the other is the data preparation and retraining channel, where the written content includes action implementation information and linkage consistency information, used as a source of historical baselines and intervention markers in the next round of data admission and threshold learning stages. Understandably, during the writing process, when certain receipt elements are missing due to collection anomalies, this section only marks the missing information without performing estimation filling, and writes the missing marker and the reason for the missing information at the write target. When there is an inconsistency between the batch receipt records and the execution log, this section adds a contradiction pointer to the effect receipt structure. The contradiction pointer is read during online retraining and triggers the recalculation process. After completing the aforementioned processing, the output field of this section is named "Effect Feedback". This effect feedback will be used as part of the historical baseline in the data admission scenario corresponding to main process S100 and input to S120 for extraction and drift verification. Simultaneously, it will be referenced by the strategy indication generated in threshold candidate generation in main process S200, and evaluated and fused in main process S300 for trigger item review and cause graph update. In summary, the technical effect of this step is: through the implementation of the execution list and the registration of feedback for multi-dimensional elements, this step establishes a unified execution-feedback link at the action layer, resource layer, and boundary layer. The output effect feedback serves as evidence input across main steps, supporting the continuous operation of subsequent data admission, threshold learning, and evaluation fusion.

[0045] Example 2: Figure 2 This diagram illustrates a structural block diagram of an intelligent production monitoring and management system for aerospace ring forgings according to an embodiment of the present invention. Figure 2 As shown, the structure may include: The micro-area grid and probe deployment and calibration module 01 is used to install environmental micro-area probes, equipment status probes, and process cycle probes, and complete coordinate binding, time anchor point registration, and channel health self-check. It provides multi-source raw data with location confidence tags to the data acquisition and access module. Specifically, it receives production line station topology, micro-area grid division, and equipment list as input, and completes the installation and numbering of environmental micro-area probes, equipment status probes, and process cycle probes according to station number and micro-area index. It establishes the coordinate binding relationship and time anchor between probes and stations, equipment, and micro-areas. Point registration and recording; before the probe enters the acquisition channel for the first time, a health self-check is performed on the sampling channel, covering channel open and short circuit detection, zero drift registration, range anomaly registration, and missing measurement marking, forming multi-source raw data with position confidence labels; when a missing measurement occurs in the acquisition channel, the missing measurement mark and start and end time are recorded and written to the self-check log, without valuation replacement; after completing the above processing, the multi-source raw data with position confidence labels is submitted to the data acquisition and access module as input, and the coordinate binding table and time anchor table are archived in this module for subsequent auditing and traceability calls.

[0046] Data acquisition and access module 02 is used to receive the multi-source raw data and perform calibration segment extraction, historical baseline association, drift verification, homogeneous deduplication, robust denoising of outlier segments, and minimum exposure screening to generate usable sequences, confidence label sets, and impact factor candidate sets. It then submits the usable sequences and impact factor candidate sets to the threshold learning and version management module. Specifically, it receives multi-source raw data with location confidence labels from the micro-area grid and probe deployment and calibration module, segments the data into calibration segments according to probe numbers and time anchors, associates historical baseline records, and completes time axis alignment. Drift verification is performed based on the stable segments of the historical baseline, and channels with shifts are labeled with drift levels and written into the drift list. For the same... Redundant entries for a physical quantity generated from multiple acquisitions undergo source deduplication, retaining entries with high source confidence and high location confidence matching. In robust denoising of abnormal segments, segments containing impact pulsations and short-term spikes are robustly filtered out while maintaining the original segment index. In minimum exposure screening, only necessary fields such as field name, dimension, time label, and location confidence label are retained, while auxiliary fields not required for judgment are removed. Usable sequences and confidence label sets are output, and candidate sets of impact factors are extracted from stable segments and critical channels. The usable sequences and candidate sets of impact factors are submitted to the threshold learning and version management module, while the confidence label set is passed to the evaluation fusion module as a constraint input, forming a chain relationship between the modules.

[0047] The threshold learning and version management module 03 is used to perform quantile adaptive threshold learning, probability switching, and rollback strategy configuration based on the available sequences and working condition labels, and to perform causal impact detection and boundary reinforcement on threshold candidates, outputting an adaptive threshold set and a threshold version mirror. The adaptive threshold set is provided to the evaluation and fusion module, and the threshold version mirror is provided to the multi-objective constraint and conflict resolution module. Specifically, it receives available sequences and influence factor candidate sets from the data acquisition and admission module, reads working condition labels registered from the production line scheduling side, performs quantile adaptive threshold learning according to learning groups and window granularity, and forms threshold candidates. In the probability switching and rollback strategy configuration, a system is established for each threshold candidate. Morphological registration, historical stable morphological citation, and rollback registration citation trigger morphological switching and rollback when insufficient sample coverage or short-term mutations occur. In causal impact detection, intervention markers returned by the execution and receipt registration modules are read, impact window controls are constructed, candidates affected by the intervention are marked, boundary reinforcement is carried out, and protection zone parameters are generated. The reinforced entries are aggregated into an adaptive threshold set within the learning group, and a threshold version mirror containing source annotations, version annotations, protection zone parameters, and rollback registration citations is generated. The adaptive threshold set is passed to the evaluation fusion module, and the threshold version mirror is passed to the multi-objective constraint and conflict resolution module. At the same time, the version genealogy is registered and stored in this module for subsequent auditing and retraining.

