Multi-base calendering process knowledge sharing and optimization method based on federated learning

Abstract

The application discloses a multi-base calendering process knowledge sharing and optimization method based on federal learning, relates to the technical field of battery manufacturing and intelligent process control, and the application converts multi-source process data into process state facts in each production base, and further constructs process knowledge units containing at least premise state, process action, result state and applicable boundary, wherein a federal coordination end only receives local knowledge update packages and does not receive original process data, process knowledge units with similar premise states are subjected to rule alignment, conflict resolution and fusion in a process similar cluster, and then a target base calls a matching knowledge subset according to a material system identifier, an equipment type identifier and a process path identifier, and executes local reasoning, so that multi-base calendering experience can be deposited into shared knowledge with applicable conditions, and directly mapped into specific parameter adjustment actions such as calendering pass arrangement, single pass reduction amount, roller temperature adjustment and line speed adjustment.

Description

Technical Field

[0001] This invention relates to the field of battery manufacturing and intelligent process control technology, and in particular to a method for knowledge sharing and optimization of multi-base calendering processes based on federated learning. Background Technology

[0002] Dry-process thick electrodes refer to thick electrodes formed using solvent-free mixing and fiberization processes, including dry-process electrode sheets and all-solid-state battery thick electrode sheets. These thick electrodes typically employ a few-pass calendering process under high load conditions to balance sheet formation, compaction state, interfacial contact, and continuous manufacturing. During the few-pass calendering process, the sheet state is simultaneously affected by the degree of fiberization, premixing uniformity, thermal history, and the force path between the rollers. When the material system, equipment type, or process path changes, the sheet is prone to edge cracks, springback fluctuations, unstable compaction, and changes in solid-solid contact. Therefore, directly reusing the same calendering experience across different production sites often results in unstable parameter adjustments, inconsistent triggering, and repeated on-site confirmation.

[0003] In existing technologies, one approach involves developing a process formula based on trial production results within a single production base, and then making local adjustments based on roll temperature, linear pressure, roll gap, linear speed, or thickness detection results. Another approach introduces federated learning under multi-base conditions, achieving collaborative modeling or parameter aggregation by storing raw data from each base locally. While these approaches can support local optimization or cross-base sharing to some extent, the shared content typically remains at the level of model parameters, global models, or fragments of empirical rules, failing to organize calendering experience into callable knowledge with preconditions, process actions, result states, and applicable boundaries. Thus, when there are differences in material systems, equipment structures, sensor apertures, and process paths across different production bases, cross-base shared results are difficult to directly correspond to specific parameter adjustments for the current batch.

[0004] Furthermore, existing multi-site sharing methods lack a unified alignment and conflict resolution mechanism for experience records with similar preconditions but different process actions. For calendering experience that only applies to certain material systems, equipment types, or process paths, existing technologies also lack differentiated expressions oriented towards applicable boundaries. As a result, after aggregation, shared content uploaded from different sites easily leads to mixed scopes of application, unclear calling priorities, and coexisting conflicting rules. This makes it difficult for the target site to determine which action to take when calling shared results, resulting in mis-triggering, frequent switching, and result fluctuations in the field execution of calendering adjustment strategies.

[0005] On the other hand, existing cross-site sharing solutions typically use on-site execution results for local model updates or experience corrections, but lack a closed-loop recording method to establish a correspondence between batch execution information, corresponding knowledge base, and post-execution results. Thus, if the post-calendering wafer state deviates from expectations, it becomes difficult to trace back the shared basis invoked in that adjustment, its applicable scope, and whether subsequent invocation priorities should be lowered. This makes it difficult for the shared results to be continuously iterated into a stable and usable basis for calendering decisions.

[0006] Therefore, under the condition that the original process data of each production base is kept locally, how to organize the scattered experience in the multi-base, few-pass rolling process into shared knowledge with applicable boundaries, and complete the conflict resolution after cross-base aggregation, generate specific rolling adjustment strategies for the target base, and continuously write back and update based on the execution results has become an urgent technical problem to be solved. Summary of the Invention

[0007] This application provides a method for knowledge sharing and optimization of multi-base calendering processes based on federated learning, which solves the problems of difficult implementation of knowledge sharing, rule conflicts, and untraceable decisions in multi-base, few-pass calendering.

[0008] This invention provides a knowledge sharing and optimization method for multi-base rolling processes based on federated learning, applicable to the scenario of few-pass rolling of dry thick electrodes, including: Each production base collects multi-source process data of its own few-pass calendering process and generates process state facts locally based on a unified state vocabulary and observation mapping rules. Each production base updates its local state extraction model based on the multi-source process data, constructs local process knowledge units based on the process state facts, and generates a local knowledge update package. The process knowledge unit includes at least the premise state, process action, result state, applicable boundary, evidence source identifier, and evidence weight. The local knowledge update package includes newly added, revised, or invalid process knowledge units and model update results. The federal coordination terminal only receives the local knowledge update package. After clustering according to the base process profile, it first performs rule alignment on process knowledge units with similar premise states within the same process similar cluster. Then, it resolves and merges conflicts according to the overlap of applicable boundaries, evidence weight, and consistency of execution results, forming a common knowledge subset, a base-specific knowledge subset, and a conflict pending set. The target base retrieves a subset of knowledge that matches the material system identifier, equipment type identifier, and process path identifier of the target base based on the current batch's process status facts, performs local reasoning, and outputs a calendering adjustment strategy; After the target base executes the rolling adjustment strategy, it writes back the batch identifier, the corresponding process knowledge unit identifier, the action execution information, and the result status to the local machine to update the local process knowledge unit and generate a new local knowledge update package, which is then used by the federated coordinator to iteratively update the shared knowledge set.

[0009] In some embodiments, the multi-source process data includes at least two of the following: calendering equipment operation data, sheet online detection data, calendering pretreatment data, and calendering post-calendering quality characterization data; The process state facts include at least two of the following: fibrous state, densification state, risk of edge cracking, springback state, solid-solid contact stability, thermal history offset, and stress offset.

[0010] In some embodiments, the process actions include at least one of the following: rolling pass arrangement, single pass reduction, roll temperature adjustment, line speed adjustment, and pretreatment callback; The applicable boundaries include at least the material system identifier, equipment type identifier, and process path identifier.

[0011] In some embodiments, the base process profile consists of a material system identifier, a pretreatment path identifier, a calendering equipment type identifier, and a sheet structure identifier; the federal coordination end divides multiple production bases into at least one process similarity cluster based on the base process profile, and forms the shared knowledge set in each process similarity cluster.

[0012] In some embodiments, the federated coordination end first performs rule alignment on process knowledge units that are in the same process similar cluster and have similar premise states. The rule alignment includes unifying the premise state expression, unifying the process action field and unifying the result state field. For process knowledge units that have completed rule alignment, the degree of overlap of applicable boundaries is calculated based on the degree of overlap of material system identifiers, equipment type identifiers, and process path identifiers, and the fusion order is determined by combining evidence weights and execution result consistency. For process knowledge units whose applicable boundary overlap reaches the corresponding trigger threshold, whose process action direction is consistent, whose risk level is consistent, and whose result state convergence intervals have an intersection, they are merged and written into the common knowledge subset. For process knowledge units whose applicable boundaries differ in at least one of the production base, equipment type, and process path, the differentiated boundaries are retained and written into the base-specific knowledge subset. For process knowledge units that conflict with process actions and meet at least one of the following conditions, write them into the conflict pending set and restrict their use to the corresponding process similar cluster: the convergence intervals of the result states do not intersect; the execution result has not converged. The conflict pending set includes at least the prerequisite status identifier, conflict action identifier, conflict cause identifier, restricted call scope identifier, and pending verification status identifier.

[0013] In some embodiments, the target base performs forward reasoning within the common knowledge subset and the base-specific knowledge subset corresponding to the base through a local reasoning engine; The process knowledge units with overlapping trigger boundaries are filtered according to the action risk level and evidence weight, and the output includes calendering adjustment strategies including calendering pass arrangement, roll temperature adjustment direction, line speed adjustment direction, single pass reduction amount, and pre-processing callback trigger conditions.

