Intelligent manufacturing production quality parameter dynamic optimization control system based on machine learning

Abstract

The application discloses a machine learning-based intelligent manufacturing production quality parameter dynamic optimization control system and relates to the technical field of intelligent manufacturing and industrial process control. A plurality of switching actions continuously appearing are arranged into a single composite switching event through event classification processing and adjacent event folding processing, the same transition section is avoided from being repeatedly divided or cross-covered, the boundary of the switching event is clearer, and unified analysis around the same transition process is facilitated. Each switching event is simultaneously traced back to a steady-state production segment in front and tracked to an initial production segment in back, so that the system not only records the switching time, but also obtains the steady-state length before the event, the beat falling mode, the air window length and the initial segment fluctuation density, thereby expanding the original single-point switching record into an interface pulse segment with the correlation of the front and back processes.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing and industrial process control technology, specifically to a dynamic optimization control system for intelligent manufacturing production quality parameters based on machine learning. Background Technology

[0002] Existing production control methods typically treat entire batches, entire processes, or fixed quantities of upstream products as processing units, using uniform rules to isolate, release, or correct parameters for products after a changeover event. This approach has significant shortcomings: Firstly, existing systems often only record the events of batch change, formula change, shift change, or restart itself, lacking fine-grained identification of the transition interval after the switch, making it difficult to distinguish between transitional parts that are truly in a non-converged state and products that have returned to a stable state. Secondly, existing systems typically rely on a single quality result or a fixed threshold to judge risk, and cannot identify small-scale target objects that are not scrapped but are gradually approaching the defect boundary from continuous trajectories such as beat drop, process fluctuation, and continuation of the trail. Third, existing parameter control methods mostly adopt unified correction of the whole segment or manual experience correction, lacking local scripted control for narrow-domain intervention components and their wake segments, and also lacking a chain verification and knowledge writing mechanism for the stable state after parameter adjustment.

[0003] The aforementioned situation and shortcomings typically arise in production environments characterized by scattered multi-source data, overlapping switching events, delayed process stabilization, and delayed parameter actions. When MES, PLC, equipment logs, and quality inspection records are not uniformly organized onto the same product sequence chain, the true state of the preceding products after a switchover can easily be obscured by the entire batch of data. When the system processes data based solely on a fixed number of units or static thresholds, some products that have already stabilized may be mistakenly included in the intervention scope, while some products that are still not converged and are constantly hovering may be missed. When local parameter tuning lacks observation window constraints and stabilization verification, situations such as continuous tuning of the same parameter, repeated disturbance of transition trails, and distorted judgment of convergence completion points can easily occur. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a machine learning-based intelligent manufacturing production quality parameter dynamic optimization control system, which solves the problems mentioned in the background technology.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solution: a machine learning-based intelligent manufacturing production quality parameter dynamic optimization control system, including a transition window weaving module, a transition part identification module, an edge-fitting risk screening module, a micro-domain parameter tuning control module, and a stabilization verification and write-back module; The transition window weaving module collects timestamps, window lengths, cycle time differences, and initial process trajectories for batch change, formula change, shift change, short stop restart, and fixture cleaning, and aligns them according to the product sequence to form candidate transition windows. The transition component identification module divides the sequence within the candidate window into steady state, transition state, and return-to-steady state, and compares it with historical return-to-steady segments to mark strong transition components, weak transition components, and convergence completion points. The edge risk screening module continues to compare quality inspection records, agent quality labels and historical boundary neighborhood samples in the identified transitional parts to distinguish between observation parts and narrow-domain intervention parts, and at the same time gives the range of influence of the wake. The micro-domain parameter tuning control module calls the corresponding local parameter tuning script based on the narrow domain intervention component and the wake range, only performs exploratory adjustments in the small interval where the target component is located, and sets a minimum observation window to limit continuous changes to the same parameters; The stabilization verification and write-back module compares the trajectory after parameter tuning with the historical stabilization template again, and writes the current event, the distribution of transition components, the parameter tuning path and the stabilization result into the knowledge base for direct use in subsequent similar switches.

[0006] Preferably, the transition window weaving module includes an interface pulse extraction unit and a sequence attachment weaving unit; The interface pulse extraction unit extracts, merges, and fragments the original switching traces from MES, PLC, equipment logs, and continuous production records into a single unit. The original switching traces include batch end time, first piece online time, formula call time, shift change time, short stop recovery time, cycle restart time, fixture cleaning end time, as well as the cycle difference between the preceding and following pieces and the process fluctuation trajectory of the first few pieces after restart. The specific process is as follows: Time base normalization processing is performed on the original switching traces to unify the timestamps of different systems to the same production line clock; then event classification processing is performed to mark batch change, formula change, shift change, short stop restart, and fixture cleaning restart as different event types. Then, perform adjacent event folding processing. If multiple switching actions occur consecutively within a short period of time, they are merged into a composite switching event to prevent the same transition interval from being repeatedly cut. Finally, for each switching event, the steady-state production segment is traced back to extract the steady-state length before the event; the initial production segment is traced back to extract the beat drop pattern, window length, and recovery initial fluctuation density from the first piece to the Kth piece, and encoded as an interface pulse segment.

[0007] Preferably, the sequence-attached weaving unit attaches the interface pulse fragments to a unified product sequence axis, forming a set of candidate transition windows. The specific steps are as follows: Read the interface pulse segment, then call the product flow record, first item online record and continuous beat record to establish a unique product sequence axis, that is, arrange each product into a continuous sequence according to the actual station passing order; Starting from the first event after the switching event, the corresponding interface pulse segment is mapped into the sequence, and extended backward according to the beat fall trajectory and process fluctuation trajectory to define an initial transition interval; during the extension process, the candidate boundary of the interval is dynamically determined based on the beat difference narrowing trend, the initial fluctuation density fall trend and the degree of retention of the switching tail of the segment after the event. The candidate boundary is determined as follows: First, a component-by-component comparison chain is established for the continuous products after the switchover, using the previous steady-state reference segment as a benchmark; then, the boundary search phase begins. When several consecutive products simultaneously meet the three conditions of "the beat difference enters the steady-state convergence range, the process fluctuation value falls back to the steady-state allowable range, and the wake retention mark changes from an active state to a decaying state", the first product of the continuous segment is set as the candidate boundary starting point; and then the verification continues. If the subsequent adjacent products still maintain the above state continuously and stably, the aforementioned candidate boundary starting point is officially determined as the candidate boundary of the transition range. The product segment between the switching start point and the candidate boundary is used as the initial transition interval, and each product in the window is given a switching sequence number, the number of the transition window to which it belongs, an interface pulse intensity label, and a pre-steady-state reference segment index. After processing, the switching traces that were originally scattered in MES, PLC, equipment logs and production records are unified and organized into a transition interval description result based on product sequence expansion, and finally output a set of candidate transition windows with timing structure and boundary markers.

[0008] Preferably, the transition component identification module includes a convergence trajectory segmentation unit and a status label determination unit; The convergence trajectory subdivision unit performs intra-window subdivision processing on the candidate transition window set. First, it expands the candidate transition windows one by one, and establishes a local time sequence chain for the product sequence in each window. Using product ranking as the main line, the cycle time difference change, process variable fluctuation and control quantity fall trajectory corresponding to continuous products within the window are divided into several continuous short segments. Each short segment retains at least the start and end ranking, fluctuation direction, fluctuation slope, fall amplitude and local stability length. After generating the short segment, using the steady-state reference segment before the switch as a baseline, the trajectory within the current window is divided into three segments to obtain the trajectory subdivision results. The process is as follows: The segments that still have obvious switching disturbances and whose trajectories deviate from the steady-state reference band are marked as transition state candidate segments; The segment where the trajectory begins to converge back toward the steady-state reference zone, but still exhibits tail sway or local slippage, is marked as a candidate segment for returning to steady state. The segments whose trajectory shape is consistent with the steady-state reference segment and whose beat and process variables have fallen back into the reference zone are marked as steady-state candidate segments.