[0048] The evaluation fusion module 04 is used to perform data consistency measurement and online estimation of time-varying weights on the candidate set of impact factors and the adaptive threshold set under confidence label constraints, and to complete robust fusion of the environmental domain, equipment domain, and production line domain. It outputs evaluation records and confidence intervals, and submits anomaly trigger-related inputs to the anomaly interpretation and handling generation module. Specifically, it receives the candidate set of impact factors and the confidence label set from the data acquisition and admission module, and the adaptive threshold set from the threshold learning and version management module. It performs consistency measurement on the items and threshold forms at the learning group and window granularity, and produces a consistency benchmark area and a reference benchmark. Within the consistency benchmark area, time-varying weights are estimated online. Based on confidence labels, stable segment registration, and boundary hit records, the contribution of entries is upgraded or downgraded to form a window-level weight snapshot. In the intra-domain fusion stage, weighted aggregation and replacement event registration are performed on entries in the environmental domain, equipment domain, and production line domain respectively to generate intra-domain fusion sequences. In the cross-domain aggregation stage, weight allocation and time-series splicing are performed on the results of the three domains based on the window-level weight snapshot to form a learning group-level aggregation sequence and confidence boundary description, and output evaluation records and confidence intervals. The evaluation records and confidence intervals are submitted to the anomaly interpretation and handling generation module to form the input source for the interpretation process.

[0049] The anomaly interpretation and handling generation module 05 is used to identify trigger items, locate factors, and construct cause maps based on the evaluation records and confidence intervals, forming an anomaly evidence package and handling suggestions, and submitting the handling suggestions to the multi-objective constraint and conflict resolution module. Specifically, it receives the evaluation records and confidence intervals output by the evaluation fusion module, performs double-boundary consistency judgment on out-of-boundary candidates, and upgrades candidates that are in the same window and direction as the adaptive threshold item boundary hit record to trigger items; it backtracks the fusion sequence and window-level weight snapshot within the domain around the trigger items, locates items with high contribution and that have been replaced or weighted, and registers the item source, working condition label, and micro-area location. With time indexing; construct a cause graph covering workstations, micro-zones, equipment, items, operating condition labels, and threshold forms, and label trigger associations, replacement associations, rollback associations, and boundary hit associations; based on the graph node and edge information, compile trigger window snapshots, item-level contribution descriptions, replacement event summaries, and rollback registration references to form an anomaly evidence package, and assemble disposal suggestions according to the requirements of the safety barrier constraint module, covering cycle adjustment, operating status switching, short-term maintenance, standby machine replacement, and amplitude limiting strategies; submit the disposal suggestions to the multi-objective constraint and conflict resolution module, and archive the anomaly evidence package for the execution and receipt registration module and the threshold learning and version management module to read.

[0050] The multi-objective constraint and conflict resolution module 06 is used to receive disposal suggestions, threshold version images, and production resource images, as well as assembly output scale constraints, power consumption constraints, quality stability constraints, maintenance window constraints, and safety barrier constraints. It performs conflict resolution and generates priority sequences and equipment scheduling schemes, which are provided to the safety barrier and execution compilation module. Specifically, it receives disposal suggestions from the anomaly interpretation and disposal generation module, receives threshold version images from the threshold learning and version management module, and reads equipment occupancy, tooling status, standby machine list, and maintenance window from the production resource image. Based on the disposal suggestions... The system addresses constraints related to assembly output scale, power consumption, quality stability, maintenance window, and safety barrier. Items that touch the protection zone are marked with shadow execution or limiting strategies. In conflict resolution, it adjudicates contention for actions on the same equipment during the same period, concurrent occupation of resources across workstations, and actions running within the maintenance window. Based on switching costs, release times, and linkage relationships, it generates a sequence expression, forming a priority sequence. On the basis of the sequence expression, it reorganizes the action blocks at the workstation and equipment dimensions, writes pre-check, action issuance, and subsequent confirmation nodes, generates an equipment scheduling plan, and passes the priority sequence and equipment scheduling plan to the safety barrier and execution compilation module.

