An agent parameter setting method and system for a multi-parameter coupling device

Abstract

The application provides an agent parameter setting method and system for multi-parameter coupled equipment, the method comprising: an online acquisition step, acquiring initial control parameters, initial measurement indicators, batch identifiers, sub-unit identifiers, parameter constraints and indicator specifications of the equipment or equipment sub-units; a training condition judgment step, judging whether conditions for generating an automatic training task are met by the agent according to sample size, coverage, failure rate and batch drift degree; a model training step, when the training conditions are met, the agent generates an automatic training task, and trains, iterates and optimizes the parameter prediction model in a training sandbox; a strategy package generation step, the agent encapsulates the trained parameter prediction model as a parameter setting strategy package. The application improves the adaptation capability and online execution safety in the batch drift scenario, and realizes the intelligentization, self-adaptation and sustainable optimization of batch multi-parameter coupled equipment parameter setting.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial equipment parameter setting, in particular to an agent parameter setting method and system for multi-parameter coupled equipment. BACKGROUND

[0002] Batch multi-parameter coupled equipment refers to equipment that needs to adjust multiple control parameters and judge whether it is qualified through multiple performance indicators during production, testing, calibration or setting. Such equipment usually has the following characteristics: (1) the same type of equipment is produced or tested in batches, and there may be differences in process, material, environment or assembly between different batches; (2) a single device may contain multiple sub-units, channels or functional areas; (3) multiple control parameters jointly affect multiple performance indicators; (4) adjusting one parameter may improve one indicator while damaging other indicators that have already passed; (5) the parameter setting process needs to consider success rate, measurement times, test time and safety boundary.

[0003] Taking an optical module as an example, an optical module can contain multiple lanes. Each lane may need to adjust control parameters such as bias_dac (bias current), mod_dac (modulation amplitude), step_dac (cross-point adjustment amount), ohm (load resistance), and measure indicators such as TxPwr (transmit power), ER (extinction ratio), CP (cross-point), MaskMargin (eye diagram margin). Only when all indicators meet the specification requirements, the lane can be determined as Pass.

[0004] Existing parameter adjustment schemes mainly include three types: (1) automatic debugging schemes for specific parameters of optical modules (such as optical power and extinction ratio); (2) statistical calibration schemes based on historical data of similar products; (3) model optimization or drift detection schemes for multi-parameter equipment.

[0005] However, the above schemes have the following shortcomings: (1) Existing schemes are mostly targeted at local links and do not form a unified parameter setting closed loop for batch multi-parameter coupled equipment; (2) Existing schemes are mostly rule-based parameter adjustment and rely on step-by-step traversal and trial, which is prone to problems such as many iterations and long time consumption; (3) Traditional step-by-step parameter adjustment requires multiple measurements, which is time-consuming and affects productivity; (4) Existing processes assume that subsequent adjustments will not damage previously passed indicators, which is not true in multi-parameter coupled scenarios. Although existing schemes do not lack the ability to judge whether candidate parameters will damage passed indicators, they may compromise for "engineering / efficiency" and not check further in this step-by-step parameter adjustment paradigm, i.e., measuring previous indicators and assuming that previous indicators remain passed.

[0006] (5) The existing scheme lacks a legal filtering, coupling risk assessment and comprehensive benefit ranking mechanism for model prediction results, and often directly outputs a single recommended result; (6) The existing scheme is not suitable for new batches and new working conditions, lacks a closed-loop mechanism for online result backflow triggering model retraining and strategy updating, and when the step-by-step rule parameter adjustment fails for new batches and working conditions, manual intervention is usually used for parameter adjustment, relying on artificial experience knowledge.

[0007] Therefore, a new parameter setting method is needed, which is automatically dominated by an agent based on online closed-loop data for model training, converts the training results into a verifiable, publishable and rollbackable strategy package, and is controlled by a safe executor for online execution. SUMMARY

[0008] The prior art solved by the present application relies on fixed rules, is difficult to adapt to new batches and new working conditions, the subsequent steps in the existing step-by-step parameter adjustment process destroy the previously passed indicators, and the step-by-step rule parameter adjustment has the problem of many iterations, affecting production capacity.

[0009] To solve the above problems, the present application provides an agent parameter setting method, system, electronic device and storage medium for a multi-parameter coupled device.

[0010] In a first aspect, the present application provides an agent parameter setting method for a multi-parameter coupled device, comprising the following steps: An online acquisition step acquires initial control parameters, initial measurement indicators, batch identifiers, subunit identifiers, parameter constraints and indicator specifications of the device or device subunits; A training condition judgment step judges whether the conditions for generating an automatic training task are met by the agent according to sample size, coverage, failure rate and batch drift degree; A model training step, when the training conditions are met, the agent generates an automatic training task, and trains, iterates and optimizes the parameter prediction model in a training sandbox; A strategy package generation step, the agent encapsulates the trained parameter prediction model into a parameter setting strategy package; A candidate generation and screening step, based on the strategy package, a plurality of candidate parameter combinations are generated, and the candidate parameter combinations are subjected to legal filtering, coupling risk assessment and comprehensive benefit ranking; A safe execution step, a safe executor performs final verification on the ranked candidate parameters, issues the verified candidate parameters to the target device, and performs a recovery operation when the candidate parameters fail or trigger a risk; A data backflow step, the execution results of the candidate parameters on the target device are collected and backflowed; a retraining trigger step, in which the agent determines whether to trigger a new round of model training according to the returned execution results.

[0011] Optionally, the parameter tuning strategy package includes applicable object information, data digest, model version information, candidate generator configuration, candidate verification rule, ranking target parameter, execution strategy, recovery strategy, and release gate information.

[0012] Optionally, the parameter prediction model includes one or more of a global parameter prediction model, a local response model, and a coupling risk model. The global parameter prediction model is used to predict a candidate parameter combination that can make all indicators eligible according to the initial state of the device. The local response model is used to estimate the impact of local changes in control parameters on performance indicators. The coupling risk model is used to assess whether a candidate parameter combination will improve one indicator while damaging other eligible indicators.

[0013] Optionally, in the candidate generation and screening step, the plurality of candidate parameter combinations are derived from one or more of the following ways: candidates directly output by the global parameter prediction model, similar distribution candidates of currently eligible devices in the batch, revised candidates derived by the local response model, legal candidates obtained by constraint search, risk-averse candidates, and rule mainline bottom-up candidates.

[0014] Optionally, in the candidate generation and screening step, the comprehensive benefit ranking is based on ranking points, which are determined by the following ways: The ranking points are positively correlated with the predicted passing probability, and negatively correlated with the coupling risk points, predicted execution cost, and uncertainty points. The coupling risk assessment is used to identify whether a candidate parameter combination is likely to damage the passed indicators. If the coupling risk of a candidate parameter combination exceeds a preset threshold, the candidate parameter combination is excluded from the ranking.

[0015] Optionally, in the safe execution step: The top-1 candidate parameter combination with the optimal comprehensive benefit ranking is executed by default. When the strategy package allows, the exploration budget is greater than zero, the candidate parameter combination passes the hard constraint check and risk threshold check, and the current test state allows additional measurement, limited top-k candidate parameter combinations are allowed to be executed.

[0016] Optionally, the recovery operation includes one or more of the following ways: fallback to a safe parameter, switch to a rule mainline tuning, return to a data collection phase, disable the current strategy package. The failure cause and recovery result are recorded and fed back for model retraining.

