Data labeling method and system applied to gold foil product detection

Abstract

The embodiment of the invention discloses a data labeling method and system applied to gold foil product detection, and the method can provide comprehensive materials for subsequent analysis through obtaining a quality description data set containing quality description statements and process parameter records; context semantic features and parameter association features of statements can be accurately captured based on feature extraction processing; based on multi-dimensional strategy analysis processing of a dynamic matching algorithm, a quality labeling result can be creatively generated, so that a mapping relation between a gold foil product defect type and a process parameter adjustment direction is clearly presented; and a process optimization strategy set is generated according to a quality labeling result and is fed back to a production control system to execute parameter regulation and control, so that the gold foil production process can be quickly and accurately optimized, the product quality is improved, the defect rate is reduced, and the intelligence and automation level of the production process is enhanced.

Description

Technical Field

[0001] The embodiments of the present invention relate to the technical field of data analysis, and particularly to a data annotation method and system for gold foil product detection. Background Art

[0002] In the field of production quality inspection of gold foil products, with the wide application of gold foil in many fields such as art decoration and electronics, the quality requirements for it are increasing day by day. Traditional production quality inspection of gold foil mainly relies on manual experience and simple physical inspection means, and the evaluation of product quality is relatively single and one-sided. Manual inspection is not only inefficient, but also prone to human errors, making it difficult to comprehensively and accurately judge the quality problems of gold foil products.

[0003] In addition, on the one hand, other existing technologies cannot effectively integrate the complex relationship between quality description information and process parameters, resulting in difficulty in deeply analyzing the root causes of quality problems. On the other hand, in the face of complex and diverse types of gold foil product defects, there is a lack of intelligent analysis methods to determine the corresponding process parameter adjustment directions, making it difficult to achieve targeted process optimization. Summary of the Invention

[0004] The embodiments of the present invention provide a data annotation method and system for gold foil product detection, which are used to solve the problems that it is difficult to comprehensively analyze quality problems and accurately optimize processes in the production quality inspection of gold foil products by the existing technologies, and improve the quality and efficiency of gold foil production.

[0005] In a first aspect, the embodiments of the present invention provide a data annotation method for gold foil product detection, which is applied to a data annotation system. The method includes: obtaining a quality description data set of gold foil products, where the quality description data set includes a plurality of quality description statements and corresponding process parameter records; performing feature extraction processing on the quality description data set to obtain the context semantic features of each quality description statement and the parameter association features of the process parameter records; performing multi-dimensional strategy analysis processing on the context semantic features and the parameter association features based on a dynamic matching algorithm to generate a quality annotation result for the quality description statement; the quality annotation result is used to indicate the mapping relationship between the gold foil product defect type and the process parameter adjustment direction; generating a process optimization strategy set according to the quality annotation result, and feeding back the process optimization strategy set to the gold foil production control system to perform parameter regulation operations.

[0006] In a second aspect, the embodiments of the present invention provide a data annotation system, including: a processor; a storage device on which a computer program is stored, When the computer program is executed by the processor, the processor implements any of the data annotation methods applied to the detection of gold foil products.

[0007] An embodiment of the present invention provides a readable storage medium, on which a program or instruction is stored, and when the program or instruction is executed by a processor, the steps of the data annotation method applied to the detection of gold foil products are implemented.

[0008] Thus, the embodiments of the present invention have the following beneficial effects: The embodiments of the present invention realize the in-depth mining and effective utilization of the quality data of gold foil products: by obtaining a quality description data set covering quality description statements and process parameter records, comprehensive materials can be provided for subsequent analysis; based on feature extraction processing, the semantic features of sentence context and parameter association features can be accurately captured; based on the multi-dimensional strategy analysis processing of the dynamic matching algorithm, the quality annotation results can be innovatively generated, so as to clearly present the mapping relationship between the defect types of gold foil products and the adjustment direction of process parameters; according to the quality annotation results, a process optimization strategy set is generated and fed back to the production control system to execute parameter regulation, which can quickly and accurately optimize the gold foil production process, improve product quality, reduce the defect rate, and enhance the intelligent and automated level of the production process. Description of the Drawings

[0009] Figure 1 It is a flowchart of a data annotation method applied to the detection of gold foil products provided by an embodiment of the present invention.

[0010] Figure 2 It is a schematic diagram of the basic structure of a data annotation system provided by an embodiment of the present invention. Detailed Embodiments

[0011] To make the above objects, features, and advantages of the present invention more obvious and understandable, the embodiments of the present invention will be further described in detail below with reference to the drawings and specific embodiments.

[0012] See Figure 1 As shown, this figure is a flowchart of a data annotation method applied to the detection of gold foil products provided by an embodiment of the present invention, and this method can be applied to a data annotation system. As Figure 1 shown, this method may include Step 110 - Step 140.

[0013] Step 110: Obtain a quality description data set of gold foil products, where the quality description data set includes a plurality of quality description statements and corresponding process parameter records.

[0014] In an embodiment of the present invention, the data annotation system of the gold foil production factory continuously collects various types of information during the production process. For example, in one production cycle, the data annotation system records multiple quality description statements, including but not limited to "fine scratches appear on the surface of the gold foil", "the thickness of the gold foil is uneven", etc. At the same time, a series of process parameters are correspondingly recorded. For example, temperature parameter A, whose value gradually changes from the initial value A1 to A2 during this cycle; pressure parameter B, fluctuating from B1 to B2; rolling speed parameter C, varying within the range from C1 to C3. The above quality description statements and the corresponding process parameter records together constitute the quality description data set.

[0015] Step 120: Perform feature extraction processing on the quality description data set to obtain the context semantic features of each quality description statement and the parameter correlation features of the process parameter records.

[0016] In an embodiment of the present invention, the system starts to carry out feature extraction work on the quality description data set. The quality description statements and the process parameter records are processed respectively to obtain the required features.

[0017] Optionally, step 120 further includes: Step 121: Perform semantic unit segmentation processing on the quality description statements to obtain multiple semantic unit sequences; each semantic unit sequence includes at least one semantic segment and the position identifier of the semantic segment in the statement.

[0018] For example, for the quality description statement "fine scratches appear on the surface of the gold foil", the system uses a commonly used semantic analysis tool in related technologies to perform semantic unit segmentation. It is segmented into semantic segments such as "gold foil", "surface", "appear", "fine", "scratch", etc. Each semantic segment is assigned a position identifier in the statement. For example, "gold foil" is in the 1st position, "surface" is in the 2nd position, etc., thus forming multiple semantic unit sequences. The above semantic unit sequences provide a basis for further in-depth analysis of the semantics of the statement, facilitating the system to further explore the relationships and potential information among various parts of the statement. Then, the segmented semantic unit sequences are passed to the next processing link, that is, calling a pre-trained language feature encoder for context encoding processing.

[0019] Step 122: Call a pre-trained language feature encoder to perform context encoding processing on the semantic unit sequences to generate the context semantic features of the quality description statements; the context semantic features include the local semantic features of each semantic segment and the context dependence relationship between adjacent semantic segments.

[0020] In a preferred embodiment, step 122 may include: Step 1221: Perform part-of-speech tagging on the semantic fragments in each semantic unit sequence to generate a tagged sequence containing part-of-speech tags.

