Reinforcement learning-based adaptive parameter optimization control method and system for industrial sewing machine

Abstract

The application discloses an industrial sewing machine adaptive parameter optimization control method and system based on reinforcement learning, relates to the technical field of intelligent control of industrial sewing machines, and comprises the following steps: carrying out defect risk budget verification and safety action projection on a candidate parameter increment action, mapping adjustment with the risks of thread breakage and needle skipping and feeding slip to a safety action space, forming a safety control instruction and an actual control state; based on the decay amplitude of a defect risk peak value in a trace forming memory state, calculating a defect accumulation inhibition score, and correcting the trace forming memory state and experience cache according to the defect accumulation inhibition score. Through the defect risk budget verification and safety action projection on the candidate parameter increment action, the application improves the advance, phase matching and execution safety of trace quality control.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for industrial sewing machines, and in particular to an adaptive parameter optimization control method and system for industrial sewing machines based on reinforcement learning. Background Technology

[0002] Industrial sewing machines are widely used in the processing of clothing, bags, and composite material products. Conventional control methods typically set target stitch length, target spindle speed, thread specifications, and seam type parameters around the sewing work order, and combine spindle encoders, tension detection, feed feedback, and vision inspection to obtain the operating status. During automated sewing, the controller adjusts the thread tension, feed displacement, presser foot pressure, and spindle speed based on the detection data to maintain the continuity of the stitch forming process and the consistency of processing.

[0003] The conventional methods mentioned above usually focus more on the quality deviation of a single needle cycle or local moment during parameter optimization, and are insufficient in expressing the cumulative evolution of tension changes, feeding deviation and stitch image deviation within continuous needle cycles. On the other hand, the parameter adjustment is not closely related to the constraints of mechanical phases such as needle insertion, loop formation, take-up coordination and feeding advancement, which can easily lead to the strategy output not being fully matched with the current mechanical action execution timing. Summary of the Invention

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

[0005] Therefore, this invention provides an adaptive parameter optimization control method for industrial sewing machines based on reinforcement learning to solve the problems of insufficient expression of continuous needle cycle defect accumulation and insufficient timing matching between parameter decision and mechanical phase execution in existing parameter optimization methods.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides an adaptive parameter optimization control method for industrial sewing machines based on reinforcement learning, which includes: collecting sewing work order data and real-time operation data of industrial sewing machines, and performing phase synchronization processing on the stitch cycle according to the spindle angle to generate a needle cycle phase data packet; Based on the needle cycle phase data package, the tension change, feeding deviation and stitch image deviation within the continuous needle cycle are compressed by sliding memory to construct a stitch forming memory state that characterizes the trend of defect accumulation, and a phase memory state vector is formed by combining the current phase label. Based on the phase memory state vector, the phase action mask is invoked to perform strategy mapping reasoning, and the parameter adjustment is restricted to the executable range of the current mechanical phase to form candidate parameter incremental actions. Defect risk budget verification and safety action projection are performed on the incremental actions of candidate parameters. Adjustments with risks of wire breakage and skipped needles and feeding slippage are mapped to the safety action space to form safety control commands and actual control states. Based on the decline of the peak defect risk in the trace forming memory state in the actual control state, the defect cumulative suppression score is calculated, and the trace forming memory state and experience cache are corrected according to the defect cumulative suppression score.

[0007] As a preferred embodiment of the reinforcement learning-based adaptive parameter optimization control method for industrial sewing machines described in this invention, the sewing work order data includes work order number, fabric type, number of fabric layers, thread specifications, spindle angle change data, target stitch distance, and target spindle speed. The real-time operating data includes spindle angle, spindle speed, wire tension, feed displacement, presser foot pressure, spindle load, and stitch image features.

[0008] As a preferred embodiment of the reinforcement learning-based adaptive parameter optimization control method for industrial sewing machines described in this invention, the step of performing phase synchronization processing on the stitch period based on the spindle angle to generate a needle period phase data packet is as follows: Using the target spindle speed as the periodic stability verification condition and the target needle pitch as the feeding matching verification condition, the spindle angle change data is divided into stitch periods to generate spindle angle stitch period data. Based on the seam type requirements, the stitch period phase division rules are determined, and the real-time running data is synchronously mapped to the corresponding phase according to the main axis angle stitch period data to generate phase synchronous running data; The phase synchronization operation data is bound to the sewing work order data to generate a needle cycle phase data package.

[0009] As a preferred embodiment of the reinforcement learning-based adaptive parameter optimization control method for industrial sewing machines described in this invention, the step of performing sliding memory compression on tension variation feed deviation and stitch image deviation within continuous needle cycles based on needle cycle phase data packets is as follows: Based on the continuous needle cycle records under the same sewing work order in the needle cycle phase data packet, obtain the tension amplitude difference between adjacent needle cycles, the deviation value between the actual feeding displacement and the target needle distance, and the offset of the center line of the stitch image, and generate a continuous needle cycle change data group. Write the continuous needle cycle change data group into the sliding memory window according to the needle cycle number, and retain the current phase label corresponding to each needle cycle to generate a sliding memory data group with phase label; Abnormal fluctuations are removed and scale is normalized from the phase-labeled sliding memory dataset to generate a standardized sliding memory dataset.

[0010] As a preferred embodiment of the reinforcement learning-based adaptive parameter optimization control method for industrial sewing machines described in this invention, the steps of constructing a stitch forming memory state representing the defect accumulation trend and forming a phase memory state vector by combining the current phase label are as follows: Based on the standardized sliding memory data set, the tension amplitude difference, feeding deviation value and the center line offset of the stitch image are statistically increased continuously within the continuous needle cycle to construct the stitch forming memory state. The line-shaped memory state is associated with the current phase label in the sliding memory data group with phase label to form a phase memory state vector.