[0051] The safety barrier and execution compilation module 07 is used to extract execution units, safety barrier parameters, and priority sequences from the equipment scheduling scheme, complete priority sequence solving, amplitude limiting configuration, and execution list compilation, output safety barrier configuration and execution list, and submit the execution list to the execution and receipt registration module. Specifically, it receives the equipment scheduling scheme and priority sequence output by the multi-objective constraint and conflict resolution module, parses the execution units from the scheme, reads the safety barrier parameters, and identifies shadow execution markers; sorts the peer set according to the priority sequence, and generates synchronization constraints for execution units with linkage references; in amplitude limiting configuration, it configures upper and lower limits of amplitude, upper limit of duration, and minimum interval for each execution unit, and writes bypass trigger conditions and rollback candidates; for standby machine switching execution units, it generates blocking entries for incompatible situations and marks them as pending review; after completing the above processing, it compiles the execution list along the workstation, equipment, and time slice sequence, writes pre-check, action issuance, subsequent confirmation, and abnormal rollback nodes, outputs the safety barrier configuration and execution list, and submits the execution list to the execution and receipt registration module.

[0052] The execution and receipt registration module 08 is used to issue instructions, perform shadow execution and rollback calls based on the execution list, register and write effect receipts to form effect receipts, and return the effect receipts to the data acquisition and admission module as a source of historical baselines. At the same time, it provides the necessary records for auditing and retraining for the threshold learning and version management module and the evaluation fusion module. Specifically, the system receives the execution list and safety barrier configuration output by the safety barrier and execution compilation module. At the start of the time slice, it reads the occupancy status of the production resource image and the maintenance window to conduct pre-checks. When the pre-check shows a conflict, it registers a rejection record and alternative time slices and writes them into the receipt draft. When the pre-check passes, it issues actions based on the execution list and records the command number, issuance time, and target amplitude. During the action implementation, it collects channel data and aligns it with the time anchor point, registering the actual effective time, boundary hit, and bypass trigger. When the rollback condition is triggered, it calls the rollback candidate and records the equipment switching, channel switching, and amplitude contraction process. At the end of the time slice, it summarizes four elements: action implementation, boundary and rollback, resources and maintenance, and linkage consistency, forming a time slice receipt record and accumulating them into batch receipt records at the workstation and equipment dimensions, and summarizing them to generate an effect receipt. The effect receipt is fed back to the data acquisition and access module as a source of historical baselines, and simultaneously written into the version audit channel of the threshold learning and version management module and the trigger review channel of the evaluation fusion module, closing the back-and-forth link between data and control.

Claims

1. A method for intelligent production monitoring and management of aerospace ring forgings, characterized in that, include: Obtain the micro-area grid, probe deployment scheme and equipment list, and perform coordinate binding, time anchor point registration, channel health self-check, drift verification, homogeneity deduplication, working condition label mapping, abnormal segment robust denoising and minimum exposure screening to generate the candidate set structure of impact factors; The structure of the candidate set of impact factors is obtained, and quantile adaptive threshold learning, threshold learning input set construction, probability switching, and backoff strategy configuration are performed. Item-level threshold candidates containing upper and lower boundaries, candidate positions, morphological labels and applicable working conditions are generated and aggregated into learning group-level threshold candidate units. Causal impact detection and boundary reinforcement processing are performed to generate an adaptive threshold set. Based on an adaptive threshold set, data consistency measurement, online estimation of time-varying weights, robust fusion, trigger item identification, factor-level localization and cause graph construction are performed to generate a disposal suggestion structure. Based on the proposed disposal structure, the system performs multi-objective constraint assembly, conflict resolution, priority sequence solving, amplitude limit configuration, execution list compilation, implementation, effect feedback registration and feedback writing, and generates an effect feedback structure.

2. The method according to claim 1, characterized in that, The microgrid and probe deployment scheme and equipment list include: The micro-grid refers to dividing the production line area into stable regions with fixed boundaries and numbers according to heating stations, forming stations, heat treatment stations, bulging stations, straightening stations, grinding stations, and inspection stations; the probe deployment scheme refers to the written configuration of the installation positions, installation methods, numbering rules, and channel mapping relationships of environmental micro-area probes, equipment status probes, and process cycle probes; the equipment list refers to the list of all monitoring-related equipment names, equipment numbers, station affiliations, unit levels, maintenance record entry points, and available sensor channels within the production line.

3. The method according to claim 1, characterized in that, The process of performing causal impact detection and boundary reinforcement also includes: Intervention markers are read from the execution records of the previous or adjacent cycles, and the corresponding sample intervals of the candidates are aligned with the intervention markers on the time axis. Parallel control groups with impact windows and non-impact windows are constructed based on the alignment relationship.

4. The method according to claim 1, characterized in that, The process of boundary reinforcement also includes: For items affected by the intervention, a guard band is applied to the upper and lower boundaries of the item-level candidates. The width of the guard band is related to the dispersion description and sample coverage within the learning group, and a constraint condition for preferentially triggering bypass backoff is set within the guard band.