[0021] Taking the semantic unit sequence of "Fine scratches appear on the gold foil surface" as an example, the system performs part-of-speech tagging on each semantic fragment therein. "Gold foil" is tagged as a noun, "surface" as a noun, "appear" as a verb, "fine" as an adjective, and "scratch" as a noun, thus generating a tagged sequence containing part-of-speech tags. This tagged sequence helps the system better understand the grammatical roles of each semantic fragment in the sentence, providing an important basis for subsequent construction of the semantic dependency tree and hierarchical encoding processing. Then, the tagged sequence is used as input for constructing the semantic dependency tree.

[0022] Step 1222: Construct a semantic dependency tree based on the tagged sequence to determine the subject-predicate relationship, modification relationship, and coordination relationship among the semantic fragments.

[0023] Optionally, based on the tagged sequence obtained from the above part-of-speech tagging, the system constructs a semantic dependency tree. In "Fine scratches appear on the gold foil surface", "gold foil" is the main body described by the whole sentence, "surface" describes the position of "gold foil", and there is a modification relationship between them; "appear" is an action, and its subject is "gold foil surface", so there is a subject-predicate relationship; "fine" modifies "scratch", which is also a modification relationship. By constructing the semantic dependency tree, the system determines the complex relationships among the semantic fragments, and these relationships will play a key role in the subsequent hierarchical encoding processing, helping the system better capture the semantic information of the sentence. After completing the construction of the semantic dependency tree, hierarchical encoding processing is carried out based on this.

[0024] Step 1223: Perform hierarchical encoding processing on the semantic unit sequence based on the semantic dependency tree: perform local encoding on each semantic fragment to generate an initial semantic vector, perform relationship aggregation on the initial semantic vector according to the subject-predicate relationship to generate an intermediate semantic vector, and perform context enhancement on the intermediate semantic vector based on the modification relationship and coordination relationship to generate an enhanced semantic vector.

[0025] Optionally, the system first performs local encoding on each semantic segment. For example, for the semantic segment "gold foil", it is converted into an initial semantic vector V1 through the encoding algorithm commonly used in related technologies. Then, according to the subject-predicate relationship, the initial semantic vectors of the "gold foil surface" related to "appearance" are aggregated in relation to obtain an intermediate semantic vector V2. Next, considering the modification relationship of "subtle" to "scratch" and the relationships between other relevant semantic segments, the intermediate semantic vector V2 is contextually enhanced to generate an enhanced semantic vector V3. This enhanced semantic vector V3 synthesizes the local information of the semantic segments and their context dependencies, and more comprehensively represents the semantic features of this semantic unit sequence. After generating the enhanced semantic vector, the system proceeds to the next step of processing.

[0026] Step 1224: Arrange the enhanced semantic vectors in chronological order according to the position identifiers to generate the context semantic features of the quality description statement.

[0027] In an embodiment of the present invention, the system arranges the generated enhanced semantic vectors in chronological order according to the previously assigned position identifiers. For example, the enhanced semantic vector of "gold foil" is ranked first, the one of "surface" is ranked second, etc. Eventually, an ordered vector sequence is formed, which is the context semantic feature of the quality description statement. This context semantic feature completely retains the information of each semantic segment in the quality description statement and their context relationships, providing strong support for subsequent analysis based on the dynamic matching algorithm. Next, the system starts to extract features from the process parameter records.

[0028] Step 123: Perform parameter parsing processing on the process parameter records to extract parameter category identifiers and parameter chronological change patterns; the parameter chronological change patterns include the fluctuation characteristics of parameter values changing over time and the associated fluctuation trends between different parameters.

[0029] In a preferred embodiment, step 123 may include: Step 1231: Divide the process parameter records into multiple parameter subsequences according to time windows, and each parameter subsequence contains a set of parameter values within a preset time range.

[0030] For example, for the process parameter records of temperature parameter A, pressure parameter B, and rolling speed parameter C, the system divides them into time windows of every 10 minutes. In the first time window, the value set of temperature parameter A is [A11, A12, A13], the value set of pressure parameter B is [B11, B12, B13], and the value set of rolling speed parameter C is [C11, C12, C13], thus forming multiple parameter subsequences. The above parameter subsequences provide basic data units for subsequent analysis of the change characteristics of parameters in different time periods. After the division, trend analysis processing is performed on each parameter subsequence.

[0031] Step 1232: Perform trend analysis processing on each parameter subsequence, and extract the rising trend segment, falling trend segment, and stable trend segment.

[0032] Taking a parameter subsequence [A11, A12, A13] of temperature parameter A as an example, the system performs trend analysis by comparing the magnitude relationship of adjacent data points. If A11 < A12 < A13, then this subsequence shows a rising trend and is extracted as the rising trend segment; if A11 > A12 > A13, it is the falling trend segment; if the values of A11, A12, and A13 fluctuate little and are in a relatively stable range, it is the stable trend segment. By performing such trend analysis on each parameter subsequence, the system can clearly understand the change trend of the parameter in different time periods and prepare for subsequent further feature extraction. After completing the trend analysis, corresponding feature calculations are performed on different trend segments.

[0033] Step 1233: Perform slope calculation processing on the rising trend segment to generate a first slope feature set; perform absolute value calculation processing on the falling trend segment to generate a second slope feature set; perform variance calculation processing on the stable trend segment to generate a fluctuation variance feature set.

[0034] For the rising trend segment of temperature parameter A, the system calculates the slope between adjacent data points. For example, if the rising trend segment is [A11, A12, A13], by calculating (A12 - A11) / (t2 - t1), (A13 - A12) / (t3 - t2), etc. (where t1, t2, t3 are the corresponding time points), a series of slope values are obtained, and these slope values form the first slope feature set. For the falling trend segment, the slope is also calculated, but the absolute value is taken to form the second slope feature set. For the stable trend segment, the variance of the parameter values is calculated to measure its fluctuation degree, and the fluctuation variance feature set is obtained. The above feature sets describe the change characteristics of the parameter in different trend segments from different perspectives. Then, these feature sets are classified according to the parameter category identifier.

[0035] Step 1234: Classify the first slope feature set, the second slope feature set, and the fluctuation variance feature set according to the parameter category identifier to generate an in-class trend feature matrix.

[0036] Among them, the system classifies the first slope feature set, the second slope feature set, and the fluctuation variance feature set of temperature parameter A into one category according to the parameter category identifier, the corresponding feature sets of pressure parameter B into another category, and those of rolling speed parameter C into yet another category. Then, the feature sets in each category are arranged in a certain order to form an in-class trend feature matrix. For example, in the in-class trend feature matrix of temperature parameter A, the elements of the first slope feature set are placed in the first row, the elements of the second slope feature set in the second row, and the elements of the fluctuation variance feature set in the third row. Such a matrix form is more convenient for the system to perform subsequent correlation calculations and feature fusion processing. After generating the in-class trend feature matrix, perform correlation calculations on the matrices between different parameter categories.

[0037] Step 1235: Perform correlation calculation processing on the in-class trend feature matrices between different parameter categories to generate a cross-class trend association matrix.

[0038] The system calculates the correlation between the in-class trend feature matrix of temperature parameter A and the in-class trend feature matrix of pressure parameter B, for example, by calculating the Pearson correlation coefficient between matrix elements. For example, if the in-class trend feature matrix of temperature parameter A is M1 and that of pressure parameter B is M2, calculate the correlation between each element in M1 and the corresponding element in M2 to obtain a series of correlation values. Organize these values into a matrix, which is a part of the cross-class trend association matrix between temperature parameter A and pressure parameter B. Use the same method to calculate the cross-class trend association matrix parts between temperature parameter A and rolling speed parameter C, and between pressure parameter B and rolling speed parameter C, and finally form a complete cross-class trend association matrix, which reflects the degree of trend association between different parameter categories. After completing the correlation calculation, perform matrix splicing processing.