[0011] As a preferred embodiment of the reinforcement learning-based adaptive parameter optimization control method for industrial sewing machines described in this invention, the step of calling the phase action mask based on the phase memory state vector to perform policy mapping inference, restricting parameter adjustment within the executable range of the current mechanical phase, and forming candidate parameter incremental actions, is as follows: Read the current phase label and trace forming memory state from the phase memory state vector, match the parameter items that the current mechanical phase can be adjusted according to the current phase label, generate a phase action mask and load it into the reinforcement learning policy model; The reinforcement learning policy model includes a phase state encoding layer, a phase action masking layer, a policy inference layer, and an action verification output layer; The phase state encoding layer encodes the current phase label, the line forming memory state, and the work order constraint information in the phase memory state vector to generate phase state features; The phase action mask layer matches executable parameter channels based on the current phase label in the phase state characteristics, and marks parameter channels that are not allowed to be adjusted in the current mechanical phase as non-output channels, thereby generating a phase action mask; The strategy reasoning layer filters executable parameter channels based on phase action masks, generates a phase-constrained action space, and performs strategy reasoning on phase state features within the phase-constrained action space to generate initial incremental actions for parameters. The action verification output layer performs parameter channel verification and amplitude boundary processing on the initial incremental action of the parameters, and retains the adjustment amount that passes the channel verification and is within the amplitude boundary to form candidate parameter incremental actions.

[0012] As a preferred embodiment of the reinforcement learning-based adaptive parameter optimization control method for industrial sewing machines described in this invention, the steps of performing defect risk budget verification and safety action projection on the incremental actions of candidate parameters are as follows: Perform defect risk budget verification on the incremental actions of candidate parameters, combine the line forming memory state to predict the comprehensive defect risk increase caused by the incremental actions of candidate parameters, and generate defect risk budget verification data. Based on the defect risk budget verification data, candidate parameter incremental actions that exceed the budget for both the breakage and skipped needle risk and the feed slip risk are marked as incremental actions to be projected.

[0013] As a preferred embodiment of the reinforcement learning-based adaptive parameter optimization control method for industrial sewing machines described in this invention, the steps for forming safety control commands and actual control states are as follows: Based on the safety action space, the incremental actions of the projection parameters are bounded and channel remapping is allowed to generate safety parameter incremental actions. The incremental actions of safety parameters are converted into safety control commands for industrial sewing machines, and the equipment response data after the execution of the safety control commands is collected to form the actual control state.

[0014] As a preferred embodiment of the reinforcement learning-based adaptive parameter optimization control method for industrial sewing machines described in this invention, the steps of calculating the defect accumulation suppression score, correcting the stitch forming memory state and experience cache based on the defect accumulation suppression score, and completing the online update of the reinforcement learning strategy model are as follows: Based on the actual control state, the corresponding needle cycle feedback data is collected and matched with the defect risk peak record in the stitch forming memory state to generate defect risk peak backtracking data. By comparing and calculating the changes in peak height and peak duration cycle in the peak retrospective data of defect risk, a defect cumulative suppression score is generated. The line forming memory state is corrected based on the defect accumulation suppression score, and the corrected line forming memory state is written into the experience cache to form the updated line forming memory state and the updated experience cache. The reinforcement learning policy model is updated online based on the updated trace-shaped memory state and the updated experience cache.

[0015] Secondly, the present invention provides an adaptive parameter optimization control system for industrial sewing machines based on reinforcement learning, comprising: a phase synchronization module, which collects sewing work order data and real-time operation data of the industrial sewing machine, and performs phase synchronization processing on the stitch cycle according to the spindle angle to generate a needle cycle phase data packet; The stitch memory module performs sliding memory compression on the tension change, feeding deviation and stitch image deviation within a continuous needle cycle based on the needle cycle phase data package, constructs a stitch forming memory state that characterizes the trend of defect accumulation, and forms a phase memory state vector by combining the current phase label. The phase decision module calls the phase action mask based on the phase memory state vector to perform strategy mapping reasoning, restricts parameter adjustment to the range that the current mechanical phase can execute, and forms candidate parameter incremental actions; The risk projection module performs defect risk budget verification and safety action projection on the incremental actions of candidate parameters, mapping the adjustments for risks of wire breakage and skipped needles and feeding slippage to the safety action space, forming safety control instructions and actual control status; The feedback update module calculates the defect accumulation suppression score by retrospectively analyzing the decline of the defect risk peak in the trace forming memory state based on the actual control state, and then corrects the trace forming memory state and experience cache according to the defect accumulation suppression score.

[0016] The beneficial effects of this invention are as follows: by performing sliding memory compression on tension change feeding deviation and stitch image deviation based on needle cycle phase data packets, a stitch forming memory state is constructed and a phase memory state vector is formed, so that stitch defects are transformed from single-point judgment to cumulative characterization within a continuous needle cycle; by performing defect risk budget verification and safety action projection on incremental actions of candidate parameters, the advance capability, phase matching and execution safety of stitch quality control are improved. Attached Figure Description

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

[0018] Figure 1 This is a flowchart of adaptive parameter optimization control for industrial sewing machines based on reinforcement learning.

[0019] Figure 2 This is a schematic diagram of an adaptive parameter optimization control system for an industrial sewing machine based on reinforcement learning.

[0020] Figure 3 A flowchart for constructing the line-shaped memory state and phase memory state vectors.

[0021] Figure 4 A flowchart for defect risk budget verification and safety action projection. Detailed Implementation

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

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

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

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides an adaptive parameter optimization control method for industrial sewing machines based on reinforcement learning, comprising the following steps: S1. Collect sewing work order data and real-time operation data of industrial sewing machines, and perform phase synchronization processing on the stitch cycle according to the spindle angle to generate a needle cycle phase data packet.

[0026] The sewing work order data includes the work order number, fabric type, number of fabric layers, thread specifications, target stitch length, seam type requirements, and target spindle speed.

[0027] Real-time operating data includes spindle angle, spindle speed, spindle angle change data, upper tension, feed displacement, presser foot pressure, spindle load, and stitch image features.

[0028] It should be noted that the sewing work order data is issued by the production management system and confirmed and entered through the equipment operation interface. The work order number, fabric type, number of fabric layers, thread specifications, target stitch length, seam type requirements, and target spindle speed serve as the initial constraint data for the sewing task. Real-time operating data is synchronously collected by the spindle encoder, servo driver, tension sensor, feed motor feedback, presser foot pressure sensor, load detection unit, and vision acquisition device during the operation of the industrial sewing machine.

