Product defect prediction method and device, electronic device, and storage medium

Abstract

The present disclosure is about a product defect prediction method and device, electronic device, storage medium, and storage medium. The method includes: obtaining parameter data of each device parameter of the target device within a preset time period; the parameter data refers to the value of each device parameter recorded at a preset interval when the product passes through the target device stage; obtaining a fusion feature characterizing the characteristics of the target device according to the parameter data; inputting the fusion feature into a preset prediction model to obtain an inference result output by the prediction model that characterizes whether the product is abnormal. This embodiment can predict whether abnormal products will be produced during the production process, so that management personnel can adjust the target equipment in advance and improve production yield.

Description

Technical Field

[0001] The present disclosure relates to the field of data processing technology, and in particular to a product defect prediction method and device, electronic equipment, and storage medium. Background Art

[0002] At present, the production line of OLED products includes several process equipment, and multiple inspection stations are set up on the entire production line. Multiple process equipment are usually set up in the inspection station. When a defect is found at the inspection station, it is necessary to trace back to the process station and locate the process equipment. The management personnel need to spend a lot of time to check each process equipment one by one. In addition, after locating the abnormal process equipment, the management personnel need to spend a lot of time to pay attention to it until it is fully confirmed that it is working abnormally. In addition, the abnormal product may have an abnormality in one of the process equipment in the site, and the abnormality is usually discovered in the last equipment in the site, that is, it takes a certain amount of time from the occurrence of the abnormality in production to the discovery of the abnormality. The above situations accumulate, resulting in a lot of time spent on locating the equipment that causes the defect. Summary of the invention

[0003] The present disclosure provides a product defect prediction method and device, electronic equipment, storage medium, and storage medium to solve the deficiencies of related technologies.

[0004] According to a first aspect of an embodiment of the present disclosure, a method for predicting product defects is provided, the method comprising:

[0005] Obtaining parameter data of each device parameter of the target device within a preset time period; the parameter data refers to the value of each device parameter recorded at preset intervals during the product's passage through the target device stage;

[0006] Acquire a fusion feature characterizing a feature of the target device according to the parameter data;

[0007] The fused features are input into a preset prediction model to obtain an inference result output by the prediction model that characterizes whether the product is abnormal.

[0008] Optionally, the prediction model includes one of the following: logistic regression, random forest, LightGBM and Xgboost.

[0009] Optionally, obtaining parameter data of each device parameter of the target device within a preset time period includes:

[0010] Get the number of parameter data obtained by different products through the target equipment stage;

[0011] The number of the parameter data is adjusted based on the number of data to obtain a preset number of parameter data.

[0012] Optionally, acquiring a fusion feature characterizing a feature of the target device according to the parameter data includes:

[0013] Obtaining statistical characteristics of the parameter data, and obtaining time-frequency characteristics of the parameter data;

[0014] The time-frequency feature is inserted into the statistical feature to fuse the statistical feature and the time-frequency feature to obtain a fused feature of the target device.

[0015] Optionally, obtaining the time-frequency characteristics of the parameter data includes:

[0016] Obtaining a time series of each device parameter from the parameter data;

[0017] Performing wavelet packet decomposition on the time series into a preset number of layer frequency bands, wherein the preset number of layers includes a designated layer;

[0018] Reconstruct the signal of the specified layer frequency band to obtain a reconstructed signal;

[0019] The energy value of each frequency band is obtained, and the maximum energy value is used as the time-frequency feature of the parameter data.

[0020] Optionally, the method further comprises:

[0021] When the inference result indicates that the product is abnormal, obtaining a second number of candidate parameters of the target device;

[0022] Acquire characteristic data representing the target parameter having the greatest impact on the product from among the second number of candidate parameters;

[0023] Output the feature data of the second number of candidate parameters and the target parameter.

[0024] Optionally, obtaining a second number of candidate parameters of the target device includes:

[0025] Obtaining a weight value of each device parameter in the target device based on the prediction model;

[0026] Sort each device parameter in the target device from high to low according to the weight value;

[0027] A second number of device parameters ranked top are determined as candidate parameters.