[0009] Preferably, based on the trajectory segmentation results, the status label determination unit completes the identification of transition component categories and the determination of convergence completion points. The specific processing method is as follows: Retrieve the historical normal switching sample library, construct the stabilization template cluster corresponding to the current event type, and then perform segment similarity matching between the stabilization candidate segments and the template cluster segment by segment. Compare the degree of fit between the current product segment and the historical normal stabilization segment in terms of beat drop pattern, process variable convergence pattern and control quantity recovery rhythm. If a segment containing a product still shows deviation, unidirectional trajectory slippage, and has not entered the stable convergence zone, then the product is marked as a strong transition component. If the trajectory has begun to approach the steady-state zone, but still retains local oscillations, wake remnants, or incomplete convergence, it is marked as a weak transition piece; If a trajectory is within the candidate window but already closely matches a historical stable segment, and is included in the window only because it is in the early stage of the sequence, it is marked as a pseudo transition piece. If the trajectory has stabilized and entered the stable reference zone, it is marked as a stable component; Continue searching along the product sequence. When it is found that a certain product starts and several subsequent products maintain a stable fit with the stabilizing template cluster, and there is no more reverse slippage of the trajectory, re-increase of process variables, or recovery of control imbalance, the product position is determined as the convergence completion point.

[0010] Preferably, the edge-following risk screening module includes a boundary neighborhood mapping unit and an object diversion and wake delimitation unit; Based on the transition component identification results, the boundary neighborhood mapping unit establishes boundary proximity relationships for each transition component within the window; First, expand each transition window one by one and extract the local feature string of each transition component. The local feature string includes trajectory shape, beat fall method, process fluctuation residue, window length combination, parameter cutback rhythm and positional relationship before and after convergence completion point. The local feature string of the current transition component is compared bidirectionally with historical boundary neighborhood samples and historical safe recovery samples. The comparison with the historical boundary neighborhood sample set is used to obtain the degree of similarity between the current transition component and the boundary neighborhood features in terms of trajectory shape, beat fall mode, window length combination, parameter cut-back speed and process fluctuation residue. The comparison with the historical safe recovery sample set is used to obtain the degree of similarity between the current transition component and the safe recovery characteristics in various aspects; Based on the comparison results of the two types of proximity, boundary proximity marks and safety stabilization proximity marks are generated for each transition component, and a boundary proximity sequence and a safety stabilization proximity sequence are formed by arranging them one by one along the product sequence.

[0011] Preferably, the object splitting and wake delimitation unit further splits the transition element based on the boundary proximity sequence and determines the wake influence range. The processing procedure is as follows: Rearrange the transition components in the window according to product order, and compare their boundary proximity sequence with the safety stabilization proximity sequence one by one; If a transitional component is in a non-convergent state, but its trajectory, rhythm recovery pattern, and parameter retracement rhythm are generally close to historical safe stabilization samples, and the boundary proximity trend continues to weaken in subsequent rankings, then it is marked as an observation component. If a transition component simultaneously approaches historical boundary neighborhood samples in multiple features such as trajectory shape, window length combination, beat fall method, and parameter cut-back speed, and the proximity relationship remains continuous in the adjacent order, then it is marked as a narrow-domain intervention component. After the diversion is completed, a trail search is performed backward around the narrow-domain intervention component: the changes in process fluctuation residue, beat difference attenuation trend, interface pulse residual intensity and boundary proximity continuity of subsequent adjacent products are checked in turn; When subsequent products still retain the same source switching trail and form a continuous edge chain with the narrow domain intervention component, the continuous section will be included in the trail's influence range. When the continuity of subsequent product contact is interrupted, process fluctuations fall back to the safe stabilization zone, and interface pulse residues enter a low-activity state, the location will be used as the termination boundary of the wake's influence range.

[0012] Preferably, the micro-domain parameter tuning control module includes a local script arrangement unit and an observation window execution and secondary adjustment unit; The local script arrangement unit is expanded in the form of candidate transition windows. The narrow-domain intervention components within the same window are rearranged according to product order, and then each narrow-domain intervention component is linked to its corresponding trail influence range to form a local intervention segment chain. Based on the current transition event type, the non-converged state type of the current object, and the boundary approach direction, the corresponding local parameter tuning script is retrieved from the historical sample library; the local parameter tuning script shall at least include the target parameter set, parameter action sequence, first round adjustment range, action start item number, action end item number, and observation trigger condition; After the script retrieval is completed, the influence range of the boundary-close sequence and the trail will be overlaid for analysis: When the boundary proximity relationship is concentrated only near a single narrow-domain intervention, the parameter action only covers the small interval where the narrow-domain intervention is located; When the boundary proximity relationship extends continuously along the adjacent sequence, and the influence range of the wake covers several subsequent items, the parameter action range is extended to the corresponding continuous segment. Subsequently, based on the determined parameter action range, a first round of tentative narrow-range adjustment commands is generated, so that the local parameter actions first act on the target transition part and its wake-affected section, and the generated parameter action plan is output.

[0013] Preferably, the observation window execution and secondary adjustment unit executes the first round of tentative narrow-range adjustment according to the parameter action sequence; after the first round of adjustment takes effect, the same parameter is not modified again, and a minimum observation window is added to the current script based on the current transition event type and historical recovery rhythm; the minimum observation window is limited by product priority and local beat, and a freeze constraint is applied to the same parameter that has been adjusted before the observation window ends, so that it does not enter the repeated modification state; The process trajectory, boundary proximity sequence changes, and tail decay of the corresponding product within the smallest observation window are collected and compared with the local state before the first round of adjustments: When the process trajectory within the observation window has deviated from the boundary approach zone, the influence range of the wake begins to shrink, and no new continuous edge-attaching chains appear in subsequent products, retain the results of the first round of adjustments and end the current local intervention. When the boundary remains close to the edge, the trail decay is insufficient, or adjacent products continue to enter the edge chain within the observation window, the freeze constraint is lifted after the observation window ends, and the subsequent trimming path in the same local parameter tuning script is called based on the current observation results to generate a second local trimming instruction.

