Electroplating additive performance prediction and dynamic optimization method based on big data analysis

Abstract

The invention discloses an electroplating additive performance prediction and dynamic optimization method based on big data analysis, particularly relates to the field of electroplating process modeling, and is used for solving the problem that the performance of an electroplating additive is difficult to accurately predict and stably control under detection lag and working condition fluctuation. An electroplating production line time axis is divided according to detection or dosing time points, cycle contour features are constructed, key process operation information is extracted, and an additive performance prediction model is trained by using historical cycle contour features, additive performance characterization values and coating quality characterization values. In the production operation stage, real-time periodic contour features are input into the model to obtain a basic performance prediction result and a process adjustment prediction result, a control scheme is generated by combining a prediction credibility index, a performance target interval and a control gear, and incremental updating is performed on the model based on subsequent detection feedback; therefore, online prediction and dynamic optimization control of the performance of the electroplating additive are realized.

Description

Technical Field

[0001] This invention relates to the field of electroplating process modeling, and more specifically, to a method for predicting and dynamically optimizing the performance of electroplating additives based on big data analysis. Background Technology

[0002] In modern electroplating production, various organic additives are widely used to adjust the brightness, leveling, and pore-filling ability of the plating layer. To reduce manual testing and experience-based chemical dosing, production lines are generally equipped with analytical devices based on cyclic voltammetry, electrochemical impedance spectroscopy, and microfluidic electrochemical detection. These devices calculate the content of different additives in the plating bath through sampling, testing, and curve fitting. An automated dosing system then replenishes the plating tank according to the analysis results, forming an automated operation mode integrating detection and control. Especially in continuous production, fast-paced environments, and complex product structures, electroplating companies heavily rely on these automated analysis and dosing systems to maintain the relative stability of the plating bath. Simultaneously, the system continuously accumulates data on process parameters, production cycle time, test records, and quality results for daily traceability and process improvement, laying a data foundation for the subsequent introduction of big data and machine learning methods.

[0003] In this operating mode combining automatic detection and automatic chemical dosing, there exists a critical yet subtle technical problem: electrochemical analysis requires a complete testing and calculation process from sampling to obtaining results. However, the additives in the electroplating tank are continuously consumed during this process. Automatic chemical dosing control relies on the state of the plating solution at the time of detection, but the actual state within the tank has already changed by the time chemical dosing is executed, creating a temporal misalignment between the detection results and the controlled object. With changes in production load, plate type, and process conditions, this misalignment manifests as the additive concentration fluctuating within the allowable range. The automatic chemical dosing action and changes in plating performance are difficult to synchronize in a timely manner, easily leading to situations where brightness, pore-filling ability, or stress performance deviates significantly at certain stages, while the detection records still show essentially normal results. Existing technologies for improving such systems mostly focus on increasing the accuracy of a single detection, shortening the time of a single detection, or adjusting the dosing threshold. They do not truly utilize the process and quality data accumulated over a long period on the production line to predict the state and performance of electroplating additives in the future and optimize dosing and process parameters accordingly. A method for predicting and dynamically optimizing the performance of electroplating additives based on big data analysis has not yet been developed, making it difficult to fundamentally alleviate the stability and consistency problems caused by the aforementioned detection lag and control misalignment.

[0004] To address the aforementioned problems, a technical solution is provided. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a method for predicting and dynamically optimizing the performance of electroplating additives based on big data analysis. This method divides the electroplating production line timeline by dividing it into detection or dosing points, constructs periodic profile features, and extracts key process operation information. It trains an additive performance prediction model using historical periodic profile features, additive performance characterization values, and coating quality characterization values. During production operation, real-time periodic profile features are input into the model to obtain basic performance prediction results and process adjustment prediction results. A control scheme is generated by combining the prediction confidence index, performance target range, and control level. The model is incrementally updated based on subsequent detection feedback, thereby achieving online prediction and dynamic optimization control of electroplating additive performance, thus solving the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: S1: The electroplating production line time axis is divided into multiple control cycles based on the detection or chemical addition time point. Within each control cycle, the process operation data and detection results are normalized and differentially processed according to the time sequence to obtain the cycle contour feature sequence. S2: Establish a correlation between the periodic profile feature sequence of each control cycle and the additive performance characterization value and coating quality characterization value at the beginning of the adjacent control cycle, and construct a training sample set with the evolution of the periodic profile as input and the trend of additive performance change as output in chronological order. S3: Train the additive performance prediction model based on the training sample set, so that when the model receives the cycle profile feature sequence of the current control cycle and the historical control cycle, it outputs the basic additive performance prediction value at the predetermined control time and the process adjustment prediction value for compensating for the detection lag effect. S4: During production, input the cycle profile feature sequence of the current control cycle into the model to obtain the prediction result. Determine the correction range of the prediction result based on the prediction result of the most recent control cycle and the time relationship of the process response. Generate a control scheme based on the corrected prediction result and execute it. Update the model based on the execution result.

[0007] Furthermore, step S1 includes the following: The electroplating production line time axis is divided into multiple control cycles according to the detection time point and the chemical addition time point. Control cycles with a time length lower than the preset lower threshold are merged with adjacent control cycles. Within each control cycle, interpolated sampling values ​​are obtained from the production load data sequence and the detection data sequence based on a unified relative time sampling point.

[0008] Furthermore, step S1 also includes the following: Within each control cycle, thresholds for production load variation range and detection variation range are set based on the historical noise range and the effective variation range. Amplitude normalization and adjacent difference are performed on the production load sampling sequence and the detection sampling sequence. The four results at each relative time sampling point are combined into a periodic profile feature sequence.

[0009] Furthermore, step S2 includes the following: Based on the obtained cycle profile feature sequence of each control cycle, the additive performance characterization value is calculated using the sorted median value method within the detection time sub-interval near the start time of the corresponding control cycle. The coating quality characterization value is calculated using the sorted median value method within the quality detection time window corresponding to the start time. A one-to-one correspondence is established between the additive performance characterization value and the coating quality characterization value and the control cycle index.

[0010] Furthermore, step S2 also includes the following: The range of span period numbers is determined based on the historical additive performance characterization value sequence. The span period number is selected based on the length of the monotonic performance change interval. The cross-cycle performance change trend value of each starting control cycle is calculated on the control cycle sequence according to the span period number. The cycle profile feature sequence of multiple consecutive control cycles, the corresponding performance change trend value, and the coating quality characterization values ​​at the start and end points are combined to form a training sample set.

[0011] Furthermore, step S3 includes the following: An additive performance prediction model containing convolutional coding mapping and mid-segment aggregation structure is constructed using a training sample set. The set of periodic contour feature sequences of continuous control cycles is encoded into a fixed-dimensional intermediate representation vector. The performance trend output branch generates cross-cycle additive performance change trend prediction values ​​based on the intermediate representation vector, and the process adjustment output branch generates process adjustment prediction values ​​based on the intermediate representation vector.

[0012] Furthermore, step S3 also includes the following: On each training sample, the error between the predicted value of performance change trend and the value of cross-cycle additive performance change trend, as well as the error between the predicted value of process adjustment and the reference value of process adjustment, are calculated. Two local losses are constructed by hyperbolic tangent function and absolute value. In a single training sample, the larger of the two local losses is taken as the comprehensive loss. The comprehensive loss is summed on all training samples as the overall loss. All parameters of the additive performance prediction model are updated by gradient-based iterative optimization method.

[0013] Furthermore, step S4 includes the following: At the end of each control cycle, a prediction window containing several control cycles spanning the cycle is constructed. The set of cycle contour feature sequences within the prediction window is input into the additive performance prediction model. Based on the additive performance characterization value obtained from the most recent real detection and the historical single-cycle average performance change prediction, the reference performance value of the starting control cycle of the prediction window and the basic additive performance prediction value of the target control cycle are calculated.

[0014] Furthermore, step S4 also includes the following: Based on the error ratio between the observed performance change and the predicted value of the additive performance change trend across cycles, a prediction confidence index and a prediction confidence threshold are constructed. Combined with the performance target range determined by the stable operation data and the control level obtained by mapping the predicted value of process adjustment, a dosing and process parameter adjustment scheme for the target control cycle is generated. After obtaining the test results covering the span of cycles, the error record and additive performance prediction model parameters are updated.