[0017] Optionally, in the retraining triggering step, the condition for triggering a new round of model training includes one or more of the following: The failure rate rises above a first threshold, the candidate legality rate falls below a second threshold, the recovery number rises above a third threshold, the model prediction deviates above a fourth threshold, the batch drift degree exceeds a fifth threshold, or a preset periodic inspection time is reached.

[0018] Optionally, before the training condition judgment step, if the collected data does not meet the condition for generating an automatic training task, the system enters an online safety collection mode; In the online safety collection mode, the safety rule mainline completes device setting, and collects limited exploration points within a preset safety budget and constraint range, which are used to provide online response samples for model training.

[0019] In a second aspect, the present application provides an intelligent agent parameter setting system for a multi-parameter coupled device, which is used to implement the method of the first aspect, and the system comprises: An online collection and constraint layer is used to collect initial state, online measurement feedback, execution result and maintenance safety constraint data; An intelligent agent automatic training arrangement layer is used to perform data quality diagnosis, batch drift identification, automatic training task generation and release gate judgment; A model training and optimization layer is used to train, iterate and offline verify the global parameter prediction model, the local response model and the coupling risk model; A strategy package and candidate decision layer is used to encapsulate the training result as a parameter setting strategy package, generate candidate parameter combinations based on the strategy package, perform legality filtering, coupling risk assessment and comprehensive benefit sorting; A safety execution and feedback layer is used to perform final safety verification and issue of the candidate parameters, perform failure recovery, and feed back the execution result to trigger model retraining.

[0020] In a third aspect, the present application provides an electronic device comprising a processor, a communication interface, a memory and a bus, wherein the processor, the communication interface and the memory communicate with each other through the bus, and the processor can invoke the logic instructions in the memory to perform the steps of the method provided in the first aspect.

[0021] In a fourth aspect, the present application provides a non-transitory computer readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the intelligent agent parameter setting method for a multi-parameter coupled device as described in the first aspect.

[0022] The agent parameter setting method for the multi-parameter coupled device has the following beneficial effects: The agent parameter setting method for the multi-parameter coupled device provided by the application collects initial control parameters, initial measurement indexes, batch identifiers, subunit identifiers, parameter constraints and index specifications of the device or device subunits through an online collection step, thereby providing a data basis for subsequent modeling; the agent determines whether the conditions for generating an automatic training task are met according to the collected data, thereby avoiding invalid training caused by insufficient or poor data, and ensuring the rational use of training resources and the quality of the model; when the conditions are met, the agent generates an automatic training task and trains, iterates and optimizes the parameter prediction model in a training sandbox, so that the model can automatically adapt to the data distribution changes of different batches and different subunits, thereby solving the problem that the prior art relies on fixed rules and is difficult to adapt to new batches and new working conditions.

[0023] The trained model is encapsulated as a parameter setting strategy package, so that the model output is decoupled from the hardware, and the model itself does not directly operate the device, thereby significantly improving the engineering safety and deployment flexibility of the system; based on the strategy package, multiple candidate parameter combinations are generated and subjected to legality filtering, coupling risk assessment and comprehensive benefit sorting, thereby effectively avoiding the safety risks and coupling damage problems that may exist in a single prediction result, and through the coupling risk assessment, it can be identified whether the candidate parameters will damage another passed index while improving one index, thereby solving the technical problem that the subsequent steps in the existing step-by-step parameter tuning process damage the pre-sequenced passed index.

[0024] The safety executor performs final verification on the sorted candidate parameters and issues them, thereby further ensuring the safety of online execution and preventing abnormal model output values from directly acting on the device; by collecting the execution results of the candidate parameters on the target device and flowing back, the agent determines whether to trigger a new round of model training according to the backflow results, thereby forming a sustainable evolving data closed loop, so that the system can automatically trigger model retraining and strategy updating when a new batch goes online, the failure rate increases or the candidate legality rate decreases, thereby continuously improving the setting success rate and efficiency of subsequent devices. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 The flowchart of the agent parameter setting method for the multi-parameter coupled device in the embodiment of the application; Figure 2 The principle diagram of the agent parameter setting method for the multi-parameter coupled device in the embodiment of the application; Figure 3 The relationship diagram of the tuning and acceptance of the optical module in the embodiment of the application; Figure 4 The architecture diagram of the agent parameter setting system for the batched multi-parameter coupled device in the embodiment of the application; Figure 5 Figure 1 is a structural diagram of an electronic device in an embodiment of the present application. DETAILED DESCRIPTION

[0026] In order to more clearly understand the objects, technical solutions and advantages of the present application, the present application will be described and explained in detail below in conjunction with the accompanying drawings and embodiments.

[0027] Unless otherwise defined, technical terms or scientific terms used in the present application shall have the general meaning understood by one of ordinary skill in the art to which the present application belongs. In the present application, "one", "a", "an", "the", "these" and similar words do not represent a quantitative limitation, but can be singular or plural. In the present application, the terms "include", "contain", "have" and any variants thereof are intended to cover non-exclusive inclusion; for example, a process, method and system, product or device containing a series of steps or modules (units) are not limited to the listed steps or modules (units), but can include steps or modules (units) not listed, or can include other steps or modules (units) inherent to the process, method, product or device. In the present application, the terms "connected", "connected", "coupled" and the like do not limit to physical or mechanical connection, but can include electrical connection, whether direct or indirect. In the present application, "multiple" means two or more. The association between the associated objects is described by "and / or", which means that there can be three relationships, for example, "A and / or B" can mean that A exists alone, A and B exist together, and B exists alone. In general, the character " / " represents an "or" relationship between the objects before and after it. In the present application, the terms "first", "second", "third" and the like are only used to distinguish similar objects, and do not represent a specific order of the objects.

[0028] The present embodiment takes batched optical module parameter setting as a specific application scenario for illustration. The present embodiment is only used to explain the specific implementation mode of the present application, and does not limit the present application to only be applied to optical modules.

[0029] One optical module can contain multiple lanes, and each lane can be regarded as a subunit to be set. Each lane can need to adjust control parameters such as bias_dac (bias current), mod_dac (modulation amplitude), step_dac (cross-point adjustment amount), ohm (load resistance), and measure indexes such as TxPwr (transmit optical power), ER (extinction ratio), CP (cross-point), and MaskMargin (eye diagram margin). Only when all indexes meet the specification requirements, the lane can be determined as Pass.

[0030] In the optical module embodiment, the candidate set is denoted as The set of legal candidates after hard constraint filtering is denoted as Hard constraint violation is marked as ,when This indicates that the candidate did not violate any hard constraints.

[0031] The set of valid candidates can be represented as: .

[0032] In this embodiment, let i be the index of the object to be tuned (such as a certain lane of an optical module), and k be the index of the candidate parameter combination. This represents the k-th candidate parameter combination for the i-th object.

[0033] Additionally, definition as a candidate The expected probability of passing, To couple risk distribution, To estimate measurement or execution costs, For uncertainty, The comprehensive ranking score can be expressed as: ; in, The weights are non-negative. The higher the expected probability of approval, the higher the priority of the candidate; the higher the coupling risk, execution cost, and uncertainty, the lower the priority of the candidate.

[0034] definition For object The maximum online trial budget, the set of online executable candidates can be represented as: ; when When, the system executes top-1; when At this time, the system can perform a finite top-k operation within the budget, but will not perform infinite trials.

[0035] In the optical module embodiment, the definition is as follows: The measured value is TxPwr. For ER measurement values, For CP measurement value, This is the MaskMargin measurement value. The corresponding acceptance constraint is... , , , The full index Pass can be represented as: ; in, Representation Object All indicators passed; indicates that at least one indicator is not met or the candidate parameter is determined to be unexecutable by the safety executor.