[0014] Preferably, the stabilization verification and write-back module includes a stabilization verification unit and a chain-like knowledge write-back unit; The stabilization verification unit establishes a feedback sequence after this parameter adjustment based on the candidate transition window. It rearranges the cycle recovery trajectory, process fluctuation decline trajectory, and parameter cut-off trajectory of each product in the observation window according to the product order and compares them with the historical stabilization template cluster. When the subsequent product trajectory closely matches the historical normal recovery segment, it is determined that the current window has entered the recovery state; If the subsequent product trajectory does not meet the continuous bonding requirement, it is determined that the current window is in an incomplete recovery state; Compare the boundary proximity sequence after parameter tuning with the edge proximity distribution before parameter tuning to check whether the edge proximity chain has been interrupted, whether the influence range of the trail has shrunk backward, and whether the boundary proximity direction has changed from a continuous edge proximity state to a decaying state. If the current feedback sequence still does not meet the stabilization template requirements, or if the boundary-close sequence still retains a continuous edge chain and the influence range of the trail has not converged, then this window will be marked as not having completed stabilization, and a feedback control result will be generated. The returned control results include at least the current script non-closed flag, the suggested shrinking parameter application area, the suggested extended observation window range, and control suggestions for continuing to freeze or change the local script. If the current feedback sequence remains stably aligned with the historical stabilization template, and the boundary-adjacent sequence has moved out of the danger zone and the influence range of the wake has ended before the predetermined boundary, then this window will be marked as "stabilization complete" and the corresponding window will be switched to the write-back state. The chain-like knowledge write-back unit reads the writeable status information, extracts the corresponding information according to the complete processing order of the transition window, and organizes the extracted information into a complete transition chain sample according to the order of "event triggering - window weaving - transition component identification - edge screening - micro-domain parameter tuning - stabilization verification", and writes it into the transition knowledge base; After the writing is completed, the transition processing chain record is then classified and associated with the existing similar switching records in the knowledge base to generate a knowledge retrieval index, which can be directly called by subsequent switching events under the same machine, same recipe type, same switching category, and same shift structure.

[0015] The information includes at least the switching event type, window length, pre-steady-state reference segment, candidate transition window boundary, distribution of strong and weak transition components, convergence completion point, boundary proximity style, narrow-domain intervention component segment, wake influence range, local parameter tuning script identifier, parameter action sequence, first or second round of adjustment path, observation window length, and final stabilization result.

[0016] This invention provides a machine learning-based intelligent manufacturing production quality parameter dynamic optimization control system, which has the following beneficial effects: (1) During system operation, multiple consecutive switching actions are organized into a single composite switching event through event classification and adjacent event folding, avoiding repeated segmentation or overlapping of the same transition section, making the boundaries of switching events clearer, and facilitating subsequent unified analysis around the same transition process. For each switching event, the steady-state production segment is traced back forward and the initial production segment is traced backward, so that the system not only records the time of switching, but also obtains the steady-state length before the event, the beat drop pattern, the window length, and the initial fluctuation density of recovery, thereby expanding the original single-point switching record into an interface pulse segment with the correlation between the preceding and following processes.

[0017] (2) By retrieving historical normal switching sample libraries and constructing a stabilization template cluster corresponding to the current event type, the stabilization candidate segments within the current window can be matched segment by segment with similar historical stabilization segments, forming a judgment path based on historical similar switching processes, rather than relying on single experience or isolated rules for judgment. By classifying products into strong transition components, weak transition components, pseudo-transition components, and stable components, the system can further subdivide a seemingly continuous segment of products within the same candidate window. In this way, products that are still in a clearly non-converged state, products that are stabilizing but still retain tail disturbances, products that have basically reached a steady state but were included in the window due to their high priority, and products that have stabilized and returned to the reference band are clearly separated, making it easier for subsequent modules to continue processing according to object differences.

[0018] (3) By generating boundary proximity markers and safety stabilization proximity markers, and forming boundary proximity sequences and safety stabilization proximity sequences along the product order, the risk judgment results are no longer scattered single-item conclusions, but become a continuous risk distribution chain unfolding according to the product order. Subsequent modules can directly use this distribution chain to identify the risk expansion direction and the starting point of local intervention. The object diversion and wake delimitation unit further divides the transitional components into observation components and narrow-domain intervention components, so that the "objects that need to be observed" and the "objects that need to be directly adjusted" are clearly separated in this stage. In this way, when adjusting parameters later, the system can concentrate local actions on the small range corresponding to the narrow-domain intervention component, rather than intervening in all non-converged components at the same time.

[0019] (4) The first round of actions adopts a tentative narrow-range adjustment path, so that the local parameter adjustment starts with small-scale, low-amplitude movements, and then the subsequent path is judged in combination with the trajectory changes in the observation window. Compared with the whole segment uniform correction or one-time large adjustment, this approach is more suitable for the transition interval after switching, and can control the parameter movement within the local object range.

[0020] By continuously collecting process trajectories, boundary proximity sequence changes, and wake attenuation within the smallest observation window, and comparing them with the local state before the first round of adjustments, the system can make judgments based on continuous phenomena such as "whether the current edge-fitting section is exiting the danger zone, whether the wake section is starting to shrink, and whether subsequent products are still forming new continuous edge-fitting chains," rather than deciding subsequent actions based solely on a single moment or a single detection point. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the block diagram of the intelligent manufacturing production quality parameter dynamic optimization control system based on machine learning of the present invention. Figure 2 This is a schematic diagram of the candidate transition window determination process of the present invention; Figure 3 This is a schematic diagram of the process for obtaining the secondary local correction instruction of the present invention. Detailed Implementation

[0022] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0023] Example 1 This invention provides a machine learning-based intelligent manufacturing production quality parameter dynamic optimization control system. Please refer to [link / reference]. Figures 1 to 3 It includes a transition window weaving module, a transition component identification module, an edge-fitting risk screening module, a micro-domain parameter tuning and control module, and a stabilization verification and write-back module. The transition window weaving module collects timestamps, window lengths, cycle time differences, and initial process trajectories for batch change, formula change, shift change, short stop restart, and fixture cleaning, and aligns them according to the product sequence to form candidate transition windows. The transition component identification module divides the sequence within the candidate window into steady state, transition state, and return-to-steady state, and compares it with historical return-to-steady segments to mark strong transition components, weak transition components, and convergence completion points. The edge risk screening module continues to compare quality inspection records, agent quality labels and historical boundary neighborhood samples in the identified transitional parts to distinguish between observation parts and narrow-domain intervention parts, and at the same time gives the range of influence of the wake. The micro-domain parameter tuning control module calls the corresponding local parameter tuning script based on the narrow domain intervention component and the wake range, only performs exploratory adjustments in the small interval where the target component is located, and sets a minimum observation window to limit continuous changes to the same parameters; The stabilization verification and write-back module compares the trajectory after parameter tuning with the historical stabilization template again, and writes the current event, the distribution of transition components, the parameter tuning path and the stabilization result into the knowledge base for direct use in subsequent similar switches.

[0024] In this embodiment, scattered records after batch change, formula change, shift change, short stop restart, and fixture cleaning restart are uniformly organized onto the same product sequence chain. This ensures that switching events are no longer isolated time point records, but rather form candidate transition windows with starting points, extended segments, and candidate boundaries, facilitating subsequent processing around specific product sequences. Instead of using a fixed first N items isolation method or uniformly judging based on the entire batch, candidate transition windows are first constructed, and then the actual non-converged segments are identified within the windows. This allows products that are close to a stable state to be processed separately from those still experiencing switching disturbances, reducing the likelihood of upstream products being included in the same control logic.

[0025] The edge-risk screening module doesn't just look at pass / fail results; it further filters out transitional parts that are "not yet scrapped but approaching the defect boundary." This separates the output of observation parts and narrow-area intervention parts, allowing the focus to be concentrated on the small, truly needing intervention. By determining the range of influence of the wake, the system can identify continuous product segments that are still affected by the same-source switching wake after the narrow-area intervention part. This prevents local risks from being seen as isolated phenomena but rather as continuously propagating segments that are described and handled.