[0028] Optionally, when the prediction model is a random forest, obtaining a second number of candidate parameters of the target device includes:

[0029] Based on the preset out-of-bag data, obtain the first out-of-bag data error of each decision tree in the random forest;

[0030] Adding noise interference to the out-of-bag data to change the value of each device parameter of the sample, and obtaining a second out-of-bag data error of each decision tree in the random forest;

[0031] Calculating the importance of each device parameter based on the second out-of-bag data error, the first out-of-bag error, and the number of decision trees in the random forest;

[0032] A second number of device parameters ranked high in importance are determined as candidate parameters of the target device.

[0033] Optionally, the method further includes the step of training a preset prediction model, specifically including:

[0034] Acquire a training sample set; the training sample set includes multiple training samples, each training sample includes a fusion feature of a process device and a sample label;

[0035] Input each training sample in the training sample set into the prediction model to be trained to obtain a trained prediction model; the number of the prediction models to be trained is a first number; the first number of prediction models includes logistic regression, random forest, LightGBM and Xgboost;

[0036] Obtaining an evaluation score of each prediction model in the first number of prediction models based on a preset evaluation index;

[0037] The prediction model with the largest evaluation score is determined as the preset prediction model.

[0038] Optionally, before obtaining the training sample set, the method further includes:

[0039] Obtain the number of defective points of products produced by the testing site within a specified time period;

[0040] Determine the sample classification of each product based on the defective points and the point threshold, the sample classification includes positive samples and negative samples; the positive sample refers to the product whose defective points are less than or equal to the point threshold, and the negative sample refers to the product whose defective points are greater than the point threshold;

[0041] Rank the importance of each process equipment to locate the equipment that causes a high incidence of bad samples.

[0042] According to a second aspect of an embodiment of the present disclosure, a device for predicting defective products is provided, the device comprising:

[0043] A parameter data acquisition module, used to acquire parameter data of each device parameter of the target device within a preset time period; the parameter data refers to the value of each device parameter recorded at preset intervals during the product passing through the target device stage;

[0044] A fusion feature acquisition module, used to acquire a fusion feature representing a feature of the target device according to the parameter data;

[0045] The inference result acquisition module is used to input the fusion feature into a preset prediction model to obtain the inference result output by the prediction model that characterizes whether the product is abnormal.

[0046] According to a third aspect of an embodiment of the present disclosure, there is provided an electronic device, including:

[0047] processor;

[0048] a memory for storing a computer program executable by the processor;

[0049] The processor is configured to execute the computer program in the memory to implement the above method.

[0050] According to a fourth aspect of an embodiment of the present disclosure, a computer-readable storage medium is provided, and when an executable computer program in the storage medium is executed by a processor, the above method can be implemented.

[0051] The technical solution provided by the embodiments of the present disclosure may have the following beneficial effects:

[0052] As can be seen from the above embodiments, the parameter data of each device parameter of the target device within a preset time period can be obtained in the disclosed embodiments; the parameter data refers to the value of each device parameter recorded at preset intervals during the product passing through the target device stage; then, the fusion feature characterizing the characteristics of the target device can be obtained based on the parameter data; thereafter, the fusion feature can be input into a preset prediction model to obtain the inference result output by the prediction model characterizing whether the product is abnormal. In this way, the present embodiment can predict whether abnormal products will be produced during the production process, so that the management personnel can adjust the target device in advance and improve the production yield.

[0053] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and, together with the description, serve to explain the principles of the present disclosure.

[0055] Figure 1 The figure is a flow chart of a method for predicting a product defect according to an exemplary embodiment.

[0056] Figure 2 The figure is a flowchart of adjusting parameter data according to an exemplary embodiment.

[0057] Figure 3 The figure is a flow chart of obtaining fusion features according to an exemplary embodiment.

[0058] Figure 4 The figure is a flow chart of obtaining time-frequency features according to an exemplary embodiment.

[0059] Figure 5 The figure is a flowchart of obtaining a preset prediction model according to an exemplary embodiment.

[0060] Figure 6 The figure is a flow chart of another method for predicting product defects according to an exemplary embodiment.

[0061] Figure 7 is a flowchart of obtaining candidate parameters according to an exemplary embodiment.

[0062] Figure 8 The figure is a block diagram of a device for predicting defective products according to an exemplary embodiment. DETAILED DESCRIPTION

[0063] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The embodiments described below by way of example do not represent all embodiments consistent with the present disclosure. Instead, they are merely examples of devices consistent with some aspects of the present disclosure as detailed in the appended claims.

