A quality defect prediction method based on data coupling micro-feature correlation analysis

Abstract

The application discloses a quality defect prediction method based on data coupling micro-feature correlation analysis, constructs a data coupling micro-feature correlation analysis method, divides intervals of quality characteristic coupling data, identifies the correlation between the quality characteristic coupling data, extracts significant defect feature data through the confidence threshold, builds a manufacturing process quality defect prediction model based on the coupling data micro-feature correlation rule, and completes the prediction of the quality defect. The method can predict the manufacturing quality in real time, analyzes and predicts the key quality characteristic data leading to the output quality defect and the manufacturing process multi-element input parameter affecting the quality characteristic, is beneficial to the quality decoupling of the multi-element quality characteristic coupling data, finds the quality defect formation mechanism, performs multi-chain tracing, has good cause explanation, and can effectively solve the problems of interval division of the quality characteristic coupling data and too many fuzzy intervals of the fuzzy algorithm in a small value range.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of manufacturing process quality prediction control, and particularly relates to a quality defect prediction method based on data coupling micro-feature correlation analysis. BACKGROUND

[0002] Various characteristics in a manufacturing process are always fluctuating, and in order to meet people's needs, the fluctuation needs to be within an acceptable range. System factors and accidental factors bring fluctuations to various quality characteristic data, and the latter often deviates from the original design of the manufacturing process and is the cause of quality defects. Under the condition of the wide application of various sensors and database systems, it is possible to comprehensively analyze quality defects by using coupled data of multiple quality characteristics. At the same time, due to the time sequence of the manufacturing process, the use of a quality prediction model can effectively reduce manufacturing costs by preventing possible quality defects in the manufacturing process. Effective decoupling of factors leading to quality defects, construction of a quality defect prediction model, and real-time prediction using key quality characteristic parameters can effectively achieve manufacturing quality control.

[0003] For quality defect analysis methods, there are methods of forming and control chart pattern recognition combining SPC represented by control charts, methods of judging quality defects by using neural networks, methods of extracting quality defect association rules based on GA, methods of combining fuzzy reasoning and Stateflow for quality defect analysis, and methods of using a case base to diagnose quality by measuring the similarity of cases based on the combination of CBR (Case Based Reasoning) and KNN (K-Nearest Neighbor Algorithm). However, most of the research focuses on the final characteristic output, and the model has poor interpretability, which is not conducive to quality decoupling, especially for neural network methods. For prediction models, there are widely used models such as Bayesian networks, random forests, LGBM (Light Gradient Boost Machine), multilayer perceptrons, RNN (Recurrent Neural Network), and deep neural networks, but the processing effect of data fuzziness is not very good. If fuzzy theory is combined for processing, it is found that too many fuzzy intervals may be generated in the actual test process, reducing the efficiency of the algorithm. SUMMARY

[0004] To solve the above technical problems, the application provides a quality defect prediction method based on data coupling micro-feature correlation analysis.

[0005] The technical scheme adopted by the application is as follows: a quality defect prediction method based on data coupling micro-feature correlation analysis, and the specific steps are as follows:

[0006] S1, construct a micro-feature correlation analysis method of coupling data, divide the quality characteristic coupling data into intervals and identify the correlation between the quality characteristic coupling data, and extract significant defect feature data from the high-frequency micro-feature item set through the confidence threshold level;

[0007] S2, build a manufacturing process quality defect prediction model based on the micro-feature correlation rules of coupling data, and complete the prediction of quality defects.

[0008] Further, in step S1, the following is specifically as follows:

[0009] Initialize k points μ = {μ1, μ2, …, μ k} as class centers.

[0010] For each quality defect coupling data micro-feature sample i, calculate the feature class it should belong to, and define the loss function:

[0011]

[0012] Where c represents the class set, μ represents the set center point, x i represents the i-th sample, c i represents the class to which x i belongs, μ ci represents the class corresponding center point, and n represents the total number of samples.

[0013] For each sample x i , assign it to the class to which the nearest center belongs:

[0014]

[0015] Where, represents the i-th class of the t-th iteration, represents the k-th center point of the t-th iteration.

[0016] For each center point μ i , iteratively calculate the micro-feature core value of the class

[0017]

[0018] Where μ represents the set center point.

[0019] In the case of k intervals of quality characteristic coupling data, the feature core value stabilizes in a small range after multiple iterations.

[0020] After determining the micro-feature interval of quality characteristic coupling data, determine the support domain value and the confidence domain value of the micro-feature correlation. According to the support domain value, generate a strong correlation rule, and according to the confidence domain value, generate a micro-feature high-frequency item set.