[0026] The micro-domain parameter tuning control module calls local parameter tuning scripts based on narrow-domain intervention components and wake segments. It performs tentative adjustments only within the target small interval, avoiding uniform corrections for the entire batch. This ensures that the parameter action range corresponds to the abnormal segment range, reducing the likelihood of normal products being affected by cascading corrections. Through a minimum observation window and same-parameter freeze constraints, the system observes subsequent trajectory changes after the first round of adjustments before determining subsequent action paths. This prevents the same parameter from being continuously modified within a short period, providing a stable observation interval for cycle recovery and process stabilization in the transition segment.

[0027] The stabilization verification and write-back module compares the feedback trajectory after parameter adjustment with the historical stabilization template again, and re-verifies it by combining the boundary proximity sequence and the range of the tail. This ensures that the local parameter adjustment does not stop at the level of "parameters have been adjusted", but further gives a clear processing result of "stabilized" or "continue to shrink the action segment and extend the observation window".

[0028] Transitional knowledge is no longer archived separately as scattered fields, but is written as a complete chain sample in the order of "event triggering - window weaving - transition component identification - edge screening - micro-domain parameter tuning - stabilization verification", so that subsequent similar switches can directly call the processing path that is closer to the structure of the machine, the formula, and the shift.

[0029] Example 2 Please refer to Figure 1 Specifically: the transition window weaving module includes an interface pulse extraction unit and a sequence attachment weaving unit; The interface pulse extraction unit extracts, merges, and fragments the original switching traces from MES, PLC, equipment logs, and continuous production records into a single unit. The original switching traces include batch end time, first piece online time, formula call time, shift change time, short stop recovery time, cycle restart time, fixture cleaning end time, as well as the cycle difference between the preceding and following pieces and the process fluctuation trajectory of the first few pieces after restart. The specific process is as follows: Time base normalization processing is performed on the original switching traces to unify the timestamps of different systems to the same production line clock; then event classification processing is performed to mark batch change, formula change, shift change, short stop restart, and fixture cleaning restart as different event types. Then perform adjacent event folding processing. If multiple switching actions occur consecutively within a short period of time, they are merged into a composite switching event. Finally, for each switching event, the steady-state production segment is traced back to extract the steady-state length before the event; the initial production segment is traced back to extract the beat drop pattern, window length, and recovery initial fluctuation density from the first piece to the Kth piece, and encoded as an interface pulse segment.

[0030] The sequence-attached weaving unit attaches interface pulse segments to a unified product sequence axis, forming a set of candidate transition windows. The specific steps are as follows: Read the interface pulse fragments, then call the product flow records, first item online records and continuous cycle records to establish a unique product sequence axis; Starting with the first event after the switching event, the corresponding interface pulse segment is mapped into the sequence, and then extended backward according to the beat fall trajectory and process fluctuation trajectory to define an initial transition interval; During the unfolding process, the beat difference, process fluctuation value and switching wake retention status of each product relative to the previous steady-state reference segment are extracted simultaneously to construct a continuous boundary judgment chain; Using the previous steady-state reference segment as the comparison benchmark, a convergence search is performed on the continuous products after the switch: When the product in a certain continuous segment shows a continuous narrowing of the cycle time difference, falls back to the steady state range in the process fluctuation, and changes from the dominant state to the residual state in the switching wake, the starting product of the continuous segment is determined as the candidate boundary starting point. Then, continuous stability verification and smoothing review are performed on adjacent products after the starting point. Only when the subsequent segments maintain convergence and continuity, there is no reverse amplification of the beat difference, the process fluctuation rises again, and the wake becomes active again, can they be officially determined as candidate boundaries. The product segment between the switching start point and the candidate boundary is used as the initial transition interval, and each product in the window is given a switching sequence number, the number of the transition window to which it belongs, an interface pulse intensity label, and a pre-steady-state reference segment index. After processing, the switching traces that were originally scattered in MES, PLC, equipment logs and production records are unified and organized into a transition interval description result based on product sequence expansion, and finally output a set of candidate transition windows with timing structure and boundary markers.

[0031] In this embodiment, scattered switching traces from MES, PLC, equipment logs, and continuous production records are uniformly extracted and unified under the same production line clock. This changes the original situation where multi-source data were fragmented and difficult to analyze directly, enabling switching events such as batch change, formula change, shift change, short stop restart, and fixture cleaning restart to be continuously identified on the same time reference.

[0032] By classifying and folding adjacent events, multiple consecutive switching actions are organized into a single composite switching event, avoiding repeated segmentation or overlapping of the same transition section. This makes the boundaries of switching events clearer and facilitates unified analysis around the same transition process. For each switching event, the system simultaneously traces back the steady-state production segment and backward the initial production segment. This allows the system to not only record the switching time but also obtain the steady-state length before the event, the beat drop pattern, the window length, and the initial fluctuation density of the recovery phase. This expands the original single-point switching record into interface pulse segments with preceding and following process relationships.

[0033] By establishing a unique product sequence axis, interface pulse segments are linked to the sequentially arranged product sequence, ensuring a one-to-one correspondence between switching events and specific products. This solves the problem in existing systems where "a switch has occurred, but it's impossible to pinpoint which products were affected." By expanding backward from the first item after the switching event and simultaneously extracting the cycle time difference, process fluctuation value, and switching tail retention status for each product, the system can construct a continuous boundary judgment chain around the actual product sequence. This eliminates the reliance on a static division method with a fixed first N items, making the transition interval's landing point closer to the actual production process.

[0034] Previously, a steady-state reference segment was used as the comparison benchmark. A convergence search was performed on consecutive products after the handover, so that the determination of the candidate boundary was based on the joint changes of continuously narrowing beat difference, decreasing process fluctuation, and attenuating handover wake. This solved the problem that it was difficult to accurately identify the termination position of the transition segment based on a single threshold or a fixed number of pieces. Through continuous stability verification and smoothing review, adjacent products after the starting point of the candidate boundary were further verified, reducing misjudgments caused by accidental stability of a single piece or local short-term stability. This made the candidate boundary position more stable, and the transition window description results more suitable as the input basis for subsequent transition piece identification.

[0035] The product segment between the switching start point and the candidate boundary is used as the initial transition interval. Each product in the window is given a switching sequence number, the number of the transition window to which it belongs, an interface pulse intensity label, and a pre-steady-state reference segment index. This ensures that each product has a clear temporal position and context marker, which facilitates subsequent modules to continue to carry out convergence state determination and edge risk screening.

[0036] Example 3 Please refer to Figure 2 Specifically: the transition component identification module includes a convergence trajectory segmentation unit and a status label determination unit; The convergence trajectory subdivision unit performs intra-window subdivision processing on the candidate transition window set. First, it expands the candidate transition windows one by one, and establishes a local time sequence chain for the product sequence in each window. Using product ranking as the main line, the cycle time difference changes, process variable fluctuations, and control quantity decline trajectories corresponding to consecutive products within the window are divided into several continuous short segments. After generating the short segment, using the steady-state reference segment before the switch as a baseline, the trajectory within the current window is divided into three segments to obtain the trajectory subdivision results. The process is as follows: The segments that still have obvious switching disturbances and whose trajectories deviate from the steady-state reference band are marked as transition state candidate segments; The segment where the trajectory begins to converge back toward the steady-state reference zone, but still exhibits tail sway or local slippage, is marked as a candidate segment for returning to steady state. The segments whose trajectory shape is consistent with the steady-state reference segment and whose beat and process variables have fallen back into the reference zone are marked as steady-state candidate segments.