[0015] The technical effects and advantages of the electroplating additive performance prediction and dynamic optimization method based on big data analysis of this invention are as follows: By uniformly collecting and organizing production load, test results, and chemical dosing behavior on a control cycle scale, this invention transforms the originally discrete and scattered process operation records into continuous cycle profile features. Combined with cross-cycle additive performance characterization values ​​and coating quality characterization values, the model simultaneously reflects the correspondence between load rhythm, consumption rhythm, and quality response. This enables the production line to make forward-looking judgments on the trend of additive performance changes over a future period, even under conditions of limited detection frequency and time lag in detection. This fundamentally alleviates the performance distortion and control timing deviation problems caused by traditional reliance on single-point detection and experience-based judgment.

[0016] By introducing the collaborative design of process adjustment prediction output, prediction confidence index, performance target range, and control level into the additive performance prediction model, this invention no longer simply converts the prediction results directly into the dosage. Instead, it establishes a unified criterion between prediction reliability, target performance range, and historical adjustment amplitude statistics, and utilizes the prediction results in a graded manner. When the performance deviation trend is clear and the prediction confidence is sufficient, it actively provides a dosage or reduction plan. When the prediction is uncertain or the performance is close to the target range, it adopts a more conservative adjustment. This allows the control behavior to respond promptly to performance degradation while reducing unnecessary fluctuations and overdosing, thereby improving the stability of coating quality and reducing additive consumption.

[0017] By continuously recording the deviation between predictions and actual detections during production operation, and using new observed performance changes and coating quality changes to incrementally correct process adjustment references and model parameters, this invention enables the parameters and prediction reliability evaluation of the additive performance prediction model to gradually converge to the current process state as the production line conditions change over a long period. When product structure, load rhythm, or formulation strategy changes, it is not necessary to re-establish rules or completely remodel; the prediction and control logic can be automatically adjusted through online learning, achieving adaptability and sustainability of electroplating additive performance prediction and dynamic optimization control throughout the entire life cycle. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the electroplating additive performance prediction and dynamic optimization method based on big data analysis according to the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1: Figure 1 This invention presents a method for predicting and dynamically optimizing the performance of electroplating additives based on big data analysis, including: S1: The electroplating production line time axis is divided into multiple control cycles based on the detection or chemical addition time point. Within each control cycle, the process operation data and detection results are normalized and differentially processed in chronological order to obtain the cycle profile feature sequence.

[0021] S2: Establish a correlation between the periodic profile feature sequence of each control cycle and the additive performance characterization value and coating quality characterization value at the beginning of the adjacent control cycle, and construct a training sample set with the evolution of the periodic profile as input and the trend of additive performance change as output in chronological order.

[0022] S3: Train the additive performance prediction model based on the training sample set, so that when the model receives the cycle profile feature sequence of the current control cycle and the historical control cycle, it outputs the basic additive performance prediction value at the predetermined control time and the process adjustment prediction value for compensating for the detection lag effect.

[0023] S4: During production, input the cycle profile feature sequence of the current control cycle into the model to obtain the prediction result. Determine the correction range of the prediction result based on the prediction result of the most recent control cycle and the time relationship of the process response. Generate a control scheme based on the corrected prediction result and execute it. Update the model based on the execution result.

[0024] In the actual operation of the electroplating production line, detection and chemical dosing actions occur sequentially along the time axis as discrete events, while changes in production load and additive consumption evolve continuously. Subsequent steps require constructing training samples on a control cycle basis. Therefore, in step S1, the time axis is first divided according to detection and chemical dosing events. Then, within each control cycle, a rhythmic feature sequence with a unified structure is extracted from the production load data and detection data. This rhythmic feature sequence needs to simultaneously characterize the production load level, detection level, and how they change over time, providing a consistent input format for constructing cross-cycle additive performance change samples in step S2.

[0025] 1.1 Control cycle division.

[0026] Record the occurrence times of all detection events and all chemical dosing events during the operation of the electroplating production line, and arrange them in chronological order to form an event time sequence. The time interval between any two adjacent event times is defined as a control cycle, corresponding to a number. Each control cycle has a start time and an end time, and the difference between the two is the length of the control cycle.

[0027] To ensure that each control cycle contains a sufficient amount of sampled data, a lower limit threshold for the control cycle length needs to be set. This lower limit threshold is limited to two time scales. The first time scale is derived from the sampling interval of the sampling device and the maximum number of samples under no-load operating conditions, corresponding to the upper bound of the time span when no load fluctuations occur. The second time scale is derived from the shortest dwell time of a single batch of products through critical processes in process engineering experience, corresponding to the shortest time span that can cover a complete load change process. By statistically analyzing the time intervals between a large number of historical events, a distribution of event time intervals is formed. The interval greater than the aforementioned upper bound of the time span but not exceeding the aforementioned shortest time span is selected as the allowable range for the lower limit threshold of the control cycle length, and the specific threshold is determined within this range.

[0028] If the duration of a control cycle is less than the lower limit threshold for control cycle duration, it is merged with the adjacent control cycle. If a control cycle precedes it, it is merged with the preceding control cycle, and the start time of the preceding control cycle and the end time of the current control cycle constitute a new control cycle time period. If the current control cycle is the first segment and its duration is less than the lower limit threshold for control cycle duration, it is merged with the following control cycle, and the start time of the current control cycle and the end time of the following control cycle constitute a new control cycle time period. After merging, the duration of the merged control cycle is recalculated. After the above processing, the time axis is divided into several control cycles, and the duration of each control cycle is not less than the preset lower limit threshold for duration.

[0029] 1.2 Extraction of raw data sequences.

[0030] Two types of data that change over time are collected independently during production line operation. The first type of data is production load data, which represents the intensity of production tasks undertaken by the production line per unit time. This can be expressed as the number of plates, processing area, or other indicators reflecting production intensity, and is recorded according to a fixed sampling period. The second type of data is detection data, which represents the detection results directly related to the state of additives, such as electrochemical detection values, potential measurement results, or online analysis concentration results, and is also recorded according to their respective sampling periods.

[0031] After data collection, two time series are generated: one for production load data and one for monitoring data. The sampling points of the two time series are generally not completely consistent. Therefore, within each control cycle, the values ​​of both time series need to be mapped to the same set of time positions to provide a foundation for constructing a feature sequence with a unified structure.

[0032] 1.3 Construction of relative time grid within the control cycle.

[0033] Within any given control cycle, taking the start time of that control cycle as zero and the length of the control cycle as the scale, the absolute time within the control cycle is converted into a relative time between zero and one. To use a uniform time position for feature sampling across all control cycles, several fixed relative time points are pre-selected within the zero-to-one interval. The first relative time point corresponds to the start time of the control cycle, the last relative time point corresponds to the end time of the control cycle, and the remaining relative time points are arranged in between, in a sequential order.

[0034] The number of relative time sampling points needs to meet two conditions. Firstly, each significant change in production load or detection result should be characterized between two adjacent relative time sampling points, preventing compression into a single sampling point and loss of shape information. Secondly, the number of relative time sampling points should not be excessive, as this would lead to an overly long periodic contour feature sequence, burdening subsequent training and inference computations. By sampling and analyzing the production load change curves and detection change curves within historical control cycles, a suitable number of relative time sampling points was selected, meeting both conditions, and this number and location distribution were uniformly adopted across all control cycles. Each relative time sampling point corresponds one-to-one with the absolute time corresponding to the start time and duration of its respective control cycle, thus determining a fixed set of absolute sampling time points within each control cycle.

[0035] 1.4 Data interpolation and sampled value generation within the control cycle.

[0036] Within each control cycle, for each of the aforementioned absolute sampling time points, the corresponding production load value is obtained from the production load data time series. If there is a time in the original sampling time that is exactly the same as the absolute sampling time point, the production load record for that time is directly used. If there is no time in the original sampling time that is exactly the same, interpolation is performed between the two sampling times immediately before and after the absolute time point. The interpolation method can be linear interpolation, that is, based on the time and corresponding production load values ​​of two adjacent sampling points, the production load value of the absolute sampling time point is calculated proportionally on the time axis. In this way, a production load sampling value is generated at each relative time sampling point within the control cycle, forming a standardized production load sampling sequence within the control cycle.