[0064] In order to solve the above technical problems, this embodiment provides a product defect prediction method, which can be applied to various process equipment or defective systems of production lines of products such as OLED, MiniLED, and LCD. The following description will be given by taking various process equipment as an example, and the process equipment executing the prediction method will be referred to as the target equipment. For the convenience of description, the product is described as an OLED product.

[0065] Figure 1 is a flowchart of a method for predicting product defects according to an exemplary embodiment. Figure 1 , a product defect prediction method, comprising steps 11 to 13, wherein:

[0066] In step 11, parameter data of each device parameter of the target device within a preset time period is obtained.

[0067] In this embodiment, the OLED product production line can produce different types of display panels, and each type of display panel can pass through different process equipment. For example, display panel A passes through process equipment a, b, c and d, and display panel B passes through process equipment a, c and d. In addition, the equipment parameters required for different types of display panels to pass through the same process equipment may also be different. For example, the temperature required for display panel A to pass through process equipment a is 36 degrees Celsius, and the temperature required for display panel B to pass through process equipment a is 37 degrees Celsius. The equipment parameters may include but are not limited to: temperature, humidity, air pressure, residence time, etc. It is understandable that during the production process, each equipment parameter is constantly changing from the time the product enters the process equipment to the time it leaves the process equipment.

[0068] In this embodiment, the target device can obtain parameter data of each device parameter in the production process, and the parameter data can be represented in a time series. Since the number of data corresponding to the time series of parameters of different dimensions in the same device is inconsistent. Taking the temperature and humidity parameters in the device parameters as an example, the temperature parameters of the production display panel are [36.1, 36.2, 36.2, 36.3, 36.1], and the humidity parameters of the production display panel A are [76.0, 76.2, 76.3, 76.3, 76.5, 76.7, 76.8].

[0069] To facilitate subsequent use, in this embodiment, the parameter data of each device parameter is preprocessed, see Figure 2 In step 21, the number of parameter data obtained by different OLED products after the target device stage is obtained. In step 22, the parameter data is adjusted based on the number of data to obtain a preset number of parameter data. That is to say, after steps 21 and 22, the parameter data of different dimensions of the same device can be adjusted to the same number. The adjustment method can be one of the following: the maximum number of time series, the average number of time series, or the minimum number of time series. In order to avoid the loss of some data, the maximum number of time series can be used in this embodiment, and the adjustment process includes: (1) first calculating the number of data in each time series; (2) selecting the maximum value of the number of time series corresponding to the same device parameter; (3) After determining the maximum value, the other time series with insufficient numbers can be padded with zeros to make the number of data in all time series the same.

[0070] In step 12, a fusion feature characterizing the characteristics of the target device is obtained according to the parameter data.

[0071] In this embodiment, after obtaining the parameter data, the target device can obtain a fusion feature according to the parameter data. The fusion feature can be used to characterize the characteristics of the target device and can be set according to the specific scenario. In one example, the fusion feature can include statistical features and time-frequency features. Obtaining the fusion feature includes: Figure 3 In step 31, the statistical features of the parameter data are obtained, and the time-frequency features of the parameter data are obtained. In step 32, the time-frequency features are inserted into the statistical features to fuse the statistical features and the time-frequency features to obtain the fused features of the target device.

[0072] Taking obtaining statistical features as an example, the statistical features include at least one of the following features: maximum value, minimum value, mean value, median, standard deviation, difference between maximum value and minimum value, subscript of maximum value, subscript of minimum value, sum of difference values ​​of downward trend, sum of difference values ​​of upward trend, slope, sum of positive values, sum of negative values, maximum value among the sum of decreasing interval data, subscript of starting data of continuous decreasing interval data and the largest interval, subscript of ending data of continuous decreasing interval data and the largest interval.

[0073] Take the time series A: [-2, 1, -1, 2, 3, -3, 4, -4] as an example to describe the above characteristics:

[0074] Minimum value: -4;

[0075] Mean: (-2+1+-1+2+3+-3+4+-4) / 8=0;

[0076] Median: The number in the middle of a set of data arranged in order. After A is sorted, it is: -4, -3, -2, -1, 1, 2, 3, 4. The median is (-1+1) / 2=0.