[0029] Using the target spindle speed as the periodic stability verification condition and the target needle pitch as the feeding matching verification condition, the spindle angle change data is divided into stitch periods to generate spindle angle stitch period data.

[0030] Furthermore, using the target spindle speed as the periodic stability verification condition, the spindle angle change data is used to verify the consistency of the angle acquisition time and adjacent angle changes, thus obtaining valid spindle angle change data that conforms to the target spindle speed fluctuation range. Using the target needle pitch as the feeding matching verification condition, the feeding displacement change corresponding to the valid spindle angle change data is matched and verified, thus obtaining valid feeding matching data that conforms to the target needle pitch requirement. Based on the angle reset mark and needle cycle number in the valid spindle angle change data, the start and end boundaries of each stitch cycle are determined, and the valid feeding matching data is bound to the corresponding stitch cycle to generate spindle angle stitch cycle data.

[0031] Based on the seam type requirements, the stitch period phase division rules are determined, and the real-time running data is synchronously mapped to the corresponding phase according to the main axis angle stitch period data to generate phase-synchronized running data.

[0032] Furthermore, based on the stitch type requirements, the analysis of the needle bar's downward insertion, thread loop formation, thread take-up and tightening, and material feeding within a stitch cycle is performed. Combined with the spindle angle stitch cycle data, the start and end boundaries of each action within the spindle angle range are determined, dividing the continuous spindle angle range into phase intervals reflecting the stitch forming process, thus forming a stitch cycle phase division rule. Based on the stitch cycle boundaries and spindle angle positions in the spindle angle stitch cycle data, the spindle angle corresponding to each acquisition moment in the real-time running data falls into the phase interval of the stitch cycle phase division rule. This establishes a correspondence between acquisition moment, spindle angle position, stitch cycle number, phase label, and real-time running data, generating phase-synchronized running data.

[0033] It should be noted that the stitch cycle phase division rule refers to mapping the continuous needle bar movement, loop formation, take-up coordination, and feeding actions that occur within a complete stitch cycle to different spindle angle ranges based on the stitch type requirements and spindle angle stitch cycle data. A corresponding phase label is configured for each spindle angle range, thereby improving the matching accuracy between real-time running data and the actual stitch forming stage, and providing more granular data basis for subsequent defect trend identification and parameter adaptive adjustment.

[0034] The phase synchronization operation data is bound to the sewing work order data to generate a needle cycle phase data packet containing work order constraint information and phase operation information.

[0035] Furthermore, the phase synchronization operation data is grouped according to the work order number and stitch cycle number to form a needle cycle phase operation record; the needle cycle phase operation record is aligned with the sewing work order data to form a correspondence between work order constraint information and phase operation information; based on the correspondence between work order constraint information and phase operation information, the operation data of each phase within the same needle cycle is encapsulated in the same data structure to generate a needle cycle phase data package containing work order constraint information and phase operation information.

[0036] S2. Based on the needle cycle phase data package, the tension change, feeding deviation and line pattern image deviation within the continuous needle cycle are subjected to sliding memory compression to construct a line pattern forming memory state that characterizes the trend of defect accumulation, and a phase memory state vector is formed by combining the current phase label.

[0037] Based on the continuous needle cycle records under the same sewing work order in the needle cycle phase data packet, obtain the tension amplitude difference between adjacent needle cycles, the deviation value between the actual feeding displacement and the target needle distance, and the offset of the center line of the stitch image, and generate a continuous needle cycle change data group.

[0038] Furthermore, based on the work order number in the needle cycle phase data packet, the needle cycle data corresponding to the same sewing work order is filtered, and the needle cycle data corresponding to the same sewing work order is continuously sorted according to the needle cycle number to form a continuous needle cycle record under the same sewing work order; the peak value of the upper thread tension in each needle cycle is read according to the continuous needle cycle record under the same sewing work order, and the difference in the peak value of the upper thread tension between adjacent needle cycles is calculated to obtain the tension amplitude difference between adjacent needle cycles; the actual feeding displacement in each needle cycle is read according to the continuous needle cycle record under the same sewing work order, and... The difference between the actual feed displacement and the target stitch distance in the sewing work order data is calculated to obtain the deviation value between the actual feed displacement and the target stitch distance. Based on the continuous needle cycle record under the same sewing work order, the stitch image features in each needle cycle are read, and the offset distance of the stitch center line relative to the target sewing path is determined according to the stitch image features to obtain the offset of the stitch image center line. The tension amplitude difference between adjacent needle cycles, the deviation value between the actual feed displacement and the target stitch distance, and the offset of the stitch image center line are packaged according to the needle cycle number to generate a continuous needle cycle change data group.

[0039] Write the continuous needle cycle change data group into the sliding memory window according to the needle cycle number, and retain the current phase label corresponding to each needle cycle to generate a sliding memory data group with phase label.

[0040] Furthermore, the continuous needle cycle change data groups are arranged in ascending order of needle cycle number to obtain a time sequence queue of the continuous needle cycle change data groups; several adjacent needle cycle records are selected from the time sequence queue of the continuous needle cycle change data groups according to a preset window length, and these adjacent needle cycle records are written into a sliding memory window; during the writing process into the sliding memory window, the current phase tag corresponding to each needle cycle number is read from the needle cycle phase data packet, and the current phase tag is bound to the tension amplitude difference, the deviation value between the actual feeding displacement and the target needle distance, and the offset of the center line of the stitch image under the same needle cycle number in the sliding memory window; when the sliding memory window moves to a new needle cycle number, the latest needle cycle record is retained and the needle cycle records exceeding the preset window length are removed, generating a sliding memory data group with phase tags.

[0041] It should be noted that the sliding memory window refers to a data buffer that stores several adjacent stitch cycle records in a continuous stitch cycle change data group according to a preset window length. New stitch cycle records are dynamically added and stitch cycle records exceeding the window length are removed as the stitch cycle number is updated. The window length is the shortest identifiable stitch cycle span based on the defect accumulation trend in the stitch forming memory state, and is jointly set by the target spindle speed in the sewing work order data and the spindle speed fluctuations in the real-time operation data.

[0042] Abnormal fluctuations are removed and scale is normalized from the phase-labeled sliding memory dataset to generate a standardized sliding memory dataset.