[0036] As shown in Figures 1 to 4 A multi-parameter coupled device-oriented agent parameter setting method provided in an embodiment of the present application includes the following steps: S1: online acquisition step, after the device or device subunit enters the setting process, the online acquisition step is first performed, which acquires the following data: Initial control parameters of the device or device subunit: in the optical module embodiment, for each lane, the system acquires the global initial state before any adjustment occurs, including the current values of bias_dac, mod_dac, step_dac, and ohm.

[0037] Initial measurement indicators: in the optical module embodiment, the current measurement values of TxPwr, ER, CP, and MaskMargin are acquired.

[0038] Batch identification and subunit identification: the production batch number to which the optical module belongs, the device number of the optical module, and the lane number of the lane to be set are acquired.

[0039] Parameter constraints: the upper and lower limits of each control parameter are acquired, such as the minimum and maximum values of bias_dac, the minimum and maximum values of mod_dac, etc.

[0040] Indicator specifications: the acceptance specifications of each performance indicator are acquired, such as the upper and lower limits of TxPwr, the upper and lower limits of ER, the upper and lower limits of CP, and the lower limit of MaskMargin.

[0041] The system stores these data in the online acquisition database in a unified format for subsequent training condition judgment and model training. For each lane that enters the setting process, the system performs the above acquisition operation.

[0042] It should be noted that before the training condition judgment step, if the acquired data does not meet the condition for generating an automatic training task, the system enters a one-stage online safety acquisition mode. In the one-stage online safety acquisition mode, the system completes device setting by the safety rule main line, and acquires limited exploration points within the preset safety budget and constraint range, which are used to provide online response samples for model training.

[0043] Exemplarily, taking the optical module as an example, when the online data sample size is insufficient, the coverage is insufficient, or the number of failed samples is too small, the system judges that it does not have training conditions, at this time the safety executor executes the setting according to the safety rule main line. The safety rule main line is, for example, first execute SetTxPwr adjustment bias_dac to make TxPwr up to standard, then execute SetER adjustment mod_dac to make ER up to standard, then execute SetCp adjustment step_dac to make CP up to standard, and finally execute ScanMaskMargin adjustment ohm to make MaskMargin up to standard. In each step of the adjustment process, the system adds limited exploration points under the premise of not exceeding the safety boundary. For example, when the influence of mod_dac on ER is uncertain, the system can select a small number of legal mod_dac candidates within the budget for measurement to estimate the local response; when a certain ohm region may affect MaskMargin, the system can collect a small number of response samples under recoverable conditions.

[0044] The exploration points are not aimed at immediate pass, but are used to generate online response samples needed to train global parameter prediction models, local response models, and coupling risk models. The number of exploration points is limited by the budget Once the risk is triggered, the measurement is abnormal, or the indicators are significantly deteriorated, the safety executor immediately stops exploration and returns to the rule main line.

[0045] The system records each parameter issuance and measurement result in a stage, including successful setting results and failed attempts, which will be used as initial samples for subsequent model training.

[0046] S2: Training condition judgment step, the agent judges whether the conditions for generating an automatic training task are met according to the sample size, coverage, failure rate, and batch drift degree; The agent judges whether the conditions for generating an automatic training task are met according to the collected data. The training conditions can include whether the sample size reaches a preset threshold, whether the parameter coverage is sufficient, whether the constraint integrity is met, whether the main supervision sample quality is qualified, whether the risk sample exists, whether the candidate legal rate is up to standard, whether the failure rate is within an acceptable range, and whether the batch drift degree is significant.

[0047] Exemplarily, taking the optical module as an example, the agent performs the training condition judgment after each new batch goes online or after processing a preset number (such as 100) of lanes. Specifically, the agent counts the number of samples in the current online collection database, checks whether different value intervals of bias_dac, mod_dac, step_dac, and ohm parameters are covered, checks whether the number of pass samples and failure samples both exceeds a preset threshold (for example, each is not less than 50), calculates the candidate pass rate and failure rate of the last preset number (for example, 50) of lanes, and calculates the difference degree between the initial state distribution of the current batch and the applicable batch distribution of the existing strategy. If the above indicators all meet the preset threshold, it is judged that the condition for generating an automatic training task is met; otherwise, continue to collect data.

[0048] S3: model training step, when the training condition is met, the agent generates an automatic training task, and trains, iterates and optimizes the parameter prediction model in a training sandbox; When it is judged that the training condition is met, the agent generates an automatic training task, determines the training data range, training target, model type to be trained, verification index and release gate.

[0049] In the optical module embodiment, the automatic training task generated by the agent specifically includes: Training data range: specify which batches, which models of optical modules, and which lanes of online collection data to use as the training set, and reserve which data as the validation set; Training target: for example, the goal of the global parameter prediction model is to predict the candidate parameter combination that may pass the global indicator Pass from the initial state, and the goal of the local response model is to estimate the influence of the local changes of bias_dac, mod_dac, step_dac, and ohm on TxPwr, ER, CP, and MaskMargin; Model type to be trained: specify which one or more of the global parameter prediction model, the local response model, and the coupling risk model needs to be trained; Verification index: specify the candidate pass rate, the predicted pass rate, the risk interception effect, the average measurement times, the recovery trigger rate, etc. as the model verification index; Release gate: specify the threshold of the verification index that needs to be met before the strategy package is released.

[0050] Specifically, the agent sends the above training task to the training sandbox in the form of structured instructions, triggering the subsequent model training process. The three types of models share the online collected data, but the prediction targets and training sample constructions are different. The training samples all come from the records naturally generated in the online setting process, without additional manual annotation and without dependence on external historical databases.

[0051] Let the object or subunit to be tuned be , its initial state vector be , the candidate parameter combination be , the candidate measurement vector after execution be , where may include performance indicators such as TxPwr, ER, CP, MaskMargin, etc. Let the parameter combination when the final full indicator of the object is qualified be , and the coupling risk label after candidate execution be , when the candidate causes the indicator to deteriorate, destroys the passed indicator, triggers a recovery operation, or enters a high-risk area, , otherwise .

[0052] It should be noted that the global parameter prediction model is used to predict a candidate parameter combination that can make all performance indicators meet the specification requirements according to the initial state of the device or device subunit. Its input features can include initial control parameters, initial measurement indicators, parameter upper and lower limits, indicator specifications, batch identifiers, product models, lane identifiers, test temperatures, or test working conditions, etc. Its output can be one or more candidate parameter combinations, such as predicted values of bias_dac, mod_dac, step_dac, ohm, etc. Parameters, or candidate parameter distribution, candidate mean, and uncertainty.

[0053] In a specific algorithm structure, the global parameter prediction model can use a multi-output regression model, and the model structure (including feature subset, output dimension, regularization method, etc.) is automatically determined by the agent according to the online data volume, feature coverage, and candidate effect. Its form can be represented as: ; wherein represents the global parameter prediction model, represents the model parameters, represents the candidate parameter combination obtained by prediction. The model can be implemented using multi-output linear regression, ridge regression, random forest regression, gradient boosting tree, neural network, or integrated model, and the specific model type is automatically selected by the agent according to the online data volume, feature coverage, batch drift degree, and candidate legality rate. If some control parameters are discrete gears, the corresponding output can also be set to classification output or sequence candidate output.