[0039] Step 1236: Splice the in-class trend feature matrix and the cross-class trend association matrix to generate the parameter time series change pattern.

[0040] The system splices the previously generated intra-category trend feature matrices and cross-category trend correlation matrices of various categories according to certain rules. For example, the intra-category trend feature matrices of temperature parameter A, pressure parameter B, and rolling speed parameter C are arranged in sequence, and then the cross-category trend correlation matrix is spliced on their right side, thus forming a complete matrix structure, which is the parameter time-series change pattern. It synthesizes the change trend characteristics of parameters in different time periods and the correlation fluctuation trends between different parameters, and comprehensively describes the dynamic change characteristics of process parameters. After generating the parameter time-series change pattern, feature fusion processing is performed based on the parameter category identifier.

[0041] Step 124: Perform feature fusion processing on the parameter time-series change pattern based on the parameter category identifier to generate the parameter correlation feature recorded by the process parameter; the parameter correlation feature includes the multi-dimensional fluctuation correlation under the same parameter category and the cross-category influence weight between different parameter categories.

[0042] Optionally, the system performs feature fusion on the parameter time-series change pattern according to the parameter category identifier. For the same parameter category, such as temperature parameter A, the system analyzes the relationship between different feature sets in its intra-category trend feature matrix to determine the multi-dimensional fluctuation correlation. For example, there may be a certain correlation between the first slope feature set and the fluctuation variance feature set, and the strength of this correlation is calculated by using the algorithms commonly used in related technologies. For different parameter categories, the cross-category influence weight is determined based on the cross-category trend correlation matrix. For example, the influence weight of the change of temperature parameter A on pressure parameter B is quantitatively determined by analyzing the correlation values in the matrix. Finally, these multi-dimensional fluctuation correlations under the same parameter category and the cross-category influence weights between different parameter categories are integrated together to form the parameter correlation feature recorded by the process parameter. The above parameter correlation feature provides comprehensive and in-depth information about the process parameter for the subsequent analysis based on the dynamic matching algorithm. After completing the feature extraction, it enters the multi-dimensional strategy analysis and processing stage.

[0043] Step 130: Perform multi-dimensional strategy analysis and processing on the context semantic feature and the parameter correlation feature based on the dynamic matching algorithm to generate the quality annotation result of the quality description statement; the quality annotation result is used to indicate the mapping relationship between the gold foil product defect type and the process parameter adjustment direction.

[0044] In an exemplary embodiment, step 130 further includes: Step 131: Map the context semantic feature to the defect feature space to generate a defect feature vector including the probability distribution of defect types.

[0045] Furthermore, the system uses a commonly used mapping function in related technologies to map the context semantic features of the previously generated quality description statements, such as the context semantic feature vector corresponding to "fine scratches appear on the gold foil surface", to the defect feature space. In this space, different regions represent different defect types. Through mapping, a defect feature vector containing the probability distribution of defect types is obtained. For example, the probability value of this vector in the "scratch defect" region is 0.8, and the probability value in the "non-uniform thickness defect" region is 0.2, indicating that the situation described by this quality description statement is more likely to be a scratch defect. After generating the defect feature vector, a similar mapping process is performed on the parameter correlation features.

[0046] Step 132: Map the parameter correlation features to the regulation feature space to generate a regulation feature vector containing the weights of parameter adjustment directions.

[0047] In this step, the system uses another commonly used feature mapping method to map the parameter correlation features recorded by the process parameters to the regulation feature space. In this space, different dimensions represent different parameter adjustment directions. For example, for the parameter correlation features of temperature parameter A, pressure parameter B, and rolling speed parameter C, a regulation feature vector is obtained after mapping. The weight of this vector in the direction of "increasing temperature parameter A" is 0.6, the weight in the direction of "decreasing pressure parameter B" is 0.3, and the weight in the direction of "keeping rolling speed parameter C unchanged" is 0.1, indicating that according to the current parameter correlation features, increasing temperature parameter A is the more likely adjustment direction. After completing the mapping, spatial alignment processing is performed.

[0048] Step 133: Perform spatial alignment processing on the defect feature vector and the regulation feature vector to determine the mapping relationship between the defect type probability distribution and the weights of the parameter adjustment directions.

[0049] Among them, step 133 may include: Step 1330: Construct a dual-channel alignment network including a defect type dimension and a parameter adjustment dimension; input the defect feature vector into the defect type channel for feature enhancement processing to generate an enhanced defect feature vector; input the regulation feature vector into the parameter adjustment channel for feature enhancement processing to generate an enhanced regulation feature vector; perform cross-attention calculation processing on the enhanced defect feature vector and the enhanced regulation feature vector to generate a defect-parameter attention weight matrix; perform weighted summation processing on the enhanced defect feature vector and the enhanced regulation feature vector according to the defect-parameter attention weight matrix to generate a joint feature vector; perform fully connected mapping processing on the joint feature vector to generate the mapping relationship between the defect type probability distribution and the weights of the parameter adjustment directions.

[0050] Specifically, the system first constructs a dual-channel alignment network, which includes two channels: the defect type dimension and the parameter adjustment dimension. The defect feature vector is input into the defect type channel, and through a series of operations such as convolution and pooling for feature enhancement, an enhanced defect feature vector is obtained. Similarly, the regulation feature vector is input into the parameter adjustment channel for enhancement processing to generate an enhanced regulation feature vector. Then, cross-attention calculation is performed on these two enhanced vectors. For example, the correlation degree between each element in the enhanced defect feature vector and each element in the enhanced regulation feature vector is calculated to obtain a defect-parameter attention weight matrix. According to this matrix, weighted summation is performed on the enhanced defect feature vector and the enhanced regulation feature vector to obtain a joint feature vector. Finally, the joint feature vector is processed through a fully connected mapping function to generate a mapping relationship between the defect type probability distribution and the parameter adjustment direction weight. For example, the obtained mapping relationship indicates a strong correlation between "scratch defect" and "increase rolling speed parameter C". After determining the mapping relationship, a dynamic matching matrix is constructed.

[0051] Step 134: Construct a dynamic matching matrix according to the mapping relationship, and perform singular value decomposition on the dynamic matching matrix to obtain a defect type principal component vector and a parameter adjustment principal component vector.

[0052] Optionally, the system constructs a dynamic matching matrix based on the mapping relationship between the defect type probability distribution and the parameter adjustment direction weight. The rows of the matrix represent defect types, the columns represent parameter adjustment directions, and the matrix elements are the corresponding correlation degree values. Then, singular value decomposition is performed on this matrix. Through an algorithm commonly used in related technologies, the matrix is decomposed into the product of three matrices, and the defect type principal component vector and the parameter adjustment principal component vector are extracted from them. The above principal component vectors can more concisely represent the main features of defect types and parameter adjustment directions. After completing the singular value decomposition, defect type annotation and parameter adjustment direction annotation are performed.

[0053] Step 135: Determine the defect type annotation of the quality description statement according to the similarity calculation between the defect type principal component vector and the preset defect type standard vector. The defect type annotation is used to characterize the defect type of the gold foil product.