[0021] The support domain value is calculated by the ratio of the number of micro-feature patterns of a certain item set X to the total number of micro-feature pattern database D, and the calculation formula is:

[0022]

[0023] Wherein, T represents the target pattern number.

[0024] The minimum support domain value is the minimum support threshold set in the micro-feature rule item set, denoted as sup_min, and the item set with a support domain value greater than or equal to sup_min is called a micro-feature high-frequency item set.

[0025] Strong association rules are found by the confidence domain value, which means the ratio of the number of micro-feature patterns containing both item sets X and Y to the number of micro-feature patterns containing only X, and the calculation expression is:

[0026]

[0027] Further, in the step S1, the quality coupling data micro-feature correlation rule analysis task is specifically as follows:

[0028] Generate high-frequency pattern item set: for all micro-feature item sets that meet the minimum support threshold requirement, extract these micro-feature item sets, and call these item sets high-frequency micro-feature item sets.

[0029] Generate micro-feature strong correlation rules: extract rules from high-frequency micro-feature item sets according to the confidence threshold.

[0030] Further, in the step S2, the following is specifically as follows:

[0031] Map the coupling data micro-feature sample data to a multi-dimensional space and perform regression to fit a continuous function, i.e. a model function f(x) = wx + b, representing the sample point (x, y) on the linear surface of the normal vector w and the offset b, with the linear function as the center, and the two sides define the "micro-distance area", the samples in the micro-distance area are not considered in the fuzzy abnormal interval; while the samples outside the micro-distance area become members of the fuzzy abnormal interval function, finally, by minimizing the width of the micro-distance area, the fuzzy abnormal interval is reduced to the minimum, thereby establishing an optimized prediction model.

[0032] Introduce two isolation variables ξ (positive micro-distance area offset constant) and ξ* (negative micro-distance area offset constant) to make the model have a certain clarity to the sample points.

[0033] Wherein, f(x) = wx + b is the model function finally required; f(x) + ε and f(x) - ε represent the upper and lower limits of the micro-distance area, and ε represents the fuzzy abnormal interval coefficient.

[0034] The formula can be expressed as:

[0035]

[0036]

[0037] wherein, ξ i represents the positive micro-distance zone offset constant corresponding to the i-th sample; y i represents the y-axis vector corresponding to the i-th sample; represents the negative micro-distance zone offset constant corresponding to the i-th sample.

[0038] For any sample x i , if it is inside the micro-distance zone or on the boundary of the micro-distance zone, then both ξ and ξ * are 0; if it is above the upper boundary of the micro-distance zone, then ξ>0, ξ * =0; if it is below the lower boundary of the micro-distance zone, then ξ=0, ξ * >0.

[0039] The mathematical description of the optimization prediction problem can be expressed as:

[0040]

[0041] f(x i )-y i ≤ε+ξ i ;

[0042]

[0043]

[0044] wherein, the key quality characteristic coupling data micro-feature correlation rule constructs an input vector, takes manufacturing quality as output, and trains a quality prediction model. The training process optimizes hyperparameters using GA, and the optimized parameters are the correlation coefficients C and fuzzy abnormal interval coefficients ε of the quality prediction model.

[0045] The beneficial effects of the present application: the method of the present application builds a data coupling micro-feature correlation analysis method, divides the quality characteristic coupling data into intervals, identifies the correlation between the quality characteristic coupling data, extracts significant defect feature data through the confidence threshold, builds a manufacturing process quality defect prediction model based on the coupling data micro-feature correlation rules, and completes the prediction of quality defects. The method of the present application can predict manufacturing quality in real time, analyze and predict key quality characteristic data leading to output quality defects and manufacturing process multi-input parameters affecting the quality characteristics, facilitate quality decoupling of multi-quality characteristic coupling data, discover the quality defect formation mechanism, perform multi-chain traceability, have good cause explanation, and effectively solve the problems of interval division of quality characteristic coupling data and excessive fuzzy intervals of fuzzy algorithm in small value domain. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 A flowchart of a quality defect prediction method based on data coupling micro-feature correlation analysis of the present application.

[0047] Figure 2 A quality defect prediction model construction diagram based on coupling data micro-feature correlation rules in an embodiment of the present application.

[0048] Figure 3 A clustering diagram of mold temperature, melting temperature and injection time in an embodiment of the present application.

[0049] Figure 4 A pressure holding time and pressure holding temperature clustering diagram in an embodiment of the present application.