[0077] Standard Deviation: Formula: The calculation is 2.93;

[0078] Range: the difference between the maximum and minimum values, i.e. 4-(-4)=8;

[0079] Index_min: the subscript of the minimum value, the subscript of -4 is 8;

[0080] Index_max: the subscript of the maximum value, the subscript of 4 is 7;

[0081] Stat_downtrend: The sum of the differences of the downward trend, that is, the difference of the interval [1,-1] is -1-1=-2, the difference of the interval [3,-3] is -3-3=-6, the difference of the interval [4,-4] is -4-4=-8, and the sum of the differences is -2-6-8=-16;

[0082] Stat_uptrend: The sum of the differences of the upward trend. The difference of the interval [-2,1] is 1-(-2)=3, the difference of the interval [-1,2] is 2-(-1)=3, the difference of the interval [2,3] is 3-2=1, and the difference of the interval [-3,4] is 4-(-3)=7, that is, 3+3+1+7=14;

[0083] Slope: Slope formula: That is -0.21428571;

[0084] Positive_sum: the sum of positive values, i.e. 1+2+3+4=10;

[0085] Positive_max: the maximum value in the sum of the ascending interval, which is [2,3], i.e. 2+3=5;

[0086] Positive_maxstart: The starting subscript of the maximum interval of the continuous rising interval. This interval is [2,3], and the subscript of 2 is 4;

[0087] Positive_maxend: The end subscript of the maximum interval of the continuous rising interval. This interval is [2,3], and the subscript of 3 is 5;

[0088] Negative_sum: the sum of negative values, i.e. -2-1-3-4=-10;

[0089] Negative_max: the maximum value among the sums of the descending intervals, the maximum value among the sums of the descending intervals - 4;

[0090] Negative_maxstart: the starting subscript of the sum of the continuous descending intervals and the maximum interval, -4 is the maximum value of the sum of the descending intervals, and its subscript is 8;

[0091] Negative_maxend: The end subscript of the sum of the continuous descending intervals and the maximum interval. -4 is the maximum value of the sum in the descending interval, and the subscript is 8.

[0092] Based on the above features, the statistical characteristics of time series A can be obtained as [-4, 4, 0, 0, 2.93, 8, 8, 7, -16, 14, -0.214, 10, 5, 4, 5, -10, -4, 8, 8].

[0093] By analogy, the target device can calculate the corresponding statistical characteristics of each device parameter.

[0094] In this embodiment, the above statistical features can only average the time series characteristics of the time series, but cannot characterize the characteristics of its time series or the characteristics of a certain time period. Therefore, in this embodiment, time-frequency characteristics can also be obtained. Figure 4 In step 41, the target device can obtain the time series of each device parameter from the parameter data. In step 42, the time series can be decomposed into a preset number of frequency bands by wavelet packets, and the preset number of layers includes the specified layer. In step 43, the signal of the specified frequency band can be reconstructed to obtain a reconstructed signal. In step 44, the energy value of each frequency band can be obtained, and the maximum energy value is used as the time-frequency feature of the parameter data.

[0095] In this embodiment, wavelet packet decomposition is adopted because the wavelet packet transform can decompose both the low-frequency signal and the high-frequency signal in the time series; and the above decomposition has neither redundancy nor omission, so the signal containing a large amount of medium and high frequency information can be better analyzed in time-frequency localization, which is more suitable for the actual needs of the scene of this embodiment. As for how to decompose and reconstruct the wavelet packet, please refer to the relevant technology, which will not be repeated here.

[0096] In an example, wavelet transform can be used, and the preset number is 3 and the designated layer is the third layer as an example to describe the detailed process of obtaining the time-frequency features, including:

[0097] (1) Obtain the time series value of a single parameter of each glass panel;

[0098] (2) Perform wavelet packet decomposition on it, the decomposition layer is 3, and use db6 wavelet packet function to decompose the third layer signal into (x 31 , x 32 , x 33 , x 34 , x 35 , x 36 , x 37 , x 38 );

[0099] (3) Reconstruct the signals of different frequency bands in the third layer to improve the time-frequency resolution. The reconstructed signal is (x 31 , x 32 , x 33 , x 34 , x 35 , x 36 , x 37 , x 38 );

[0100] (4) Construct an energy feature vector for the reconstructed signal. The energy of each frequency band is N is the number of signal sampling points, and then normalization is performed. Take its maximum value, i.e. the main frequency, as the time-frequency feature of the parameter data.