[0054] In addition, the training samples of the global parameter prediction model can be constructed from the records of the full indicator qualification in the online tuning process, represented as: ; wherein is the initial state of lane before any adjustment occurs, the parameter combination of the lane when it finally passes the overall indicator, indicates that the sample is an overall indicator pass sample. Take the optical module lane as an example: the input is the initial state of the lane, including the initial bias_dac, mod_dac, step_dac, ohm, initial TxPwr, ER, CP, MaskMargin, batch number, product model, lane number, test working condition, parameter upper and lower limits, and indicator specifications, etc.; the output label is a set of parameters corresponding to the lane when it finally passes the overall indicator, for example, the final qualified bias_dac, mod_dac, step_dac, and ohm. That is, the question answered by the global supervision sample is: what is the state of a lane when it just enters the setting process, and what should it fall into when it finally succeeds. Only the record of the final successful setting and overall indicator pass is suitable as such a sample; failure records are generally not used as positive labels, but can be used for risk samples or subsequent model verification.

[0055] During training, the parameter prediction error can be used as the main loss function, for example, the weighted mean square error: ; wherein, is the parameter weight matrix, which is used to reflect the differences in the dimensions, step sizes, or safety importance of different control parameters. For discrete control parameters, cross-entropy loss can be used; for the case of outputting multiple candidates at the same time, the minimum error between the optimal candidate and can be used as the loss, or the candidate ranking loss can be used, so that the candidate closer to the final qualified parameter obtains a higher ranking.

[0056] In addition, in further embodiments, the loss function of the global parameter prediction model can also include a legality penalty term and a risk penalty term, for example: ; wherein, is used to penalize the prediction parameters that exceed the parameter upper and lower limits or the device executable range, is used to penalize the prediction candidates that may damage the passed indicators, and are weight coefficients. In this way, the model not only learns the final qualified parameters, but also learns to generate safer and more executable candidate parameters.

[0057] The local response model is used to estimate the influence of the local change of the control parameter on the performance indicator. Its input can include the current state , the current parameter combination , and the candidate parameter change , parameter constraints, index specifications and context features; its output can be the performance index change , or the predicted measurement value after candidate execution .

[0058] In a specific algorithm structure, the local response model can be represented as: ; Or: ; Wherein, represents the local response model, represents the model parameters. The model can be implemented by local linear model, multi-output regression tree, gradient boosting tree, Gaussian process regression, neural network regression or parameterized response model. For example, in the optical module embodiment, the influence of bias_dac on TxPwr and ER, the influence of mod_dac on ER and MaskMargin, the influence of step_dac on CP and eye pattern, and the influence of ohm on MaskMargin can be learned respectively.

[0059] The training sample of the local response model can be constructed by recording the candidate parameters after online execution, represented as: ; Or, when the measurement values before and after the parameter change are recorded, it is constructed as: ; Wherein, represents the change of performance index before and after the candidate execution, is a certain candidate parameter combination, is the measurement vector obtained after executing the candidate, and the example is . The model learns the influence of candidate parameter change on measurement indexes such as TxPwr, ER, CP and MaskMargin.

[0060] Taking the optical module lane as an example: the input includes the current parameters and current measurement values ​​before deployment, such as the current bias_dac, mod_dac, step_dac, ohm, and the current TxPwr, ER, CP, and MaskMargin; it also records the changes of the candidate parameters relative to the current parameters, such as how much bias_dac increases and mod_dac decreases; the output label is the change in measurement values ​​after deployment, such as how much TxPwr increases or decreases, how much ER changes, how much CP changes, and how much MaskMargin changes. In other words, the local response sample answers the question: in the current state, if a parameter is adjusted slightly in a certain direction, what impact will it have on the various measurement indicators? These samples come not only from successful tuning processes but also from limited exploration points in a phase one of secure acquisition. Even if a candidate parameter does not ultimately pass all lane indicators, as long as it is executed within safety constraints and produces valid measurement results, it can still be used as a local response sample because it provides the relationship between parameter changes and indicator changes.

[0061] Weighted mean squared error or Huber loss can be used during training to reduce the impact of outlier measurements on the model. ; or: ; in, This is a weight matrix for performance metrics, used to reflect the importance of different performance metrics. For example, MaskMargin or passed metrics can be assigned higher weights to make the model pay more attention to metric responses that may lead to changes in pass / fail ratings.

[0062] Furthermore, in a further implementation, the local response model can also incorporate a directional consistency loss. If a change in a candidate parameter causes the measured index to rise or fall, the direction of change predicted by the model should be consistent with the measured direction. The directional consistency loss can be expressed as: ; The training objective of the final local response model can be expressed as: ; in, The weights are used for directional consistency loss. Through this model, the system can predict the performance metric changes that a change in a candidate parameter might bring without directly traversing a large number of parameters, and generate corrected candidate parameters or evaluate candidate returns accordingly.

[0063] The coupling risk model is used to assess whether a combination of candidate parameters might improve one performance metric while degrading other already satisfactory performance metrics. Its input may include the initial state. Currently, the indicator set and candidate parameter combination have been approved. The output of the local response model includes predicted measurements or predicted changes, candidate sources, historical failure types, batch identifiers, and parameter constraints; its output can also be the coupled risk probability. It can also be a multi-label risk score for different risk types.

[0064] In a specific algorithmic structure, the coupled risk model can adopt a binary classification or multi-label classification model, as follows: ; in, This indicates a coupled risk model. Indicates model parameters, This indicates the mask or status of the indicator that has been passed. This represents the probability that a candidate breach has passed an indicator or triggered a recovery operation. The model can be implemented using logistic regression, random forest classifiers, gradient boosting tree classifiers, neural network classifiers, or ensemble classifiers.

[0065] It should also be noted that the training samples for the coupled risk model can be constructed from failure records, recovery records, and metric deterioration records after candidate execution, as follows: ; in, This indicates that after the candidate is executed, at least one passed indicator becomes unqualified, the indicator deteriorates beyond the preset threshold, recovery is triggered, or it is judged as high risk by the safety executor. This indicates that the aforementioned risks were not triggered after the candidate was executed.

[0066] Taking an optical module lane as an example: the input includes the state before executing the candidate, the state of the passed indicators, the combination of parameters for this candidate, the candidate source, batch information, parameter constraints, and indicator specifications. If, after executing the candidate, previously passed indicators become unqualified, or an indicator significantly deteriorates, or the system triggers recovery, or the safety actuator determines that the candidate has entered a high-risk area, then the sample is marked as risky. If executing the candidate does not damage passed indicators or trigger recovery, then it is marked as risk-free. In other words, the question answered by coupled risk samples is: although a candidate parameter may improve currently unqualified indicators, will it re-degrade already passed indicators? For example, if a candidate increases TxPwr but causes ER or MaskMargin to change from qualified to unqualified, this record will be constructed as a coupled risk sample and marked as risky.

[0067] Binary cross-entropy loss can be used during training: ; When there are few risk samples or the positive and negative samples are imbalanced, weighted cross-entropy or focal loss can be used: ; in, , Used to balance risky and non-risky samples. This is used to improve the model's attention to difficult samples.

[0068] In a further implementation, the coupled risk model may also employ ranking loss, ensuring that the candidate risk score for actually triggered risks is higher than the candidate risk score for non-triggered risks. This applies to a pair of candidate risks for the same object or in similar states. and If candidates Triggering risk and candidate If no risk is triggered, the following ranking loss can be added: ; in, The minimum risk interval is defined. The final loss of the coupled risk model can be expressed as: ; This model enables the system to identify high-risk candidates before they are issued. If a candidate's risk score exceeds a preset threshold, it is removed from the candidate set. If a candidate's risk is in the medium range but its expected return is high, its overall ranking score is reduced, and it is only allowed to enter a limited top-k execution range when the budget allows and there are no better low-risk candidates.