[0014] In some embodiments, when generating process status facts, each production base first establishes a state mapping table between observation items and a unified state vocabulary based on local sensor configuration. The state mapping table includes at least the observation item identifier, corresponding state name, time window width, time anchor point type, window aggregation method, dimensionless reference interval, direction attribute, reference detection position, position conversion coefficient, and missing data handling method. For cases where the same state name corresponds to multiple observations, time alignment is first performed according to the time window width and time anchor type, then a unified dimensional expression is completed according to the dimensional normalization reference interval and direction attribute, and finally, the observation data of different detection positions are converted into a unified state expression according to the reference detection position and position conversion coefficient. The unified state representation includes at least a state name, state level, time interval, location interval, state source, and state confidence flag; For missing observations, a missing marker is written and the partial state facts formed by existing observations are retained. For abnormal fluctuation observations, an abnormal marker is written and the evidence weight of the process knowledge unit in which they participate is reduced.

[0015] In some embodiments, when the number of local effective process knowledge units of the target base is lower than a preset threshold, the target base calls an initial knowledge subset from the shared knowledge set according to the material system identifier, equipment type identifier and process path identifier, and generates an initial rolling adjustment strategy within a preset safety range; When the write-back result status reaches the preset consistency condition, the initial knowledge subset is replaced with the target knowledge subset containing local process knowledge units.

[0016] In some embodiments, the resulting state includes at least two of the following: sheet integrity state, compaction state, and solid-solid contact state. When writing back the result status, each production base simultaneously writes the batch identifier, equipment identifier, action execution time, and corresponding process knowledge unit identifier to form a knowledge traceability chain.

[0017] In some embodiments, when the inconsistency between the result status of the target base write-back and the result status corresponding to the process knowledge unit exceeds a preset threshold, the process knowledge unit is marked as pending verification, and the calling priority of the process knowledge unit is reduced. When multiple batches consecutively meet the preset consistency conditions, the applicable boundaries and evidence weights of the process knowledge unit are updated.

[0018] Through the above technical solution, the present invention can achieve at least the following beneficial effects: This invention transforms multi-source process data into process state facts locally at each production base, and further constructs process knowledge units that include at least prerequisite states, process actions, result states, and applicable boundaries. The federated coordination end only receives local knowledge update packages and not the original process data. Within process similarity clusters, process knowledge units with similar prerequisite states are aligned by rules, conflict resolution, and fusion. Then, the target base calls the matching knowledge subset according to material system identifier, equipment type identifier, and process path identifier and performs local inference. This enables the multi-base calendering experience to be precipitated into shared knowledge with applicable conditions, and directly mapped to specific parameter adjustment actions such as calendering pass arrangement, single pass reduction, roll temperature adjustment, and line speed adjustment. After execution, the batch identifier, corresponding process knowledge unit identifier, action execution information, and result state are written back to the local machine and participate in the iterative update of the shared knowledge set. This forms a closed-loop traceability chain from knowledge call, action execution, to result feedback, thereby addressing the problems of difficult implementation of shared knowledge, rule conflicts, and untraceable decisions in multi-base, low-pass calendering.

[0019] By unifying the state vocabulary and observation mapping rules, observation data with different dimensions, sampling frequencies, and detection locations in different production bases are converged into a unified state expression. This reduces the triggering deviation caused by differences in sensor aperture to knowledge sharing and retrieval, and enables process states such as fibrous state, densification state, edge crack risk, springback state, solid-solid contact stability, thermal history offset, and stress offset to have a consistent expression caliber under cross-base conditions. This provides a reusable input basis for subsequent knowledge alignment, conflict resolution, and local reasoning.

[0020] By introducing applicable boundaries, evidence source identifiers, and evidence weights into process knowledge units, and combining the overlap of applicable boundaries, evidence weights, and consistency of execution results at the federal coordination end to perform conflict resolution, it is possible to separate and store calendering experience that is only valid under certain material systems, certain equipment types, or certain process paths from calendering experience that is valid across multiple bases. For knowledge units with conflicting process actions and unconverged results, a conflict pending set is retained and calls are restricted to the corresponding process similar clusters. This can reduce the risk of miscalling, mismerging, and rule jitter when directly reusing the same set of parameter tuning rules across bases.

[0021] By calling matching knowledge subsets based on material system identifiers, equipment type identifiers, and process path identifiers at the target base, and updating the evidence weight, applicable boundaries, and calling priority of process knowledge units according to the result status after execution, shared knowledge can be transformed from a static set of experience into a dynamic decision-making basis that is continuously corrected with the field results. For newly imported bases or bases with scarce labels, an initial knowledge subset can be called first to generate an initial rolling adjustment strategy constrained by safety boundaries, and then gradually replaced with a target knowledge subset containing local process knowledge units as the local result status is written back. This can reduce the dependence on complete labels and long-term trial production accumulation when importing new lines.

[0022] By associating batch identifiers, process knowledge unit identifiers, action execution information, and result statuses with the write-back link, a correspondence can be established between each rolling adjustment strategy and its triggering basis, execution content, and post-execution result. When the result status is inconsistent with the result status corresponding to the process knowledge unit, the corresponding process knowledge unit can be transferred to the pending verification state and the calling priority can be reduced. When multiple batches continuously meet the consistency conditions, the applicable boundary and evidence weight are updated. This can suppress rule mis-triggers caused by contact state fluctuations, changes in edge crack risk, and changes in equipment force path in low-pass rolling, and maintain the write-back capability, verifiability, and sustainable optimization of the shared knowledge set in multi-site continuous operation. Attached Figure Description

[0023] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation on the scope of this application.

[0024] Figure 1 This is a flowchart of the knowledge sharing and optimization method for multi-base calendering process based on federated learning in the embodiments. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0026] All terms used in this application (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0027] To facilitate understanding of the relevant objects in this embodiment, the terms used in the embodiment are defined as follows: Dry-process thick electrodes refer to thick electrodes formed using solvent-free mixing and fiberization methods, including dry-process electrode sheets and all-solid-state battery thick electrode sheets. A unified state lexicon is a set of state terms used to uniformly express the states of the calendering process at different production bases. Observation mapping rules are mapping relationships that convert observation data from different sources, with different dimensions, different sampling frequencies, and different detection locations into a unified state expression. Process state facts are structured records generated based on the unified state lexicon and observation mapping rules, used to characterize the current calendering process state, and at least include state name, state level, state source, time interval, batch identifier, and state confidence marker. The fact record table is a record table storing process state facts, and at least includes batch identifier, state name, state level, time interval, location interval, state source, state confidence marker, and version identifier. The version index table is an index table binding state mapping versions, threshold versions, and location conversion versions, using batch identifiers and process knowledge unit identifiers as alignment keys.

[0028] The Federated Knowledge Blackboard refers to a shared knowledge carrier maintained by the federated coordinator, used to receive local knowledge update packages uploaded by each production base and perform clustering, conflict resolution, fusion, indexing, and version updates. A process knowledge unit refers to a knowledge record that stores rolling experience in a structured form. A local knowledge update package refers to an updated data set consisting of newly added, revised, or invalidated process knowledge units and their corresponding version information. The knowledge update record table is a record table storing local knowledge update packages, and includes at least the process knowledge unit identifier, update type, update time, evidence source identifier, evidence weight, applicable boundary field, and invalidation status identifier. A base process profile refers to a structured set of identifiers used to characterize the process similarity of production bases. A process similarity cluster refers to a set of production bases with similar base process profiles. The shared knowledge set refers to a set of process knowledge that can be called by the target base after fusion by the federated coordinator, and includes at least a common knowledge subset, a base-specific knowledge subset, a conflict-pending set, a version identifier, and a calling scope identifier. The applicable boundary overlap refers to the degree of overlap between two process knowledge units in terms of material system identifier, equipment type identifier, and process path identifier. Execution result consistency refers to the degree of convergence of the result state after multiple batches of execution under similar preconditions for the same process knowledge unit. The target base refers to the production base currently invoking the shared knowledge set and executing the rolling adjustment strategy. The rolling adjustment strategy refers to the set of process actions output by local inference. Example 1:

[0029] like Figure 1As shown, this embodiment revolves around the processing chain of transforming shared results into specific parameter adjustment actions during multi-site, low-pass calendering. It forms a closed-loop connection between cross-site knowledge fusion and on-site action issuance by locally generating process state facts, locally constructing process knowledge units, executing rule alignment and conflict resolution on the federated side, and the target site performing local inference and writing back the result state. Based on this processing chain, cross-site calendering experience is organized into shared knowledge with applicable boundaries and evidence weights, and a correspondence is established with the material system identifier, equipment type identifier, and process path identifier of the current batch at the target site. This enables the calendering adjustment strategy to have a directly implementable, rewritable, and traceable implementation method.