[0101] Based on the above steps, the time-frequency characteristics of time series A can be obtained as [0.721].

[0102] Then, the target device can fuse the statistical features and time-frequency features. Taking time series A as an example, its statistical features are [-4, 4, 0, 0, 2.93, 8, 8, 7, -16, 14, -0.214, 10, 5, 4, 5, -10, -4, 8, 8], its time-frequency features are [0.721], and its fusion features are [-4, 4, 0, 0, 2.93, 8, 8, 7, -16, 14, -0.214, 10, 5, 4, 5, -10, -4, 8, 8, 0.721].

[0103] In one embodiment, in order to eliminate the dimensionality effect between the statistical features and the time-frequency features, the fused features need to be normalized so that different indicators are comparable. In this embodiment, the Z-score can be used, and its formula is expressed as:

[0104]

[0105] Among them, x is a specific value, μ is the mean, and σ is the standard deviation.

[0106] When there are too many parameters, the dimension of the feature will be too large, so the feature needs to be reduced in dimension. In this example, the Pearson correlation coefficient is used to calculate the degree of correlation between the features. The formula is as follows:

[0107]

[0108] Among them, Cov(X, Y) is the covariance of X and Y, σ x is the variance of X, σ Y is the variance of Y. The value of the correlation coefficient is between -1 and +1, that is, -1≤r≤+1. It is generally divided into: |r|<0.5 is a low correlation; 0.5≤|r| is a significant correlation.

[0109] That is, this embodiment can achieve the purpose of feature dimensionality reduction by calculating the Pearson correlation coefficient between features and eliminating significantly correlated features.

[0110] In step 13, the fusion feature is input into a preset prediction model to obtain an inference result output by the prediction model that characterizes whether the OLED product is abnormal.

[0111] In this embodiment, the target device may store a preset prediction model, which is used to infer whether the product produced is a normal product or a defective product based on the parameter data of each device parameter of the target device. The prediction model is a type of machine learning and can be implemented using one of the following models: logistic regression, random forest, LightGBM, and Xgboost. In addition, the prediction model can be trained in the following ways, see Figure 5 ,include:

[0112] In step 51, a training sample set is obtained; the training sample set includes multiple training samples, and each training sample includes a fusion feature and a sample label of a process device. In this step, in the actual production process of OLED products, when the process equipment that causes product defects is located, the parameter data of each device parameter within a period of time can be obtained, and the fusion feature of each device parameter can be obtained according to steps 11 and 12, thereby obtaining sample data. Then, according to the proportion of defective points of the produced OLED products, it is determined whether the device parameter belongs to a positive sample or a negative sample. When it belongs to a positive sample, the sample label of the sample data is set to 1, and when it belongs to a negative sample, the sample label of the sample data is set to -1, thereby obtaining a training sample including the fusion features and sample labels of each device parameter of the process equipment.

[0113] It should be noted that the process equipment for locating defective products in this step can be obtained in the following ways: a defect analysis system can be set up on the production line, and the defect analysis system can first create samples for a certain time period, specify the detection site, defects and other related information, and divide the positive and negative sample data, that is, the proportion of defective points or the ratio of defective points on each display panel is used as the threshold for the overall sample division. If it is greater than the threshold, it is a negative sample, and if it is less than the threshold, it is a positive sample. Afterwards, an intelligent mining template can be set up on the production line, and the intelligent mining module can be used to obtain the process equipment data passed during the production process, and use a decision tree algorithm (such as ID3, C4.5, CART, etc.) for analysis, and rank the importance of each process equipment, so as to locate the equipment that causes a high incidence of defective samples as the target equipment.

[0114] Of course, after locating the device, the management personnel can continue to trace back until the device causing the high incidence of bad samples is completely determined. After locating the device, the device can be used as the target device and step 51 is performed again.