5. The method according to claim 1, characterized in that, The process of generating an adaptive threshold set also includes: The item-level hardening results are aggregated within the learning group to generate a learning group-level threshold item set. The references to the probability switching registration domain and rollback registration domain used in the current period are recorded at the set layer. For all learning groups and all impact factor items, a hardened threshold form set is formed. The source label, version label, impact registration domain, temporary output registration domain, and protection band parameter description are recorded at the set layer to construct an adaptive threshold set. At the same time, a threshold version mirror structure is generated. The snapshot of the threshold form, item source, form label, protection band parameter, rollback registration reference, and probability switching registration reference of all items in the current period is recorded by period as the organizational unit. The pointer of the previous period's mirror is written into the mirror to form a continuous spectrum.

6. The method according to claim 1, characterized in that, The process of performing data consistency measurement, online estimation of time-varying weights, robust fusion, trigger identification, factor-level localization, and causal graph construction also includes: The correspondence between the aggregated sequence and the confidence interval is read from the evaluation record. When the aggregated sequence crosses the boundary between any window and the upper or lower boundary of the confidence interval, a trigger candidate is registered. When the boundary crossing relationship and the boundary hit record of the adaptive threshold entry appear in the same window and are in the same direction, the trigger candidate is upgraded to a trigger item, and a source label and window index are attached to the trigger item. Within the trigger window, the intra-domain fusion sequence and window-level weight snapshot are backtracked to retrieve the entry that has the highest contribution to cross-domain convergence within the window and has experienced a replacement event or weight jump in intra-domain fusion. Possible causal factors are registered at the item level.

7. The method according to claim 1, characterized in that, The process of constructing the causal map also includes: A multi-layered node graph covering workstations, micro-areas, equipment, items, working condition labels, and threshold patterns is constructed. The edge types include trigger association, replacement association, fallback association, and boundary hit association. Each edge is accompanied by a trigger window, direction description, and source label. When multiple trigger items appear consecutively in adjacent windows within the same learning group, the cause graph establishes a time sequence among these trigger items and records the cluster identifier.

8. The method according to claim 1, characterized in that, The process of generating the disposal recommendation structure includes: Based on the trigger items and the node and edge information in the graph, abnormal evidence elements are compiled, including the correspondence between trigger window snapshot, trigger item source label, item-level contribution description, replacement event summary, rollback registration reference and adaptive threshold form, and the above elements are organized into abnormal evidence packages; each package corresponds to a trigger item or a group of consecutive trigger items, and contains indexes and references for review.

9. The method according to claim 1, characterized in that, The process of generating the abnormal evidence package and handling recommendation structure also includes: The candidate strategy modules include reducing workstation load, switching equipment operating status, adjusting cycle time parameters, requesting short-term maintenance, using backup machines, and limiting operating boundaries. The assembly logic of the candidate strategy modules reads the weight parameters, the source and version labels of the adaptive threshold entry protection band and the threshold version mirror, and forms a structured suggestion that can be read by the executable system.

10. An intelligent production monitoring and management system for aerospace ring forgings, applied to the method described in any one of claims 1-9, characterized in that, include: The micro-area grid and probe deployment and calibration module is used to install environmental micro-area probes, equipment status probes and process cycle probes and complete coordinate binding, time anchor point registration and channel health self-check, and provide multi-source raw data with location confidence labels to the data acquisition and access module; The data acquisition and admission module is used to receive raw data from multiple sources and perform calibration fragment extraction, historical baseline association, drift verification, homogeneous deduplication, robust denoising of outlier segments and minimum exposure screening to generate usable sequences, confidence label sets and impact factor candidate sets. The threshold learning and version management module is used to perform quantile adaptive threshold learning, probability switching, and rollback strategy configuration based on available sequences and operating condition labels. The evaluation fusion module is used to perform data consistency measurement and online estimation of time-varying weights on the candidate set of impact factors and the adaptive threshold set under confidence label constraints. The anomaly interpretation and handling generation module is used to identify triggers, locate factors, and construct cause maps based on assessment records and confidence intervals, forming an anomaly evidence package and handling recommendations. The multi-objective constraint and conflict resolution module is used to receive disposal suggestions, threshold version images and production resource images, assembly output scale constraints, power consumption constraints, quality stability constraints, maintenance window constraints and safety barrier constraints, perform conflict resolution and generate priority sequences and equipment scheduling schemes, which are then provided to the safety barrier and execution programming module. The safety barrier and execution compilation module is used to extract execution unit and safety barrier parameters and priority sequences from the equipment scheduling scheme, complete priority sequence solving, limit configuration and execution list compilation, and output safety barrier configuration and execution list; The execution and receipt registration module is used to issue instructions, perform shadow execution and rollback calls based on the execution list, and register and write effect receipts to form effect receipts.

Images