[0069] For example, in this embodiment, the training sandbox first constructs features for model training. Taking the optical module lane as an example, the features may include: Initial parameter tuning characteristics: bias_dac, mod_dac, step_dac, ohm; Initial measurement characteristics: TxPwr, ER, CP, MaskMargin; Specification constraints: TxPwr upper and lower limits, ER upper and lower limits, CP upper and lower limits, MaskMargin lower limit, and upper and lower limits of each parameter tuning value; Contextual features: batch identifier, product model, lane number, test temperature or test condition; Process statistical characteristics: number of candidates tried, candidate validity rate, number of recovery triggers, and failure type distribution; Coverage and uncertainty features: the distance of the current sample relative to the distribution of trained samples, and the distribution difference between the current batch and the batches to which the strategy applies.

[0070] The agent can automatically select the feature set to be included in the training based on data quality, missing rate, feature stability, and candidate performance. For example, if the local response of mod_dac to ER changes significantly in a certain batch, the agent can increase the weight or priority of features related to mod_dac, ER, and batch identifier in the local response model.

[0071] The three models mentioned above can be trained separately, or multiple candidate models can be trained in parallel by an agent in a training sandbox. The agent can automatically adjust training parameters such as feature subsets, sample weights, tree depth, learning rate, regularization strength, risk threshold, number of candidates, and ranking weights, and select the optimal model or model combination based on offline validation results.

[0072] Model validation metrics can include: parameter prediction error, metric response prediction error, candidate validity rate, expected pass rate, coupling risk recall rate, false interception rate, average number of measurements, recovery trigger rate, new batch validation performance, and policy execution time. Only when the model combination meets the release gating requirements will the agent encapsulate the model version, candidate generator, risk threshold, ranking weights, execution budget, and recovery strategy into a parameter tuning policy package; otherwise, it will continue collecting online samples or trigger the next round of training.

[0073] For example, the agent performs the following automated training process in the training sandbox: Input:online_records,constraints,current_strategy Step1:diagnose_data_quality(online_records) Step2:detect_batch_drift(online_records,current_strategy) Step3:ifnottraining_gate_passed:continuedata_collection Step4:buildD_global,D_response,D_risk Step5:trainglobal_parameter_predictoronD_global Step6:trainlocal_response_modelonD_response Step7:traincoupling_risk_modelonD_risk Step8:generatecandidate_generators Step9:validatecandidatelegality,risk,passprobabilityandcost Step10:packagemodels,validators,rankingpolicyandrecoverypolicy Step11:publishonlyifrelease_gate_passed This pseudocode is only used to illustrate the process of automatic training of agents and generation of policy packages, and is not limited to specific programming languages, model algorithms or data structures.

[0074] If the verification result does not meet the release gate, the policy package will not enter the online activation state, and the system will continue to collect data or trigger a new training task so that the features, sample weights and model parameters can be adjusted in the next round of training.

[0075] S4: Strategy package generation step, the agent encapsulates the trained parameter prediction model into a parameter tuning strategy package; The strategy package specifically includes the following: Applicable object information: including device type, batch range, sub-unit or channel range. In the optical module embodiment, the applicable optical module model, batch number range, and lane number range are specified.

[0076] Data summary: Includes training sample size, coverage (coverage of each parameter's value range), number of failed samples, and number of risky samples.

[0077] Model version information: including the version number, training time, and training data cutoff time of the global parameter prediction model, local response model, and coupling risk model.

[0078] Candidate generators include: global prediction candidates, local response candidates, constraint search candidates, current batch distribution candidates, risk avoidance candidates, and rule-based candidates. Candidate validation rules: hard constraints, risk thresholds, and coupling violation judgment; Ranking target parameters: expected pass probability, expected measurement cost, coupling risk cost, uncertainty; Execution strategy: Specify the default to execute top-1; if the strategy package allows, test the budget. When all candidates pass hard constraint verification and risk threshold verification, and the current production line or testing status allows for additional measurements, a limited top-k operation is permitted.

[0079] Recovery strategy: Specify the fallback to security parameters, rule mainline, data collection stage, or disable the current policy package after failure.

[0080] Publish gating information: specify whether grayscale is allowed, whether full-scale is allowed, when to disable or trigger retraining.

[0081] After training, the agent does not directly distribute model predictions, but instead generates a policy package for tuning the aforementioned parameters. The policy package is stored in a structured data format (such as JSON or Protocol Buffer) for easy loading and parsing by the system.

[0082] S5: Candidate generation and screening step, generating multiple candidate parameter combinations based on the strategy package, and performing legality filtering, coupled risk assessment and comprehensive benefit ranking on the candidate parameter combinations; Candidate parameter combinations can be derived from one or more of the following methods: Candidate parameters directly output by the global parameter prediction model: The initial state of the current lane is input into the global parameter prediction model, and the model outputs one or more recommended parameter combinations. In the optical module embodiment, the model outputs recommended bias_dac, mod_dac, step_dac, and ohm values.

[0083] Candidate distribution of similar qualified equipment in the current batch: Select lanes with initial states similar to the current lane from the lanes that have passed in the current batch, and combine their final pass parameters as candidates.

[0084] Correction candidates derived from the local response model: Based on the relationship between the current candidate parameters and the measured index, the expected index change after fine-tuning a certain parameter is predicted using the local response model. If the predicted result is better than the current value, a correction candidate is generated.

[0085] Valid candidates obtained from constraint search: Perform grid search or random sampling in the parameter space according to a preset step size, and retain only candidates that satisfy hard constraints.

[0086] Risk-averse candidates: Candidates with lower risk scores in the coupled risk model assessment.

[0087] Candidate main rules: Candidate parameter combinations generated according to the traditional step-by-step parameter tuning process, serving as a safety fallback.

[0088] In the optical module embodiment, the system uses the above six methods in parallel, each method generates several candidate parameter combinations, which are then merged to form an initial candidate set C.

[0089] Candidate parameters are first filtered using hard constraints. Hard constraints include: whether each control parameter is within its upper and lower limits, and whether it meets the device's executable range. Taking an optical module as an example, this involves checking whether bias_dac is within the range of [bias_min, bias_max], mod_dac is within the range of [mod_min, mod_max], step_dac is within the range of [step_min, step_max], and ohm is within the range of [ohm_min, ohm_max]. Candidates that do not meet the hard constraints are directly eliminated, and the set of valid candidates is denoted as . .

[0090] Subsequently, a coupling risk assessment is performed on the legitimate candidates. The system then calculates a comprehensive ranking score for candidates that pass the legitimacy filtering and coupling risk assessment, according to the candidate ranking scoring strategy. The overall ranking score is determined as follows: the ranking score is positively correlated with the expected pass probability and is coupled with the risk score. Expected execution costs and uncertainty Negative correlation.

[0091] Specifically: .

[0092] I. Methods for estimating the probability of passing It is expected that the probability will be used to represent the likelihood that the target device or device subunit will meet all the required indicators after a certain combination of candidate parameters is executed.

[0093] In the optical module embodiment, the expected pass probability can be estimated in one or more of the following ways: 1: Statistical estimation of similar samples The system searches for samples in historical online tuning records that are similar to the initial state of the current lane, such as initial TxPwr, ER, CP, MaskMargin, initial parameter values, batch information, product model, and lane number. Then, it calculates the proportion of these similar samples that ultimately pass after using the same or similar combinations of candidate parameters, and uses this proportion as the expected pass probability for that candidate.