[0037] The interpolation process for the detection data is consistent with that for the production load data. Within each control cycle, for each absolute sampling time point, if a detection value exists in the time series of the detection data at the same moment, that detection value is directly used; otherwise, linear interpolation is performed between the two detection sampling times immediately before and after that time point to calculate the detection value corresponding to that absolute time point. In this way, a detection sampling value is formed at each relative time sampling point within the control cycle, constituting a standardized detection sampling sequence within the control cycle.

[0038] After interpolation and resampling, each control cycle has a production load sampling sequence and a detection sampling sequence of the same length. Each position in the sequence corresponds one-to-one with a unified relative time sampling point, providing a consistent data structure for subsequent normalization and rhythm difference calculation.

[0039] 1.5 Normalization of production load data amplitude within the control cycle.

[0040] Within each control cycle, the maximum and minimum production load values ​​are identified from the production load sampling sequence. The difference between these two values ​​is taken as the production load variation range. The production load variation range reflects the fluctuation amplitude of the load curve on the vertical axis within the control cycle. If the fluctuation amplitude is too small, it often only contains measurement noise and does not represent the actual operating condition change. Therefore, it is necessary to introduce a production load variation range threshold to identify effective changes and noise changes.

[0041] The threshold range for production load variation is determined based on historical data. First, time segments in the historical data are selected where the production line is either idle or under stable process conditions with a relatively constant production load. The difference between the maximum and minimum production load within each time segment is calculated. These differences primarily reflect the impact of measurement noise on the production load record. The maximum value among these differences is then selected as the upper limit of the noise variation range. Second, time segments in the historical data containing significant production load switching behaviors, such as shift changes, capacity changes, and product changes, are selected. The difference between the maximum and minimum production load within each time segment is calculated, resulting in a set of variation range values. The minimum value among these is then selected as the lower limit of the effective production load variation range. The allowable range for the production load variation threshold is limited to between the upper limit of the noise variation range and the lower limit of the effective variation range. The specific threshold is selected by process engineers within this range.

[0042] For a given control cycle, if the production load variation range within that control cycle is less than the aforementioned threshold, the production load curve within that control cycle is considered not to contain sufficient true load variation. During normalization, the normalized production load result for the entire control cycle is directly set to a zero-value sequence. For control cycles where the production load variation range is not less than the aforementioned threshold, amplitude normalization is performed on the sampled production load values ​​within that control cycle. The normalization method uses the average of the maximum and minimum values ​​as a reference zero point. Each sampled production load value is subtracted from this reference zero point, and then a division operation is performed using the production load variation range as the normalization coefficient to obtain a dimensionless normalized production load sequence. After this processing, the production load curves within different control cycles are comparable in terms of vertical dimensions and unfold symmetrically around zero.

[0043] 1.6 Normalization of the amplitude of detection data within the control cycle.

[0044] Within each control cycle, the maximum and minimum detection values ​​are identified from the detection sampling sequence. The difference between these two values ​​is taken as the range of detection data variation. This range measures the degree of fluctuation in the detection quantity within the control cycle, and it is also necessary to distinguish between noise-driven fluctuations and actual changes in the additive's state.

[0045] The threshold range for the variation range of detection data is also determined using historical data. First, time segments where the additive is in a long-term stable state without replenishment or formulation adjustments are selected. Within each time segment, the difference between the maximum and minimum detection values ​​is calculated, and the maximum value from this set of differences is selected as the upper limit of the detection noise range. Second, time segments where the additive exhibits significant performance changes are selected, such as periods when additives are replenished or when the additive is nearing its expiration date. Similarly, the difference between the maximum and minimum detection values ​​is calculated within each time segment, and the minimum value from this set of differences is selected as the lower limit of the effective detection variation range. The upper limit of the detection noise range and the lower limit of the effective detection variation range form an interval, limiting the threshold value for the variation range of detection data to within this interval. The specific threshold is selected by process engineers based on experience.

[0046] When the variation range of the detection data within a certain control cycle is less than the aforementioned threshold, it is considered that the detection data within that control cycle does not reflect a significant change in the additive state. During normalization, all detection sample values ​​within that control cycle are normalized to a zero-value sequence. When the variation range of the detection data is not less than the aforementioned threshold, the detection sample values ​​within that control cycle undergo amplitude normalization in the same form as the production load normalization operation. Specifically, the average of the maximum and minimum detection values ​​is used as the reference zero point. Each detection sample value is subtracted from the reference zero point, and a division operation is performed using the variation range of the detection data as the normalization coefficient, resulting in a dimensionless detection normalized sequence. After this processing, the detection data can be compared between different control cycles through shape and sign distribution.

[0047] 1.7 Rhythm difference calculation and periodic profile feature construction.

[0048] After completing the production load normalization and detection normalization, to reflect the rhythmic information of time-varying changes within the control cycle, it is necessary to perform adjacent difference calculations on the normalized sequence. Specifically, within the control cycle, starting from the second relative time sampling point, the normalized production load value at each sampling point is subtracted from the normalized production load value at the previous sampling point to obtain the adjacent difference value of the production load; the difference value of the first sampling point is set to zero since there is no previous sampling point. Similarly, within the control cycle, the same operation is performed on the detection normalization sequence. Starting from the second relative time sampling point, the normalized detection value at each sampling point is subtracted from the normalized detection value at the previous sampling point to obtain the adjacent difference value of the detection; the difference value of the first sampling point is set to zero.

[0049] At each relative time sampling point within the control cycle, four pieces of information at that location are combined: the normalized production load value, the adjacent difference in production load, the normalized detection value, and the adjacent difference in detection, forming a four-dimensional feature vector. Arranged sequentially along the relative time sampling points, all feature vectors within the control cycle are connected into a feature sequence. This feature sequence is the periodic profile feature sequence of the control cycle, used to describe the common rhythm of production load changes and detection changes within that control cycle.

[0050] All control cycles generate cycle profile feature sequences using the method described above, and maintain a one-to-one correspondence with the control cycle number, providing a unified input format for cross-cycle correlation of additive performance and coating quality in subsequent steps.

[0051] Step S1 divides the electroplating production line into time axes by detection time points and chemical addition time points, and constructs a unified relative time grid and normalized differential features within each control cycle. This ensures that subsequent modeling uses a cyclical contour feature sequence that reflects changes in production load and the rhythm of additive detection response, rather than scattered and isolated instantaneous detection points. This allows the training samples to truly reflect the load evolution and additive consumption process over a period of time before the detection action, achieving a data expression method that matches the contradiction in the background technology where detection results lag behind the actual plating solution state. This provides a structurally consistent and information-rich input foundation for subsequent prediction and dynamic optimization based on control cycles.

[0052] Step S1 has divided the time axis into multiple control cycles according to the detection time point and the dosing time point, and generated a cycle profile feature sequence within each control cycle. The cycle profile feature sequence records the normalized value of production load, the change in production load, the normalized value of detection, and the change in detection on a unified relative time grid, which is used to characterize the dynamic behavior within a single control cycle. However, the decay and recovery of additive performance usually spans multiple control cycles. Using the profile information of a single control cycle alone cannot reflect the overall direction of performance change over a period of time. It is necessary to establish a mapping relationship from the evolution of the cycle profile to the trend of additive performance change between control cycles. Therefore, step S2 constructs a training sample set around the cycle profile feature sequence, the additive performance characterization value, and the coating quality characterization value, so that the subsequent model can learn the performance change pattern across cycle scales.

[0053] 2.1 Calculation logic for additive performance characterization values.

[0054] Within each control cycle, the start time has been determined by step S1. To obtain representative additive performance characteristics near the start time of the current control cycle, a start detection time sub-interval is constructed within each control cycle. The start time of the start detection time sub-interval is the start time of the control cycle, and the end time is the start time of the control cycle plus an start detection time length.

[0055] The initial detection time is determined based on historical detection response data and the control cycle length. First, the time required from the start of detection to the point where the detected value reaches a stable state is statistically analyzed from historical detection response curves. An upper limit time is selected from the statistical values ​​of multiple detection curves as the upper limit of the detection response stabilization time. Second, the product of the control cycle length and a preset proportional coefficient is taken as another time length. This proportional coefficient, ranging from zero to one, is used to limit the initial detection time sub-interval from covering the entire control cycle. Finally, the initial detection time length is selected from the upper limit of the detection response stabilization time and the product mentioned above, with the selected value not exceeding the control cycle length.