[0050] Figure 5 A GA-MFCCD and GA-ELM comparison diagram in an embodiment of the present application. DETAILED DESCRIPTION

[0051] The present application will be further described below in combination with the drawings and embodiments.

[0052] First, in Example 1, as shown in the flowchart of a quality defect prediction method based on data coupling micro-feature correlation analysis of the present application, the specific steps are as follows: Figure 1

[0053] S1, build a data coupling micro-feature correlation analysis method, divide the quality characteristic coupling data into intervals, identify the correlation between the quality characteristic coupling data, and extract significant defect feature data from the high-frequency micro-feature item set through the confidence threshold;

[0054] S2, build a manufacturing process quality defect prediction model based on the coupling data micro-feature correlation rules, and complete the prediction of quality defects. ​

[0055] In the embodiment, the step S1 is specifically as follows:

[0056] Initialize k points μ = {μ1, μ2, …, μ k} as class centers.

[0057] For each quality defect coupling data micro-feature example i, calculate the feature class it should belong to, define the loss function:

[0058]

[0059] Where c represents the class set, μ represents the set center point, x i represents the ith sample, c i represents the class to which x i belongs, μc represents the class corresponding center point, and n represents the total number of samples.

[0060] For each sample x i , it is assigned to the class to which the nearest center belongs:

[0061]

[0062] Where, μc represents the ith class in the tth iteration, μk represents the kth center point in the tth iteration.

[0063] For each center point μ i , iteratively calculate the micro-feature core value of the class:

[0064]

[0065] Where μ represents the set center point.

[0066] In the case where the quality characteristic coupling data has k intervals, the micro-feature core value is stable in a small range after multiple iterations.

[0067] After determining the micro-feature interval of the quality characteristic coupling data, the support value (support) and the confidence value (confidence) of the micro-feature correlation are determined. The strong rule is generated according to the support value, and the micro-feature frequent item set is generated according to the confidence value.

[0068] The support value is calculated by the ratio of the number of micro-feature patterns of a certain item set X to the total number of micro-feature pattern databases D, and the calculation formula is:

[0069]

[0070] wherein T represents the target mode number.

[0071] The minimum support domain value is the minimum support threshold set in the micro-feature rule set, denoted as sup_min, and the item set with a support domain value greater than or equal to sup_min is called a micro-feature high-frequency item set.

[0072] Strong association rules are found by the confidence domain value, which means the ratio of the number of micro-feature patterns containing both item sets X and Y to the number of micro-feature patterns containing only X, and the calculation expression is:

[0073]

[0074] In the present embodiment, in the step S1, the quality coupling data micro-feature correlation rule analysis task is specifically as follows:

[0075] High-frequency pattern item sets are generated: for all micro-feature item sets reaching the minimum support threshold requirement, the micro-feature item sets are extracted, and the item sets are called high-frequency micro-feature item sets.

[0076] Strong micro-feature correlation rules are generated: according to the confidence threshold, rules are extracted from the high-frequency micro-feature item sets.

[0077] In the present embodiment, in the step S2, the following is specifically as follows:

[0078] As shown in Figure 2 w represents the vector form of w, the coupling data micro-feature sample data is mapped to a multi-dimensional space, and regression is performed, so as to fit a continuous function, that is, a model function f(x) = wx + b, indicating that the sample point (x, y) is on the linear surface of the normal vector w and the offset b, and the linear function is the center, and the two sides define the "micro-distance area", and the sample points in the micro-distance area are not considered as the fuzzy abnormal interval; and the sample points outside the micro-distance area become members of the fuzzy abnormal interval function, and finally, by minimizing the width of the micro-distance area, the fuzzy abnormal interval is minimized, so as to establish an optimized prediction model.

[0079] Due to the fuzzy correlation of the quality coupling data, the micro-feature pattern sample points are all in the micro-distance area, in order to further clarify the interval, two isolation variables ξ (positive micro-distance area offset constant) and ξ * (negative micro-distance area offset constant) are introduced to make the model have a certain clarity to the sample points.

[0080] wherein f(x) = wx + b represents the model function finally required; f(x) + ε and f(x) - ε represent the upper and lower limits of the micro-distance area; and ε represents a fuzzy abnormal interval coefficient.

[0081] The formula can be expressed as:

[0082]

[0083]

[0084] wherein, ξ i represents the positive micro-distance zone offset constant corresponding to the i-th sample; y i represents the y-axis vector corresponding to the i-th sample; represents the negative micro-distance zone offset constant corresponding to the i-th sample;

[0085] For any sample x i , if it is inside the micro-distance zone or on the boundary of the micro-distance zone, then ξ and ξ * are both 0; if it is above the upper boundary of the micro-distance zone, then ξ>0, ξ * =0; if it is below the lower boundary of the micro-distance zone, then ξ=0, ξ * >0.