[0115] In step 52, each training sample in the training sample set is input into the prediction model to be trained to obtain the trained prediction model; the number of the prediction models to be trained is the first number. In this step, the number of prediction models to be trained is the first number, for example, the first number is 3. In one example, the prediction models to be trained include logistic regression, random forest, LightGBM and Xgboost. The target device can divide the training sample set into a training set, a validation set and a test set in a ratio of 8:1:1. In this way, each training sample in the training set can be input into each prediction model, and the training samples in the validation set and the test set can be used for verification or testing. After the prediction model passes the verification, it can be determined that the prediction model has completed the training. Among them, the process of training the prediction model using the training sample set can refer to the relevant technology, which will not be repeated here.

[0116] In step 53, an evaluation score of each prediction model in the first number of prediction models is obtained based on a preset evaluation index. In this step, the evaluation index of the model can be preset in the target device, and the evaluation index includes but is not limited to: accuracy, precision, recall, etc., which are used to analyze the credibility of each prediction model.

[0117] Table 1 Prediction results of the prediction model

[0118] Prediction: Positive Sample Prediction: Negative Samples Actual: Positive sample TP FN Actual: Negative samples FP TN

[0119] Note: TP (True Positive) indicates the number of positive classes predicted as positive classes; FN (False Negative) indicates the number of positive classes predicted as negative classes; FP (False Positive) indicates the number of negative classes predicted as positive classes; TN (True Negative) indicates the number of negative classes predicted as negative classes.

[0120] The accuracy calculation formula is as follows:

[0121] Accuracy=(TP+TN) / (TP+FN+FP+TN);

[0122] The accuracy calculation formula is as follows:

[0123] Precision = TP / (TP+FP);

[0124] The recall calculation formula is as follows:

[0125] Recall = TP / (TP+FN).

[0126] In this step, after the evaluation index is determined, the target device can calculate the evaluation score of each prediction model.

[0127] In step 54, the prediction model with the largest evaluation score is determined as the preset prediction model. In this step, the target devices can be sorted from large to small according to the evaluation scores, and the prediction model with the largest evaluation score is determined as the preset prediction model.

[0128] In this embodiment, after obtaining the fusion feature, the fusion feature can be input into the above-mentioned prediction model, and the prediction model outputs an inference result. The inference result can characterize the inference result of whether the OLED product is abnormal, or in other words, the inference result of whether the production of OLED products in the subsequent process is abnormal based on the current working status of the target device. The inference result is convenient for management personnel to determine whether the target device is working normally, and timely maintenance can be carried out when it is working abnormally.

[0129] In one embodiment, after locating the abnormal operation of the target device, it is also necessary to clearly locate the device parameters that cause the problem, see Figure 6 :

[0130] In step 61, when the inference result indicates that the OLED product is abnormal, a second number of candidate parameters of the target device are obtained; the candidate parameters are used to assist in analyzing the cause of the abnormality of the OLED product.

[0131] The second number of candidate parameters can be obtained in the following manner:

[0132] See also Figure 7 , in step 71, the target device can obtain the weight value of each device parameter based on the prediction model. Taking logistic regression as an example, the weight value of each feature can be obtained by calculating the absolute value of the regression coefficient of each variable. In practical applications, a weight value calling function (feature_importances_) can be set for each prediction model. After the training is completed, the weight value or importance of each feature can be obtained by directly calling the above weight value calling function. Take the importance of a feature X in a random forest as an example: 1. For each decision tree in the random forest, use the corresponding OOB (out-of-bag data) data to calculate its out-of-bag data error, recorded as errOOB1, that is, the first out-of-bag data error. 2. Randomly add noise interference to the feature X of all samples of the out-of-bag data OOB. At this time, the value of the sample at the feature X can be randomly changed; and its out-of-bag data error is calculated again, recorded as errOOB2, that is, the first out-of-bag data error. 3. Assuming that there are Ntree trees in the random forest, the importance of feature X = ∑(errOOB2-errOOB1) / Ntree.

[0133] With reference to the importance of feature X, all features of the target device may be ranked by importance to obtain a second number of features with higher importance, that is, a second number of device parameters with greater importance and greater impact on the target device.

[0134] It should be noted that the reason why this expression can be used as a measure of the importance of the corresponding feature in this example is that if the out-of-bag accuracy is greatly reduced after noise is randomly added to a feature, it means that this feature has a great influence on the classification result of the sample, that is, its importance is relatively high.