[0056] Within each control cycle, based on the initial detection time sub-interval, all detection data falling within that sub-interval are extracted from the detection data time series to form a set of initial detection values. These initial detection values ​​are sorted in ascending order. When the number of initial detection values ​​is odd, the middle value after sorting is taken as the additive performance characterization value for this control cycle. When the number of initial detection values ​​is even, the arithmetic mean of the two middle values ​​after sorting is taken as the additive performance characterization value for this control cycle. Through this process, each control cycle index corresponds to a unique additive performance characterization value. This value has a suppressive effect on occasional abnormal detection points and is constructed around the start time of the control cycle.

[0057] 2.2 Calculation logic for coating quality characterization values.

[0058] Electroplated products experience a fixed or nearly fixed transmission time from entering the electroplating station to completing electroplating and undergoing quality inspection. Therefore, the plating quality inspection results corresponding to the start time of the control cycle need to be traced back along the timeline. To obtain a set of plating quality inspection values ​​that match the start time of the control cycle, a quality inspection time window is constructed on the timeline for each control cycle. The end point of the quality inspection time window is the control cycle start time minus the average transmission time, and the start point is the end point minus the quality inspection window length.

[0059] The average transfer time is determined using historical production data. Specifically, multiple batches of products are selected, and the time intervals from entering the electroplating station to completing quality inspection are statistically analyzed. The average or median value of these time intervals is taken as the average transfer time. The quality inspection window length is obtained by analyzing the time required for a batch of products to continuously complete quality inspection. The shortest time span that can completely cover the inspection process of a batch of products is selected from the inspection time spans of multiple batches as the quality inspection window length.

[0060] Within the quality inspection time window corresponding to each control cycle, all coating quality inspection values ​​are collected to form a set of quality inspection values. The quality inspection values ​​are sorted in ascending order. When the number of quality inspection values ​​is odd, the middle value after sorting is taken as the coating quality characterization value for that control cycle. When the number of quality inspection values ​​is even, the arithmetic mean of the two middle quality inspection values ​​after sorting is taken as the coating quality characterization value for that control cycle. Through this process, each control cycle index is associated with an additive performance characterization value and a coating quality characterization value, which respectively describe the additive state and the finished product quality state at the start of the control cycle.

[0061] 2.3 Logic for selecting the number of span cycles.

[0062] To describe the performance change trend of additives across multiple control periods, a specific span period number needs to be selected in the control period sequence. The span period number represents the number of periods spanned between the starting and ending control periods when calculating the performance change trend.

[0063] The number of control cycles is determined using a historical sequence of additive performance characterization values. First, a continuous sequence of control cycles is selected from the historical operating data. The additive performance characterization values ​​corresponding to each control cycle are arranged chronologically to form a performance sequence. Then, monotonic performance variation intervals are identified within the performance sequence. The identification principle is as follows: when the performance characterization value of a subsequent control cycle is consistently greater than that of the previous control cycle, it is determined to be a monotonic performance increase interval; when the performance characterization value of a subsequent control cycle is consistently less than that of the previous control cycle, it is determined to be a monotonic performance decrease interval. When several consecutive cycles meet the above conditions, this continuous cycle segment is numbered as a monotonic performance variation interval.

[0064] For each monotonic performance variation interval, the number of control cycles contained within that interval is counted, resulting in a set of performance variation interval length values. The minimum value among these length values ​​is taken as the lower limit for the number of cycles spanned by a single performance variation process, and the maximum value is taken as the upper limit for the number of cycles spanned by a complete performance variation process. The arithmetic mean of the lower and upper limits is calculated, and the integer obtained by rounding down the arithmetic mean is taken as the number of cycles spanned. Multiple different integers can also be selected as needed and used in different model training stages. The value of the number of cycles spanned should be sufficient to cover a significant performance variation process without spanning excessively long time periods that would cause the trend information to become overly smoothed.

[0065] 2.4 Calculation logic for cross-cycle performance change trend value.

[0066] After determining the number of span periods, the performance change trend value of the additive across periods can be calculated on the control period index sequence. For each control period index, as long as there is a control period with an interval equal to the number of span periods following it, a combination of a starting control period and an ending control period is constructed. The difference between the additive performance characterization value of the starting control period and the additive performance characterization value of the ending control period is used as the performance change trend value of this combination. A positive difference indicates that the additive performance has improved after several control periods, a negative difference indicates that the additive performance has decreased, and a zero difference indicates that the performance level remains basically unchanged within the span period.

[0067] To facilitate the subsequent use of this trend value to distinguish between noise-driven minor fluctuations and actual performance changes, a performance change trend threshold needs to be determined from historical data. The range of the performance change trend threshold is given through the following steps. First, select time segments in the historical data where the process formulation, production load, and operating conditions remain stable. Within these time segments, calculate the absolute values ​​of the performance change trend values ​​corresponding to all possible starting control cycles and spanning cycle numbers. Take the maximum value from these absolute values ​​as the upper limit of noise change. Second, select time segments in the historical data where there is a clear addition of additives or significant additive failure. Within these time segments, calculate the absolute values ​​of the performance change trend values ​​corresponding to all possible starting control cycles and spanning cycle numbers. Take the minimum value from these absolute values ​​as the lower limit of effective performance change. The allowable range of the performance change trend threshold is limited to the above-mentioned upper limit of noise change and lower limit of effective performance change. Within this range, process engineers select a specific threshold based on experience.

[0068] For each combination of start-point and end-point control cycles, when the absolute value of the performance change trend value is less than the performance change trend threshold, this combination is marked as a sample with approximately stable performance in the training data. When the absolute value of the performance change trend value is greater than or equal to the performance change trend threshold, this combination is marked as a sample with significant performance change. The sign and magnitude of the performance change trend value are retained for supervising model learning.

[0069] 2.5 Construction of training samples for periodic contour features and performance trends.

[0070] In step S1, there is a one-to-one correspondence between the control cycle index and the cycle contour feature sequence, with each control cycle having a fixed-length cycle contour feature sequence. In step S2, to construct training samples that can characterize cross-cycle performance change behavior, it is necessary to combine the cycle contour feature sequences of multiple consecutive control cycles into an input segment and establish a mapping relationship with the corresponding cross-cycle performance change trend value.

[0071] For each starting control cycle index that satisfies the condition of having several span cycles following the starting control cycle, a training sample is constructed. The input of the training sample consists of several cycle contour feature sequences arranged chronologically, starting from the starting control cycle, with the number of sequences equal to the number of span cycles. The output of the training sample is the performance change trend value between the starting control cycle and the span end control cycle. Furthermore, the coating quality characterization values ​​for both the starting control cycle and the span end control cycle are recorded in the training sample to facilitate examining the correspondence between performance change trends and quality changes during model training or subsequent analysis.

[0072] By traversing all control cycle indices, a training sample is constructed as long as there are several control cycles with a complete span following the starting control cycle, forming a training sample set. Each sample in the training sample set contains a cycle contour feature sequence of multiple consecutive control cycles, the corresponding cross-cycle performance change trend value, and the coating quality characterization value at the starting and ending points, providing sufficient data support for training the additive performance prediction model in step S3.

[0073] Step S2 extracts additive performance characterization values ​​and coating quality characterization values ​​near the start time of each control cycle, calculates the cross-cycle performance change trend by combining the number of span cycles, and pairs the continuous cycle profile feature sequence with the performance change trend to form a training sample set. This allows the data to directly reflect the cumulative effect of load and detection rhythm on multiple control cycles and the direction of additive performance decay or recovery, supporting the prediction model to learn around the problems of detection lag and rapid additive consumption, and reducing the impact of single detection error on control decisions.