[0030] Specifically, this embodiment proposes a knowledge sharing and optimization method for multi-base rolling processes based on federated learning, applied to the scenario of dry thick electrode rolling with few passes, including the following process: Step S1: Each production base collects multi-source process data of the few-pass calendering process at its own base, and generates process state facts locally based on a unified state vocabulary and observation mapping rules. Step S2: Each production base updates its state extraction model or knowledge evaluation model locally based on multi-source process data, and writes the locally learned state extraction results, evidence weight revision results, and applicable boundary revision results into the local knowledge update package. Based on the facts of the process status, a local process knowledge unit is constructed. The process knowledge unit includes at least the premise status, process action, result status, applicable boundary, evidence source identifier and evidence weight. A local knowledge update package is generated, which is an update data package containing newly added, revised or invalid process knowledge units. Step S3: The federal coordination end only receives local knowledge update packages. After clustering according to the base process profile, it first performs rule alignment on process knowledge units with similar premise states within the same process similar cluster. Then, it resolves and merges conflicts according to the overlap of applicable boundaries, evidence weight, and consistency of execution results, forming a common knowledge subset, a base-specific knowledge subset, and a conflict pending set. Step S4: The target base calls the knowledge subset that matches the material system identifier, equipment type identifier, and process path identifier of the target base based on the current batch's process status facts, performs local reasoning, outputs the calendering adjustment strategy, and writes back the result status. Step S5: After the target base executes the rolling adjustment strategy, it writes back the batch identifier, the corresponding process knowledge unit identifier, the action execution information and the result status to the local machine to update the local process knowledge unit and generate a new local knowledge update package, which is then used by the federated coordination end to iteratively update the shared knowledge set. In this embodiment, the multi-source process data includes at least two of the following: calendering equipment operation data, sheet online detection data, calendering pretreatment data, and calendering post-calendering quality characterization data. The process state facts include at least two of the following: fibrous state, densification state, edge crack risk, springback state, solid-solid contact stability, thermal history offset, and stress offset; solid-solid contact stability refers to the state item used to characterize the contact continuity between active particles, solid electrolyte particles, and conductive phase.

[0031] In this embodiment, calendering equipment operation data refers to process observation data generated during the operation of the calendering equipment, including at least roll temperature, linear speed, roll gap position, linear pressure, pass number, and operating time period. Sheet online inspection data refers to data obtained by online observation of the sheet state before, during, or after calendering, including at least thickness, width, surface defect markings, edge condition, length position, and sheet uniformity characterization values. Pre-calendering treatment data refers to data related to sheet forming before calendering, including at least batch identification of mixing materials, fiberization treatment records, premixing state records, pre-compression records, and heat treatment records. Post-calendering quality characterization data refers to data obtained by characterizing the sheet quality after calendering, including at least compaction state, springback state, integrity state, solid-solid contact state, and offline inspection result identification.

[0032] In this embodiment, the fibrous state refers to the state item characterizing the degree of fibrosis of polytetrafluoroethylene; the densification state refers to the state item characterizing the degree of sheet compaction and pore shrinkage; the edge crack risk refers to the state item characterizing the tendency for cracks to occur at the edges of the sheet; the springback state refers to the state item characterizing the thickness recovery and deformation recovery of the sheet after calendering; the solid-solid contact stability refers to the state item characterizing the contact continuity between active particles, solid electrolyte particles, and conductive phase; the thermal history offset refers to the state item characterizing the degree of deviation of the actual thermal history from the process path reference thermal history; and the stress offset refers to the state item characterizing the degree of deviation of the inter-roll stress state from the target stress path. Each type of process state fact corresponds to a state source, state level, and time interval to support subsequent knowledge generation, fusion, and write-back.

[0033] Specifically, the status data is categorized and recorded according to the following: calendering equipment operation data, wafer online inspection data, pre-calendering processing data, and post-calendering quality characterization data. Time intervals are uniformly recorded using batch clocks, and location intervals are uniformly recorded along the wafer length and width directions. Process status facts use batch identifiers, status names, and time anchors as fact index keys. When missing or abnormal markers exist, the corresponding process status facts are still retained, but the status confidence flag is set to low confidence, and its independent triggering of process knowledge unit calls is restricted.

[0034] In this embodiment, the process knowledge unit also includes evidence source identifier and evidence weight; The process actions must include at least one of the following: calendering pass arrangement, single pass reduction, roll temperature adjustment, line speed adjustment, and pretreatment revert; the applicable boundaries must include at least the material system identifier, equipment type identifier, and process path identifier. In this embodiment, the process knowledge unit includes at least a set of prerequisite states, a set of process actions, a set of result states, a set of applicable boundaries, an evidence source identifier, an evidence weight, a generation time identifier, a last update time identifier, and a failure state identifier. The set of process actions includes at least the calendering pass arrangement, single-pass reduction, roll temperature adjustment direction, linear speed adjustment direction, and pre-processing callback actions. The set of applicable boundaries includes at least a material system identifier, equipment type identifier, process path identifier, sheet structure identifier, and calling scope identifier. The evidence weight is updated as the execution result is written back; when the written-back result is consistently consistent with the result state, the evidence weight of the corresponding process knowledge unit is increased; when the written-back result is consistently inconsistent with the result state, the evidence weight of the corresponding process knowledge unit is decreased, and the failure state identifier is updated.

[0035] In this embodiment, the base process profile consists of material system identifier, pretreatment path identifier, calendering equipment type identifier, and sheet structure identifier; the federal coordination end divides multiple production bases into at least one process similarity cluster based on the base process profile, and forms a shared knowledge set in each process similarity cluster. In this embodiment, the base process profile consists of multiple fields characterizing the base manufacturing conditions and process path features. The base process profile includes at least the material system identifier, pretreatment path identifier, calendering equipment type identifier, roll system structure identifier, sheet structure identifier, online detection configuration identifier, and production line operation constraint identifier. The material system identifier characterizes the cathode system, anode system, solid electrolyte system, or composite system category. The pretreatment path identifier characterizes the combination of mixing, fiberization, pre-pressing, and heat treatment paths. The calendering equipment type identifier characterizes the roll system structure, heating method, and loading method. The sheet structure identifier characterizes the sheet thickness level, target load level, and interlayer structure category.

[0036] The federal coordination unit categorizes multiple production bases based on the consistency and similarity of fields in their process profiles. Multiple production bases with identical key fields in their process profiles are grouped into the same process similarity cluster; production bases with significant differences in key fields are grouped into different process similarity clusters, each with its own differentiated knowledge index. Process similarity clusters are used to define the scope of knowledge fusion and knowledge retrieval, ensuring that shared knowledge sets are preferentially accessed under similar material systems, equipment conditions, and process paths.

[0037] In this embodiment, the federated coordination end first performs rule alignment on process knowledge units that are in the same process similar cluster and have similar premise states. The rule alignment includes unifying the premise state expression, unifying the process action field and unifying the result state field. For process knowledge units that have completed rule alignment, the degree of overlap of applicable boundaries is calculated based on the degree of overlap of material system identifiers, equipment type identifiers, and process path identifiers, and the fusion order is determined by combining evidence weights and execution result consistency. For process knowledge units whose applicable boundary overlap reaches the corresponding trigger threshold, whose process action direction is consistent, whose risk level is consistent, and whose result state convergence intervals have an intersection, they are merged and written into the common knowledge subset. For process knowledge units whose applicable boundaries differ in at least one of the production base, equipment type, and process path, the base-specific knowledge subset is written after retaining the differences in the boundaries. For process knowledge units that have conflicting process actions and whose result states convergence intervals do not intersect, write them into the conflict pending set and restrict their use to the corresponding process similar clusters; For process knowledge units that have conflicting process actions and whose execution results have not converged, write them into the conflict pending set and restrict their use to the corresponding process similar clusters. The conflict pending set must include at least the prerequisite status identifier, conflict action identifier, conflict cause identifier, restricted call scope identifier, and pending verification status identifier; In this embodiment, the sheet structure identifier is used for dividing process-similar clusters or to further limit the overlap of applicable boundaries. The conflict pending set includes at least a prerequisite state identifier, a conflict action identifier, a conflict cause identifier, a limited call scope identifier, and a status identifier to be verified.