[0054] Optionally, the system calculates the similarity between the defect type principal component vector and various preset defect type standard vectors. For example, "scratch defect standard vector", "thickness uneven defect standard vector", etc. are preset. By calculating methods such as cosine similarity, the similarity between the defect type principal component vector and these standard vectors is compared. If the similarity with the "scratch defect standard vector" is the highest, which is 0.9, then it is determined that the defect type annotation of this quality description statement is "scratch defect", and this annotation clarifies the possible defect type of the gold foil product. After determining the defect type annotation, parameter adjustment direction annotation is performed.

[0055] Step 136: Adjust the direction matching process of the principal component vector and the historical optimization strategy vector according to the parameter, and determine the parameter adjustment direction annotation, where the parameter adjustment direction annotation is used to characterize the process parameter adjustment direction.

[0056] Optionally, the system matches the parameter adjustment principal component vector with the historical optimization strategy vector. The historical optimization strategy vector records the successful parameter adjustment directions in the past in similar situations. By calculating methods such as the angle between vectors, the direction consistency between the parameter adjustment principal component vector and the historical optimization strategy vector is judged. If it is found that the direction of the historical optimization strategy vector related to "increasing the rolling speed parameter C" is consistent and the similarity is 0.8, then the parameter adjustment direction annotation is determined to be "increasing the rolling speed parameter C", and this annotation provides a direction guidance for the process parameter adjustment. After completing the annotation, the results are combined.

[0057] Step 137: Combine the defect type annotation and the parameter adjustment direction annotation to generate the quality annotation result of the quality description statement.

[0058] Among them, the system combines the determined defect type annotation "scratch defect" and the parameter adjustment direction annotation "increasing the rolling speed parameter C" to form the quality annotation result of the quality description statement. This result clearly indicates the mapping relationship between the defect type of the gold foil product and the process parameter adjustment direction, providing a key basis for generating subsequent process optimization strategies. After generating the quality annotation result, enter the stage of generating the process optimization strategy set and performing parameter regulation operations.

[0059] Step 140: Generate a process optimization strategy set according to the quality annotation result, and feedback the process optimization strategy set to the gold foil production control system to perform parameter regulation operations.

[0060] In an alternative embodiment, generating the process optimization strategy set according to the quality annotation result in Step 140 includes: Step 141: Extract the defect type priority list and the parameter adjustment direction priority list from the quality annotation result.

[0061] Optionally, the system conducts an in-depth analysis of the generated quality labeling results. For example, for the quality labeling results of a series of quality description statements, the frequency of occurrence of different defect types and the severity assessment values ​​are counted, and they are arranged in order from high to low to form a defect type priority list. For example, among multiple quality labeling results, "scratch defects" appear more frequently and have a higher assessed severity, so they are ranked higher in the defect type priority list; "uneven thickness defects" appear relatively less frequently and have a slightly lower severity, so they are ranked later. At the same time, based on the importance assessment of the parameter adjustment direction in the quality labeling results and the degree of correlation with the defect type, a parameter adjustment direction priority list is generated. For example, "increasing the rolling speed parameter C", which is closely related to the "scratch defect" and appears multiple times, is at the front of the parameter adjustment direction priority list. The above priority list provides a basis for the subsequent determination of the optimization weight. After the extraction is completed, the optimization weight is determined based on the list.

[0062] Step 142: Determine a first optimization weight according to the defect type priority list, and determine a second optimization weight according to the parameter adjustment direction priority list.

[0063] Among them, for the defect type priority list, the system adopts the weight allocation algorithm commonly used in related technologies. For example, the first optimization weight is determined based on the proportion of the frequency of the defect type to the total frequency and the comprehensive consideration of the severity assessment value. For example, "scratch defect" accounts for 40% of the total frequency, and the severity assessment value is 8 points (out of 10 points). Through the existing weight normalization algorithm, its first optimization weight is 0.6. For the parameter adjustment direction priority list, the second optimization weight is also determined based on factors such as the closeness of its association with the defect type and the stability that appears in multiple quality labeling results. For example, "increasing the rolling speed parameter C" is closely related to the "scratch defect" and appears stable, so its second optimization weight is determined to be 0.5. After determining the optimization weight, weighted fusion processing is performed.

[0064] Step 143: Perform weighted fusion processing on the first optimization weight and the second optimization weight to generate a comprehensive optimization weight.

[0065] Among them, the system uses a weighted fusion algorithm to fuse the first optimization weight and the second optimization weight. For example, using a simple weighted average method, set the weight coefficient of the first optimization weight to 0.6, and the weight coefficient of the second optimization weight to 0.4, then the comprehensive optimization weight = 0.6 × first optimization weight + 0.4 × second optimization weight = 0.6 × 0.6 + 0.4 × 0.5 = 0.56. This comprehensive optimization weight comprehensively considers the importance of defect type and parameter adjustment direction, and provides a unified measurement standard for screening historical process parameter adjustment records. After the comprehensive optimization weight is generated, screening is performed.

[0066] Step 144: Screen and process the historical process parameter adjustment records according to the comprehensive optimization weights to obtain a set of candidate adjustment strategies.

[0067] Optionally, the system traverses the historical process parameter adjustment records. For each record, it evaluates according to the defect type and parameter adjustment direction involved, in combination with the comprehensive optimization weights. For example, a historical record adopted an adjustment strategy of "increasing rolling speed parameter C" for "scratch defect", and calculates and evaluates the score according to the comprehensive optimization weights and the historical effect of this record in defect resolution. Select the records with higher scores to form a set of candidate adjustment strategies. The strategies in this set are initially screened based on historical experience and the current quality annotation results, and have certain feasibility and optimization potential. After the screening is completed, a feasibility evaluation process is performed on the set of candidate adjustment strategies.

[0068] Step 145: Perform a feasibility evaluation process on the set of candidate adjustment strategies to generate a result list containing evaluation weights.

[0069] Specifically, Step 145 includes: Step 1450: Split each candidate adjustment strategy into a sequence of parameter adjustment steps. Each parameter adjustment step includes an adjustment target value and an adjustment time range; perform a conflict detection process on the sequence of parameter adjustment steps to identify conflict step pairs with time overlap or contradictory target values; generate a conflict evaluation score according to the number and severity of the conflict step pairs; perform a simulation execution process on the sequence of parameter adjustment steps to predict the parameter change curve and defect type probability change curve after execution; generate a parameter matching score according to the matching degree between the parameter change curve and the target parameter interval; generate a defect improvement score according to the difference degree between the defect type probability change curve and the target defect threshold; perform a weighted summation process on the conflict evaluation score, the parameter matching score, and the defect improvement score to generate the result list of the evaluation weights of the candidate adjustment strategies.

[0070] Taking a candidate adjustment strategy as an example, this strategy involves adjustments to the temperature parameter A, the pressure parameter B, and the rolling speed parameter C. It is split into a sequence of parameter adjustment steps, such as "During the time period t1 - t2, adjust the temperature parameter A from the current value to the target value A3", "During the time period t2 - t3, adjust the pressure parameter B from the current value to the target value B4", etc. The system first performs conflict detection on this sequence to check for situations of time overlap or conflicting target values. If the operations of "adjusting the temperature parameter A to A3" and "adjusting the pressure parameter B to B4" overlap partially in time and these two adjustments may affect each other, it is determined as a pair of conflicting steps. Based on the number of pairs of conflicting steps and the degree of potential serious impact on the production process, a conflict assessment score is generated. For example, if there are 3 pairs of conflicting steps and the impact is relatively serious, the conflict assessment score is 0.3 (out of a full score of 1). Then, the sequence of parameter adjustment steps is processed by simulated execution. Using the simulation model of the production process, the parameter change curve and the defect type probability change curve after execution are predicted. For example, after simulated execution, the parameter change curve shows that the temperature parameter A can better approach the target parameter range, and a parameter matching score is generated according to its matching degree with the target range, such as 0.8. The defect type probability change curve shows that the probability of "scratch defect" decreases from 0.8 to 0.5, and a defect improvement score is generated according to the difference from the target defect threshold (set to 0.3), such as 0.6. Finally, according to the preset weight distribution, for example, the weight of the conflict assessment score is 0.2, the weight of the parameter matching score is 0.3, and the weight of the defect improvement score is 0.5, a weighted sum is performed: 0.2×0.3 + 0.3×0.8 + 0.5×0.6 = 0.54, to obtain the evaluation weight value of this candidate adjustment strategy. Such an evaluation is performed for each strategy in the set of candidate adjustment strategies, generating a result list containing the evaluation weight values. After the evaluation is completed, the strategy selection is carried out.