[0074] Step S2 calculates the additive performance characterization value and coating quality characterization value for each control cycle. It constructs a cross-cycle additive performance change trend value based on the number of cycles spanned, and organizes this into a training sample set. The input of the training samples is a set of cycle profile feature sequences from multiple consecutive control cycles, and the output is the cross-cycle performance change trend value and the coating quality characterization values ​​at the start and end points. However, predicting the additive performance change trend solely through a single output is insufficient to reflect the actual needs of process adjustment under detection lag, and it is also difficult to demonstrate the coupling relationship between performance changes and coating quality changes. Therefore, in step S3, it is necessary to construct an additive performance prediction model containing a robust coding part and dual output branches based on the training sample set formed in step S2. Through targeted loss design, the model learns not only the mapping relationship between cycle profiles and performance trends, but also the adjustment direction and magnitude when performance changes have already affected coating quality, providing directly callable prediction and adjustment results for dynamic optimization control.

[0075] 3.1 Confirmation of training sample structure.

[0076] In step S2, a training sample set has been formed. Each training sample in the set is numbered starting from a certain control cycle. For any training sample, its input part is composed of a chronologically connected sequence of periodic contour feature sequences spanning several consecutive control cycles. Specifically, starting from the starting control cycle, the chronological contour feature sequences of the starting control cycle and several subsequent control cycles are taken sequentially until the number of spanning cycles is exhausted, thus obtaining a set of chronologically arranged chronological contour feature sequences.

[0077] The output of the training samples contains three types of numerical information. The first type is the trend value of additive performance changes across cycles, representing the additive performance characterization value at the end of the control cycle minus the additive performance characterization value at the beginning of the control cycle. This characterizes the overall upward or downward change in additive performance within the time range corresponding to the number of control cycles. The second type is coating quality information, including the coating quality characterization values ​​at the beginning and end of the control cycle, used to describe the changes in product quality within the same span. The third type is the process adjustment reference quantity constructed in step S3 based on the performance change trend and coating quality changes, used to constrain the process adjustment output branch during the training phase.

[0078] In this way, each training sample contains an input feature set, cross-cycle additive performance change trend values, starting and ending coating quality characterization values, and process adjustment reference values. The training sample set consists of all samples with feasible starting control cycles, providing a unified data foundation for subsequent model training.

[0079] 3.2 Construction of reference quantities for process adjustment.

[0080] During the training phase, a target quantity is needed to represent the adjustment requirements, so that the process adjustment output branch, during training, not only considers the performance change itself, but also the actual impact of the performance change on the coating quality. Therefore, a process adjustment reference quantity is constructed for each training sample.

[0081] First, the coating quality change is calculated on the training samples. The coating quality change is defined as the coating quality characterization value at the end of the span control period minus the coating quality characterization value at the beginning of the span control period. A positive value indicates that the overall quality level increases during the span, a negative value indicates that the overall quality level decreases, and a zero value indicates that the quality level remains basically unchanged during the span.

[0082] Secondly, a threshold value is set for judging the change in coating quality. The threshold range is given by historical operating data. A type of time segment is selected where process parameters, additive status, and production load remain stable. From these time segments, the absolute value of the coating quality change is calculated for all feasible starting point control cycles and span cycle combinations. The maximum value is taken as the upper bound of the quality change amplitude caused by detection noise and natural fluctuations. Another type of time segment is selected where the production process exhibits persistent quality deviations, such as consecutive over-plating, consecutive under-plating, or a significantly high defect rate across multiple control cycles. From these time segments, the absolute value of the coating quality change is calculated for all feasible starting point control cycles and span cycle combinations. The minimum value is taken as the lower bound of the effective quality change. The coating quality change threshold is limited to between the upper bound of noise change and the lower bound of effective quality change. Within this numerical range, a specific value is selected empirically as a benchmark for judging whether quality changes require attention during training.

[0083] The threshold for the trend of additive performance change has been given in a similar manner in step S2, and is used uniformly to distinguish between performance fluctuations caused by detection noise and actual performance changes.

[0084] After obtaining the additive performance change trend value, coating quality change amount, and corresponding threshold, a process adjustment reference value is constructed for each training sample. The construction rule consists of three conditions. First, when the absolute value of the additive performance change trend value across cycles is not less than the additive performance change trend threshold, the performance change is considered to have reached a level requiring attention. Second, when the absolute value of the coating quality change amount is not less than the coating quality change threshold, the quality change is considered to have also reached a level requiring attention. Third, when the additive performance change trend value and the coating quality change amount have the same sign, the performance change direction is considered to be consistent with the quality change direction, indicating that the performance change has had a positive impact on product quality.

[0085] Only when all three conditions mentioned above are met is the process adjustment reference value set to the additive performance change trend value, inheriting both the direction and magnitude of the performance change. If any condition is not met, the process adjustment reference value is set to zero. Through this construction, the training samples only provide a non-zero adjustment reference value to the model when the cross-cycle additive performance change trend and coating quality change jointly indicate a stable and significant shift. This allows the process adjustment output branch to focus on learning the scenarios that truly require adjustment, while keeping the output close to zero in scenarios dominated by noise or with inconsistent directions.

[0086] 3.3 Structure of the additive performance prediction model.

[0087] The additive performance prediction model takes a set of periodic profile feature sequences from continuous control cycles as input and outputs two quantities: a predicted value of the cross-cycle additive performance change trend and a predicted value of process adjustment. The model consists of an encoding mapping part and two output branches.

[0088] The encoding mapping section transforms the input set of periodic contour feature sequences into a fixed-dimensional intermediate representation vector. First, the periodic contour feature sequences of all control cycles within the span are concatenated sequentially from front to back according to the chronological order of the control cycles on the time axis, resulting in a long sequence. Each position in the long sequence contains four components: the normalized production load value, the adjacent difference value of the production load, the normalized detection value, and the adjacent difference value of the detection. The total length of the long sequence is equal to the product of the span number of cycles and the number of relative time sampling points within a single control cycle.

[0089] After obtaining the long sequence, several convolutional kernels are set along the time direction, each covering a fixed number of consecutive time positions. Each convolutional kernel is associated with a set of convolutional coefficients and a bias value. During convolution calculation, at each allowed time position, the four components of each time position within the coverage of the convolutional kernel are multiplied by their corresponding convolutional coefficients, summed, and then added to the bias value before being input into a piecewise linear activation function. The piecewise linear activation function outputs zero when the input is negative and outputs the input itself when the input is non-negative, thus obtaining the output value of the convolutional kernel at that time position. By performing sliding convolutions along the entire time direction, a time output sequence corresponding to each convolutional kernel can be obtained. The number of convolutional kernels and the coverage width of the convolutional kernels in the time direction are selected from a preset set of integers. By training the model separately on independent validation data and comparing the final loss function values, the combination with the smallest loss function value is selected as the final configuration.

[0090] To robustly aggregate convolutional outputs over time, each convolutional kernel is processed separately. For a given convolutional kernel, the absolute values ​​of its outputs at each time point are taken to form a non-negative sequence, which is then sorted in ascending order. After sorting, a dropout ratio parameter is introduced to mitigate the impact of occasional spikes on the aggregation result. The dropout ratio parameter is selected between zero and half, with the specific value determined by training the model under multiple candidate ratios and comparing the convergence results of the validation set loss function. Based on the dropout ratio parameter, a portion of the minimum values ​​are removed from the beginning of the sorted sequence, and a portion of the maximum values ​​are removed from the end of the sorted sequence. The arithmetic mean is calculated on the remaining portion, and this average is used as the aggregation response of the convolutional kernel to the current training sample. The above operation is repeated for all convolutional kernels, and the aggregation responses of each convolutional kernel are grouped into a fixed-length vector according to the kernel number order, serving as an intermediate representation vector. The resulting intermediate representation vector centrally reflects the rhythmic features of the periodic contour feature sequence within a span of periods at different time scales and in different combination directions, and the impact of extreme local anomalies is reduced through the dropout logic at both ends.

[0091] The performance trend output branch takes the intermediate representation vector as input and outputs a predicted value for the cross-cycle additive performance change trend. The performance trend output branch first maps the intermediate representation vector to a hidden vector through a linear transformation and piecewise linear activation. The linear transformation is performed using a set of first-layer weight matrices and a set of performance branch bias vectors. The resulting linear combination is processed component-by-component by a piecewise linear activation function to form the performance branch hidden layer vector. The dimensions of the performance branch hidden layer are selected from multiple candidate integers. The model is trained under different dimension configurations, and the performance change trend prediction errors are compared on validation data. The dimension with the smallest error is selected as the final configuration. Subsequently, the performance branch hidden layer vector is multiplied by the performance branch output weight vector, and the performance branch output bias scalar is added to obtain a scalar result, which serves as the predicted value for the cross-cycle additive performance change trend.