[0037] Based on the trajectory segmentation results, the status label determination unit completes the identification of transition component categories and the determination of convergence completion points. The specific processing method is as follows: Retrieve the historical normal switching sample library, construct the stabilization template cluster corresponding to the current event type, and then perform segment similarity matching between the stabilization candidate segments and the template cluster segment by segment. Compare the degree of fit between the current product segment and the historical normal stabilization segment in terms of beat drop pattern, process variable convergence pattern and control quantity recovery rhythm. If a segment containing a product still shows deviation, unidirectional trajectory slippage, and has not entered the stable convergence zone, then the product is marked as a strong transition component. If the trajectory has begun to approach the steady-state zone, but still retains local oscillations, wake remnants, or incomplete convergence, it is marked as a weak transition piece; If a trajectory is within the candidate window but already matches a historical stable segment, and is included in the window only because it is in the early stage, it is marked as a pseudo transition piece. If the trajectory has stabilized and entered the stable reference zone, it is marked as a stable component; Continue searching along the product sequence. When it is found that a certain product starts and several subsequent products maintain a stable fit with the stabilizing template cluster, and there is no more reverse slippage of the trajectory, re-increase of process variables, or recovery of control imbalance, the product position is determined as the convergence completion point.

[0038] In this embodiment, a local time-series chain is first established for the continuous product sequence within the candidate transition window. Then, the cycle time difference change, process variable fluctuation, and control quantity fallback trajectory are continuously segmented into short segments according to the product order. This prevents the front-end products after the switch from being treated as a whole and treated roughly, but rather decomposed into trajectory segments that can be analyzed segment by segment, making it easier to identify the state differences within the transition segment. Using the steady-state reference segment before the switch as a unified comparison benchmark, the trajectory within the window is divided into three segments: transition state candidate segment, return-to-steady-state candidate segment, and steady-state candidate segment. This allows the system to distinguish three different states from the continuous change process: "still in disturbance," "currently converging," and "already back in the steady-state band." This changes the situation in the existing method where only a single-point threshold or a fixed number of front-end items are processed.

[0039] By retrieving historical normal handover sample libraries and constructing a stabilization template cluster corresponding to the current event type, the stabilization candidate segments within the current window can be matched segment by segment with similar historical stabilization segments, forming a decision path based on historical similar handover processes, rather than relying on single experience or isolated rules. By classifying products into strong transitions, weak transitions, pseudo-transitions, and stable components, the system can further subdivide seemingly continuous segments of products within the same candidate window. In this way, products that are still in a clearly non-converged state, products that are stabilizing but still retain tail disturbances, products that have basically reached steady state but were included in the window due to their high priority, and products that have stabilized back to the reference band are clearly separated, facilitating subsequent modules to continue processing according to object differences.

[0040] The distinction between strong and weak transition components allows subsequent risk screening to focus on products closer to the non-converged core area, rather than processing the entire candidate window. The identification of pseudo-transition components and stable components can separate products that do not require further intervention from the subsequent processing scope, reducing the mixing of normal and abnormal products in the early stages.

[0041] By continuing to search for the convergence completion point along the product sequence, and using "a number of consecutive pieces maintaining stable contact with the stabilization template cluster, and no further reverse slippage of the trajectory, re-increase of process variables, or recovery of control imbalance" as the judgment basis, the termination position of the transition interval is established on the continuous stabilization process, rather than on a fixed number of pieces, a fixed time window, or the instantaneous state of a single piece, making the convergence completion point closer to the actual working conditions.

[0042] Example 4 Please refer to Figure 1 Specifically: the edge-following risk screening module includes a boundary neighborhood mapping unit and an object diversion and wake delimitation unit; Based on the transition component identification results, the boundary neighborhood mapping unit establishes boundary proximity relationships for each transition component within the window; First, expand each transition window one by one and extract the local feature string of each transition component. The local feature string includes trajectory shape, beat fall method, process fluctuation residue, window length combination, parameter cutback rhythm and positional relationship before and after convergence completion point. The local feature string of the current transition component is compared bidirectionally with historical boundary neighborhood samples and historical safe recovery samples. The comparison with the historical boundary neighborhood sample set is used to obtain the degree of similarity between the current transition component and the boundary neighborhood features in terms of trajectory shape, beat fall mode, window length combination, parameter cut-back speed and process fluctuation residue. The comparison with the historical safe recovery sample set is used to obtain the degree of similarity between the current transition component and the safe recovery characteristics in various aspects; Based on the comparison results of the two types of proximity, boundary proximity marks and safety stabilization proximity marks are generated for each transition component, and a boundary proximity sequence and a safety stabilization proximity sequence are formed by arranging them one by one along the product sequence.

[0043] The object splitting and wake delimitation unit, based on the boundary proximity sequence, further splits the transition element and determines the wake influence range. The process is as follows: Rearrange the transition components in the window according to product order, and compare their boundary proximity sequence with the safety stabilization proximity sequence one by one; If a transitional component is in a non-convergent state, but its trajectory, rhythm recovery pattern, and parameter retracement rhythm are generally close to historical safe stabilization samples, and the boundary proximity trend continues to weaken in subsequent rankings, then it is marked as an observation component. If a transition component simultaneously approaches historical boundary neighborhood samples in multiple features such as trajectory shape, window length combination, beat fall method, and parameter cut-back speed, and the proximity relationship remains continuous in the adjacent order, then it is marked as a narrow-domain intervention component. After the diversion is completed, a trail search is performed backward around the narrow-domain intervention component: the changes in process fluctuation residue, beat difference attenuation trend, interface pulse residual intensity and boundary proximity continuity of subsequent adjacent products are checked in turn; When subsequent products still retain the same source switching trail and form a continuous edge chain with the narrow domain intervention component, the continuous section will be included in the trail's influence range. When the continuity of subsequent product contact is interrupted, process fluctuations fall back to the safe stabilization zone, and interface pulse residues enter a low-activity state, the location will be used as the termination boundary of the wake's influence range.

[0044] In this embodiment, based on the results of the transition component identification, a boundary proximity relationship is further established, so that the system no longer stops at the level of "which products are in a non-converged state", but further identifies "which non-converged products have begun to approach the defect boundary". In this way, the risk differences in the front-end products are further expanded, and the subsequent processing object is no longer all transition components in general.

[0045] By extracting local feature strings such as trajectory morphology, cycle time retracement, residual process fluctuations, window length combinations, parameter cut-back rhythm, and positional relationships before and after convergence completion points, each transition component is analyzed within the same local feature framework. This changes the approach of judging solely based on a single quality inspection value, a single process value, or a single alarm signal, allowing risk assessment to be based on continuous process behavior. By simultaneously comparing the current transition component with historical boundary neighborhood samples and historical safe stabilization samples, the system can obtain the boundary neighborhood proximity and safe stabilization proximity respectively, thus distinguishing between the two adjacent but different states of "approaching the defective boundary" and "currently stabilizing normally." This reduces the possibility of misjudging normally stabilizing products as intervention targets and also reduces the possibility of simply treating products that are close to the edge as ordinary transition components.

[0046] By generating boundary proximity markers and safety stabilization proximity markers, and forming boundary proximity sequences and safety stabilization proximity sequences along the product sequence, the risk assessment results are no longer scattered, single-item conclusions, but rather a continuous risk distribution chain unfolding according to the product sequence. Subsequent modules can directly utilize this distribution chain to identify the risk expansion direction and the starting point of local intervention. The object diversion and wake delimitation unit further divides the transitional components into observation components and narrow-domain intervention components, clearly separating the "objects that need to be observed" from the "objects that need to be directly tuned" at this stage. In this way, the system can concentrate local actions on the small range corresponding to the narrow-domain intervention component during subsequent parameter tuning, rather than intervening in all non-converged components at once.