[0071] Step 146: Sort the result list according to the evaluation weight values, and select the candidate adjustment strategies that exceed the preset score threshold to generate the process optimization strategy set.

[0072] Optionally, the system sorts the result list in descending order according to the evaluation weight values. For example, the candidate adjustment strategy with the highest evaluation weight value is ranked first and arranged in sequence. A score threshold is preset, such as 0.5. The system selects the candidate adjustment strategies whose evaluation weight values exceed this threshold, combines these strategies together to form a process optimization strategy set. The strategies in this set are determined after multiple rounds of screening and evaluation, having high feasibility and optimization effects, and can provide effective parameter regulation guidance for the gold foil production control system. After generating the process optimization strategy set, it is fed back to the gold foil production control system to perform parameter regulation operations.

[0073] Based on the above, the feedback of the process optimization strategy set to the gold foil production control system in step 140 to perform parameter regulation operations includes: Step 1471: Convert the process optimization strategy set into a control instruction sequence, where each control instruction includes a parameter identifier, a target value, and an execution timestamp.

[0074] Among them, the system parses and converts each strategy in the process optimization strategy set. For example, a strategy is "During the time period from t1 to t2, adjust the rolling speed parameter C from the current value C1 to the target value C2", and the converted control instruction is: the parameter identifier is "rolling speed parameter C", the target value is "C2", and the execution timestamp is "t1" (indicating the start execution time). All strategies in the process optimization strategy set are converted in this way to form a control instruction sequence, which clarifies the specific requirements and execution time arrangements for each parameter adjustment, facilitating the gold foil production control system to receive and execute. After the conversion is completed, timing verification processing is performed.

[0075] Step 1472: Perform timing verification processing on the control instruction sequence to ensure that the execution timestamp meets the operation interval limit of the production equipment.

[0076] Optionally, the gold foil production equipment sets corresponding operation interval limits. For example, there needs to be a certain time interval between two adjustment operations for the same parameter to ensure the stable operation of the equipment. The system checks the execution timestamps of each control instruction in the control instruction sequence to determine whether they meet the operation interval limit of the equipment. For example, for two consecutive control instructions for the rolling speed parameter C, the execution timestamp of the first instruction is t1, and the second is t2, and the required operation interval time of the equipment is Δt. If t2 - t1 < Δt, it does not meet the requirements. Mark the non-compliant timestamps for subsequent adjustment. After the timing verification is completed, timestamp adjustment processing is performed.

[0077] Step 1473: Perform timestamp adjustment processing on the control instruction sequence according to the verification result to generate an optimized control instruction sequence.

[0078] Among them, for the timestamps found not to meet the operation interval limit during the verification process, the system makes adjustments. For example, adjust the execution timestamp t2 of the second control instruction for the rolling speed parameter C to t1 + Δt. Through such adjustments, ensure that all execution timestamps in the control instruction sequence meet the operation interval limit of the production equipment, generate an optimized control instruction sequence, which can ensure the orderly execution of parameter regulation operations within the acceptable range of the production equipment, and avoid equipment failures or production anomalies caused by unreasonable time arrangements. After the adjustment is completed, send the optimized control instruction sequence to the gold foil production control system.

[0079] Step 1474: Send the optimized control instruction sequence to the parameter execution module of the gold foil production control system, and instruct the parameter execution module to perform parameter regulation operations according to the optimized control instruction sequence.

[0080] Among them, the system sends the optimized control instruction sequence to the parameter execution module of the gold foil production control system through the communication interface. After receiving the instruction sequence, the parameter execution module adjusts the corresponding process parameters in sequence according to the parameter identifier, target value, and execution timestamp in the instruction. For example, after receiving the instruction "parameter identifier is 'rolling speed parameter C', target value is 'C2', and execution timestamp is 't1'", the rolling speed parameter C is adjusted to C2 at time t1. In this way, precise regulation of process parameters in the gold foil production process is achieved to improve the quality of gold foil products. After sending the instruction, the execution status is monitored in real time.

[0081] Step 1475: Monitor the execution status of the parameter regulation operation in real time. When an execution anomaly is detected, generate a rollback instruction or an adaptive adjustment instruction according to the anomaly type to update the optimized control instruction sequence.

[0082] Furthermore, the system sets up a monitoring module in the gold foil production control system to track the execution of the parameter regulation operation in real time. For example, monitor whether the rolling speed parameter C is successfully adjusted to the target value C2, and whether the operating status of the equipment is normal during the adjustment process. If an anomaly occurs during the adjustment process, such as the equipment feedback indicates that the adjustment fails or the parameter value does not reach the expected range, the monitoring module detects the anomaly type. If the anomaly is caused by incorrect parameter settings, the system generates a rollback instruction to restore the parameter to the state before adjustment; if the anomaly is caused by external environmental factors, the system generates an adaptive adjustment instruction according to the preset adaptive adjustment strategy, such as appropriately adjusting the target value or the execution time. Add the generated instruction to the optimized control instruction sequence to update the sequence to ensure that subsequent parameter regulation operations can proceed smoothly, and ensure the stability of the gold foil production process and the product quality. After performing the parameter regulation operation, enter the subsequent monitoring and optimization stage.

[0083] In a non-limiting embodiment, after the process optimization strategy set is fed back to the gold foil production control system to perform parameter regulation operations, it further includes: Step 210: Collect in real time the actual value set of process parameters and the corresponding gold foil quality inspection data set after performing the parameter regulation operation; compare each item of the actual value set of process parameters with the target parameter values in the process optimization strategy set, and calculate the real-time regulation deviation degree of each parameter; generate a dynamic compensation strategy set according to the real-time regulation deviation degree, where the dynamic compensation strategy set includes the compensation adjustment amount and priority for each parameter; perform conflict detection processing on the superposition execution of the dynamic compensation strategy set, identify parameter adjustment conflict pairs and generate conflict resolution weights; correct the dynamic compensation strategy set according to the conflict resolution weights to generate an optimized compensation instruction sequence; feedback the optimized compensation instruction sequence to the gold foil production control system to perform compensation regulation operations, and simultaneously update the process optimization strategy set.