[0086] The mathematical description of the optimization prediction problem can be represented as:

[0087]

[0088] f(x i )-y i ≤ε+ξ i ;

[0089]

[0090]

[0091] wherein, the key quality characteristic coupling data micro-feature correlation rule constructs an input vector, takes the manufacturing quality as the output, and trains the quality prediction model. The training process optimizes the hyperparameters using GA (Genetic Algorithm), and the optimized parameters are the correlation coefficient C of the quality prediction model and the fuzzy abnormal interval coefficient ε.

[0092] The application also provides Example 2 for analyzing quality defects in the injection molding process, which is specifically as follows:

[0093] In the injection molding process of plastics, product defect analysis, product quality monitoring, and timely improvement of product parameters are performed in advance to meet the quality requirements of enterprises.

[0094] According to the actual injection molding production experience feedback, the following five are the key parameters affecting the quality of injection molding process: mold temperature T1 (℃), melt temperature T2 (℃), injection time S1 (s), holding time S2 (s), holding pressure P (%). High-density polyethylene HDPE (High-density Polyethylene) is used as the experimental material, and part of the original data is shown in Table 1:

[0095] Table 1

[0096]

[0097] As Figure 3 , Figure 4 shown, the original data contains normal quality products, concave products, and warped products. Therefore, according to the coupling data micro-feature correlation analysis method, the influencing factors are clustered to obtain the general classification of the influencing factors.

[0098] Among them, Figure 3 the mold temperature is the x-axis, the melt temperature is the y-axis, and the injection time data classification is the z-axis; Figure 4 the holding time is the x-axis, and the holding pressure data classification is the y-axis.

[0099] Since the classification result is not the most suitable result, based on the implementation results of the following algorithm, the interval range is adjusted, and the adjustment results are shown in Table 2:

[0100] Table 2

[0101]

[0102] From Table 2, there are a total of 12 data intervals, so the data interval representation is shown in Table 3:

[0103] Table 3

[0104]

[0105] According to the coupling data micro-feature correlation analysis method, the correlation rules between different intervals and the final quality results are obtained, and part of the association rules are shown in Table 4:

[0106] Table 4

[0107]

[0108] Therefore, according to the above three rules obtained by the coupling data micro-feature correlation analysis method, we can know that:

[0109] (1) For products with normal quality, the quality rule is:

[0110] (20.1≤T1≤58.3)∩(123.0≤T2≤256.2)∩(0.21≤S1≤1.19)∩(3.1≤S2≤10.2)∩(61.3≤P≤78.1)

[0111] (2) For the recessed product, the quality rule is:

[0112] (T1>62.1)∩(T2>273.3)∩(P<58.1)

[0113] (3) For the warped product, the quality rule is:

[0114] (T1>68.5)∩(S2>17.5)∩(P>84.4)

[0115] Therefore, in the production process of the injection molded product, attention should be paid to controlling these influencing factors within a reasonable range, so as to realize efficient management.

[0116] Without considering accidental factors such as interference, 100 groups of data generated in the production process are selected as the training set and the test set to predict the process quality, and finally the predicted value is obtained.

[0117] The prediction results and the prediction error percentage compared with the original data are obtained by the MFCCD model. Here, the GA-ELM model with optimized hidden layer adaptive growth is compared, and the specific comparison is shown in Table 4 and Table 5: Figure 5

[0118] Table 5

[0119]

[0120] RMSE (Root Mean Square Error) is the root mean square error, and the calculation formula is:

[0121]

[0122] Wherein, n represents the total number of samples, prediction represents the predicted value, and target represents the true value.

[0123] As shown in the figure and the table, the prediction error of GA-MFCCD is relatively small. In the prediction result of the GA-MFCCD method, the maximum error is 3.58%, and the minimum is 0.87%. Overall, the method of the present application can accurately predict the injection molded recessed diameter.

[0124] ​In summary, the method of the present application builds a data coupling micro-feature correlation analysis method, divides the quality characteristic coupling data into intervals, identifies the correlation between the quality characteristic coupling data, extracts significant defect feature data from the high-frequency micro-feature item set through the confidence threshold, builds a manufacturing process quality defect prediction model based on the coupling data micro-feature correlation rule, and completes the prediction of the quality defect. The method of the present application can predict the manufacturing quality in real time, analyze and predict the key quality characteristic data leading to the output quality defect and the manufacturing process multi-input parameters affecting the quality characteristic, is beneficial to the quality decoupling of multi-quality characteristic coupling data, discovers the quality defect formation mechanism, performs multi-chain tracing, has good cause explanation, and can effectively solve the problems of interval division of quality characteristic coupling data and too many fuzzy intervals of fuzzy algorithm in a small value domain.