[0135] In step 72, the target device may sort each device parameter in the target device from high to low according to the weight value.

[0136] In step 73, a second number of device parameters ranked high are determined as candidate parameters. In one example, the second number is 10, which can be adjusted according to specific scenarios.

[0137] In step 62, characteristic data representing the target parameter having the greatest impact on the OLED product among the second number of candidate parameters may be obtained.

[0138] In step 63, the characteristic data of the second number of candidate parameters and the target parameters can be output to assist the user in determining the cause of the abnormality of the OLED product. In this way, the manager can prioritize the second number of candidate parameters to facilitate the rapid location of the equipment parameters causing the problem. In addition, the manager can also evaluate the results of the prediction model based on the results of the second number of candidate parameters. In other words, this embodiment not only provides a prediction model for predicting product quality, but also increases the interpretability of the prediction model by providing a second number of candidate parameters, so that the manager can adjust the equipment parameters in time to avoid losses; and adjust the prediction model to improve the accuracy of the inference results.

[0139] At this point, the parameter data of each device parameter of the target device within a preset time period can be obtained in the disclosed embodiment; the parameter data refers to the value of each device parameter recorded at preset intervals during the OLED product passing through the target device stage; then, the fusion feature characterizing the characteristics of the target device can be obtained based on the parameter data; thereafter, the fusion feature can be input into a preset prediction model to obtain the inference result output by the prediction model characterizing whether the OLED product is abnormal. In this way, the embodiment can predict whether abnormal OLED products will be produced during the production process, so that the management personnel can adjust the target device in advance and improve the production yield.

[0140] Figure 8 is a block diagram of a device for predicting product defects according to an exemplary embodiment. Figure 8 , a product defect prediction device, the device comprising:

[0141] The parameter data acquisition module 81 is used to acquire the parameter data of each device parameter of the target device within a preset time period; the parameter data refers to the value of each device parameter recorded at a preset interval when the OLED product passes through the target device stage;

[0142] A fusion feature acquisition module 82, used to acquire a fusion feature representing a feature of the target device according to the parameter data;

[0143] The inference result acquisition module 83 is used to input the fusion feature into a preset prediction model to obtain the inference result output by the prediction model, which characterizes whether the OLED product is abnormal.

[0144] It is understandable that the device provided in the embodiment of the present disclosure corresponds to the above method. For the specific content, reference can be made to the content of each embodiment of the method, which will not be repeated here.

[0145] In an exemplary embodiment, there is also provided an electronic device, comprising:

[0146] processor;

[0147] a memory for storing a computer program executable by the processor;

[0148] The processor is configured to execute the computer program in the memory to implement the following Figure 1 The steps of the method.

[0149] In an exemplary embodiment, there is also provided an executable computer readable storage medium, such as a memory including instructions, wherein the executable computer program can be executed by a processor to implement the following Figure 1 The steps of the method. The readable storage medium may be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, etc.

[0150] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the disclosure disclosed herein. The present disclosure is intended to cover any variations, uses, or adaptations that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The description and examples are to be considered exemplary only, and the true scope and spirit of the present disclosure are indicated by the following claims.

[0151] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A method for predicting product defects, characterized in that: The method comprises: Obtaining parameter data of each device parameter of the target device within a preset time period; the parameter data refers to the value of each device parameter recorded at preset intervals during the product's passage through the target device stage; Acquire a fusion feature characterizing a feature of the target device according to the parameter data; Inputting the fused features into a preset prediction model to obtain an inference result output by the prediction model indicating whether the product is abnormal; Acquiring a fusion feature characterizing a feature of the target device according to the parameter data includes: Obtaining statistical features of the parameter data; the statistical features are used to average the time series features of the time series of the parameter data; Acquire the time-frequency features of the parameter data; the time-frequency features are used to characterize the features of the time series of the parameter data or the features of a certain time period; The statistical features and the time-frequency features are normalized by using a Z-score method, and the time-frequency features are inserted into the statistical features to fuse the statistical features and the time-frequency features to obtain the fused features of the target device.

2. The method according to claim 1, characterized in that: The prediction model includes one of the following: logistic regression, random forest, LightGBM and Xgboost.