[0047] When certain transitional components, although not yet fully stabilized, show trajectory trends, rhythm recovery patterns, and parameter retracement rhythms that closely resemble historical safe stabilization samples, and the trend of boundary convergence continues to weaken, they are marked as observation components. This allows the scheme to identify objects that are "still within the transition window but stabilizing along a normal path." This approach offers a more refined classification compared to the current situation where "front-end products are uniformly included in the same handling scope."

[0048] When certain transitional components converge towards historical boundary neighborhood samples in multiple features such as trajectory morphology, window length combination, beat retracement mode, and parameter cut-back speed, and this proximity relationship occurs consecutively in adjacent sequences, they are marked as narrow-domain interventional components. This provides a clear target for subsequent control actions. Compared to the approach of relying on a fixed number of components, fixed thresholds, or manual experience to define intervention targets in the background, the target segment is more focused.

[0049] By performing a trail search backward around the narrow-domain intervention component, the system can continue to identify the continuation of the same-source switching trail in subsequent adjacent products. This allows local anomalies to be described as continuous impact segments propagating along the product sequence, rather than being viewed as single-item problems. This is more suitable for handling situations where "preceding local anomalies propagate to subsequent products."

[0050] By sequentially examining residual process fluctuations, the attenuation trend of beat differences, the residual intensity of interface pulses, and the continuity of boundary proximity, the termination boundary of the wake's influence range no longer depends on the fixed number of components for truncation, but is determined based on the continuous descent process. In this way, the endpoint of the wake segment more closely approximates the actual working conditions, allowing the subsequent micro-domain parameter tuning control module to narrow the range of parameter actions.

[0051] Example 5 Please refer to Figure 3 Specifically: the micro-domain parameter tuning control module includes a local script arrangement unit and an observation window execution and secondary adjustment unit; The local script arrangement unit is expanded in the form of candidate transition windows. The narrow-domain intervention components within the same window are rearranged according to product order, and then each narrow-domain intervention component is linked to its corresponding trail influence range to form a local intervention segment chain. Based on the current transition event type, the current object's non-converged state type, and the boundary approach direction, retrieve the corresponding local parameter tuning script from the historical sample library; After the script retrieval is completed, the influence range of the boundary-close sequence and the trail will be overlaid for analysis: When the boundary proximity relationship is concentrated only near a single narrow-domain intervention, the parameter action only covers the small interval where the narrow-domain intervention is located; When the boundary proximity relationship extends continuously along the adjacent sequence, and the influence range of the wake covers several subsequent items, the parameter action range is extended to the corresponding continuous segment. Subsequently, based on the determined parameter action range, the first round of tentative narrow-range adjustment commands is generated, and the generated parameter action plan is output.

[0052] The observation window execution and secondary adjustment unit executes the first round of tentative narrow-range adjustments according to the parameter action sequence; after the first round of adjustments takes effect, the same parameter is not modified again, and a minimum observation window is added to the current script based on the current transition event type and historical recovery rhythm; The process trajectory, boundary proximity sequence changes, and tail decay of the corresponding product within the smallest observation window are collected and compared with the local state before the first round of adjustments: When the process trajectory within the observation window has deviated from the boundary approach zone, the influence range of the wake begins to shrink, and no new continuous edge-attaching chains appear in subsequent products, retain the results of the first round of adjustments and end the current local intervention. When the boundary remains close to the edge, the trail decay is insufficient, or adjacent products continue to enter the edge chain within the observation window, the freeze constraint is lifted after the observation window ends, and the subsequent trimming path in the same local parameter tuning script is called based on the current observation results to generate a second local trimming instruction.

[0053] The stabilization verification and write-back module includes a stabilization verification unit and a chain-based knowledge write-back unit; The stabilization verification unit establishes a feedback sequence after this parameter adjustment based on the candidate transition window. It rearranges the cycle recovery trajectory, process fluctuation decline trajectory, and parameter cut-off trajectory of each product in the observation window according to the product order and compares them with the historical stabilization template cluster. When the subsequent product trajectory closely matches the historical normal recovery segment, it is determined that the current window has entered the recovery state; If the subsequent product trajectory does not meet the continuous bonding requirement, it is determined that the current window is in an incomplete recovery state; Compare the boundary proximity sequence after parameter tuning with the edge proximity distribution before parameter tuning to check whether the edge proximity chain has been interrupted, whether the influence range of the trail has shrunk backward, and whether the boundary proximity direction has changed from a continuous edge proximity state to a decaying state. If the current feedback sequence still does not meet the stabilization template requirements, or if the boundary-close sequence still retains a continuous edge chain and the influence range of the trail has not converged, then this window will be marked as not having completed stabilization, and a feedback control result will be generated. If the current feedback sequence remains stably aligned with the historical stabilization template, and the boundary-adjacent sequence has moved out of the danger zone and the influence range of the wake has ended before the predetermined boundary, then this window will be marked as "stabilization complete" and the corresponding window will be switched to the write-back state. The chain-like knowledge write-back unit reads the writeable status information, extracts the corresponding information according to the complete processing order of the transition window, and organizes the extracted information into a complete transition chain sample according to the order of "event triggering - window weaving - transition component identification - edge screening - micro-domain parameter tuning - stabilization verification", and writes it into the transition knowledge base; After the writing is completed, the transition processing chain record is classified and associated with the existing similar switching records in the knowledge base to generate a knowledge call index, which can be directly called by subsequent switching events under the same machine, same recipe type, same switching category and same shift structure.

[0054] In this embodiment, the narrow-domain intervention components and the influence range of the trail within the same window are linked and organized using the candidate transition window as a unit to form a local intervention segment chain. This makes the parameter actions no longer directed towards the entire batch or the entire front-end product, but directly fall on the edge segment and its subsequent trail segment, making the processing objects more concentrated and the action boundaries clearer.

[0055] The local script orchestration unit incorporates the current transition event type, non-converged state type, and boundary proximity direction into the script retrieval conditions. This establishes a correspondence between the local parameter tuning script and the current switching condition, current object state, and current risk direction, changing the previous approach of uniformly applying fixed parameter tuning rules or relying solely on manual experience. By overlaying the boundary proximity sequence with the influence range of the wake, the system can distinguish between "edge-fitting relationships concentrated only near a single narrow-domain intervention component" and "edge-fitting relationships that continuously extend along adjacent sequences." Based on this, it determines whether the parameter action covers a single small interval or a continuous segment, allowing the parameter's effective range to change with the actual edge-fitting distribution, no longer relying on a fixed number of components or a fixed length segment to define the parameter tuning range.

[0056] The initial adjustment process employs a tentative, narrow-range approach, starting with small-scale, low-amplitude adjustments to local parameters. This is then combined with observations of trajectory changes within the monitoring window to determine subsequent paths. Compared to uniform corrections or large-scale adjustments all at once, this method is more suitable for the transitional period after a switchover, keeping parameter adjustments within the local object's scope.

[0057] By continuously collecting process trajectories, boundary proximity sequence changes, and wake attenuation within the smallest observation window, and comparing them with the local state before the first round of adjustments, the system can make judgments based on continuous phenomena such as "whether the current edge-fitting section is exiting the danger zone, whether the wake section is starting to shrink, and whether subsequent products are still forming new continuous edge-fitting chains," rather than deciding subsequent actions based solely on a single moment or a single detection point.