[0084] After the parameter regulation operation is executed, the system starts to collect data in real time. For example, at regular time intervals (such as every 5 minutes), collect the actual value set of process parameters, including the actual value A4 of temperature parameter A, the actual value B5 of pressure parameter B, the actual value C3 of rolling speed parameter C, etc. At the same time, conduct quality inspection on the produced gold foil to obtain the corresponding gold foil quality inspection data set, such as data on the number of scratches on the gold foil surface, thickness deviation, etc. Then, compare the actual value set of process parameters with the target parameter values in the process optimization strategy set. For example, the target value of temperature parameter A in the process optimization strategy set is A3, and the actual value is A4. Calculate the real-time regulation deviation degree = |A4 - A3| / A3. Generate a dynamic compensation strategy set according to the real-time regulation deviation degree of each parameter. For example, for temperature parameter A, if the deviation degree is large, determine the compensation adjustment amount as ΔA, and determine a higher priority according to its influence on the gold foil quality; for rolling speed parameter C, the deviation degree is small, the compensation adjustment amount is ΔC, and the priority is low. Then perform conflict detection processing on the superposition execution of the dynamic compensation strategy set. For example, there are two compensation strategies, one is to increase temperature parameter A, and the other is to decrease temperature parameter A. These two strategies conflict, which is identified as a parameter adjustment conflict pair, and generate conflict resolution weights according to the severity of the conflict. For example, the strategy of increasing temperature parameter A is more beneficial to improving the gold foil quality, the conflict resolution weight is 0.8, and the weight of decreasing temperature parameter A is 0.2. Correct the dynamic compensation strategy set according to these weights to generate an optimized compensation instruction sequence. Finally, feedback the optimized compensation instruction sequence to the gold foil production control system to perform compensation regulation operations, and simultaneously update the process optimization strategy set to enable the strategy set to be continuously optimized according to the actual situation.

[0085] In a non-limiting embodiment, after the process optimization strategy set is feedback to the gold foil production control system to perform parameter regulation operations, it further includes: Step 220: Collect the defect type distribution data of the gold foil product after parameter regulation and the process parameter adjustment records, and construct an incremental training dataset; perform incremental feature extraction processing on the quality description statements in the incremental training dataset to generate incremental context semantic features; perform incremental parameter parsing processing on the adjusted process parameter records to generate incremental parameter association features; perform feature fusion processing on the incremental context semantic features and the historical context semantic features to generate a global semantic feature matrix; perform time-series splicing processing on the incremental parameter association features and the historical parameter association features to generate a global parameter feature matrix; perform incremental training on the dynamic matching algorithm based on the global semantic feature matrix and the global parameter feature matrix, and update the mapping relationship between the defect feature space and the regulation feature space; perform real-time annotation processing on the new quality description data set according to the updated dynamic matching algorithm, generate incremental quality annotation results, and update the process optimization strategy set.

[0086] In this embodiment, the system collects the data after parameter regulation, such as the defect type distribution data of the gold foil product shows that the proportion of "scratch defects" decreases and the proportion of "non-uniform thickness defects" changes, etc., and the process parameter adjustment records, such as the temperature parameter A is adjusted from A1 to A2, the pressure parameter B is adjusted from B1 to B2, etc. Use these data to construct an incremental training dataset. Perform incremental feature extraction processing on the quality description statements in the dataset. For example, for the newly emerged quality description statement "the edge of the gold foil has slight curling", according to the previous feature extraction method, generate incremental context semantic features. Perform incremental parameter parsing processing on the adjusted process parameter records to obtain incremental parameter association features. Fuse the incremental context semantic features and the historical context semantic features. For example, splice or perform weighted summation operations on the newly generated feature vector and the previously accumulated feature vector to generate a global semantic feature matrix. Similarly, perform time-series splicing on the incremental parameter association features and the historical parameter association features to form a global parameter feature matrix. Perform incremental training on the dynamic matching algorithm based on these two matrices, and update the mapping relationship between the defect feature space and the regulation feature space by adjusting the parameters of the algorithm, etc. Finally, use the updated algorithm to perform real-time annotation processing on the new quality description data set, generate incremental quality annotation results, and update the process optimization strategy set according to these results, so that the system can continuously adapt to new data and situations, and improve the ability to detect the quality of gold foil products and optimize the process.

[0087] In a non-limiting embodiment, after the process optimization strategy set is fed back to the gold foil production control system to perform parameter regulation operations, it further includes: Step 230: Obtain the quality evaluation data of multiple batches of gold foil products after the parameter regulation operation. The evaluation data includes the defect type statistical value, the process parameter fluctuation range, and the production energy consumption index; perform normalization processing on the defect type statistical value to generate a defect improvement rate, perform standardization processing on the process parameter fluctuation range to generate a parameter stability coefficient, and perform dimensionless conversion on the production energy consumption index to generate an energy consumption efficiency factor; perform weighted summation on the defect improvement rate, the parameter stability coefficient, and the energy consumption efficiency factor according to the preset weight configuration to generate a comprehensive optimization evaluation value; if the comprehensive optimization evaluation value is lower than the preset optimization threshold, perform multi-objective optimization processing on the process optimization strategy set, and adjust the parameter adjustment direction weight to balance the conflict between the defect improvement rate and the energy consumption efficiency factor; resend the multi-objective optimized process optimization strategy set to the gold foil production control system, and iteratively execute the parameter regulation operation and the evaluation process.

[0088] Optionally, the system obtains quality assessment data for multiple batches of gold foil products. For example, count the number of various defects such as "scratch defects" and "non-uniform thickness defects" in multiple batches of gold foil products, calculate the statistical values of defect types; record the fluctuation range of process parameters after parameter adjustment operations, such as the temperature parameter A fluctuates between A2 - A3; count the energy consumption indicators during the production process, such as the total power consumption. Normalize the statistical values of defect types. For example, compare the number of defects in the current batch with the number of defects before adjustment, and calculate the defect improvement rate = (number of defects before adjustment - number of current defects) / number of defects before adjustment. Standardize the fluctuation range of process parameters, and convert the fluctuation range into a parameter stability coefficient through the conversion formula commonly used in related technologies. Perform dimensionless conversion on the production energy consumption indicators, such as converting the total power consumption into an energy consumption efficiency factor. According to the preset weight configuration, set the weight of the defect improvement rate to 0.4, the weight of the parameter stability coefficient to 0.3, and the weight of the energy consumption efficiency factor to 0.3, and perform weighted summation to obtain the comprehensive optimization evaluation value. If the comprehensive optimization evaluation value is lower than the preset optimization threshold, it indicates that the current set of process optimization strategies is not ideal and multi-objective optimization processing is required. For example, when it is found that the defect improvement rate increases but the energy consumption efficiency factor decreases, adjust the weight of the parameter adjustment direction, such as appropriately reducing the adjustment intensity of the parameter that has a greater impact on energy consumption and increasing the weight of the parameter adjustment that is more effective in improving defects. Resend the set of process optimization strategies after multi-objective optimization to the gold foil production control system, and then perform parameter adjustment operations and evaluation processing again, and iterate in this cycle. After the execution of the parameter adjustment operation in the new round, the system obtains the quality assessment data of multiple batches of gold foil products again, repeats the above steps of normalization, standardization, dimensionless conversion, and weighted summation, and calculates the new comprehensive optimization evaluation value. If the new evaluation value still does not reach the preset optimization threshold, continue to perform multi-objective optimization processing on the set of process optimization strategies and further adjust the weight of the parameter adjustment direction. For example, if it is found through analysis that "increasing the rolling speed parameter C" can effectively reduce the probability of "scratch defects" but causes a significant increase in production energy consumption, during multi-objective optimization, appropriately reduce the weight of "increasing the rolling speed parameter C", and at the same time increase the weight of other parameter adjustment directions that can both improve defects and reduce energy consumption, such as the weight of "fine-tuning the pressure parameter B to optimize the rolling process".