[0038] In this embodiment, the federal coordination end first performs rule alignment on process knowledge units with similar premise states within the same process similar cluster, and then determines the fusion order based on the overlap of applicable boundaries, evidence weight and consistency of execution results, and writes the fusion results into the common knowledge subset, the base-specific knowledge subset or the conflict pending set respectively.

[0039] The steps for conflict resolution based on the overlap of applicable boundaries, the weight of evidence, and the consistency of execution results include: Within the same cluster of similar processes, the federal coordination unit first identifies process knowledge units with similar preconditions and competing process actions as objects to be resolved. Similar preconditions can be determined based on core state items in a unified state lexicon: among fibrous state, densification state, edge crack risk, and springback state, if the number of differences does not exceed two, and the state level deviation of each difference item does not exceed one level, then the preconditions are considered similar. Competing process actions refer to different calendering pass arrangements but the same target result state, or at least one of the following: single-pass reduction, roll temperature adjustment direction, or linear speed adjustment direction is opposite.

[0040] Process knowledge unit for the object to be digested and process knowledge unit First, calculate the applicable boundary overlap: ; in, For process knowledge unit With process knowledge unit The degree of overlap of applicable boundaries; For the overlap of the material system; Overlap of calendering equipment types; This refers to the overlap of process paths; , , These are the weighting coefficients corresponding to the degree of overlap, and .

[0041] The overlap of material systems is calibrated using a layered approach. For example: 1 for completely identical material system identifiers; 0.7 for identical main material families but different proportion ranges; 0.4 for identical active component families; and 0 for the rest. The overlap of calendering equipment types is calibrated using a layered approach based on equipment model, pressure rating, and roll surface structure: 1 for identical equipment models; 0.8 for identical equipment types and pressure ratings; 0.5 for identical equipment types; and 0 for the rest. The overlap of process paths is calibrated based on the combination of pretreatment and calendering paths: 1 for completely identical paths; 0.6 for only a single different path segment; 0.3 for two or more different segments; and 0 for the rest. In the few-pass calendering of dry thick electrodes, the material system exerts a stronger constraint on the final state. 0.5 is acceptable. 0.3 is acceptable. 0.2 is acceptable. The value is limited to between 0 and 1. The reference trigger threshold can be 0.65. If it is lower than this value, no direct conflict processing will be performed, and only the calling range will be reserved according to their respective applicable boundaries. The applicable boundary overlap trigger threshold and the conflict separation amount judgment threshold are only used when the threshold version identifier of the corresponding process similar cluster is consistent. When there are fewer than 10 valid write-back batches in the corresponding process similar cluster, or when it is a new material introduction trial production batch, the applicable boundary overlap trigger threshold is increased to 0.75, the conflict separation amount judgment threshold is increased to 0.15, and the overlapping boundary part is stopped from being written to the common knowledge subset. Only the base-specific knowledge subset or the conflict pending set is allowed to be written.

[0042] The weight of evidence for a single process knowledge unit is calculated by combining the support of batch number, the stability of results, and the freshness of update time to form a comprehensive measure of evidence. ; in, For process knowledge unit Weight of evidence; Supported by batch number; For the stable quantity of the result; To update the freshness of the data; These are the weighting coefficients for the corresponding quantities, and .

[0043] The batch number support value is calibrated according to the effective batch saturation method. For example, when the cumulative number of effective batches reaches 20, it is 1; when it is less than 20 batches, it is linearly proportional, but not less than 0.2. The result stability value is calibrated according to the result status fluctuation of the most recent 8 effective batches. When the fluctuation of the sheet integrity state, compaction state, and solid-solid contact state all fall within 5% of their respective allowable bandwidths, it is calibrated according to 1; when the fluctuation reaches 10% of the allowable bandwidth, it drops to 0.6; and when it exceeds 10% of the allowable bandwidth, it is not higher than 0.4. The update time freshness value is calibrated according to the most recent update time. It is calibrated according to 1 for updates within 30 days, 0.7 for updates between 31 and 90 days, and 0.4 for updates over 90 days. In the evidence weights, the batch number support value reflects repeatability, the result stability value reflects the degree of fluctuation convergence, and the update time freshness value reflects the adaptability to the current working conditions. The values ​​can be 0.45, 0.35, and 0.20 respectively. This evidence weight is updated once after each result status write-back; when 20 new valid write-back batches are added within the same process-similar cluster, the weight can be adjusted accordingly. Recalibrate once, with each adjustment controlled within 0.05. If equipment malfunctions, key observation items are missing, or new materials are introduced into the trial batch, pause the coefficient recalibration and update the corresponding process knowledge unit. The voltage limit is as low as 0.4.

[0044] The formula for calculating the consistency of execution results is: ; in, For process knowledge unit Consistency of execution results; For process knowledge unit The number of valid batches entering the statistical window under similar preconditions; For process knowledge unit In the Deviation in result status within each valid batch; For process knowledge unit The corresponding result state deviates from the upper limit, according to the allowable bandwidth of the sheet integrity state, compaction state, and solid-solid contact state, and adopts the same as... Synthesized using the same weighted caliber.

[0045] The deviation from the expected state is calculated using a weighted sum of normalized deviations from the sheet integrity state, compaction state, and solid-solid contact state. The reference benchmark is the average of the expected state values ​​from the most recent stable window under similar preconditions for that process knowledge unit. For example, The most recent 8 valid batches can be used. If there are fewer than 5 valid batches, the process knowledge unit will not be directly written into the common knowledge subset, and... The upper limit is set at 0.55. If the deviation of three consecutive valid batches increases in the same direction, it is judged that stable convergence has not been formed, and [the following will be implemented / restricted]. The upper limit is further limited to 0.4. The value of is limited to between 0 and 1. When the calculation result is less than 0, it is taken as 0, and when the calculation result is greater than 1, it is taken as 1. The closer it is to 1, the more convergent the results are after multiple batches of execution under similar conditions.

[0046] After obtaining the evidence weights and the consistency of the execution results, a conflict retention score is generated for each process knowledge unit: ; in, For process knowledge unit Conflict retention score; This refers to the evidence weighting coefficient; The consistency coefficient of the execution results, and .

[0047] At the conflict pair level, the conflict separation quantity is then calculated: ; in, For process knowledge unit With process knowledge unit The amount of conflict separation.

[0048] The convergence interval of the result state is the interval enclosed by the mean of the result states of the most recent stable window under similar preconditions and the corresponding allowable bandwidth of the process knowledge units. The intersection determination is only performed when the number of valid batches in the most recent stable window is not less than 5 batches; when there are fewer than 5 batches, the common merging relationship within the overlapping boundary is not determined based on this, but the corresponding process knowledge units are treated as a conflict pending set.

[0049] The risk level of an action is determined based on the risk classification threshold corresponding to the adjustment range of single-pass reduction, roll temperature, and line speed, and is divided into three levels: low risk, medium risk, and high risk. The risk classification threshold is determined within the same threshold version identifier within the corresponding process similar cluster, and is based on the median value of each action adjustment range within the most recent stable window under the corresponding process path. An adjustment range not exceeding this benchmark value is classified as low risk; an adjustment range exceeding the benchmark value but not more than twice it is classified as medium risk; and an adjustment range exceeding twice it is classified as high risk. If the number of valid batches in the most recent stable window is less than 8, the risk classification threshold of the previous version is used; if no previous version exists, the preset initial risk classification threshold of the corresponding process path is used.

[0050] Therefore, conflict resolution can be performed according to the following rules. When When the threshold is triggered by the overlap of the applicable boundaries corresponding to the threshold version identifier, it is determined that the applicable boundaries of the two are separate and do not constitute a direct conflict. Both can be retained, but are only restricted to being called within their respective applicable boundaries. Not lower than the threshold version identifier corresponding to the applicable boundary overlap trigger threshold and When the conflict separation threshold corresponding to the threshold version identifier is not lower than the threshold, it is determined that the conflict has become significantly different. The process knowledge unit with the higher score is retained as the dominant knowledge unit within the overlapping boundary, and the other process knowledge unit is removed from the overlapping boundary. The remaining applicable boundary after removal can still be split and written into the base-specific knowledge subset. When Not lower than the threshold version identifier corresponding to the applicable boundary overlap trigger threshold and When the conflict separation amount is below the threshold value corresponding to the version identifier, if the two processes have the same direction of operation, the same risk level, and the convergence interval of the result state overlaps, then no elimination process is performed. Instead, the overlapping boundary part is written into the common knowledge subset, and the non-overlapping boundary parts are retained in the base-specific knowledge subsets respectively. Not lower than the threshold version identifier corresponding to the applicable boundary overlap trigger threshold and When the conflict separation amount is below the threshold value corresponding to the version identifier, if the two process actions are opposite, or the risk levels are different, or the convergence intervals of the result states do not intersect, then the two will be written into the conflict pending set and restricted to being called within the corresponding process similar cluster, and will not be entered into cross-site common calls.