[0058] When the process trajectory within the observation window has deviated from the boundary proximity zone, the influence range of the wake begins to shrink, and subsequent products no longer form new continuous edge-fitting chains, the system retains the results of the first round of adjustments and ends the current local intervention. When the edge-fitting relationship continues to extend, the wake attenuation is insufficient, or subsequent products continue to enter the edge-fitting chain, the system calls the subsequent adjustment path of the same script after the observation window ends. This phased action method gives the local parameter tuning process a clear extension logic, and it is no longer a simple one-time parameter tuning or repeated manual adjustments.

[0059] While the current window is still in an incomplete stabilization state, the system outputs a feedback control result and sends it back to the micro-domain parameter tuning control module, allowing it to further shrink the parameter application area, extend the observation window, or switch to a subsequent adjustment path. In this way, micro-domain tuning and stabilization verification form a closed loop, rather than being two separate steps.

[0060] When the current window enters the stable state, the chain-style knowledge write-back unit does not only record the result of a single parameter tuning, but writes the complete processing chain into the transition knowledge base in the order of "event triggering - window weaving - transition component identification - edge screening - micro-domain parameter tuning - stable verification". In this way, the knowledge recording object is no longer a series of scattered fields, but a chain sample that can replay the entire switching process.

[0061] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended technical solutions and their equivalents.

Claims

1. A machine learning-based intelligent manufacturing production quality parameter dynamic optimization control system, characterized in that: This includes a transition window weaving module, a transition component identification module, an edge-fitting risk screening module, a micro-domain parameter tuning and control module, and a stabilization verification and write-back module; The transition window weaving module collects timestamps, window lengths, cycle time differences, and initial process trajectories for batch change, formula change, shift change, short stop restart, and fixture cleaning, and aligns them according to the product sequence to form candidate transition windows. The transition component identification module divides the sequence within the candidate window into steady state, transition state, and return-to-steady state, and compares it with historical return-to-steady segments to mark strong transition components, weak transition components, and convergence completion points. The edge risk screening module continues to compare quality inspection records, agent quality labels and historical boundary neighborhood samples in the identified transitional parts to distinguish between observation parts and narrow-domain intervention parts, and at the same time gives the range of influence of the wake. The micro-domain parameter tuning control module calls the corresponding local parameter tuning script based on the narrow domain intervention component and the wake range, only performs exploratory adjustments in the small interval where the target component is located, and sets a minimum observation window to limit continuous changes to the same parameters; The stabilization verification and write-back module compares the trajectory after parameter tuning with the historical stabilization template again, and writes the current event, the distribution of transition components, the parameter tuning path and the stabilization result into the knowledge base for direct use in subsequent similar switches.

2. The intelligent manufacturing production quality parameter dynamic optimization control system based on machine learning according to claim 1, characterized in that: The transition window weaving module includes an interface pulse extraction unit and a sequence attachment weaving unit; The interface pulse extraction unit extracts, merges, and fragments the original switching traces from MES, PLC, equipment logs, and continuous production records into a single unit. The original switching traces include batch end time, first piece online time, formula call time, shift change time, short stop recovery time, cycle restart time, fixture cleaning end time, as well as the cycle difference between the preceding and following pieces and the process fluctuation trajectory of the first few pieces after restart. The specific process is as follows: Time base normalization processing is performed on the original switching traces to unify the timestamps of different systems to the same production line clock; then event classification processing is performed to mark batch change, formula change, shift change, short stop restart, and fixture cleaning restart as different event types. Then perform adjacent event folding processing. If multiple switching actions occur consecutively within a short period of time, they are merged into a composite switching event. Finally, for each switching event, the steady-state production segment is traced back to extract the steady-state length before the event; the initial production segment is traced back to extract the beat drop pattern, window length, and recovery initial fluctuation density from the first piece to the Kth piece, and encoded as an interface pulse segment.

3. The machine learning-based intelligent manufacturing production quality parameter dynamic optimization control system according to claim 2, characterized in that: The sequence-attached weaving unit attaches interface pulse segments to a unified product sequence axis, forming a set of candidate transition windows. The specific steps are as follows: Read the interface pulse fragments, then call the product flow records, first item online records and continuous cycle records to establish a unique product sequence axis; Starting with the first event after the switching event, the corresponding interface pulse segment is mapped into the sequence, and then extended backward according to the beat fall trajectory and process fluctuation trajectory to define an initial transition interval; During the unfolding process, the beat difference, process fluctuation value and switching wake retention status of each product relative to the previous steady-state reference segment are extracted simultaneously to construct a continuous boundary judgment chain; Using the previous steady-state reference segment as the comparison benchmark, a convergence search is performed on the continuous products after the switch: When the product in a certain continuous segment shows a continuous narrowing of the cycle time difference, falls back to the steady state range in the process fluctuation, and changes from the dominant state to the residual state in the switching wake, the starting product of the continuous segment is determined as the candidate boundary starting point. Then, continuous stability verification and smoothing review are performed on adjacent products after the starting point. Only when the subsequent segments maintain convergence and continuity, there is no reverse amplification of the beat difference, the process fluctuation rises again, and the wake becomes active again, can they be officially determined as candidate boundaries. The product segment between the switching start point and the candidate boundary is used as the initial transition interval, and each product in the window is given a switching sequence number, the number of the transition window to which it belongs, an interface pulse intensity label, and a pre-steady-state reference segment index. After processing, the switching traces that were originally scattered in MES, PLC, equipment logs and production records are unified and organized into a transition interval description result based on product sequence expansion, and finally output a set of candidate transition windows with timing structure and boundary markers.

4. The machine learning-based intelligent manufacturing production quality parameter dynamic optimization control system according to claim 3, characterized in that: The transition component identification module includes a convergence trajectory segmentation unit and a status label determination unit; The convergence trajectory subdivision unit performs intra-window subdivision processing on the candidate transition window set. First, it expands the candidate transition windows one by one, and establishes a local time sequence chain for the product sequence in each window. Using product ranking as the main line, the cycle time difference changes, process variable fluctuations, and control quantity decline trajectories corresponding to consecutive products within the window are divided into several continuous short segments. After generating the short segment, using the steady-state reference segment before the switch as a baseline, the trajectory within the current window is divided into three segments to obtain the trajectory subdivision results. The process is as follows: The segments that still have obvious switching disturbances and whose trajectories deviate from the steady-state reference band are marked as transition state candidate segments; The segment where the trajectory begins to converge back toward the steady-state reference zone, but still exhibits tail sway or local slippage, is marked as a candidate segment for returning to steady state. The segments whose trajectory shape is consistent with the steady-state reference segment and whose beat and process variables have fallen back into the reference zone are marked as steady-state candidate segments.