[0089] Resend the adjusted set of process optimization strategies to the gold foil production control system to perform parameter adjustment operations. During this process, the system continuously monitors the execution status of the parameter adjustment operations in real time to ensure that each control instruction can be executed accurately and without error. If any abnormal situation occurs during the execution, such as the equipment feedback that a certain parameter cannot be adjusted to the target value, the system will immediately generate corresponding processing instructions according to the type of abnormality, such as a rollback instruction to restore the parameter to the state before adjustment, or generate an adaptive adjustment instruction to try to adjust the target value or execution time, etc., to ensure the continuity and stability of the production process.

[0090] In addition, in each iteration process, the system will also collect the actual values of process parameters and the corresponding gold foil quality inspection data set after performing the parameter regulation operation in real time in the manner of step 210 and step 220, generate and process the dynamic compensation strategy, and construct an incremental training data set to perform incremental training on the dynamic matching algorithm. By continuously collecting new data, updating the algorithm, and adjusting the strategy, the system can more accurately grasp the relationship between defect types and process parameters in the gold foil production process, gradually optimize the process parameters, improve the quality of gold foil products, reduce production energy consumption, and enhance the stability and efficiency of the entire production system.

[0091] In the above continuous iteration process, each parameter regulation and evaluation is based on actual production data, fully considering various factors such as the quality of gold foil products, the stability of process parameters, and production energy consumption. Through continuous adjustment and optimization, the process optimization strategy set will become more and more perfect and can better adapt to various changes and requirements in the gold foil production process. For example, due to factors such as the aging of production equipment and the fluctuation of raw material quality, the system can reasonably adjust the process parameters in a timely manner according to the collected data to ensure that the gold foil products always maintain a high quality level.

[0092] Moreover, with the continuous accumulation of data and the continuous optimization of the algorithm, the quality annotation results of the dynamic matching algorithm for quality description statements will be more accurate, and can more precisely indicate the mapping relationship between the defect types of gold foil products and the adjustment direction of process parameters. This will further improve the effectiveness and pertinence of the process optimization strategy set, enabling the gold foil production control system to perform parameter regulation operations more scientifically and reasonably, thereby realizing the intelligentization, high efficiency, and high quality of the gold foil production process. The entire process forms a virtuous cycle, continuously promoting the development of the gold foil production process to a higher level and meeting the market's demand for high-quality gold foil products.

[0093] It is worth mentioning that when implementing the above technical solutions, those skilled in the art can perform normalization processing on the process parameter features based on the feature standardization methods in the prior art (such as Z-score or Min-Max standardization). By converting the slope features and variance features of different dimension parameters such as temperature, pressure, and rolling speed into dimensionless standardized values, the influence of magnitude differences on the calculation of the cross-category trend correlation matrix is eliminated, ensuring the calculation effectiveness of statistical quantities such as the Pearson correlation coefficient.

[0094] For the dynamic matching algorithm, the cross-attention mechanism in the Transformer architecture can be used for reference to implement a dual-channel alignment network. Specifically, the BERT pre-trained model is used as the language feature encoder and fine-tuned based on domain corpora. When constructing the semantic dependency tree, the Stanford CoreNLP tool is introduced for syntax analysis. At the same time, the LSTM time-series network is used to simulate the production process evolution after process parameter adjustment. The accurate prediction of the parameter change curve is realized by defining physical models such as the rolling force-temperature coupling equation.

[0095] For the further optimization of the logical closed-loop, a multi-task joint optimization framework can be constructed to integrate the generation of quality annotation results and the evaluation of energy consumption efficiency into a unified loss function. The NSGA-II multi-objective optimization algorithm is used to balance the contradiction between the defect improvement rate and the energy consumption index, and a synchronous update mechanism for parameter adjustment weights and feature space mapping is designed. The dynamic adaptation of the policy weights during the incremental training process is realized by using the model hot update interface of TensorFlow.

[0096] In the process simulation and verification link, a finite element model of the gold foil rolling process can be established based on the ANSYS simulation platform. By setting the material constitutive equation and contact boundary conditions, the stress distribution and surface deformation after parameter adjustment are accurately simulated. Combining with the OpenCV image processing technology, scratch detection and thickness measurement are carried out on the microstructures generated by the simulation, so as to verify the quality improvement effect.

[0097] In addition, the real-time interaction between the optimization control instruction and the production equipment PLC can be realized by integrating the OPC UA industrial communication protocol. The Modbus TCP protocol is used to collect sensor data and feedback it to the compensation strategy generation module, and finally a complete technical closed-loop from quality description analysis to parameter dynamic regulation is formed.

[0098] In summary, through innovative technical means, the above technical solutions can deeply explore the potential relationship between quality description and process parameters, generate accurate quality annotation results and process optimization strategies, effectively solve the problems that the existing technologies are difficult to comprehensively analyze quality problems and accurately optimize the process in the production quality inspection of gold foil products, and improve the quality and efficiency of gold foil production.

[0099] Specifically, the embodiments of the present invention achieve in-depth mining and effective utilization of the quality data of gold foil products: by obtaining a quality description data set covering quality description statements and process parameter records, comprehensive materials can be provided for subsequent analysis; based on feature extraction processing, the semantic features of sentence context and parameter correlation features can be accurately captured; based on the multi-dimensional strategy analysis processing of the dynamic matching algorithm, quality annotation results can be innovatively generated, thus clearly presenting the mapping relationship between the defect types of gold foil products and the adjustment directions of process parameters; according to the quality annotation results, a process optimization strategy set is generated and fed back to the production control system to execute parameter regulation, which can quickly and accurately optimize the gold foil production process, improve product quality, reduce the defect rate, and enhance the intelligent and automated level of the production process.

[0100] See Figure 2 As shown, this figure is a schematic diagram of the basic structure of a data annotation system 200 provided by an embodiment of the present invention. The data annotation system 200 includes: A processor 201; A storage device 202, on which a computer program 2020 is stored; When the computer program 2020 is executed by the processor 201, the processor 201 implements any of the data annotation methods applied to gold foil product detection.

[0101] On this basis, a readable storage medium is provided. Programs or instructions are stored on the readable storage medium, and when the programs or instructions are executed by a processor, the steps of the above method are implemented.

[0102] It should be noted that the embodiments in this specification are described in a progressive manner. The key point of each embodiment is the difference from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the systems or devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method part.

Claims

1. A data annotation method applied to the detection of gold foil products, characterized in that, Including: Obtain a quality description data set of gold foil products, where the quality description data set includes multiple quality description statements and corresponding process parameter records; Perform feature extraction processing on the quality description data set to obtain the context semantic features of each quality description statement and the parameter association features of the process parameter records; Based on a dynamic matching algorithm, perform multi-dimensional strategy analysis processing on the context semantic features and the parameter association features to generate a quality annotation result for the quality description statement; The quality annotation result is used to indicate the mapping relationship between the gold foil product defect type and the process parameter adjustment direction; Generate a process optimization strategy set according to the quality annotation result, and feedback the process optimization strategy set to the gold foil production control system to perform parameter regulation operations.