[0043] Furthermore, the sliding memory data set with phase labels is checked for integrity according to the needle cycle number. Missing tension amplitude differences, deviations between actual feeding displacement and target needle distance, and offsets of the center line of the stitch image are filled in by interpolation between adjacent needle cycles. Abnormal fluctuations that exceed the range of changes between adjacent needle cycles are removed. The data after filling in the missing values ​​are scaled and normalized using the same value range mapping method to generate a standardized sliding memory data set.

[0044] Based on the standardized sliding memory data set, the tension amplitude difference, feeding deviation value, and the center line offset of the stitch image are statistically analyzed to continuously increase within the continuous needle cycle, and the stitch forming memory state is constructed.

[0045] Furthermore, based on the standardized sliding memory data set, the tension amplitude difference, feed deviation value, and stitch image centerline offset are read, and the changes in these values ​​between adjacent needle cycles are compared. The number and magnitude of needle cycles continuously exceeding a preset small change threshold are accumulated to obtain the degree of continuous increase of these values ​​within consecutive needle cycles. This degree of continuous increase is then bound to the current phase label of the corresponding needle cycle to construct a stitch forming memory state. By preserving the evolutionary relationship between tension changes, feed deviation, and stitch image deviation in consecutive needle cycles, the stitch forming memory state transforms single-needle quality judgment into a traceable defect accumulation process, thereby improving the advance of parameter adaptive adjustment and the stability of stitch quality control.

[0046] It should be noted that the threshold for minute changes is set based on the normal fluctuation range of tension amplitude difference, feed deviation value and stitch image centerline offset during the stable sewing stage under the same sewing work order, through the quantile boundary of historical qualified needle cycle data.

[0047] The line-shaped memory state is associated with the current phase label in the sliding memory data group with phase label to form a phase memory state vector.

[0048] Furthermore, the current phase label corresponding to the needle cycle at the end of the sliding memory window in the phase-labeled sliding memory data set is used as the phase reference. The degree of continuous increase, peak position, and span of consecutive needle cycles in the stitch forming memory state are aligned with the current phase label according to the needle cycle number. The current phase label is then phase-encoded, and the phase-encoding result is concatenated with the stitch forming memory state in a fixed field order to form a phase memory state vector that simultaneously represents the defect accumulation process and the current mechanical phase. This phase memory state vector fuses the defect accumulation process in the stitch forming memory state with the corresponding current phase label, limiting the defect evolution information to a specific mechanical phase, thereby improving the accuracy of matching subsequent parameter decisions with the actual stitch forming stage.

[0049] S3. Based on the phase memory state vector, call the phase action mask to perform strategy mapping reasoning, restrict parameter adjustment to the current mechanical phase executable range, and form candidate parameter incremental actions.

[0050] Read the current phase label and trace forming memory state from the phase memory state vector, match the parameter items that the current mechanical phase can be adjusted according to the current phase label, generate a phase action mask and load it into the reinforcement learning policy model.

[0051] Furthermore, after determining the current mechanical phase based on the current phase label, the allowed execution spindle angle range, parameter activation delay, and corresponding mechanical action of each parameter item are matched with the current mechanical phase (different mechanical action stages of an industrial sewing machine within a stitch cycle, such as needle insertion, loop formation, take-up coordination, and feed advancement, divided by the spindle angle position). When the allowed execution spindle angle range of a parameter item does not cover the current spindle angle position, the parameter activation delay crosses the current mechanical phase boundary, the parameter adjustment object does not participate in the current mechanical action, or the parameter change interferes with the established sequence of needle insertion, loop formation, take-up coordination, and feed advancement, the parameter item is determined to be an unexecutable parameter item of the current mechanical phase and assigned a non-output flag in the action channel sequence. Together with the output flag of the executable parameter item, a phase action mask is generated and loaded into the reinforcement learning policy model. Phase action masking improves the real-time executability and safety of parameter optimization control by limiting the parameters output by reinforcement learning to the actual executable parameter channels of the current mechanical phase, thus avoiding mismatched control actions in different stages such as needle insertion, wire loop formation, wire take-up coordination, and feeding.

[0052] The reinforcement learning policy model includes a phase state encoding layer, a phase action masking layer, a policy inference layer, and an action verification output layer.

[0053] The phase state encoding layer encodes the current phase label, stitch forming memory state, and work order constraint information in the phase memory state vector to generate phase state features. These phase state features encode the current phase label, stitch forming memory state, and work order constraint information into a single decision representation, enabling defect accumulation information to participate in parameter decision-making under corresponding mechanical phases and sewing constraints. This improves the accuracy of matching candidate parameter incremental actions with the actual sewing process.

[0054] Furthermore, after receiving the phase memory state vector, the phase state encoding layer separates the current phase label, stitch forming memory state, and work order constraint information according to the field order; it performs one-hot encoding on the current phase label and superimposes the corresponding spindle angle interval encoding to obtain the phase position encoding; it normalizes and maps the continuous increase degree, peak position, and continuous needle cycle span in the stitch forming memory state to obtain the defect accumulation encoding; it performs category encoding and numerical normalization on the fabric type, number of fabric layers, thread specifications, target stitch length, seam type requirements, and target spindle speed in the work order constraint information to obtain the work order constraint encoding; it concatenates the phase position encoding, defect accumulation encoding, and work order constraint encoding according to a fixed field order, and retains the defect accumulation features that match the current phase label through gating fusion to generate phase state features.

[0055] The phase action mask layer matches executable parameter channels based on the current phase label in the phase state characteristics, and marks parameter channels that are not allowed to be adjusted in the current mechanical phase as non-output channels, thus generating a phase action mask.

[0056] Furthermore, the phase action mask layer first reads the current phase label from the phase state features and matches it with a preset mechanical phase execution table to determine the spindle angle range, participating mechanism, and allowed effective time corresponding to the current mechanical phase. Then, it compares the actuator, allowed adjustment phase, parameter effective delay, and adjustment amplitude range corresponding to each parameter channel with the current mechanical phase. When the allowed adjustment phase of a parameter channel does not include the current phase label, or the parameter effective delay exceeds the current mechanical phase boundary, or the actuator corresponding to the parameter channel does not participate in the current mechanical phase action, the parameter channel that is not allowed to be adjusted in the current mechanical phase is marked as a non-output channel. The remaining parameter channels that meet the comparison conditions are marked as output channels and combined according to the parameter channel order to form a phase action mask. The phase action mask filters out parameter channels that do not conform to the spindle angle range and execution timing through the current mechanical phase, limiting candidate parameter decisions to the executable action range, thereby improving the timing matching and safety of parameter optimization control and the line forming process.