5. The machine learning-based intelligent manufacturing production quality parameter dynamic optimization control system according to claim 4, characterized in that: Based on the trajectory segmentation results, the status label determination unit completes the identification of transition component categories and the determination of convergence completion points. The specific processing method is as follows: Retrieve the historical normal switching sample library, construct the stabilization template cluster corresponding to the current event type, and then perform segment similarity matching between the stabilization candidate segments and the template cluster segment by segment. Compare the degree of fit between the current product segment and the historical normal stabilization segment in terms of beat drop pattern, process variable convergence pattern and control quantity recovery rhythm. If a segment containing a product still shows deviation, unidirectional trajectory slippage, and has not entered the stable convergence zone, then the product is marked as a strong transition component. If the trajectory has begun to approach the steady-state zone, but still retains local oscillations, wake remnants, or incomplete convergence, it is marked as a weak transition piece; If a trajectory is within the candidate window but already matches a historical stable segment, and is included in the window only because it is in the early stage, it is marked as a pseudo transition piece. If the trajectory has stabilized and entered the stable reference zone, it is marked as a stable component; Continue searching along the product sequence. When it is found that a certain product starts and several subsequent products maintain a stable fit with the stabilizing template cluster, and there is no more reverse slippage of the trajectory, re-increase of process variables, or recovery of control imbalance, the product position is determined as the convergence completion point.

6. The machine learning-based intelligent manufacturing production quality parameter dynamic optimization control system according to claim 5, characterized in that: The edge-following risk screening module includes a boundary neighborhood mapping unit and an object diversion and wake delimitation unit; Based on the transition component identification results, the boundary neighborhood mapping unit establishes boundary proximity relationships for each transition component within the window; First, expand each transition window one by one and extract the local feature string of each transition component. The local feature string includes trajectory shape, beat fall method, process fluctuation residue, window length combination, parameter cutback rhythm and positional relationship before and after convergence completion point. The local feature string of the current transition component is compared bidirectionally with historical boundary neighborhood samples and historical safe recovery samples. The comparison with the historical boundary neighborhood sample set is used to obtain the degree of similarity between the current transition component and the boundary neighborhood features in terms of trajectory shape, beat fall mode, window length combination, parameter cut-back speed and process fluctuation residue. The comparison with the historical safe recovery sample set is used to obtain the degree of similarity between the current transition component and the safe recovery characteristics in various aspects; Based on the comparison results of the two types of proximity, boundary proximity marks and safety stabilization proximity marks are generated for each transition component, and a boundary proximity sequence and a safety stabilization proximity sequence are formed by arranging them one by one along the product sequence.

7. The intelligent manufacturing production quality parameter dynamic optimization control system based on machine learning according to claim 6, characterized in that: The object splitting and wake delimitation unit, based on the boundary proximity sequence, further splits the transition element and determines the wake influence range. The process is as follows: Rearrange the transition components in the window according to product order, and compare their boundary proximity sequence with the safety stabilization proximity sequence one by one; If a transitional component is in a non-convergent state, but its trajectory, rhythm recovery pattern, and parameter retracement rhythm are generally close to historical safe stabilization samples, and the boundary proximity trend continues to weaken in subsequent rankings, then it is marked as an observation component. If a transition component simultaneously approaches historical boundary neighborhood samples in multiple features such as trajectory shape, window length combination, beat fall method, and parameter cut-back speed, and the proximity relationship remains continuous in the adjacent order, then it is marked as a narrow-domain intervention component. After the diversion is completed, a trail search is performed backward around the narrow-domain intervention component: the changes in process fluctuation residue, beat difference attenuation trend, interface pulse residual intensity and boundary proximity continuity of subsequent adjacent products are checked in turn; When subsequent products still retain the same source switching trail and form a continuous edge chain with the narrow domain intervention component, the continuous section will be included in the trail's influence range. When the continuity of subsequent product contact is interrupted, process fluctuations fall back to the safe stabilization zone, and interface pulse residues enter a low-activity state, the location will be used as the termination boundary of the wake's influence range.

8. The intelligent manufacturing production quality parameter dynamic optimization control system based on machine learning according to claim 7, characterized in that: The micro-domain parameter tuning control module includes a local script arrangement unit and an observation window execution and secondary adjustment unit; The local script arrangement unit is expanded in the form of candidate transition windows. The narrow-domain intervention components within the same window are rearranged according to product order, and then each narrow-domain intervention component is linked to its corresponding trail influence range to form a local intervention segment chain. Based on the current transition event type, the current object's non-converged state type, and the boundary approach direction, retrieve the corresponding local parameter tuning script from the historical sample library; After the script retrieval is completed, the influence range of the boundary-close sequence and the trail will be overlaid for analysis: When the boundary proximity relationship is concentrated only near a single narrow-domain intervention, the parameter action only covers the small interval where the narrow-domain intervention is located; When the boundary proximity relationship extends continuously along the adjacent sequence, and the influence range of the wake covers several subsequent items, the parameter action range is extended to the corresponding continuous segment. Subsequently, based on the determined parameter action range, the first round of tentative narrow-range adjustment commands is generated, and the generated parameter action plan is output.

9. The intelligent manufacturing production quality parameter dynamic optimization control system based on machine learning according to claim 8, characterized in that: The observation window execution and secondary adjustment unit executes the first round of tentative narrow-range adjustments according to the parameter action sequence; after the first round of adjustments takes effect, the same parameter is not modified again, and a minimum observation window is added to the current script based on the current transition event type and historical recovery rhythm; The process trajectory, boundary proximity sequence changes, and tail decay of the corresponding product within the smallest observation window are collected and compared with the local state before the first round of adjustments: When the process trajectory within the observation window has deviated from the boundary approach zone, the influence range of the wake begins to shrink, and no new continuous edge-attaching chains appear in subsequent products, retain the results of the first round of adjustments and end the current local intervention. When the boundary remains close to the edge, the trail decay is insufficient, or adjacent products continue to enter the edge chain within the observation window, the freeze constraint is lifted after the observation window ends, and the subsequent trimming path in the same local parameter tuning script is called based on the current observation results to generate a second local trimming instruction.

10. The machine learning-based intelligent manufacturing production quality parameter dynamic optimization control system according to claim 9, characterized in that: The stabilization verification and write-back module includes a stabilization verification unit and a chain-based knowledge write-back unit; The stabilization verification unit establishes a feedback sequence after this parameter adjustment based on the candidate transition window. It rearranges the cycle recovery trajectory, process fluctuation decline trajectory, and parameter cut-off trajectory of each product in the observation window according to the product order and compares them with the historical stabilization template cluster. When the subsequent product trajectory closely matches the historical normal recovery segment, it is determined that the current window has entered the recovery state; If the subsequent product trajectory does not meet the continuous bonding requirement, it is determined that the current window is in an incomplete recovery state; Compare the boundary proximity sequence after parameter tuning with the edge proximity distribution before parameter tuning to check whether the edge proximity chain has been interrupted, whether the influence range of the trail has shrunk backward, and whether the boundary proximity direction has changed from a continuous edge proximity state to a decaying state. If the current feedback sequence still does not meet the stabilization template requirements, or if the boundary-close sequence still retains a continuous edge chain and the influence range of the trail has not converged, then this window will be marked as not having completed stabilization, and a feedback control result will be generated. If the current feedback sequence remains stably aligned with the historical stabilization template, and the boundary-adjacent sequence has moved out of the danger zone and the influence range of the wake has ended before the predetermined boundary, then this window will be marked as "stabilization complete" and the corresponding window will be switched to the write-back state. The chain-like knowledge write-back unit reads the writeable status information, extracts the corresponding information according to the complete processing order of the transition window, and organizes the extracted information into a complete transition chain sample according to the order of "event triggering - window weaving - transition component identification - edge screening - micro-domain parameter tuning - stabilization verification", and writes it into the transition knowledge base; After the writing is completed, the transition processing chain record is classified and associated with the existing similar switching records in the knowledge base to generate a knowledge call index, which can be directly called by subsequent switching events under the same machine, same recipe type, same switching category and same shift structure.

Images