[0051] With this approach, the overlap of applicable boundaries determines whether conflict resolution is required, the weight of evidence determines which piece of knowledge has more historical support, and the consistency of execution results determines which piece of knowledge is more likely to form a stable result under similar conditions. All three are linked within the same decision chain.

[0052] For example, the applicable boundary overlap, evidence weight, consistency of execution results, conflict retention score, and conflict separation quantity are managed using the same threshold version identifier. The corresponding threshold can be adjusted based on historical sample quantiles, validation set tuning results, or upper and lower limits of process specifications. Indicators whose values ​​are limited to closed intervals do not take boundary extreme values ​​to reduce misjudgments caused by boundary jitter. The threshold version remains frozen within a single batch and is updated after the result status is written back; if there are insufficient effective write-back batches, the previous version is used.

[0053] In this embodiment, the target base performs forward reasoning within a common knowledge subset and a base-specific knowledge subset corresponding to the base through a local reasoning engine. The process knowledge units with overlapping trigger boundaries are filtered according to the action risk level and evidence weight, and the output includes calendering adjustment strategies including calendering pass arrangement, roll temperature adjustment direction, line speed adjustment direction, single pass reduction amount and pre-processing callback trigger conditions. In this embodiment, the local inference engine refers to an inference execution unit deployed locally at the target site, used to invoke a shared knowledge set and generate a calendering adjustment strategy based on the current batch's process status facts. The input to the local inference engine includes at least the set of process status facts for the current batch, the target site identifier, the material system identifier, the equipment type identifier, the process path identifier, and the safe operating boundary. The calendering adjustment strategy refers to the set of executable process actions output by the local inference engine. The calendering adjustment strategy includes at least the action type, action direction, action sequence, triggering condition, execution period, and write-back object identifier. The safe operating boundary refers to the equipment and sheet safety constraints allowed for calendering actions, including at least the upper limit of single-pass reduction, the upper limit of roll temperature adjustment range, and the upper limit of line speed adjustment range.

[0054] In this embodiment, the local inference engine first retrieves candidate process knowledge units from the shared knowledge set based on the target base identifier and applicable boundaries. Then, it performs forward inference based on the matching results between the current batch process status facts and the premise states of the candidate process knowledge units. When multiple candidate process knowledge units simultaneously meet the triggering conditions, they are sorted according to evidence weight, action risk level, and recent execution consistency. The action risk level is used to characterize the impact level of the corresponding process action on the sheet integrity state, compaction state, and solid-solid contact state. For process actions with high priority and not exceeding the safe operating boundary, they are written into the calendering adjustment strategy and issued for execution; for process actions exceeding the safe operating boundary, their triggering records are retained but not written into the calendering adjustment strategy.

[0055] In this embodiment, when the number of local effective process knowledge units of the target base is lower than a preset threshold, the target base calls an initial knowledge subset from the shared knowledge set according to the material system identifier, equipment type identifier and process path identifier, and generates an initial rolling adjustment strategy within a preset safety range; When the write-back result status meets the preset consistency condition, the initial knowledge subset is replaced with the target knowledge subset containing local process knowledge units; In this embodiment, a valid process knowledge unit refers to a process knowledge unit that simultaneously satisfies the following conditions: applicable boundary matching, failure state identification as valid, evidence weight greater than the lower bound of invocation, and corresponding result state traceability. The initial knowledge subset refers to the set of process knowledge units retrieved from the shared knowledge set and used for initial rolling decisions when the target base lacks sufficient local valid process knowledge units. The initial knowledge subset includes at least a common knowledge unit identifier, an applicable boundary field, an action field, a result state field, and a invocation priority field. The target knowledge subset refers to the stable invocation set composed of the target base's local process knowledge units and the matching knowledge units in the shared knowledge set after subsequent execution and write-back. The preset consistency condition refers to the target consistency condition that the result states of multiple consecutive batches satisfy the sheet integrity state, compaction state, and solid-solid contact state. The lower bound of invocation refers to the minimum evidence weight threshold allowed for a process knowledge unit to enter local inference.

[0056] In this embodiment, when the target base is a newly imported base or a base with scarce tags, an initial knowledge subset is first retrieved from the shared knowledge set based on the material system identifier, equipment type identifier, and process path identifier. Then, an initial rolling adjustment strategy is generated by combining the current batch process status facts. The initial rolling adjustment strategy is limited to execution within the safe operating boundary of the target base. As the target base continuously writes back the result status and forms new local process knowledge units, the number of local effective process knowledge units gradually increases. When the preset consistency conditions are met, the target knowledge subset is used to replace the initial knowledge subset, thereby realizing the switch from the shared import stage to the local stable operation stage.

[0057] Furthermore, the initial knowledge subset is filtered sequentially based on material system identifier, equipment type identifier, process path identifier, evidence weight, and recent execution consistency. When multiple candidate process knowledge units exist, process knowledge units with lower action risk levels are prioritized for retention. During the period when the target base has not met the preset consistency conditions, a conservative output caliber is adopted for single-pass reduction adjustment, rapid roll temperature adjustment, and rapid line speed adjustment; when the corresponding batch shows a pending verification status or continuous abnormal records, the calendering adjustment strategy of the previous batch is maintained, and the result status continues to be written back.

[0058] In this embodiment, the result state includes at least two of the following: sheet integrity state, compaction state, and solid-solid contact state. When writing back the result status, each production base simultaneously writes the batch identifier, equipment identifier, action execution time, and corresponding process knowledge unit identifier to form a knowledge traceability chain. In this embodiment, the knowledge traceability chain refers to a structured record chain that associates and records the entire process of a process knowledge unit from generation, invocation, execution to write-back. The knowledge traceability chain includes at least batch identifier, equipment identifier, material system identifier, process path identifier, process knowledge unit identifier, action execution time, action execution content, write-back result status, evidence source identifier, and version identifier. Through the knowledge traceability chain, it is possible to locate the process knowledge unit invoked by a certain rolling mill adjustment strategy, its corresponding prerequisite status, the result status after execution, and the version change process of that process knowledge unit within the shared knowledge set.

[0059] In this embodiment, each production base synchronously writes the knowledge traceability chain according to batch, equipment, and process knowledge unit dimensions when writing back the result status. After receiving the local knowledge update package, the federated coordinator writes the corresponding version change information into the knowledge traceability chain. When the target base subsequently calls the shared knowledge set, it can filter process knowledge units with continuous and consistent execution records based on the knowledge traceability chain and restrict the call to process knowledge units with continuous abnormal records.

[0060] In this embodiment, when the inconsistency between the result status written back from the target base and the result status corresponding to the process knowledge unit exceeds a preset threshold, the process knowledge unit is marked as pending verification, and the calling priority of the process knowledge unit is reduced. When multiple batches consecutively meet the preset consistency conditions, update the applicable boundaries and evidence weights of the process knowledge unit; In this embodiment, inconsistency refers to the degree of deviation between the result state written back from the target base and the result state in the process knowledge unit. Inconsistency is based on at least two of the following: sheet integrity state, compaction state, and solid-solid contact state. The state to be verified refers to the state identifier of the corresponding process knowledge unit that is not a priority call target at the current stage but retains the qualification for subsequent re-verification. Call priority refers to the priority order identifier used by the local inference engine when sorting multiple candidate process knowledge units.

[0061] In this embodiment, when the inconsistency between the write-back result status and the final result status of a certain process knowledge unit in multiple consecutive batches continuously increases, the process knowledge unit is marked as pending verification, and its evidence weight and invocation priority are simultaneously reduced. When the write-back result status and the final result status of multiple consecutive batches meet the consistency condition again, the pending verification status is lifted and the invocation priority is restored. For process knowledge units that have been lifted from the pending verification status, the federated coordination end also updates their applicable boundaries, so that they can only be invoked within the scope of re-verified material systems, equipment types, and process paths.