[0057] The strategy reasoning layer filters executable parameter channels based on phase action masks, generates a phase-constrained action space, and performs strategy reasoning on phase state features within the phase-constrained action space to generate initial incremental actions for parameters.

[0058] Furthermore, the strategy inference layer breaks down the phase state features into current phase information, stitch forming memory state, and work order constraint information. It uses the current phase information to determine the mechanical action stage of the current needle cycle, the stitch forming memory state to identify the main sources of defect accumulation, and the work order constraint information to determine the acceptable range for parameter adjustment. When the stitch forming memory state shows that the tension amplitude difference continuously increases within consecutive needle cycles, it prioritizes generating initial incremental parameter actions within the phase constraint action space to reduce the upper tension or reduce the spindle speed. When the stitch forming memory state shows that the feeding deviation value continuously increases within consecutive needle cycles, it prioritizes generating initial incremental parameter actions within the phase constraint action space to correct the feeding displacement or reduce the spindle speed. When the stitch forming memory state shows that the centerline offset of the stitch image continuously increases within consecutive needle cycles, it prioritizes generating initial incremental parameter actions within the phase constraint action space to correct the feeding direction or adjust the presser foot pressure. The strategy inference layer determines the adjustment direction based on the matching degree of the defect accumulation source corresponding to each output channel, and determines the adjustment magnitude based on the continuous increase degree and work order constraint information, generating initial incremental parameter actions.

[0059] The action verification output layer performs parameter channel verification and amplitude boundary processing on the initial incremental action of the parameters, and retains the adjustment amount that passes the channel verification and is within the amplitude boundary to form candidate parameter incremental actions.

[0060] Furthermore, after receiving the initial incremental parameter actions, the action verification output layer compares each adjustment amount in the initial incremental parameter actions with the output channels in the phase action mask. Adjustments not falling into an output channel are set to zero, while those falling into an output channel continue to undergo consistency verification with the corresponding parameter items in the phase-constrained action space. The adjustment amounts verified by the parameter channels are processed according to the minimum allowable increment, maximum allowable increment, and maximum change per unit needle cycle of the corresponding parameter item, resulting in amplitude boundary processing. Adjustments exceeding the amplitude boundary are compressed to the boundary value, while adjustments within the amplitude boundary are retained, forming candidate parameter incremental actions. Candidate parameter incremental actions, by providing executable parameter adjustment directions and adjustment amplitudes within the current mechanical phase allowable range, transform the cumulative trend of line defects into a controllable quantity, thereby improving the real-time performance and controllability of the parameter optimization process.

[0061] It should be noted that the reinforcement learning policy model is trained using historical sewing samples and controlled trial sewing samples. The phase memory state vector is used as the state input, the phase action mask is used as the action channel constraint, and the safety parameter incremental action and the defect accumulation suppression score are used as the feedback basis. When the defect accumulation suppression score increases and the safety action projection amplitude is small, the selection probability of the corresponding candidate parameter incremental action is increased. When the defect accumulation suppression score decreases or the increase in comprehensive defect risk exceeds the risk budget, the selection probability of the corresponding candidate parameter incremental action is decreased, and a reinforcement learning policy model that can be directly involved in policy mapping inference is obtained.

[0062] S4. Perform defect risk budget verification and safety action projection on the incremental actions of candidate parameters, and map the adjustments for risks of wire breakage and skipped needles and feeding slippage to the safety action space to form safety control instructions and actual control status.

[0063] Defect risk budget verification is performed on the incremental actions of candidate parameters. Combined with the line forming memory state, the comprehensive defect risk increase caused by the incremental actions of candidate parameters is predicted, and defect risk budget verification data is generated.

[0064] The expression for predicting the increase in comprehensive defect risk is: ; in, This indicates an increase in the overall defect risk; This represents the defect risk type index, where = This indicates a risk of disconnection. = This indicates a risk of skipped needles. = Indicates the risk of feed slippage; Indicates the first Sensitivity coefficient for class defect risk; Indicates the first The transpose of the single-parameter influence vector corresponding to the class of defect risk; A vector representing the absolute values ​​of the incremental actions of the candidate parameters; This represents the incremental action of the candidate parameter, and the incremental adjustment value of each adjustable parameter within the incremental action of the candidate parameter is represented by a vector in the formula. Indicates the first The parameter coupling matrix corresponding to class defect risks; Indicates the first The memory amplification factor corresponding to the class of defect risks; The first state in the line forming memory state The cumulative memory of class defect risks; Indicates the first Class defect risk budget; This indicates that the positive part is truncated, meaning that if the value inside the parentheses is less than zero, it is taken as zero, and if it is greater than zero, the original value is retained.

[0065] Furthermore, the incremental actions of candidate parameters are represented as an incremental adjustment set of each adjustable parameter, and the individual effects of the incremental adjustment set on the risks of thread breakage, skipped needles, and feed slippage, as well as the coupled effects of simultaneous adjustment of multiple parameters, are calculated. Then, combined with the cumulative degree of corresponding defect risks in the stitch forming memory state, the individual effects and coupled effects are amplified by memory and compared with the corresponding risk budget, and only the risk increase exceeding the risk budget is retained. The risk increases of various types exceeding the risk budget are logarithmically aggregated to obtain the comprehensive defect risk increase caused by the incremental actions of candidate parameters, and defect risk budget verification data is generated.

[0066] It should be noted that the sensitivity coefficient for defect risk is set based on the allowable over-limit ranges corresponding to the breakage risk budget, skipped needle risk budget, and feed slippage risk budget, through a mapping relationship where the allowable over-limit range and the sensitivity coefficient change inversely. The memory amplification coefficient for defect risk is set based on the proximity between the cumulative defect memory amount and the peak defect risk value in the stitch forming memory state, through a mapping relationship where the cumulative defect memory amount is proportional to the peak defect risk value. The parameter coupling matrix for defect risk is obtained by performing a quadratic correlation fitting on the change in defect risk caused by the synchronous adjustment of multiple parameters, based on the correspondence between the incremental actions of candidate parameters in the empirical buffer, the actual control state, and the defect accumulation suppression score.