Claims

1. A quality defect prediction method based on data coupling micro-feature correlation analysis, the specific steps being as follows: S1. Constructing a data coupling micro-feature correlation analysis method, dividing intervals of quality characteristic coupling data and identifying the correlation between the quality characteristic coupling data, extracting significant defect feature data from a high-frequency micro-feature item set through a confidence threshold level; wherein The product production process is an injection molding process, and the quality characteristic coupling data are mold temperature, melt temperature, injection time, holding time, and holding pressure; S2. Building a manufacturing process quality defect prediction model based on coupling data micro-feature correlation rules to complete the prediction of quality defects; In the step S1, the following is specifically performed: Initialize k points As class centers; For each quality defect coupling data micro-feature sample i, the feature class to which it should belong is calculated, and a loss function is defined: ； wherein c denotes a class set, denotes a set center point, denotes the i-th sample, denotes the class to which the sample belongs, denotes the class corresponding center point, and n denotes the total number of samples; For each sample assign it to the class the nearest center belongs to: ； wherein, denotes the i-th class at the t-th iteration, denotes the k-th center point at the t-th iteration; For each center point , iteratively calculate the microfeature core value of the class : ； wherein denotes the set center point; In the case that the quality characteristic coupling data has k intervals, the feature core value is iterated multiple times and stabilized in a small range; After determining the micro-feature interval of the quality characteristic coupling data, the support domain value and the confidence domain value of the micro-feature correlation are determined; the micro-feature high-frequency item set is generated according to the support domain value and the strong correlation rules are generated according to the confidence domain value; The support domain value is calculated by the ratio of the micro-feature mode number of an item set X to the total number of the micro-feature mode database D, and the calculation formula is: ； Wherein, T represents the target mode number; The minimum support domain value is the minimum support threshold value set in the micro-feature rule item set, denoted as sup_min, and the item set with a support domain value greater than or equal to sup_min is called a micro-feature high-frequency item set; The strong correlation rules are discovered through the confidence domain value, and the meaning is the ratio of the micro-feature mode number containing both item sets X and Y to the micro-feature mode number containing only X, and the calculation expression is: 。 2. The method of claim 1, wherein, In the step S1, in the quality coupling data micro-feature correlation rule analysis task, the following is specifically performed: Generating a high-frequency mode item set: for all micro-feature item sets that meet the minimum support threshold requirement, the micro-feature item sets are extracted, and the item sets are called high-frequency micro-feature item sets; Generating micro-feature strong correlation rules: rules are extracted from the high-frequency micro-feature item set according to the confidence threshold level.

3. The method of claim 1, wherein the method is characterized by: In the step S2, the following is specifically performed: The coupling data micro-feature sample data is mapped to a multi-dimensional space and is regressed to fit a continuous function, i.e. a model function , representing a sample point (x, y) on a linear plane of a normal vector w and an offset b, centered on a linear function, with two sides defining a "micro-distance area", samples in the micro-distance area are not considered as fuzzy abnormal interval; while samples outside the micro-distance area become members of the fuzzy abnormal interval function, finally, by minimizing the width of the micro-distance area, the fuzzy abnormal interval is minimized to establish an optimized prediction model; Introduce two isolated variables, positive micro-distance area offset constant and negative micro-distance area offset constant So that the model has a certain clarity to the sample points; wherein is the final model function sought; with denotes the upper and lower limits of the macro region, denotes the fuzzy abnormal interval coefficient; The formula can be expressed as: ； ； wherein, represents a positive microzone offset constant corresponding to the i-th sample; represents a y-axis vector corresponding to the i-th sample; represents a negative microzone offset constant corresponding to the i-th sample; For any sample If it is inside the macro region or on the boundary of the macro region, then and Both are 0; if it is above the boundary of the macro region, then >0, =0; if it is below the lower boundary of the macro region, then =0, >0; The mathematical description of the optimal prediction problem can be expressed as: ； The key quality characteristic coupling data micro-feature correlation rule constructs an input vector to take manufacturing quality as an output, trains a quality prediction model, and uses GA to optimize hyperparameters in a training process, and the optimized parameters are correlation coefficients of the quality prediction model and a fuzzy abnormal interval coefficient .

Images