[0094] For example, if the initial state of the current lane is similar to that of the past 50 samples, and 35 of them pass all indicators under similar candidate parameter combinations, then the expected pass probability of this candidate can be estimated as 35 / 50.

[0095] 2: Model Output Estimation The system can also directly output the pass probability of candidate parameter combinations from the global parameter prediction model. The model input includes the current initial state, candidate parameter combinations, parameter constraints, indicator specifications, batch information, etc., and the output is the probability that all indicators will pass after the candidate is executed.

[0096] 3: Combination Estimation When the sample size is sufficient, the statistical results of similar samples and the model output can be combined. For example, the pass probability of the model output can be weighted and fused with the pass rate of similar samples to obtain a more stable predicted pass probability.

[0097] II. Calculation Method of Coupling Risk Score Coupling risk score represents the risk that a combination of candidate parameters may improve one metric while damaging other already approved metrics.

[0098] In the optical module embodiment, the coupling risk score can be output by the coupling risk model. The inputs of the coupling risk model include the current state, the status of the passed indicators, the candidate parameter combination, the changes in the predicted indicators before and after the candidate execution, the candidate source and batch information, etc.; the output is the probability of the candidate damaging the passed indicators, causing the indicators to deteriorate significantly or triggering recovery.

[0099] For example, if a candidate may enable TxPwr to meet the target, but the model predicts that it will worsen the already qualified ER or MaskMargin, then the coupling risk score of the candidate is high.

[0100] In one implementation, the coupling risk score can take values ​​between 0 and 1: The closer it is to 0, the lower the coupling risk; The closer it is to 1, the more likely the candidate is to break the passed indicators or trigger recovery.

[0101] If the coupling risk score exceeds the preset risk threshold, the candidate can be directly eliminated and will no longer be included in the comprehensive ranking; if the coupling risk is in the medium range, the candidate will be retained but its comprehensive ranking priority will be reduced.

[0102] III. Calculation Method of Estimated Execution Costs The expected execution cost represents the testing cost required to execute a particular combination of candidate parameters.

[0103] In the optical module embodiment, the expected execution cost may include one or more of the following factors: Number of times candidate parameters are issued; The number of metrics that need to be measured after executing this candidate; The time required to measure metrics such as TxPwr, ER, CP, and MaskMargin; Do we need to perform additional time-consuming tests such as eye diagram scanning and MaskMargin scanning? The recovery costs that may result from a failed candidate; The greater the change between the candidate parameter and the current parameter, the more likely it is to require confirmation or recovery costs.

[0104] For example, candidates that only require one parameter delivery and one routine measurement are expected to have lower execution costs; candidates that require additional MaskMargin scanning or may trigger recovery are expected to have higher execution costs.

[0105] To facilitate calculation with other components, the execution cost is expected to be normalized to between 0 and 1. The higher the cost, the lower the candidate's overall ranking score.

[0106] Furthermore, if a candidate requires one parameter delivery and one routine measurement, estimated to take 30 seconds, and the system sets the maximum allowed execution time for a single candidate to be 100 seconds, then the time cost of that candidate can be normalized to 0.30. If another candidate requires an additional MaskMargin scan, estimated to take 70 seconds, then its cost can be normalized to 0.70. The system can also comprehensively consider the number of delivery steps, the number of measurement steps, the number of scan steps, and the possibility of recovery to obtain the final execution cost.

[0107] IV. Calculation Method of Uncertainty Score The uncertainty score represents the degree of confidence the system has in a particular candidate prediction result.

[0108] In one implementation, the uncertainty component can be determined in the following way: 1: Model confidence interval width If the model can output a prediction confidence interval, the wider the confidence interval, the more uncertain the model is about the candidate prediction, and the higher the uncertainty score.

[0109] 2: Training sample spatial coverage density The system calculates the distance between the current sample and the training sample space. If the initial state, batch features, or candidate parameter combinations of the current lane appear frequently and have sufficient coverage in the training samples, the uncertainty is low; if the current sample is in a sparse region of the training samples, or belongs to a new batch or new operating condition, the uncertainty is high.

[0110] 3: Divergence in predictions across multiple models When an agent trains multiple models in parallel, it can compare the predictions of different models for the same candidate. If the outputs of multiple models are similar, the uncertainty is low; if the predictions of multiple models differ significantly, the uncertainty is high.

[0111] The uncertainty score can also be normalized to between 0 and 1. The higher the score, the more uncertain the system is, and the candidate should be given lower priority when ranking, or should only be executed if the limited top-k budget allows.

[0112] V. Methods for Determining Weights The weights in the overall ranking reflect business objectives and security preferences. These weights can be determined in the following ways: In the initial stage of the system, weights can be pre-configured based on business priorities. For example, if the pass rate is given more importance, the weight of the expected pass probability is increased; if security and avoiding damage to already passed metrics are given more importance, the weight of the coupling risk score is increased; if test cycle time is given more importance, the weight of the expected execution cost is increased.

[0113] Furthermore, the specific configuration of the weights can be dynamically adjusted according to the production stage.

[0114] For example, in the early stages of trial production or the initial launch of a new batch, the system prioritizes security and can be configured with a higher coupling risk weight, for example: The expected weights are: probability weight: 0.45, coupling risk weight: 0.35, execution cost weight: 0.10, and uncertainty weight: 0.10. Once the production line is stable and the model has been sufficiently validated, if more emphasis is placed on testing efficiency, the execution cost weight can be appropriately increased, for example: The expected weights are: probability weight: 0.50, coupling risk weight: 0.25, execution cost weight: 0.20, and uncertainty weight: 0.05.

[0115] The aforementioned weights can be configured in the policy package and adjusted through validation using historical samples in the training sandbox. During validation, the pass rate, average number of measurements, average tuning time, candidate validity rate, and recovery trigger rate under different weight combinations are compared, and the weight combination with the best overall performance is written into the policy package.

[0116] To facilitate understanding of how the comprehensive ranking score is calculated, an example is provided below. This example is only for illustrating the calculation logic and does not limit the actual implementation of this invention to using the same numerical values, weights, or normalization methods.

[0117] Suppose that a certain optical module lane needs to select candidate parameter combinations, and the system has already generated three candidates: Candidate A: bias_dac=620, mod_dac=410, step_dac=128, ohm=45 Candidate B: bias_dac=635, mod_dac=420, step_dac=126, ohm=48 Candidate C: bias_dac=610, mod_dac=405, step_dac=130, ohm=44 The system calculates the expected pass probability, coupling risk score, expected execution cost, and uncertainty score for each candidate. Each score can be normalized to between 0 and 1, with a higher expected pass probability being better, and lower coupling risk score, expected execution cost, and uncertainty score being better.

[0118] For example, the system calculates the results shown in Table 1 below:

[0119] The expected pass probability of candidate A is 0.82, which means that the system, based on the statistics of similar samples or the output of the model, believes that this candidate has an approximately 82% probability of passing all indicators of the lane. Candidate B has the highest expected pass probability of 0.90, but its coupling risk score is also high of 0.35, indicating that although this candidate is more likely to make the currently unpassed indicators meet the standards, it is also more likely to undermine the already qualified indicators. Candidate C has a low expected probability of passing, at 0.72, but its risk, cost, and uncertainty are all low.

[0120] Assume the weights configured in the current strategy package are: expected pass probability weight: 0.50, coupling risk weight: 0.25, expected execution cost weight: 0.15, and uncertainty weight: 0.10.