[0062] Specifically, the version identifier and batch identifier in the knowledge traceability chain serve together as the write-back alignment key. When a version switch occurs within the same process knowledge unit, parallel records of the old and new versions are retained without overwriting the original record. Process knowledge units in the pending verification state only participate in low-risk action screening within their applicable boundaries; when subsequent write-back results meet preset consistency conditions, the pending verification state is lifted and regular calls are restored; when subsequent write-back results do not meet preset consistency conditions, the restricted call scope is maintained.

[0063] Example 2: Based on Example 1, this example focuses on a unified generation chain for process state facts under inconsistent observation criteria across different production bases. It indexes observation items using a state mapping table and sequentially performs time alignment, dimension normalization, detection location conversion, state support value synthesis, and missing value weighting, converging observation data from different sampling frequencies, dimensions, and detection locations into a unified state representation. Based on this processing chain, process state facts maintain a consistent recording and retrieval standard across different bases, thus providing consistent input for subsequent process knowledge unit construction, federated fusion, and local inference.

[0064] In this embodiment, when generating process status facts, each production base first establishes a state mapping table between observation items and a unified state vocabulary based on local sensor configuration. The state mapping table includes at least the observation item identifier, corresponding state name, time window width, time anchor point type, window aggregation method, dimensionally normalized reference interval, direction attribute, reference detection position, position conversion coefficient, and missing data handling method. For cases where the same state name corresponds to multiple observations, time alignment is first performed according to the time window width and time anchor point type, then a unified dimensional expression is completed according to the dimension-normalized reference interval and direction attribute, and finally, the observation data of different detection positions are converted into a unified state expression according to the reference detection position and position conversion coefficient. A unified state representation should include at least the state name, state level, time interval, location interval, state source, and state confidence flag. For missing observations, write a missing marker and retain the partial state facts formed by existing observations; for abnormal fluctuation observations, write an abnormal marker and reduce the evidence weight of the process knowledge units they participate in generating. In this embodiment, the state mapping table includes at least the observation item identifier, corresponding state name, time window width, time anchor point type, window aggregation method, dimension-normalized reference interval, direction attribute, reference detection location, location conversion coefficient, missing item handling method, and state confidence tag generation rules. For cases where the same state name corresponds to multiple observation items, time alignment is first performed based on the time window width, time anchor point type, and window aggregation method. Then, a unified dimension expression is completed according to the dimension-normalized reference interval and direction attribute. Finally, based on the reference detection location and location conversion coefficient, observation data from different detection locations are converted into a unified state expression under the same reference detection location and the same time caliber. For missing observation items, a missing tag is written, and the partial state facts formed by existing observation items are retained. For abnormal fluctuation observation items, an abnormal tag is written, and the state confidence tag of the corresponding state expression is reduced. The process knowledge unit generated by this state expression is weighted lower according to the state confidence tag in subsequent evidence weight calculations. The state confidence tag uses a hierarchical record of high confidence, medium confidence, and low confidence, or a continuous scaling record can be used. When there are multiple observations for the same state name, observations with overlapping time intervals and consistent location intervals are retained first; if both conditions cannot be met simultaneously, a single state expression is output using window aggregation and a merge mark is written into the fact record table.

[0065] In this embodiment, the unified state representation includes at least a state name, state level, time interval, location interval, state source, and state confidence marker. The state source identifies whether the state was generated from equipment operation data, online monitoring data, preprocessing data, or quality characterization data. The state confidence marker characterizes the completeness and reliability of the state representation. Using this unified state representation, heterogeneous observation data from different production bases can be aligned at the process state fact layer to support subsequent knowledge generation, federated fusion, and local inference.

[0066] In this embodiment, when generating process status facts, each production base first establishes an observation item-state mapping table based on its local sensor configuration. The state mapping table uses each state term from a unified state lexicon as an index, and writes in the corresponding observation item identifier, time window width, time anchor point type, window aggregation method, dimensionally normalized reference interval, direction attribute, reference detection position, position conversion coefficient, missing tolerance ratio, and minimum number of valid observation items for each item. The state mapping table also records the state mapping version identifier, the normalized reference interval version identifier, the position conversion version identifier, and the threshold version identifier. Each version identifier is only used to bind the process status facts, process knowledge unit call process, and result status write-back process of the current batch to the same parameter version. The time anchor type is used to define the alignment benchmark for data from different sources. The calendering equipment operation data uses the action execution time as the anchor, the sheet online detection data uses the detection trigger time as the anchor, and the pre-calendering processing data and post-calendering quality characterization data use the batch completion time as the anchor. The direction attribute is used to characterize whether the corresponding state is enhanced or weakened when the observed value increases. The reference detection position is used to convert data from different detection positions to a unified position caliber. The missing tolerance ratio and the minimum number of valid observations are used to constrain the deweighting and prohibition conditions for missing observations.

[0067] For observation data with different sampling frequencies, the target base performs time window alignment according to the corresponding time window width in the state mapping table. For data mapped to state entries... Observation items Samples are extracted within the valid window preceding the corresponding time anchor point to form aligned observations: ; in, For observation items Corresponding status terms Time window alignment value; This represents the number of samples that fall within the valid window. For the first The original observations of each sample; For the first The time weight of each sample increases with its proximity to the time anchor. All time weights are positive, and the time weight of a sample closer to the time anchor is no less than that of a sample farther from the time anchor. For high-frequency observations, a weighted mean or weighted quantile is used as the output window. For low-frequency observations, the most recent valid value is allowed to be used within its effective retention period; otherwise, a missing value is added. Within the same batch, the time window width, time anchor type, and window aggregation method are frozen at the beginning of the batch and are not adjusted with changes in observations within the batch. The effective retention period for low-frequency observations does not exceed twice the nominal sampling period of the observation; if it exceeds this period, a missing value is directly added, and the most recent valid value is no longer used. The time window for equipment operation data such as calender roll temperature, roll pressure, and line speed can be 3 to 10 seconds; the time window for batch-level offline quality characterization data is calculated for the entire batch. After this processing, data from different sampling frequencies are compressed into a unified time caliber within the same batch and under the same status term.

[0068] For observation data with different dimensions, after time window alignment, they are converted to a unified dimension expression based on the dimension-normalized reference interval in the state mapping table. The dimension-normalized reference interval can be given by the upper and lower limits of the process specification, or it can be defined by the 10th and 90th quantiles of the most recent 30 stable batches. Stable batches are defined as those whose material system identifier, equipment type identifier, and process path identifier are consistent with the current batch, and whose dimension-normalized reference interval version identifier and location conversion version identifier are the same. If there are fewer than 10 valid batches meeting the constraints, the dimension-normalized reference interval is not updated; instead, the previous version's dimension-normalized reference interval is used. If no previous version exists, the upper and lower limits of the process specification are used as the dimension-normalized reference interval. For observation items... In the status entry The normalized value can be calculated using the following formula: ; in, For observation items Corresponding status terms Dimensional normalized value; This is a limiting function used to restrict the output to between 0 and 1; For observation items Corresponding status terms The lower bound of the reference; For observation items Corresponding status terms The upper bound of the reference. If the direction attribute in the state map table indicates that the observation is inversely related to state enhancement, then the upper bound is adopted. As a result of normalization, observations outside the reference interval are no longer extrapolated, but are saturated at 0 or 1 to avoid individual outliers having too high a weight in the state construction.

[0069] For observation data from different detection locations, a detection location conversion needs to be performed before dimension normalization. For observations acquired at the edge, center, or different lateral coordinates, the reference detection location preset in the state mapping table is used as a unified standard to convert the alignment value at the actual detection location into the reference location value: ; in, For observation items Convert to status terms The observed values ​​after the reference detection position; For observation items Time window alignment value; For observation items Corresponding status terms Position conversion factor; For status entries Reference detection location; For observation items The actual detection location. For observations requiring location conversion, use... Replace the previous formula Entering dimensionless mode. The position conversion coefficient is obtained by first-order linear fitting from the most recent 20 valid batches of observations with multiple position pairs, and is restricted to non-negative values; it is updated once every 20 new valid batches. The newly fitted position conversion coefficient is only effective if the average conversion residual on the most recent 8 paired samples is not higher than 90% of the average conversion residual of the previous version; otherwise, it reverts to the position conversion coefficient of the previous version and writes a conversion anomaly mark; if there is no previous version, the position conversion coefficient is 0, and the conversion anomaly mark is only used to record fitting failure and trigger rollback; when there are fewer than 8 paired samples, abnormal edge crack propagation, or sheet deviation exceeding the limit, the update is paused and the position conversion coefficient of the previous version is used. The correction magnitude caused by position conversion is limited to within 20% of the width of the reference interval; if it exceeds this range, it is truncated by 20%.