[0067] Based on the defect risk budget verification data, candidate parameter incremental actions that exceed the budget for both the breakage and skipped needle risk and the feed slip risk are marked as incremental actions to be projected.

[0068] Furthermore, the risk budget for thread breakage and skipped stitches is set based on the fabric type, thread specifications, target spindle speed, and tension risk peak in the stitch forming memory state from the sewing work order data. The risk budget for feed slippage is set based on the number of fabric layers, target stitch length, seam type requirements, and feed deviation risk peak in the stitch forming memory state from the sewing work order data. The comprehensive defect risk increase corresponding to the incremental action of the candidate parameter is read according to the defect risk budget verification data, and the comprehensive defect risk increase is compared with the risk budget for thread breakage and skipped stitches and the risk budget for feed slippage, respectively. When the comprehensive defect risk increase exceeds either risk budget, the corresponding incremental action of the candidate parameter is marked as the incremental action of the parameter to be projected.

[0069] Based on the safety action space, the incremental actions of the projection parameters are bounded and channel remapping is allowed to generate safety parameter incremental actions.

[0070] Furthermore, the permissible channels, minimum permissible increments, maximum permissible increments, and maximum changes per unit needle cycle of each adjustable parameter in the safety action space are read. Adjustments in the incremental action of the parameter to be projected that are still within the permissible channels but exceed the amplitude boundary are compressed to the corresponding boundary values. Adjustments with a change rate exceeding the maximum change per unit needle cycle are compressed to within the maximum change per unit needle cycle, thus completing boundary compression. For adjustments in the incremental action of the parameter to be projected that do not belong to the permissible channels, alternative permissible channels with the same control target as the adjustments that do not belong to the permissible channels and are within the current mechanical phase's executable range are found. These alternative channels are then remapped according to the alternative direction with the lowest risk increase in the defect risk budget verification data. If no alternative permissible channel exists, the adjustments that do not belong to the permissible channels are set to zero. The adjusted amounts after boundary compression and the adjusted amounts after permissible channel remapping are then recombined to generate the incremental action of the safety parameter.

[0071] The incremental actions of safety parameters are converted into safety control commands for industrial sewing machines, and the equipment response data after the execution of the safety control commands is collected to form the actual control state.

[0072] Furthermore, the incremental actions of safety parameters are processed into commands according to the control interfaces of each actuator of the industrial sewing machine. The incremental spindle speed is converted into a spindle servo drive command, the incremental feed displacement is converted into a feed motor control command, the incremental thread tension is converted into an electronic thread clamp adjustment command, and the incremental presser foot pressure is converted into a presser foot drive control command. The issuance time of each safety control command is determined according to the needle cycle number and the current phase label. After the safety control command is executed, feedback values ​​of spindle angle, spindle speed, thread tension, feed displacement, presser foot pressure, spindle load, and stitch image features are collected. The feedback values ​​are then bound with the safety control command, needle cycle number, and current phase label to form the actual control state.

[0073] S5. Based on the actual control state, trace back the decline of the peak defect risk in the trace forming memory state, calculate the defect cumulative suppression score, and correct the trace forming memory state and experience cache according to the defect cumulative suppression score to complete the online update of the reinforcement learning strategy model.

[0074] Based on the actual control state, the corresponding needle cycle feedback data is collected and matched with the defect risk peak record in the stitch forming memory state to generate defect risk peak backtracking data.

[0075] Furthermore, based on the actual control state, the needle cycle number and current phase label corresponding to the executed safety control command are determined, and the defect risk peak record in the stitch forming memory state is read under the same needle cycle number and current phase label; the peak height, peak occurrence needle cycle, and peak duration cycle in the defect risk peak record are time-aligned and phase-aligned with the tension change, feeding deviation, and stitch image deviation in the corresponding needle cycle feedback data to obtain the change trajectory of the same defect risk peak before and after control; the change trajectory is established with the safety control command according to the peak occurrence needle cycle to generate defect risk peak backtracking data.

[0076] By comparing and calculating the changes in peak height and peak duration cycle in the peak defect risk backtracking data, a defect cumulative suppression score is generated, expressed as: ; in, This represents the defect cumulative suppression score; the higher the value, the more significant the suppression effect of safety control instructions on the peak defect risk. This represents the peak trajectory of defect risk after the execution of safety control commands, where the subscript... Indicates the state after control is applied; This represents the peak trajectory of defect risks before the execution of safety control instructions, where the subscript... Indicates the state before control; This indicates the peak height of defect risk before the execution of safety control instructions, where the subscript... Indicates the state before control; Indicates the needle cycle number or needle cycle time point; Indicates the preset risk decay period; It represents the dynamic time warp distance, used to measure the morphological difference between the actual defect risk peak trajectory and the target decay curve; This represents a very small constant to prevent the denominator from being zero; This represents the exponential decay factor, used to describe the target trend of the peak defect risk gradually decreasing as the needle cycle progresses. The larger the value, the smaller the factor. The larger the value, the slower the descent.

[0077] Furthermore, the peak value data of defect risk is backtracked according to the needle cycle number to organize the changes in peak height and peak duration cycle before and after control, forming the peak value trajectory of defect risk before the execution of the safety control command and the peak value trajectory of defect risk after the execution of the safety control command, respectively. Based on the peak value height of defect risk before the execution of the safety control command and the preset risk attenuation cycle, a target attenuation curve is generated, and the dynamic time warping distance between the trajectory before control and the target attenuation curve and the trajectory after control and the target attenuation curve are calculated, respectively. The degree of decrease of the distance after control relative to the distance before control is converted into a percentage score, and a minimum constant is used to avoid the denominator being zero. At the same time, the score is limited to a preset value range to generate a defect accumulation suppression score.