[0092] The process adjustment output branch is also constructed based on the intermediate representation vector, but it undergoes expansion processing at the input. Specifically, the intermediate representation vector itself is used as the first half, and the absolute values ​​of its components are arranged in the same order as the second half, concatenating them to form an expanded vector twice the original length. This expanded vector contains both the sign and magnitude information of the convolutional kernel responses. Subsequently, a linear transformation is performed on the expanded vector using a set of first-layer weight matrices and a set of bias vectors for the process adjustment branch, and then each component is passed through a piecewise linear activation function to obtain the hidden layer vector of the process adjustment branch. The dimension of the hidden layer of the process adjustment branch is also selected from several candidate integers, with the final value chosen by comparing the process adjustment prediction error on the validation data.

[0093] After obtaining the hidden layer vector of the process adjustment branch, it is multiplied by the output weight vector of the process adjustment branch, and then the output bias scalar of the process adjustment branch is added. Finally, nonlinear compression is performed using the hyperbolic tangent function to obtain the predicted process adjustment value. The output range of the hyperbolic tangent function is a finite symmetric interval centered at zero, which makes the predicted process adjustment value automatically fall into the finite interval, more closely approximating the numerical distribution of the process adjustment reference value in the training data.

[0094] Through the above structure, the additive performance prediction model encodes the set of periodic profile feature sequences of continuous control cycles into an intermediate representation vector, and then generates cross-cycle additive performance change trend prediction values ​​and process adjustment prediction values ​​through two output branches.

[0095] 3.4 Loss function construction and training process.

[0096] To simultaneously constrain the performance trend prediction output and the process adjustment prediction output, performance error and process adjustment error are defined for each training sample. The performance error is the predicted performance trend value minus the corresponding cross-cycle additive performance trend value in the training sample, and the process adjustment error is the predicted process adjustment value minus the process adjustment reference value in the training sample.

[0097] For a single training sample, the performance error is first nonlinearly compressed using the hyperbolic tangent function, and then its absolute value is taken to obtain the performance loss for that training sample. The same operation is performed on the process adjustment error to obtain the process adjustment loss for that training sample. The hyperbolic tangent function compresses the error within a finite interval when the amplitude is large, thereby mitigating the impact of extreme outliers on the training process, while maintaining sufficient gradient changes within a moderate error interval so that the model is sensitive to moderate-intensity errors.

[0098] For each training sample, the performance loss and the process adjustment loss are compared, and the maximum of the two is taken as the overall loss for that training sample. In this way, if either the performance prediction error or the process adjustment prediction error is large, the overall loss will increase, thereby driving the model to focus on both output branches during training, rather than making only one branch reach a good level.

[0099] The overall loss function is defined as the sum of the combined losses over all training samples. The training process employs a gradient-based iterative optimization method. In each training iteration, a subset of training samples is selected from the training sample set. Performance predictions and process adjustment predictions are calculated according to the aforementioned structure to obtain the combined loss. Then, the derivative of the combined loss with respect to each parameter in the model is calculated, and all parameters, including convolution kernel coefficients, biases, hidden layer weights, and output layer weights, are adjusted in the opposite direction of the derivative. Through multiple iterations, training stops when the change in the overall loss function over adjacent iterations enters a pre-defined convergence range, resulting in a stable set of model parameters.

[0100] After training, when the additive performance prediction model receives a set of periodic profile feature sequences within any span of periods, it can output the corresponding cross-period additive performance change trend prediction value and process adjustment prediction value, providing core computing power for generating basic additive performance prediction and adjustment scheme for the impact of detection lag based on the current control cycle information in subsequent steps.

[0101] By performing convolutional encoding and robust aggregation on the periodic contour feature sequence of continuous control cycles, the additive performance prediction model is trained to simultaneously output cross-cycle additive performance change trend predictions and process adjustment predictions. This establishes a direct link between the performance learning objective and coating quality changes and control actions, rather than simply fitting a single detection index. This is beneficial for extracting stable rhythm patterns and adjustment rules from historical operating data in scenarios where detection lag and additive consumption are intertwined. It lays the foundation for directly using the model output for performance prediction and dosing rhythm optimization in the subsequent online control stage.

[0102] Step S3 trains the additive performance prediction model using a training sample set, enabling the model to simultaneously output predicted values ​​for cross-cycle additive performance changes and process adjustment when inputting a set of cycle profile feature sequences for continuous control cycles. However, the aforementioned model remains in the offline learning stage and has not yet established a closed operational link with the real-time control cycle rolling, detection lag, and process feedback in the production site. Step S4 requires constructing a prediction window consistent with the training stage around the already trained additive performance prediction model during the production operation phase. The model output is then converted into the basic additive performance prediction values ​​for the target control cycle. A control scheme is generated by combining the prediction confidence and performance target range. Furthermore, the prediction error records and model parameters are continuously corrected after new detection results and coating quality detection results arrive, ensuring consistency and usability of additive performance prediction and dynamic process optimization in long-term operation.

[0103] 4.1 Construction of online prediction window and model inference.

[0104] During the production operation phase, the control cycle division rules are consistent with step S1, and the control cycle index increases sequentially according to time. At the end of each control cycle, following the processing logic of step S1, production load sampling sequences and detection sampling sequences are generated from the production load data and detection data collected within that control cycle, based on unified relative time sampling points. Amplitude normalization and adjacent difference operations are performed on each sampling sequence, and they are combined in a predetermined order to form the cycle contour feature sequence of the current control cycle.

[0105] When the current control cycle index is not less than the span cycle number, the current control cycle and several control cycles preceding it are grouped together, with the number of control cycles within the group equal to the span cycle number. This group of control cycles is called the prediction window. The starting control cycle of the prediction window is the control cycle after counting the span cycle number forward from the current control cycle minus one, and the ending control cycle is the current control cycle. The input feature set of the prediction window consists of the chronological sequence of cycle contour features from the starting control cycle to the ending control cycle.

[0106] The input feature set of the prediction window is fed into the additive performance prediction model trained in step S3. Inference is performed using the fixed parameters in the model, yielding two outputs. The first output is the predicted value of the cross-cycle additive performance change trend, corresponding to the additive performance change between the starting control cycle of the prediction window and the control cycle one span ahead. The second output is the predicted value of process adjustment, corresponding to the adjustment direction and magnitude reference given by the model based on historical learning results within this prediction window. To use the cross-cycle performance change prediction results at the control cycle granularity, the predicted value of the cross-cycle additive performance change trend is divided by the number of span cycles to obtain the single-cycle average performance change prediction, which is used to construct a virtual performance change trajectory at the control cycle level.

[0107] 4.2 Construction of reference performance values ​​and basic performance prediction values.

[0108] To obtain the predicted basic additive performance values ​​over the target control period, a reference performance value needs to be constructed for the control period starting from the prediction window. The reference performance value is derived from two parts of information: one part is the additive performance characterization value obtained from the most recent actual test, and the other part is the cumulative value of the predicted average performance change per control period from the most recent actual test to the control period starting from the prediction window.

[0109] First, record the control cycle index of the most recent completed additive performance test and its corresponding additive performance characterization value. When the control cycle index of the prediction window start point is the same as the control cycle index of the most recent actual test, the reference performance value is directly taken as the additive performance characterization value of the most recent actual test.

[0110] When the index of the control period starting at the prediction window is greater than the index of the control period at the most recent actual detection, starting from the most recent actual detection control period, the predicted average performance change for each control period is read sequentially in ascending order of the control period index. This range covers all control periods from the most recent actual detection control period to the start of the prediction window. Within this index range, the predicted average performance change for each control period is summed to obtain the total predicted performance change from the most recent actual detection control period to the start of the prediction window control period. The reference performance value is equal to the additive performance characterization value at the most recent actual detection time plus this total predicted performance change. In this way, even when there is a detection lag, the performance level of the control period at the start of the prediction window can be derived on the timeline based on historical predictions and execution data.