[0070] When mapped to a status term When the available observations are empty, the state support value is not calculated; instead, the state terms are calculated directly. A missing marker is written, and it is prohibited from being used as a core trigger condition for the current batch in the process knowledge unit call. After completing time window alignment, dimension normalization, and detection position conversion, the target base synthesizes multiple observation items into a unified state expression.

[0071] Same state term Next, the mapping weights corresponding to each observation item are normalized so that the sum of the mapping weights is 1, thereby defining the contribution boundary of different observation items to the unified state expression. For state terms... , retrieve all corresponding observation item sets from the state mapping table and calculate the state support value according to the mapping weight: ; Among them, is the state support value of the state entry ; is the set of observation items mapped to the state entry ; is the observation item corresponding to the state entry mapping weight; is the normalization value of the observation item corresponding to the state entry ; is an available flag, taking 1 when available and 0 when missing. After calculating the state support value, write out the state level according to the grading threshold in the state mapping table; for example, for risk entries such as edge crack risk and springback state, can be recorded as low, recorded as medium, recorded as high; for process entries such as fibrosis state and densification state, a grading expression of weak, medium, and strong can be used. The unified state expression formed in this way includes at least state entries, state levels, state support values, and missing flags, thus converging the observation data with different sampling frequencies, different dimensions, and different detection positions into structured records that can be directly involved in subsequent knowledge construction and reasoning.

[0072] For missing observation items, the target base performs evidence weight correction according to the missing tolerance ratio in the state mapping table. The missing-corrected evidence weight of the corresponding process knowledge unit can be attenuated according to the following formula: ; Among them, is the evidence weight of the process knowledge unit after missing correction; the original evidence weight of the process knowledge unit ; is the number of required observation items missing in the current batch for the process knowledge unit ; is the total number of required observation items required for the process knowledge unit to be triggered; is used to limit the lower limit of the attenuated evidence weight to 30% of the original evidence weight.

[0073] When there are missing observation items in the current batch, use to replace in the aforementioned conflict retention score calculation formula.It is used for local forward inference, conflict retention score calculation, and call priority sorting; when there are no missing observations in the current batch, it maintains the current approach. When a corresponding process knowledge unit is marked with a no-adjustment flag, that process knowledge unit will not participate in the conflict retention scoring comparison in the current batch, nor will it be considered the dominant knowledge unit within the overlapping boundary.

[0074] Weight reduction is performed once each time a process state fact is generated. This occurs when an observation marked as core in the state mapping table corresponding to a certain state term is missing, or... When the missing process knowledge unit is in use, a prohibition flag is set, indicating that the unit is prohibited from participating in the current batch's forward inference. When the missing rate is between 0.2 and 0.5, the unit is only allowed to participate in low-risk action recommendations and is excluded from high-risk strategy outputs such as single-pass reduction adjustment or rapid roll temperature adjustment. The prohibition flag is removed and regular inference is resumed after three consecutive batches have recovered to a missing rate of no more than 0.2 and the core observations are available again. This process ensures that missing observations do not directly interrupt the state construction process but are incorporated into a conservative strategy through evidence weight decay and call scope contraction.

[0075] Example 3: This embodiment provides an application in a scenario of cross-site collaborative parameter adjustment in a dry-process thick electrode few-pass rolling production line.

[0076] The application targets thick electrode sheets produced using a solvent-free mixing and fiberization process. Production is organized with two or more production bases operating in parallel. Each base has the capability to record calendering equipment operation data, online sheet inspection data, pre-calendering treatment data, and post-calendering quality characterization data. Environmental boundaries are defined by storing original process data locally at each production base, with the federal coordination terminal only receiving local knowledge update packages. Differences in equipment type, material system, and process path are permitted between each production base, but all bases employ a few-pass calendering. In terms of engineering specifications, single-batch calendering processes are numbered by roll or by continuous length segment. Equipment operation data within a batch is recorded using a 3-10 s time window, offline quality characterization data is recorded batch by batch, and cross-base process sharing uses a unified state vocabulary and observation mapping rules to maintain consistent state representation.

[0077] In this embodiment, each production base completes local initialization before calendering begins. Specifically, the production base reads the material system identifier, equipment type identifier, and process path identifier for this batch, establishes a batch identifier, and loads a state mapping table based on a unified state vocabulary. The state mapping table is used to convert roll temperature, linear speed, roll gap position, linear pressure, thickness, width, surface defect markers, edge state, premixed state records, fiberization treatment records, and heat treatment records into a unified state expression. Each observation item is written with a time anchor point, location interval, and state source. When the same state name corresponds to multiple observation items, they are first aligned according to the time window width and time anchor point type, then a unified dimensional expression is completed according to the dimensional normalization reference interval, and a process state fact is formed according to the reference detection position. For the target base importing for the first time, when the number of local effective process knowledge units is lower than a preset threshold, an initial knowledge subset is first called from the shared knowledge set based on the material system identifier, equipment type identifier, and process path identifier, and the call result is limited within the safe operation boundary to form an initial calendering adjustment strategy.

[0078] During continuous operation, each production base executes the following closed loop cyclically, either in batches or along continuous length segments. First, it collects multi-source process data for the current batch and generates at least two process state facts locally, including fibrous state, densification state, edge crack risk, springback state, solid-solid contact stability, thermal history shift, and stress shift. Then, it matches existing local process knowledge units with the current process state facts. For process knowledge units that meet the prerequisite states, it reads their process actions, result states, applicable boundaries, evidence source identifiers, and evidence weights, and generates a local knowledge update package. After receiving the local knowledge update package uploaded by each production base, the federated coordination terminal first divides the process into similar clusters based on the base's process profile. Then, it performs rule alignment on process knowledge units with similar prerequisite states within the same similar cluster, and resolves and merges conflicts according to the overlap of applicable boundaries, evidence weights, and consistency of execution results, generating a common knowledge subset, a base-specific knowledge subset, and a conflict-pending set. After receiving the updated shared knowledge set, the target base performs forward reasoning through the local reasoning engine within the common knowledge subset and the base-specific knowledge subset corresponding to this base. It outputs the rolling pass arrangement, single-pass reduction amount, roll temperature adjustment direction, line speed adjustment direction, and preprocessing callback trigger conditions, and performs the current batch of few-pass rolling accordingly.

[0079] In this embodiment, after the current batch of calendering is completed, the target base writes back at least two of the following result states: sheet integrity, compaction, and solid-solid contact. During the write-back, the batch identifier, equipment identifier, action execution time, process knowledge unit identifier, and version identifier are simultaneously written, forming a knowledge traceability chain. If the write-back result state is consistent with the result state of the corresponding process knowledge unit, the evidence weight of that process knowledge unit is increased; if the inconsistency between the write-back result state and the corresponding result state exceeds a preset threshold, the process knowledge unit is marked as pending verification, and its call priority is reduced. The process knowledge unit after the write-back re-enters the local knowledge update package and is used by the federated coordination end for the next round of shared knowledge set iteration, thus forming a closed-loop operation chain of acquisition—state extraction—knowledge call—action execution—result write-back—federated update.

[0080] In this embodiment, missing observations and abnormal fluctuation observations are treated as implementation-related anomalies. Specifically, when some observations corresponding to a certain state name are missing, the partial state facts formed by the existing observations are retained, and a missing marker and a state confidence marker are written into the fact record. The process knowledge unit generated by this state fact is downweighted according to the state confidence marker. If the missing ratio is within the allowable range defined by the engineering caliber, the process knowledge unit can still participate in low-risk action recommendation; if the missing ratio exceeds the call limit, the process knowledge unit is marked with a no-call marker and does not participate in the forward inference of the current batch. For abnormal fluctuation observations, after writing the abnormal marker, the original record is retained and it participates in alignment according to the unified state expression, but its corresponding process knowledge unit is not used as the dominant knowledge unit within the overlapping boundary. In subsequent batches, when the core observations are continuously restored to be available and the missing ratio falls back to the allowable range, the no-call marker is removed and normal calls are restored. The existing batch identifier and process knowledge unit identifier are not rewritten, thereby completing the recovery, retransmission, merging, and deduplication after the anomaly.