2. The method according to claim 1, characterized in that, The performing feature extraction processing on the quality description data set to obtain the context semantic features of each quality description statement and the parameter association features of the process parameter records includes: Perform semantic unit segmentation processing on the quality description statement to obtain multiple semantic unit sequences; each semantic unit sequence contains at least one semantic segment and the position identifier of the semantic segment in the statement; Call a pre-trained language feature encoder to perform context encoding processing on the semantic unit sequence to generate the context semantic features of the quality description statement; the context semantic features include the local semantic features of each semantic segment and the context dependence relationship between adjacent semantic segments; Perform parameter parsing processing on the process parameter records to extract parameter category identifiers and parameter time series change patterns; the parameter time series change patterns include the fluctuation characteristics of parameter values changing over time and the associated fluctuation trends between different parameters; Based on the parameter category identifiers, perform feature fusion processing on the parameter time series change patterns to generate the parameter association features of the process parameter records; the parameter association features include the multi-dimensional fluctuation correlation under the same parameter category and the cross-category influence weights between different parameter categories.

3. The method according to claim 2, characterized in that, The calling a pre-trained language feature encoder to perform context encoding processing on the semantic unit sequence to generate the context semantic features of the quality description statement includes: Perform part-of-speech tagging processing on the semantic segments in each semantic unit sequence to generate a tagging sequence containing part-of-speech tags; Construct a semantic dependency tree according to the tagging sequence to determine the subject-predicate relationship, modification relationship, and parallel relationship between the semantic segments; Perform hierarchical encoding processing on the semantic unit sequence based on the semantic dependency tree: perform local encoding on each semantic segment to generate an initial semantic vector, perform relationship aggregation on the initial semantic vector according to the subject-predicate relationship to generate an intermediate semantic vector, and perform context enhancement on the intermediate semantic vector based on the modification relationship and parallel relationship to generate an enhanced semantic vector; Arrange the enhanced semantic vectors in time series according to the position identifier to generate the context semantic features of the quality description statement.

4. The method according to claim 2, characterized in that, The performing parameter parsing processing on the process parameter records to extract parameter category identifiers and parameter time series change patterns includes: Divide the process parameter records into multiple parameter subsequences according to time windows, and each parameter subsequence contains a set of parameter values within a preset time range; Perform trend analysis on each parameter subsequence, and extract the rising trend segment, falling trend segment and stable trend segment; Perform slope calculation on the rising trend segment to generate a first slope feature set; perform absolute value calculation on the falling trend segment to generate a second slope feature set; perform variance calculation on the stable trend segment to generate a fluctuation variance feature set; According to the parameter category identifier, classify the first slope feature set, the second slope feature set and the fluctuation variance feature set to generate an intra-category trend feature matrix; Perform correlation calculation on the intra-category trend feature matrices between different parameter categories to generate an inter-category trend correlation matrix; Concatenate the intra-category trend feature matrix and the inter-category trend correlation matrix to generate the parameter time series change pattern.

5. The method according to claim 1, characterized in that, The multi-dimensional strategy analysis of the context semantic feature and the parameter association feature based on the dynamic matching algorithm generates the quality annotation result of the quality description statement, including: Map the context semantic feature to the defect feature space to generate a defect feature vector containing the probability distribution of defect types; Map the parameter association feature to the regulation feature space to generate a regulation feature vector containing the weight of the parameter adjustment direction; Perform spatial alignment on the defect feature vector and the regulation feature vector to determine the mapping relationship between the probability distribution of defect types and the weight of the parameter adjustment direction; Construct a dynamic matching matrix according to the mapping relationship, and perform singular value decomposition on the dynamic matching matrix to obtain the defect type principal component vector and the parameter adjustment principal component vector; According to the similarity calculation between the defect type principal component vector and the preset defect type standard vector, determine the defect type annotation of the quality description statement, and the defect type annotation is used to characterize the defect type of the gold foil product; According to the direction matching between the parameter adjustment principal component vector and the historical optimization strategy vector, determine the parameter adjustment direction annotation, and the parameter adjustment direction annotation is used to characterize the process parameter adjustment direction; Combine the defect type annotation and the parameter adjustment direction annotation to generate the quality annotation result of the quality description statement.

6. The method according to claim 5, characterized in that, The spatial alignment of the defect feature vector and the regulation feature vector to determine the mapping relationship between the probability distribution of defect types and the weight of the parameter adjustment direction includes: Construct a two-channel alignment network including the defect type dimension and the parameter adjustment dimension; Input the defect feature vector into the defect type channel for feature enhancement to generate an enhanced defect feature vector; Input the regulation feature vector into the parameter adjustment channel for feature enhancement to generate an enhanced regulation feature vector; Perform cross-attention calculation on the enhanced defect feature vector and the enhanced regulation feature vector to generate a defect-parameter attention weight matrix; Perform a weighted summation process on the enhanced defect feature vector and the enhanced regulation feature vector according to the defect-parameter attention weight matrix to generate a joint feature vector; Perform a fully connected mapping process on the joint feature vector to generate a mapping relationship between the defect type probability distribution and the parameter adjustment direction weight.

7. The method according to claim 1, wherein The generating of the process optimization strategy set according to the quality annotation result includes: Extract a defect type priority list and a parameter adjustment direction priority list from the quality annotation result; Determine a first optimization weight according to the defect type priority list and a second optimization weight according to the parameter adjustment direction priority list; Perform a weighted fusion process on the first optimization weight and the second optimization weight to generate a comprehensive optimization weight; Perform a screening process on the historical process parameter adjustment records according to the comprehensive optimization weight to obtain a candidate adjustment strategy set; Perform a feasibility evaluation process on the candidate adjustment strategy set to generate a result list including evaluation weights; Perform a sorting process on the result list according to the evaluation weights, and select candidate adjustment strategies exceeding a preset score threshold to generate the process optimization strategy set.

8. The method according to claim 7, wherein The performing of the feasibility evaluation process on the candidate adjustment strategy set to generate a result list including evaluation weights includes: Split each candidate adjustment strategy into a parameter adjustment step sequence, and each parameter adjustment step includes an adjustment target value and an adjustment time range; Perform a conflict detection process on the parameter adjustment step sequence to identify conflict step pairs with time overlap or target value contradictions; Generate a conflict evaluation score according to the number and severity of the conflict step pairs; Perform a simulated execution process on the parameter adjustment step sequence to predict the parameter change curve and the defect type probability change curve after execution; Generate a parameter matching score according to the matching degree between the parameter change curve and the target parameter interval; generate a defect improvement score according to the difference degree between the defect type probability change curve and the target defect threshold; Perform a weighted summation process on the conflict evaluation score, the parameter matching score, and the defect improvement score to generate a result list of the evaluation weights of the candidate adjustment strategies.

9. The method according to claim 1, wherein The feeding back of the process optimization strategy set to the gold foil production control system to perform parameter regulation operations includes: Convert the process optimization strategy set into a control instruction sequence, and each control instruction includes a parameter identifier, a target value, and an execution timestamp; Perform a timing verification process on the control instruction sequence to ensure that the execution timestamp meets the operation interval limit of the production equipment; Perform a timestamp adjustment process on the control instruction sequence according to the verification result to generate an optimized control instruction sequence; Send the optimized control instruction sequence to the parameter execution module of the gold foil production control system, and instruct the parameter execution module to perform parameter regulation operations according to the optimized control instruction sequence; Real-time monitor the execution status of the parameter regulation operation, and when an execution anomaly is detected, generate a rollback instruction or an adaptive adjustment instruction according to the anomaly type to update the optimized control instruction sequence.

10. A data annotation system, wherein including: a processor; A storage device on which a computer program is stored. When the computer program is executed by the processor, the processor implements the data annotation method applied to the detection of gold foil products as described in any one of claims 1-9.

Images