[0121] The comprehensive ranking score can be understood as follows: The overall ranking score is calculated as follows: expected pass probability contribution - coupling risk penalty - execution cost penalty - uncertainty penalty.

[0122] Substituting the above values, we obtain the comprehensive ranking results shown in Table 2 below:

[0123] Therefore, the ranking of the three candidates is: Candidate A > Candidate C > Candidate B.

[0124] Although candidate B has the highest expected success rate, its overall ranking score is lower than candidate A due to its higher coupling risk and uncertainty. Candidate C has lower risk and cost, but also a lower expected success rate, so its overall ranking score is slightly lower than candidate A. The system defaults to prioritizing candidate A.

[0125] If there is another candidate D, whose expected pass probability is 0.95, but whose coupling risk score is 0.80, and the risk elimination threshold set in the strategy package is 0.60, then even if candidate D has a higher expected pass probability, it will be directly eliminated because the coupling risk exceeds the threshold and will not enter the comprehensive ranking or safe execution stage.

[0126] The system sorts all legitimate candidates according to a comprehensive ranking. Sort them from highest to lowest, and the candidates with the highest overall returns are those ranked first.

[0127] S6: Safety execution steps. The safety actuator performs a final verification of the candidate parameters after ranking the comprehensive benefits. The verification includes parameter upper and lower limit verification, equipment executable range verification, candidate risk threshold verification, and trial budget verification. The candidate parameters that pass all verifications are sent to the target equipment in order of sorting. If the candidate parameter fails to execute or triggers a risk, a recovery operation is performed. The safety actuator performs a final verification on the sorted candidate parameters and sends the verified candidate parameters to the target device. Before sending the candidates, the safety actuator performs a final verification, including parameter upper and lower limit verification (even if the candidates have been filtered by hard constraints, they are still verified again), device executable range verification (different device models may have different executable ranges), candidate risk threshold verification, and trial budget verification.

[0128] By default, the system selects the top-1 candidate parameter combinations that yield the best overall return. In other words, the system only issues the parameters by default. The highest-ranking candidate.

[0129] When the strategy package allows, test the budget. A finite number of top-k candidate parameter combinations are allowed to be executed when all candidate parameter combinations pass hard constraint verification and risk threshold verification, and the current test state allows for additional measurements. Specifically, when At that time, the system can execute sequentially within the budget. Sort the top k candidates (k≥2), and execute them in the following order: Execution proceeds from highest to lowest. If the measurement results after executing a candidate satisfy all metrics (Pass), execution of subsequent candidates is stopped. The finite top-k algorithm avoids infinite trial and stops once the budget limit is reached or a risk is triggered.

[0130] In the optical module embodiment, if the current lane's trial budget Bi = 2 measurements, and the policy package allows top-2 execution, the system can execute the first and second ranked candidates in sequence, measuring once after each candidate is executed. If the first candidate has passed all indicators, the second candidate will not be executed.

[0131] If the measurement obtained after the candidate is executed If the full-indicator acceptance criteria are met, then record... Successful samples will be returned; if they fail, or if it is found after execution that the candidate has damaged the previously passed indicators, the failure indicator, candidate source, risk score and recovery path will be recorded.

[0132] When a candidate parameter fails to execute or triggers a risk, the safety executor performs a recovery operation. The recovery operation includes one or more of the following methods: Rollback to safe parameters: Restores device parameters to a known safe combination of parameters, such as the initial or default state.

[0133] Switch to main rule tuning: Abandon model-based candidate execution and switch to the traditional security rule main line (step-by-step parameter tuning process) to complete the tuning.

[0134] Return to data collection phase: Stop the execution of the current strategy, mark the current lane as needing to collect data again, and return to the first-stage online secure collection mode.

[0135] Disable the current policy package: If the candidate validity rate decreases or the recovery count increases in multiple lanes in the same batch, the current policy package will be disabled and will no longer be used for tuning of subsequent devices, and model retraining will be triggered.

[0136] In the optical module implementation, if TxPwr fails and bias_dac is close to the boundary, it falls back to the main line of safety rules; if ER passes but MaskMargin deteriorates significantly, it is determined that the candidate has coupling risk and the sample is marked as a risk sample; if the candidate validity rate drops in multiple consecutive lanes in the same batch, the current policy package is disabled and model retraining is triggered; if the initial state distribution of the new batch deviates significantly from the applicable range of the existing policy, the global parameter prediction model and local response model are retrained or tuned, and the reasons for failure and recovery results are recorded and fed back for model retraining.

[0137] S7: Data feedback step, collecting the execution results of the candidate parameters on the target device (including measurement feedback, failure reasons and recovery results) and feeding them back to the online acquisition database for subsequent model retraining; S8: Retraining trigger step, the agent determines whether to trigger a new round of model training based on the execution result of the reflow.

[0138] The system collects the execution results of candidate parameters on the target device and feeds them back. Specifically, the system feeds back the candidate execution results, measurement feedback, failure reasons, and recovery results to the online acquisition database.

[0139] The agent determines whether to trigger a new round of model training based on the execution result of the reflow. The conditions for triggering a new round of model training include one or more of the following: Failure rate rises above the first threshold: for example, the tuning failure rate of the most recent 50 lanes exceeds 10%.

[0140] The candidate validity rate drops below the second threshold: for example, the proportion of candidates in the most recent 50 lanes that pass the hard constraint validation is less than 80%.

[0141] Recovery count increases beyond the third threshold: for example, more than 15 times in the last 50 lanes that trigger recovery.

[0142] Model predictions deviate from the fourth threshold: for example, the average deviation between the model's predicted measurements and the actual measurements exceeds a preset tolerance.

[0143] Batch drift exceeds the fifth threshold: for example, the KL divergence or Wasserstein distance between the initial state distribution of a new batch and the distribution of batches applicable to the policy package exceeds a preset threshold.

[0144] Reach the preset cycle inspection time: for example, automatically trigger model retraining once every 500 lanes processed or every 24 hours.

[0145] When any of the above conditions are met, the agent regenerates the automatic training task and enters a new round of model training, iteration, and tuning. After the new training is completed, the agent generates an updated policy package, which is then deployed after gating and canary-scale verification.

[0146] Through the above process, the system can transform each success or failure into a training sample for the next round, enabling the model to continuously evolve and adapt to new batches, new operating conditions, and new failure modes.

[0147] This invention also provides an intelligent agent parameter tuning system for multi-parameter coupled devices. This system is used to implement the above-described method embodiments, and details already described will not be repeated. The terms "module," "unit," and "subunit," etc., used below refer to combinations of software and / or hardware that perform a predetermined function. Although the apparatus described in the following embodiments is preferably implemented in software, hardware implementation or a combination of software and hardware is also possible and contemplated.

[0148] The system comprises five functional layers: Online Acquisition and Constraint Layer: This layer is used to acquire initial states, online measurement feedback, candidate execution results, failure and recovery records, and maintain specifications, parameter upper and lower limits, and safety constraint data. It connects to the device's test interface to receive measurement and status data from the device.

[0149] The agent-based automatic training orchestration layer is used for data quality diagnosis, batch drift identification, coverage gap identification, automatic training task generation, training plan formulation, and deployment gating. This layer is the core of the system's decision-making, controlling the initiation and progression of the entire closed loop. The agent performs group modeling, group sorting, or group risk assessment based on batch identifiers and sub-unit identifiers.

[0150] Model Training and Tuning Layer: Used to isolate and perform training, iteration, parameter tuning, model selection, and offline validation of the global parameter prediction model, local response model, and coupled risk model within a training sandbox. The training sandbox is isolated from the online execution environment, ensuring that the training process does not affect online production.