[0078] It should be noted that the risk decay cycle is based on the number of needle cycles that the peak defect risk under the same sewing work order takes to fall from its peak height to a stable fluctuation range. It is set by the statistical quantile value of the actual control status and the backtracking data of the peak defect risk in the experience cache.

[0079] The line forming memory state is corrected based on the defect accumulation suppression score, and the corrected line forming memory state is written into the experience cache to form the updated line forming memory state and the updated experience cache.

[0080] Furthermore, based on the comparison between the defect cumulative suppression score and the preset suppression score threshold, the peak height, peak duration, and continuous increase of the defect risk in the trace forming memory state are corrected. When the defect cumulative suppression score reaches the preset suppression score threshold, the peak height and continuous increase of the defect risk in the trace forming memory state are reduced according to the suppression ratio corresponding to the defect cumulative suppression score, and the peak duration is shortened. When the defect cumulative suppression score is lower than the preset suppression score threshold, the peak height and peak duration of the defect risk in the trace forming memory state are retained, and the corresponding defect cumulative memory amount is increased. The corrected peak height, peak duration, continuous increase, and defect cumulative memory amount are repackaged into the corrected trace forming memory state, and the corrected trace forming memory state, together with the actual control state, the defect cumulative suppression score, and the incremental action of the safety parameters, are written into the experience cache to form the updated trace forming memory state and the updated experience cache.

[0081] It should be noted that the experience cache refers to a data set that stores the correspondence between the phase memory state vector, safety parameter incremental action, actual control state, defect accumulation suppression score and corrected trace forming memory state according to the needle cycle number.

[0082] The reinforcement learning policy model is updated online based on the updated trace-shaped memory state and the updated experience cache.

[0083] Furthermore, based on the updated stitch forming memory state, the current defect risk peak height, peak duration period, and direction of increase are determined. The corresponding records of phase memory state vectors, safety parameter incremental actions, actual control states, and defect accumulation suppression scores under the same sewing work order and current phase label are retrieved from the updated experience cache. When the defect accumulation suppression score increases, the policy mapping weight between the safety parameter incremental action and the phase memory state vector is strengthened. When the defect accumulation suppression score decreases, the policy mapping weight between the safety parameter incremental action and the phase memory state vector is weakened, and the risk penalty weight is increased. The corrected policy mapping weights are then written into the reinforcement learning policy model, completing the online update of the reinforcement learning policy model.

[0084] It should be noted that the strategy mapping weight is based on the correspondence between the phase memory state vector, safety parameter incremental action, actual control state and defect accumulation suppression score in the updated experience cache, and is set by positive and negative feedback mapping of the defect accumulation suppression score to the safety parameter incremental action.

[0085] This embodiment also provides an adaptive parameter optimization control system for industrial sewing machines based on reinforcement learning, including: a phase synchronization module, which collects sewing work order data and real-time operation data of the industrial sewing machine, and performs phase synchronization processing on the stitch cycle according to the spindle angle to generate a needle cycle phase data packet; The stitch memory module performs sliding memory compression on the tension change, feeding deviation and stitch image deviation within a continuous needle cycle based on the needle cycle phase data package, constructs a stitch forming memory state that characterizes the trend of defect accumulation, and forms a phase memory state vector by combining the current phase label. The phase decision module calls the phase action mask based on the phase memory state vector to perform strategy mapping reasoning, restricts parameter adjustment to the range that the current mechanical phase can execute, and forms candidate parameter incremental actions; The risk projection module performs defect risk budget verification and safety action projection on the incremental actions of candidate parameters, mapping the adjustments for risks of wire breakage and skipped needles and feeding slippage to the safety action space, forming safety control instructions and actual control status; The feedback update module calculates the defect accumulation suppression score by retrospectively analyzing the decline of the defect risk peak in the trace forming memory state based on the actual control state, and then corrects the trace forming memory state and experience cache according to the defect accumulation suppression score.

[0086] In summary, this invention improves the predictability, phase matching, and execution safety of stitch quality control by: performing sliding memory compression on tension change feeding deviation and stitch image deviation based on needle cycle phase data packets, constructing stitch forming memory state and forming phase memory state vector, so that stitch defects are transformed from single-point judgment to cumulative representation within continuous needle cycles; and by performing defect risk budget verification and safety action projection on incremental actions of candidate parameters.

[0087] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for adaptive parameter optimization control of industrial sewing machines based on reinforcement learning, characterized in that, include: Collect sewing order data and real-time operation data of industrial sewing machines, and perform phase synchronization processing on the stitch cycle according to the spindle angle to generate a needle cycle phase data packet; Based on the needle cycle phase data package, the tension change, feeding deviation and stitch image deviation within the continuous needle cycle are compressed by sliding memory to construct a stitch forming memory state that characterizes the trend of defect accumulation, and a phase memory state vector is formed by combining the current phase label. Based on the phase memory state vector, the phase action mask is invoked to perform strategy mapping reasoning, and the parameter adjustment is restricted to the executable range of the current mechanical phase to form candidate parameter incremental actions. Defect risk budget verification and safety action projection are performed on the incremental actions of candidate parameters. Adjustments with risks of wire breakage and skipped needles and feeding slippage are mapped to the safety action space to form safety control commands and actual control states. Based on the decline of the peak defect risk in the trace forming memory state in the actual control state, the defect cumulative suppression score is calculated, and the trace forming memory state and experience cache are corrected according to the defect cumulative suppression score.

2. The adaptive parameter optimization control method for industrial sewing machines based on reinforcement learning as described in claim 1, characterized in that: The sewing work order data includes work order number, fabric type, number of fabric layers, thread specifications, spindle angle variation data, target stitch length, and target spindle speed; The real-time operating data includes spindle angle, spindle speed, wire tension, feed displacement, presser foot pressure, spindle load, and stitch image features.

3. The adaptive parameter optimization control method for industrial sewing machines based on reinforcement learning as described in claim 2, characterized in that, The steps for performing phase synchronization processing on the stitch period based on the spindle angle to generate a needle period phase data packet are as follows: Using the target spindle speed as the periodic stability verification condition and the target needle pitch as the feeding matching verification condition, the spindle angle change data is divided into stitch periods to generate spindle angle stitch period data. Based on the seam type requirements, the stitch period phase division rules are determined, and the real-time running data is synchronously mapped to the corresponding phase according to the main axis angle stitch period data to generate phase synchronous running data; The phase synchronization operation data is bound to the sewing work order data to generate a needle cycle phase data package.