[0081] In this embodiment, another type of implementation-related anomaly is the persistent existence of a conflict-pending set. Specifically, when two process knowledge units within the same process similarity cluster conflict in their process actions within the overlapping area of ​​their applicable boundaries, and the convergence intervals of their result states do not intersect or the execution results do not converge, the federated coordination end writes them into the conflict-pending set and restricts their use to the corresponding process similarity cluster. When the target base calls the shared knowledge set in subsequent batches, it first filters common knowledge subsets and base-specific knowledge subsets; only when there are insufficient local valid process knowledge units, and there are records in the conflict-pending set that are identical to the material system identifier, equipment type identifier, and process path identifier of the current batch, are the process knowledge units in the conflict-pending set treated as low-priority candidates within the safe operating boundary. After the corresponding batch is executed, the result state is written back and the evidence weight and applicable boundary are updated; when multiple consecutive batches meet the preset consistency conditions, the process knowledge unit is restored from the pending verification state to the regular call state, and its applicable boundary is narrowed to the range of re-verification success.

[0082] By implementing the above scenarios, under multi-base, low-pass calendering conditions, the calendering experience scattered across each production base can be transformed into shareable, interpretable, and traceable process knowledge, and executable calendering adjustment strategies can be directly output at the target base. For situations involving inconsistent sensor calibers, different equipment types, introduction of new bases, missing observations, abnormal fluctuation observations, and conflicting pending sets, a unified state vocabulary, observation mapping rules, base process profiles, and result state write-back can maintain consistency in state expression, knowledge retrieval, and retrieval priority updates. This ensures that shared results are continuously applied to specific parameter adjustment actions and maintains the continuous operation of the calendering decision chain.

[0083] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0084] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are meant to be within the scope of this application and form different embodiments. For example, all the embodiments above can be used in any combination. The information disclosed in this background section is intended only to enhance the understanding of the general background of this application and should not be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art.

Claims

1. A knowledge sharing and optimization method for multi-base rolling processes based on federated learning, applied to the scenario of few-pass rolling of dry thick electrodes, characterized in that... include: Each production base collects multi-source process data of its own few-pass calendering process and generates process state facts locally based on a unified state vocabulary and observation mapping rules. Each production base updates its local state extraction model based on the multi-source process data, constructs local process knowledge units based on the process state facts, and generates a local knowledge update package. The process knowledge unit includes at least the premise state, process action, result state, applicable boundary, evidence source identifier, and evidence weight. The local knowledge update package includes newly added, revised, or invalid process knowledge units and model update results. The federal coordination terminal only receives the local knowledge update package. After clustering according to the base process profile, it first performs rule alignment on process knowledge units with similar premise states within the same process similar cluster. Then, it resolves and merges conflicts according to the overlap of applicable boundaries, evidence weight, and consistency of execution results, forming a common knowledge subset, a base-specific knowledge subset, and a conflict pending set. The target base retrieves a subset of knowledge that matches the material system identifier, equipment type identifier, and process path identifier of the target base based on the current batch's process status facts, performs local reasoning, and outputs a calendering adjustment strategy; After the target base executes the rolling adjustment strategy, it writes back the batch identifier, the corresponding process knowledge unit identifier, the action execution information, and the result status to the local machine to update the local process knowledge unit and generate a new local knowledge update package, which is then used by the federated coordinator to iteratively update the shared knowledge set.

2. The knowledge sharing and optimization method for multi-base calendering processes based on federated learning according to claim 1, characterized in that, The multi-source process data includes at least two of the following: calendering equipment operation data, sheet online detection data, calendering pretreatment data, and calendering post-calendering quality characterization data. The process state facts include at least two of the following: fibrous state, densification state, risk of edge cracking, springback state, solid-solid contact stability, thermal history offset, and stress offset.

3. The knowledge sharing and optimization method for multi-base calendering processes based on federated learning according to claim 1, characterized in that, The process actions include at least one of the following: rolling pass arrangement, single pass reduction, roll temperature adjustment, line speed adjustment, and pretreatment callback; The applicable boundaries include at least the material system identifier, equipment type identifier, and process path identifier.

4. The method for knowledge sharing and optimization of multi-base calendering process based on federated learning according to claim 1, characterized in that, The base process profile consists of material system identifiers, pretreatment path identifiers, calendering equipment type identifiers, and sheet structure identifiers. The federal coordination end divides multiple production bases into at least one process similarity cluster based on the base process profile, and forms the shared knowledge set in each process similarity cluster.

5. The knowledge sharing and optimization method for multi-base calendering processes based on federated learning according to claim 4, characterized in that, The federal coordination end first performs rule alignment on process knowledge units that are in the same process similar cluster and have similar premise states. The rule alignment includes unifying the premise state expression, unifying the process action field and unifying the result state field. For process knowledge units that have completed rule alignment, the degree of overlap of applicable boundaries is calculated based on the degree of overlap of material system identifiers, equipment type identifiers, and process path identifiers, and the fusion order is determined by combining evidence weights and execution result consistency. For process knowledge units whose applicable boundary overlap reaches the corresponding trigger threshold, whose process action direction is consistent, whose risk level is consistent, and whose result state convergence intervals have an intersection, they are merged and written into the common knowledge subset. For process knowledge units whose applicable boundaries differ in at least one of the production base, equipment type, and process path, the differentiated boundaries are retained and written into the base-specific knowledge subset. For process knowledge units that conflict with process actions and meet at least one of the following conditions, write them into the conflict pending set and restrict their use to the corresponding process similar cluster: the convergence intervals of the result states do not intersect; The execution result did not converge; The conflict pending set includes at least the prerequisite status identifier, conflict action identifier, conflict cause identifier, restricted call scope identifier, and pending verification status identifier.

6. The method for knowledge sharing and optimization of multi-base calendering process based on federated learning according to claim 1, characterized in that, The target base performs forward reasoning within the common knowledge subset and the base-specific knowledge subset corresponding to this base through the local reasoning engine; The process knowledge units with overlapping trigger boundaries are filtered according to the action risk level and evidence weight, and the output includes calendering adjustment strategies including calendering pass arrangement, roll temperature adjustment direction, line speed adjustment direction, single pass reduction amount, and pre-processing callback trigger conditions.

7. The method for knowledge sharing and optimization of multi-base calendering process based on federated learning according to claim 1, characterized in that, When generating process status facts, each production base first establishes a state mapping table between observation items and a unified state vocabulary based on local sensor configuration. The state mapping table includes at least the observation item identifier, corresponding state name, time window width, time anchor point type, window aggregation method, dimensionless reference interval, direction attribute, reference detection position, position conversion coefficient, and missing data handling method. For cases where the same state name corresponds to multiple observations, time alignment is first performed according to the time window width and time anchor type, then a unified dimensional expression is completed according to the dimensional normalization reference interval and direction attribute, and finally, the observation data of different detection positions are converted into a unified state expression according to the reference detection position and position conversion coefficient. The unified state representation includes at least a state name, state level, time interval, location interval, state source, and state confidence flag; For missing observations, a missing marker is written and the partial state facts formed by existing observations are retained. For abnormal fluctuation observations, an abnormal marker is written and the evidence weight of the process knowledge unit in which they participate is reduced.

8. The method for knowledge sharing and optimization of multi-base calendering process based on federated learning according to claim 1, characterized in that, When the number of local effective process knowledge units at the target base is lower than a preset threshold, the target base calls an initial knowledge subset from the shared knowledge set according to the material system identifier, equipment type identifier, and process path identifier, and generates an initial rolling adjustment strategy within a preset safety range. When the write-back result status reaches the preset consistency condition, the initial knowledge subset is replaced with the target knowledge subset containing local process knowledge units.

9. The knowledge sharing and optimization method for multi-base calendering processes based on federated learning according to claim 1, characterized in that, The resulting state includes at least two of the following: sheet integrity state, compaction state, and solid-solid contact state. When writing back the result status, each production base simultaneously writes the batch identifier, equipment identifier, action execution time, and corresponding process knowledge unit identifier to form a knowledge traceability chain.

10. The method for knowledge sharing and optimization of multi-base calendering process based on federated learning according to claim 1, characterized in that, When the inconsistency between the result status written back from the target base and the result status corresponding to the process knowledge unit exceeds a preset threshold, the process knowledge unit is marked as pending verification, and the calling priority of the process knowledge unit is reduced. When multiple batches consecutively meet the preset consistency conditions, the applicable boundaries and evidence weights of the process knowledge unit are updated.

Images