[0151] Strategy Package and Candidate Decision Layer: These layers encapsulate model training results into parameter tuning strategy packages, generate multi-source candidates based on these packages, and perform validity filtering, coupled risk assessment, and comprehensive benefit ranking. Each strategy package records the applicable batch range, sub-unit range, model version, candidate generation method, validity filtering rules, ranking objectives and weights, execution budget, recovery conditions and fallback paths, as well as policy enabling and disabling conditions.

[0152] Safety Execution and Backflow Layer: This layer is used by the safety actuator to perform final safety checks, parameter distribution, measurement reception, pass / fail determination, finite top-k trial and recovery operations, and backflow the results for model retraining.

[0153] In this system, the intelligent agent does not directly distribute hardware parameters. The intelligent agent is responsible for automatic training and orchestration, policy package generation, deployment judgment, and review; the safety executor is responsible for final legality verification and online parameter distribution.

[0154] like Figure 5 As shown in the figure, an electronic device provided by an embodiment of the present invention includes: a processor 610, a communication interface 620, a memory 630, and a communication bus 640, wherein the processor 610, the communication interface 620, and the memory 630 communicate with each other through the communication bus 640. The processor 610 can call logical instructions in the memory 630 to execute the following method: The online data acquisition process involves collecting the initial control parameters, initial measurement indicators, batch identifiers, sub-unit identifiers, parameter constraints, and indicator specifications of the equipment or equipment sub-units. The training condition judgment step involves the agent determining whether the conditions for generating an automatic training task are met based on the sample size, coverage, failure rate, and batch drift. The model training steps involve the agent generating an automatic training task when the training conditions are met, and training, iterating, and optimizing the parameter prediction model in a training sandbox. In the strategy package generation step, the agent encapsulates the trained parameter prediction model into a parameter tuning strategy package. The candidate generation and screening steps involve generating multiple candidate parameter combinations based on the strategy package, and performing legality filtering, coupled risk assessment, and comprehensive benefit ranking on the candidate parameter combinations. The safety execution steps involve the safety actuator performing a final verification on the sorted candidate parameters, sending the verified candidate parameters to the target device, and performing a recovery operation when the candidate parameters fail to execute or trigger a risk. The data feedback step involves collecting the execution results of the candidate parameters on the target device and feeding them back. In the retraining trigger step, the agent determines whether to trigger a new round of model training based on the execution results of the reflow.

[0155] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0156] This invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, is implemented to perform the methods provided in the above embodiments.

[0157] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments or some parts of embodiments.

[0158] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for tuning intelligent agent parameters in multi-parameter coupled devices, characterized in that, Includes the following steps: The online data acquisition process involves collecting the initial control parameters, initial measurement indicators, batch identifiers, sub-unit identifiers, parameter constraints, and indicator specifications of the equipment or equipment sub-units. The training condition judgment step involves the agent determining whether the conditions for generating an automatic training task are met based on the sample size, coverage, failure rate, and batch drift. The model training steps involve the agent generating an automatic training task when the training conditions are met, and training, iterating, and optimizing the parameter prediction model in a training sandbox. In the strategy package generation step, the agent encapsulates the trained parameter prediction model into a parameter tuning strategy package. The candidate generation and screening steps involve generating multiple candidate parameter combinations based on the strategy package, and performing legality filtering, coupled risk assessment, and comprehensive benefit ranking on the candidate parameter combinations. The safety execution steps involve the safety actuator performing a final verification on the sorted candidate parameters, sending the verified candidate parameters to the target device, and performing a recovery operation when the candidate parameters fail to execute or trigger a risk. The data feedback step involves collecting the execution results of the candidate parameters on the target device and feeding them back. In the retraining trigger step, the agent determines whether to trigger a new round of model training based on the execution results of the reflow.

2. The method according to claim 1, characterized in that, The parameter tuning strategy package includes: applicable object information, data summary, model version information, candidate generator configuration, candidate verification rules, sorting target parameters, execution strategy, recovery strategy, and release gating information.

3. The method according to claim 1, characterized in that, The parameter prediction model includes one or more of the following: global parameter prediction model, local response model, and coupled risk model; The global parameter prediction model is used to predict candidate parameter combinations that will make all indicators qualified based on the initial state of the device. The local response model is used to estimate the impact of local changes in control parameters on performance indicators; The coupling risk model is used to assess whether a combination of candidate parameters will improve one metric while damaging other qualified metrics.

4. The method according to claim 3, characterized in that, In the candidate generation and screening step, the combination of multiple candidate parameters comes from one or more of the following methods: The candidates directly output by the global parameter prediction model, the similar distribution candidates of qualified equipment in the current batch, the corrected candidates derived by the local response model, the legal candidates obtained by constraint search, the risk avoidance candidates, and the fallback candidates of the rule main line.

5. The method according to claim 1, characterized in that, In the candidate generation and screening steps, the overall benefit ranking is based on a ranking score, which is determined in the following way: The ranking score is positively correlated with the expected pass probability and negatively correlated with the coupling risk score, expected execution cost, and uncertainty score. The coupling risk assessment is used to identify whether candidate parameter combinations may undermine the passed indicators. If the coupling risk of a candidate parameter combination exceeds a preset threshold, the candidate parameter combination is removed from the ranking.

6. The method according to claim 1, characterized in that, In the aforementioned secure execution steps: By default, the top-1 candidate parameter combinations with the best overall return are selected. A finite number of top-k candidate parameter combinations are allowed to be executed when the policy package allows, the trial budget is greater than zero, all candidate parameter combinations pass hard constraint verification and risk threshold verification, and the current test state allows for additional measurements.

7. The method according to claim 1, characterized in that, The recovery operation includes one or more of the following methods: rollback to safety parameters, switch to rule mainline tuning, return to the data acquisition stage, or disable the current policy package; The reasons for failure and the recovery results are recorded and fed back for model retraining.

8. The method according to claim 1, characterized in that, In the retraining triggering step, the conditions for triggering a new round of model training include one or more of the following: Failure rate increases above the first threshold, candidate validity rate decreases above the second threshold, recovery attempts increase above the third threshold, model prediction deviation exceeds the fourth threshold, batch drift exceeds the fifth threshold, or the preset cycle inspection time is reached.

9. The method according to claim 1, characterized in that, Before the training condition judgment step, if the collected data does not meet the conditions for generating an automatic training task, the system enters a first-stage online secure data collection mode. In the first-stage online secure acquisition mode, the device is calibrated by the main security rules and a limited number of exploration points are collected within the preset security budget and constraints. These exploration points are used to provide online response samples for model training.

10. A system for tuning intelligent agent parameters for multi-parameter coupled devices, used to execute the method as described in any one of claims 1 to 9, characterized in that, The system includes: The online acquisition and constraint layer is used to acquire initial state, online measurement feedback, execution results, and maintain safety constraint data. The intelligent agent automatic training orchestration layer is used for data quality diagnosis, batch drift identification, automatic training task generation, and deployment gating judgment. The model training and tuning layer is used for training, iterating, and offline validating the global parameter prediction model, the local response model, and the coupled risk model. The strategy package and candidate decision layer are used to encapsulate the training results into a parameter tuning strategy package, and generate candidate parameter combinations based on the strategy package to perform legality filtering, coupled risk assessment and comprehensive benefit ranking. The secure execution and reflow layer is used to perform final secure verification and distribution of candidate parameters, recover from execution failures, and reflow the execution results to trigger model retraining.

Images