4. The adaptive parameter optimization control method for industrial sewing machines based on reinforcement learning as described in claim 1, characterized in that, The sliding memory compression of tension variation, feeding deviation, and stitch image deviation within consecutive needle cycles based on needle cycle phase data packets is as follows: Based on the continuous needle cycle records under the same sewing work order in the needle cycle phase data packet, obtain the tension amplitude difference between adjacent needle cycles, the deviation value between the actual feeding displacement and the target needle distance, and the offset of the center line of the stitch image, and generate a continuous needle cycle change data group. Write the continuous needle cycle change data group into the sliding memory window according to the needle cycle number, and retain the current phase label corresponding to each needle cycle to generate a sliding memory data group with phase label; Abnormal fluctuations are removed and scale is normalized from the phase-labeled sliding memory dataset to generate a standardized sliding memory dataset.

5. The adaptive parameter optimization control method for industrial sewing machines based on reinforcement learning as described in claim 4, characterized in that, The steps for constructing a trace-forming memory state that characterizes the trend of defect accumulation, and combining it with the current phase label to form a phase memory state vector, are as follows: Based on the standardized sliding memory data set, the tension amplitude difference, feeding deviation value and the center line offset of the stitch image are statistically increased continuously within the continuous needle cycle to construct the stitch forming memory state. The line-shaped memory state is associated with the current phase label in the sliding memory data group with phase label to form a phase memory state vector.

6. The adaptive parameter optimization control method for industrial sewing machines based on reinforcement learning as described in claim 1, characterized in that, The step of calling the phase action mask based on the phase memory state vector to perform strategy mapping reasoning, restricting parameter adjustment to the executable range of the current mechanical phase, and forming candidate parameter incremental actions, is as follows: Read the current phase label and trace forming memory state from the phase memory state vector, match the parameter items that the current mechanical phase can be adjusted according to the current phase label, generate a phase action mask and load it into the reinforcement learning policy model; The reinforcement learning policy model includes a phase state encoding layer, a phase action masking layer, a policy inference layer, and an action verification output layer; The phase state encoding layer encodes the current phase label, the line forming memory state, and the work order constraint information in the phase memory state vector to generate phase state features; The phase action mask layer matches executable parameter channels based on the current phase label in the phase state characteristics, and marks parameter channels that are not allowed to be adjusted in the current mechanical phase as non-output channels, thereby generating a phase action mask; The strategy reasoning layer filters executable parameter channels based on phase action masks, generates a phase-constrained action space, and performs strategy reasoning on phase state features within the phase-constrained action space to generate initial incremental actions for parameters. The action verification output layer performs parameter channel verification and amplitude boundary processing on the initial incremental action of the parameters, and retains the adjustment amount that passes the channel verification and is within the amplitude boundary to form candidate parameter incremental actions.

7. The adaptive parameter optimization control method for industrial sewing machines based on reinforcement learning as described in claim 1, characterized in that, The steps for performing defect risk budget verification and safety action projection on incremental actions of candidate parameters are as follows: Perform defect risk budget verification on the incremental actions of candidate parameters, combine the line forming memory state to predict the comprehensive defect risk increase caused by the incremental actions of candidate parameters, and generate defect risk budget verification data. Based on the defect risk budget verification data, candidate parameter incremental actions that exceed the budget for both the breakage and skipped needle risk and the feed slip risk are marked as incremental actions to be projected.

8. The adaptive parameter optimization control method for industrial sewing machines based on reinforcement learning as described in claim 7, characterized in that, The steps for generating safety control commands and actual control states are as follows: Based on the safety action space, the incremental actions of the projection parameters are bounded and channel remapping is allowed to generate safety parameter incremental actions. The incremental actions of safety parameters are converted into safety control commands for industrial sewing machines, and the equipment response data after the execution of the safety control commands is collected to form the actual control state.

9. The adaptive parameter optimization control method for industrial sewing machines based on reinforcement learning as described in claim 1, characterized in that, The calculation of the defect cumulative suppression score, the correction of the stylus shaping memory state and experience cache based on the defect cumulative suppression score, and the completion of the online update of the reinforcement learning policy model are as follows: Based on the actual control state, the corresponding needle cycle feedback data is collected and matched with the defect risk peak record in the stitch forming memory state to generate defect risk peak backtracking data. By comparing and calculating the changes in peak height and peak duration cycle in the peak retrospective data of defect risk, a defect cumulative suppression score is generated. The line forming memory state is corrected based on the defect accumulation suppression score, and the corrected line forming memory state is written into the experience cache to form the updated line forming memory state and the updated experience cache. The reinforcement learning policy model is updated online based on the updated trace-shaped memory state and the updated experience cache.

10. A reinforcement learning-based adaptive parameter optimization control system for industrial sewing machines, based on the reinforcement learning-based adaptive parameter optimization control method for industrial sewing machines according to any one of claims 1 to 9, characterized in that, include: The phase synchronization module collects sewing work order data and real-time operation data of industrial sewing machines, and performs phase synchronization processing on the stitch cycle according to the spindle angle to generate a needle cycle phase data packet. The stitch memory module performs sliding memory compression on the tension change, feeding deviation and stitch image deviation within a continuous needle cycle based on the needle cycle phase data package, constructs a stitch forming memory state that characterizes the trend of defect accumulation, and forms a phase memory state vector by combining the current phase label. The phase decision module calls the phase action mask based on the phase memory state vector to perform strategy mapping reasoning, restricts parameter adjustment to the range that the current mechanical phase can execute, and forms candidate parameter incremental actions; The risk projection module performs defect risk budget verification and safety action projection on the incremental actions of candidate parameters, mapping the adjustments for risks of wire breakage and skipped needles and feeding slippage to the safety action space, forming safety control instructions and actual control status; The feedback update module calculates the defect accumulation suppression score by retrospectively analyzing the decline of the defect risk peak in the trace forming memory state based on the actual control state, and then corrects the trace forming memory state and experience cache according to the defect accumulation suppression score.

Images