[0111] After constructing the reference performance values, the predicted values ​​of additive performance changes across cycles within the current prediction window are extrapolated over a cross-cycle scale to obtain the predicted values ​​of basic additive performance for the target control period. The index of the target control period equals the starting control period index of the prediction window plus the number of cycles spanned, and the predicted value of basic additive performance equals the reference performance value of the starting control period of the prediction window plus the predicted value of additive performance changes across cycles within this prediction window. This allows for the provision of basic performance predictions for control decision-making even before detection results are obtained for the target control period.

[0112] 4.3 Construction of prediction confidence index and threshold setting.

[0113] Over long-term operation, the distribution of prediction errors generated by the model varies under different operating conditions, necessitating a quantitative evaluation of the reliability of the current prediction results. Therefore, a prediction reliability index is calculated to reflect the overall level of prediction errors over a recent period.

[0114] At the starting control cycle index of a certain prediction window, when time progresses to a control cycle that spans one span of cycles, the difference between the additive performance characterization values ​​of the starting control cycle and the ending control cycle can be calculated using the rules in step S2 to obtain the observed performance change corresponding to that prediction window. Subtracting the observed performance change from the model's predicted trend of additive performance change across cycles given in that prediction window yields the performance change prediction error for that prediction window.

[0115] To construct a dimensionless error index, the absolute value of the performance change prediction error is divided by the sum of the absolute value of the observed performance change and the minimum performance change scale. The minimum performance change scale is determined using historical stable operating data. Specifically, over a period of time when process parameters are not adjusted, production load is within the set range, and additive dosing strategy remains fixed, the absolute values ​​of the observed performance changes corresponding to the starting control cycle of all prediction windows are statistically analyzed to form a performance change sample set. This set is sorted, and the value in the middle position is selected as a candidate minimum performance change scale. Then, based on the experience of process engineers, a specific value is selected as the minimum performance change scale within a finite numerical interval centered on the candidate value, ensuring that it is neither lower than the natural fluctuation level dominated by detection noise nor overshadows the slight changes in actual performance. The minimum performance change scale obtained in this way is used to calculate the stability error ratio when the observed performance change is close to zero.

[0116] For each prediction window where observations have been completed, a prediction error ratio is obtained by following the steps described above. Under the current control period, to calculate the prediction confidence index, several prediction windows whose starting control period index is closest to the current control period are selected from the prediction windows where observations have been completed in the most recent period, forming an error evaluation window set. The number of error evaluation windows is selected from a finite set of integers; candidate values ​​can correspond to dozens or hundreds of predictions. By comparing the correlation between the prediction confidence index and actual quality violation events under different number configurations in historical data playback, a number that makes the correspondence between the two clear and statistically stable is selected.

[0117] Within the error assessment window set, all prediction error ratios are collected, sorted, and the value at the middle of the sorted range is taken as the median prediction error. The value closest to the three-quarters mark is taken as the upper quartile prediction error. The median prediction error reflects the central level of the error distribution, while the upper quartile reflects the larger portions of the error distribution.

[0118] The prediction confidence index is given by a monotonically decreasing transformation. Specifically, it is calculated by adding one to the sum of the median and upper quartile values ​​of the prediction error as the denominator, and using one as the numerator, resulting in a value between zero and one. The closer the median and upper quartile values ​​of the prediction error are to zero, the closer the denominator is to one, and the closer the prediction confidence index is to one, indicating that the overall prediction error level in the recent period is relatively low. Conversely, the larger the two error statistics, the larger the denominator, and the closer the prediction confidence index is to zero, indicating that the recent prediction error is relatively large.

[0119] The prediction reliability threshold was determined through data statistics during the trial operation phase. During this phase, the model output was not directly used for substantive control; instead, prediction results were recorded in parallel with subsequent actual detection and coating quality performance. For each prediction, the corresponding prediction reliability index was calculated, and whether any events occurred where the coating quality deviated from the target range within the time range corresponding to the number of spanning cycles were recorded. By segmenting all predictions according to the prediction reliability index and statistically analyzing the frequency of quality violation events within each segment, the distribution of the frequency of quality violation events with respect to the prediction reliability index can be obtained. As the prediction reliability index gradually increases, the frequency of quality violation events generally shows a decreasing trend. The first value in the prediction reliability index value space that stably keeps the frequency of quality violation events below the preset control target was selected as the prediction reliability threshold. The prediction reliability threshold ranges from zero to one.

[0120] 4.4 Generation of performance target range, control level and control scheme.

[0121] After obtaining the basic additive performance prediction value and prediction confidence index for the target control cycle, it is necessary to construct the performance target range by combining historical stable quality operation data, determine the control level based on the process adjustment prediction value, and then generate a specific control scheme based on this information.

[0122] The performance target range is constructed based on historical long-term stable operating segments. Within these segments, each control cycle corresponds to a set of additive performance characterization values ​​and acceptable coating quality. The additive performance characterization values ​​from all these control cycles are collected to form a performance sample set. This performance sample set is then sorted, and a value at a low quantile is selected as the lower limit of the performance target, and a value at a high quantile is selected as the upper limit. The performance level within the performance target range corresponds to historically stable coating quality. Process engineers can make minor adjustments to the lower and upper limits of the performance target based on experience to ensure the performance target range has sufficient safety margin without leading to excessive additive use.

[0123] During the target control period, the predicted performance values ​​of the basic additives are compared with the target performance range. The lower deviation is obtained by subtracting the lower limit of the performance target from the predicted performance value; a lower deviation indicates that the predicted performance level is below the target range. The upper deviation is obtained by subtracting the predicted performance value from the upper performance target; a lower deviation indicates that the predicted performance level is above the target range. The lower and upper deviations form the basis for determining the performance prediction position.

[0124] The predicted values ​​for process adjustments need to be mapped to control commands with a finite number of levels. To this end, the absolute values ​​of all non-zero process adjustment reference values ​​are collected from the training samples constructed in steps S2 and S3, as well as the samples generated during the trial operation phase, forming an adjustment range sample set. The adjustment range sample set is then sorted numerically, and values ​​at the lower middle range are selected as the first threshold for adjustment range, while values ​​at the higher middle range are selected as the second threshold. The range of values ​​for the first and second thresholds is limited by the minimum and maximum non-zero values ​​in the adjustment range sample set.

[0125] For the predicted process adjustment value in the current prediction window, its absolute value is calculated and compared with the first and second thresholds of the adjustment range. When the absolute value is less than the first threshold, the control level is defined as level zero, indicating no adjustment or only a very small adjustment. When the absolute value is not less than the first threshold and less than the second threshold, the control level is defined as level one, indicating a medium-amplitude dosing or reducing operation. When the absolute value is not less than the second threshold, the control level is defined as level two, indicating a larger dosing or reducing operation. The specific dosing amount or process parameter change range corresponding to each level is predetermined in the process specification, forming a fixed mapping from level to execution amount.

[0126] When generating a control scheme based on the above preparatory work, the joint criterion of prediction confidence index and performance target range is calculated.

[0127] When the prediction confidence index is not less than the prediction confidence threshold and the predicted performance value of the basic additive is outside the performance target range, the model is considered to have sufficient reliability in predicting performance under the current operating conditions and there is a deviation that needs to be adjusted. The decision logic in this case is illustrated in the code below.

[0128] When the predicted value of the basic additive performance is lower than the lower limit of the performance target, if the predicted value of the process adjustment is positive, it means that the model judges that the additive needs to be added or the process parameters related to the additive need to be strengthened. In this case, the corresponding level of incremental addition or parameter enhancement is directly executed according to the current control level. If the predicted value of the process adjustment is negative, it means that the model's judgment direction is inconsistent with the fact that the performance is low. In this case, the control level is only allowed to take values ​​between zero and one, and the adjustment is made with a limited range in conjunction with the subsequent test results for correction.

[0129] When the predicted value of the basic additive performance is higher than the upper limit of the performance target, if the predicted value of the process adjustment is negative, it means that the model judges that the additive needs to be reduced or the process parameters related to the additive need to be weakened. In this case, the corresponding level of dosage reduction or parameter reduction shall be executed according to the current control level. If the predicted value of the process adjustment is positive, it means that the predicted direction is not completely consistent with the fact that the performance is too high. In this case, the limit level shall not exceed one level, and a milder adjustment intensity shall be used in conjunction with subsequent testing for correction.