3. The method according to claim 1 or 2, characterized in that: Get the parameter data of each device parameter of the target device within a preset time period, including: Get the number of parameter data obtained by different products through the target equipment stage; The number of the parameter data is adjusted based on the number of data to obtain a preset number of parameter data.

4. The method according to claim 1, characterized in that: Acquiring the time-frequency characteristics of the parameter data includes: Obtaining a time series of each device parameter from the parameter data; Performing wavelet packet decomposition on the time series into a preset number of layer frequency bands, wherein the preset number of layers includes a designated layer; Reconstruct the signal of the specified layer frequency band to obtain a reconstructed signal; The energy value of each frequency band is obtained, and the maximum energy value is used as the time-frequency feature of the parameter data.

5. The method according to claim 1 or 2, characterized in that: The method further comprises: When the inference result indicates that the product is abnormal, obtaining a second number of candidate parameters of the target device; Acquire characteristic data representing the target parameter having the greatest impact on the product from among the second number of candidate parameters; Output the feature data of the second number of candidate parameters and the target parameter.

6. The method according to claim 5, characterized in that Acquiring a second number of candidate parameters of the target device, including: Obtaining a weight value of each device parameter in the target device based on the prediction model; Sort each device parameter in the target device from high to low according to the weight value; A second number of device parameters ranked top are determined as candidate parameters.

7. The method according to claim 5, characterized in that When the prediction model is a random forest, obtaining a second number of candidate parameters of the target device includes: Based on the preset out-of-bag data, obtain the first out-of-bag data error of each decision tree in the random forest; Adding noise interference to the out-of-bag data to change the value of each device parameter of the sample, and obtaining a second out-of-bag data error of each decision tree in the random forest; calculating the importance of each device parameter based on the second out-of-bag data error, the first out-of-bag data error, and the number of decision trees in the random forest; A second number of device parameters ranked high in importance are determined as candidate parameters of the target device.

8. The method according to claim 1 or 2, characterized in that: The method further comprises the step of training a preset prediction model, specifically comprising: Acquire a training sample set; the training sample set includes multiple training samples, each training sample includes a fusion feature of a process device and a sample label; Input each training sample in the training sample set into the prediction model to be trained to obtain a trained prediction model; the number of the prediction models to be trained is a first number; the first number of prediction models includes logistic regression, random forest, LightGBM and Xgboost; Obtaining an evaluation score of each prediction model in the first number of prediction models based on a preset evaluation index; The prediction model with the largest evaluation score is determined as the preset prediction model.

9. The method according to claim 8, characterized in that Before acquiring the training sample set, the method further includes: Obtain the number of defective points of products produced by the testing site within a specified time period; Determine the sample classification of each product based on the defective points and the point threshold, the sample classification includes positive samples and negative samples; the positive sample refers to the product whose defective points are less than or equal to the point threshold, and the negative sample refers to the product whose defective points are greater than the point threshold; The importance of each process equipment is ranked to locate the equipment that causes a high incidence of defective samples as the target equipment.

10. A product defect prediction device, characterized in that: The device comprises: A parameter data acquisition module, used to acquire parameter data of each device parameter of the target device within a preset time period; the parameter data refers to the value of each device parameter recorded at preset intervals during the product passing through the target device stage; A fusion feature acquisition module, used to acquire a fusion feature representing a feature of the target device according to the parameter data; An inference result acquisition module, used for inputting the fusion feature into a preset prediction model to obtain an inference result output by the prediction model indicating whether the product is abnormal; The fusion feature acquisition module acquires the fusion feature representing the target device feature according to the parameter data, including: Obtaining statistical features of the parameter data; the statistical features are used to average the time series features of the time series of the parameter data; Acquire the time-frequency features of the parameter data; the time-frequency features are used to characterize the features of the time series of the parameter data or the features of a certain time period; The statistical features and the time-frequency features are normalized by using a Z-score method, and the time-frequency features are inserted into the statistical features to fuse the statistical features and the time-frequency features to obtain the fused features of the target device.

11. An electronic device, characterized in that: include: processor; a memory for storing a computer program executable by the processor; The processor is configured to execute the computer program in the memory to implement the method according to any one of claims 1 to 9.

12. A computer-readable storage medium, characterized in that: When the executable computer program in the storage medium is executed by a processor, the method as claimed in any one of claims 1 to 9 can be implemented.

Images