[0130] When the prediction confidence index is less than the prediction confidence threshold, or when the predicted performance value of the basic additive is within the performance target range, a conservative control strategy is adopted. If the performance prediction is within the target range, the control level is typically set to zero, maintaining quality stability by keeping existing process parameters and dosing rhythm. If the prediction confidence index is below the threshold but the performance prediction deviates from the target range, the control level can be set to zero or one, prioritizing adjustments with limited magnitude. Simultaneously, the system awaits the results of a new round of testing to verify the prediction direction, and the control strategy is adjusted in subsequent control cycles based on the verification results.

[0131] Once the control scheme is determined, at the actual execution time corresponding to the target control cycle, the specific dosage or process parameter modification is performed according to the control level mapping relationship. The control cycle index, control level, dosage, process parameter change value, and execution time are recorded on-site to provide basic data for subsequent feedback analysis.

[0132] 4.5 Feedback phase error record update and model incremental adjustment.

[0133] After the control scheme is implemented, as production progresses, additive performance test results and coating quality test results will be obtained successively for the target control cycle and several control cycles covering the subsequent span.

[0134] When all detection results within the span corresponding to a certain prediction window are available, the observed performance change and observed coating quality change between the starting control period and the ending control period after spanning a certain number of periods can be recalculated according to the rules given in step S2. The new performance change prediction error is calculated using the observed performance change and the currently stored cross-period additive performance change trend prediction value. The error ratio sequence is then updated according to the method for constructing the prediction error ratio in step S4, providing the latest data for the next calculation of the prediction confidence index.

[0135] For parameter updates of the additive performance prediction model, an incremental fine-tuning approach is adopted. After each update of the detection results covering the entire span, the input feature set corresponding to the prediction window, the observed performance changes, and the process adjustment reference values ​​reconstructed based on the new observations are combined into a new training sample and added to a small-scale sample set dedicated to online fine-tuning. On this small-scale sample set, several rounds of gradient descent iterations are performed according to the comprehensive loss function defined in step S3. The model parameters are then finely adjusted using a pre-set learning step size, allowing the model to gradually absorb the performance and quality changes under the latest operating conditions while maintaining the overall patterns formed during the offline training phase.

[0136] As more prediction windows complete observation and feedback, error records are continuously updated, and model parameters gradually converge to the specific operating conditions of the current production line through incremental adjustments. This makes the basic additive performance prediction values ​​and process adjustment prediction values ​​of subsequent control cycles more closely match the actual operating performance, thereby improving the effectiveness of dynamic optimization control.

[0137] Step S4 constructs a prediction window consistent with the training phase during the production operation phase. The cycle profile feature sequence of the continuous control cycle is input into the additive performance prediction model to obtain the basic additive performance prediction value for the target control cycle. At the same time, the prediction confidence index, performance target range, and control level are introduced. The prediction results under the detection lag condition are used in a graded manner. The model output is transformed into decisions on dosing and process parameter adjustment. Incremental correction is performed by combining subsequent detection results and coating quality detection results, so that the additive performance prediction and the dynamic control of the electroplating production line form a closed link, reducing the impact of single-point detection errors and fluctuations in human experience on the stability of coating quality.

[0138] Specifically, the above description is only a preferred embodiment of this application and is not intended to limit this application.

[0139] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0140] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for predicting and dynamically optimizing the performance of electroplating additives based on big data analysis, characterized in that, Including the following steps: S1: The electroplating production line time axis is divided into multiple control cycles based on the detection or chemical addition time point. Within each control cycle, the process operation data and detection results are normalized and differentially processed according to the time sequence to obtain the cycle contour feature sequence. S2: Establish a correlation between the periodic profile feature sequence of each control cycle and the additive performance characterization value and coating quality characterization value at the beginning of the adjacent control cycle, and construct a training sample set with the evolution of the periodic profile as input and the trend of additive performance change as output in chronological order. S3: Train the additive performance prediction model based on the training sample set, so that when the model receives the cycle profile feature sequence of the current control cycle and the historical control cycle, it outputs the basic additive performance prediction value at the predetermined control time and the process adjustment prediction value for compensating for the detection lag effect. S4: During production, input the cycle profile feature sequence of the current control cycle into the model to obtain the prediction result. Determine the correction range of the prediction result based on the prediction result of the most recent control cycle and the time relationship of the process response. Generate a control scheme based on the corrected prediction result and execute it. Update the model based on the execution result.

2. The method for predicting and dynamically optimizing the performance of electroplating additives based on big data analysis according to claim 1, step S1 includes the following: The electroplating production line time axis is divided into multiple control cycles according to the detection time point and the chemical addition time point. Control cycles with a time length lower than the preset lower threshold are merged with adjacent control cycles. Within each control cycle, interpolated sampling values ​​are obtained from the production load data sequence and the detection data sequence based on a unified relative time sampling point.

3. The method for predicting and dynamically optimizing the performance of electroplating additives based on big data analysis according to claim 2, step S1 further includes the following: Within each control cycle, thresholds for production load variation range and detection variation range are set based on the historical noise range and the effective variation range. Amplitude normalization and adjacent difference are performed on the production load sampling sequence and the detection sampling sequence. The four results at each relative time sampling point are combined into a periodic profile feature sequence.

4. The method for predicting and dynamically optimizing the performance of electroplating additives based on big data analysis according to claim 3, step S2 includes the following: Based on the obtained cycle profile feature sequence of each control cycle, the additive performance characterization value is calculated using the sorted median value method within the detection time sub-interval near the start time of the corresponding control cycle. The coating quality characterization value is calculated using the sorted median value method within the quality detection time window corresponding to the start time. A one-to-one correspondence is established between the additive performance characterization value and the coating quality characterization value and the control cycle index.

5. The method for predicting and dynamically optimizing the performance of electroplating additives based on big data analysis according to claim 4, step S2 further includes the following: The range of span period numbers is determined based on the historical additive performance characterization value sequence. The span period number is selected based on the length of the monotonic performance change interval. The cross-cycle performance change trend value of each starting control cycle is calculated on the control cycle sequence according to the span period number. The cycle profile feature sequence of multiple consecutive control cycles, the corresponding performance change trend value, and the coating quality characterization values ​​at the start and end points are combined to form a training sample set.

6. The method for predicting and dynamically optimizing the performance of electroplating additives based on big data analysis according to claim 5, step S3 includes the following: An additive performance prediction model containing convolutional coding mapping and mid-segment aggregation structure is constructed using a training sample set. The set of periodic contour feature sequences of continuous control cycles is encoded into a fixed-dimensional intermediate representation vector. The performance trend output branch generates cross-cycle additive performance change trend prediction values ​​based on the intermediate representation vector, and the process adjustment output branch generates process adjustment prediction values ​​based on the intermediate representation vector.

7. The method for predicting and dynamically optimizing the performance of electroplating additives based on big data analysis according to claim 6, step S3 further includes the following: On each training sample, the error between the predicted value of performance change trend and the value of cross-cycle additive performance change trend, as well as the error between the predicted value of process adjustment and the reference value of process adjustment, are calculated. Two local losses are constructed by hyperbolic tangent function and absolute value. In a single training sample, the larger of the two local losses is taken as the comprehensive loss. The comprehensive loss is summed on all training samples as the overall loss. All parameters of the additive performance prediction model are updated by gradient-based iterative optimization method.

8. The method for predicting and dynamically optimizing the performance of electroplating additives based on big data analysis according to claim 7, step S4 includes the following: At the end of each control cycle, a prediction window containing several control cycles spanning the cycle is constructed. The set of cycle contour feature sequences within the prediction window is input into the additive performance prediction model. Based on the additive performance characterization value obtained from the most recent real detection and the historical single-cycle average performance change prediction, the reference performance value of the starting control cycle of the prediction window and the basic additive performance prediction value of the target control cycle are calculated.

9. The method for predicting and dynamically optimizing the performance of electroplating additives based on big data analysis according to claim 8, step S4 further includes the following: Based on the error ratio between the observed performance change and the predicted value of the additive performance change trend across cycles, a prediction confidence index and a prediction confidence threshold are constructed. Combined with the performance target range determined by the stable operation data and the control level obtained by mapping the predicted value of process adjustment, a dosing and process parameter adjustment scheme for the target control cycle is generated. After obtaining the test results covering the span of cycles, the error record and additive performance prediction model parameters are updated.

Images