A method for establishing an integrated circuit package failure prediction model and a failure prediction method

Abstract

The application relates to the technical field of a prediction model establishment method, in particular to an integrated circuit package fault prediction model establishment method and a fault prediction method; the method comprises the following steps: preprocessing multi-source data, screening features related to fault prediction, and constructing a standardized training data set; adopting k-fold cross-validation, equally dividing the data set into k subsets, selecting one subset as a test set each time, and taking the remaining k-1 subsets as training sets; splitting the training set into input features and target outputs, inputting a machine learning algorithm to construct an initial fault prediction model; splitting the test set into test features and test target outputs, inputting the model to obtain a prediction result, comparing and calculating the precision, recall and F1 value; repeating the above process k times, and averaging each index; if the average value is higher than a set threshold, the model construction is completed; otherwise, the data is expanded or the features are re-screened. The application can realize fault prediction based on real-time data and overcome the limitation of subjective judgment.

Description

Technical Field

[0001] This invention relates to the technical field of prediction model establishment methods, specifically to a method for establishing an integrated circuit packaging fault prediction model and a fault prediction method. Background Technology

[0002] Production failures have a significant impact on a company's normal operations and production efficiency. Failures not only cause production line downtime and delays, but can also lead to equipment damage and safety hazards. This demonstrates that predicting and promptly maintaining production failures is crucial for companies to improve production efficiency and reduce costs.

[0003] Traditional experience- and rule-based fault prediction and maintenance strategies are often limited by subjective judgment and cannot meet the needs of complex production environments. Summary of the Invention

[0004] The purpose of this invention is to solve the technical problem that traditional fault prediction based on experience and rules is limited by subjective judgment and limitations, and to provide a method for establishing a fault prediction model for integrated circuit packaging and a fault prediction method.

[0005] To solve the above-mentioned technical problems, the technical solution provided by the present invention is as follows: A method for establishing an integrated circuit packaging fault prediction model includes the following steps: S1. Obtain multi-source data generated by various devices in the integrated circuit packaging process during operation, and preprocess the multi-source data to obtain a standardized dataset; Filter out features related to fault prediction from the standardized dataset and use them as the training dataset; S2. Divide the training dataset into k equal training subsets, and use random sampling to select any one of the training subsets as the test set, and the remaining k-1 training subsets as the training set. S3. Divide the data in the training set into input features and target output; Input features and target outputs are fed into a machine learning algorithm, which learns the mapping relationship between the input features and the target output to obtain an initial fault prediction model. S4. Divide the data in the test set into test input features and test target output; The test input features are input into the initial fault prediction model, so that the initial fault prediction model outputs the corresponding prediction target output based on the test input features. The output of the predicted target is compared with the output of the test target by the initial fault prediction model, and the precision, recall and F1 score of the initial fault prediction model are calculated. S5. Repeat steps S2 to S4k times to complete the training and testing of the initial fault prediction model; S6. Calculate the mean of precision, recall and F1 score of all initial fault prediction models, and determine whether the mean of precision, recall and F1 score all exceed the threshold. If any of the mean values ​​of precision, recall, or F1 score do not exceed the threshold, then increase the training data or re-filter the features related to fault prediction, and repeat step S1. If the mean of precision, recall, and F1 score all exceed the threshold, the fault prediction model is established.

[0006] Furthermore, in step S1, the specific method for preprocessing the multi-source data to obtain standardized data is as follows: Filter multi-source data to remove invalid, incomplete, outlier, or noisy data; Calculate the mean and standard deviation of all multi-source data; Divide the difference between each multi-source data point and the mean by the standard deviation, and remove multi-source data points that do not fall within the (0,1) interval to obtain a standardized dataset.

[0007] Furthermore, in step S1, the standardized dataset includes historical operating data, historical process parameters, historical operating environment parameters, historical fault reports, real-time operating data, real-time process parameters, real-time operating environment parameters, and real-time fault reports.

[0008] Furthermore, in step S1, the minimum and maximum values ​​are extracted from all multi-source data; Divide the difference between each multi-source data point and the minimum value by the difference between the maximum value and the minimum value, and remove the data that does not fall within the (0,1) interval to obtain the standardized dataset.

[0009] Furthermore, step S3 specifically includes: The historical operating data, historical process parameters, historical operating environment parameters, real-time operating data, real-time process parameters, and real-time operating environment parameters in the training set are used as input features, and historical fault reports and real-time fault reports are used as target outputs. The input features and target output are fed into a machine learning algorithm, which learns the mapping relationship between the input features and the target output to obtain an initial fault prediction model; wherein, the machine learning algorithm is a combination of one or more algorithms such as decision tree, support vector machine or neural network.

[0010] Furthermore, step S4 specifically includes: The test set includes historical operating data, historical process parameters, historical operating environment parameters, real-time operating data, real-time process parameters, and real-time operating environment parameters as test input features, and historical fault reports and real-time fault reports as test target outputs. The test input features in the test set are input into the initial fault prediction model. The fault prediction model outputs the corresponding prediction target output based on the test input features. At the same time, the fault prediction model compares the prediction target output with the test target output and constructs a confusion matrix. The precision, recall and F1 score of the fault prediction model are calculated through the basic statistics in the confusion matrix.

[0011] Furthermore, in step S4, the accuracy P = TP / (TP + FP); Recall rate R = TP / (TP + FN); F1 value = 2 * P * R / (P + R); In the formula, TP, FP, TN, and FN are the basic statistics of the confusion matrix, where TP represents true positives, FP represents false positives, TN represents true negatives, and FN represents false negatives.

[0012] Furthermore, in step S6, the thresholds for the precision, recall, and mean F1 score of all initial fault prediction models are 70%, 90%, and 0.85, respectively.

[0013] Furthermore, in step S1, the equipment in the integrated circuit packaging process includes chip mounting equipment, bonding equipment, molding packaging equipment, electroplating equipment, dicing equipment, testing equipment, and cleaning equipment; wherein: Chip placement equipment includes high-speed pick and place machines and flip-chip bonding machines. Its fault prediction-related features include placement pressure, positioning accuracy, temperature profile and solder paste amount. Typical fault prediction scenarios are chip misalignment, insufficient solder paste and cold solder joints. Bonding equipment includes wire bonding machines and carrier tape automatic bonding equipment. Its fault prediction-related features include bonding pressure, ultrasonic power, bonding time, and gold wire tension. Typical fault prediction scenarios are wire breakage, insufficient bonding strength, and poor solder joint. Molding and packaging equipment includes transfer molding machines and compression molding machines. Its failure prediction-related characteristics include molding pressure, temperature distribution, curing time and resin flowability. Typical failure prediction scenarios are bubble defects, incomplete filling and stress cracking. Electroplating equipment includes chemical plating equipment and electroplating production lines. Its fault prediction features include current density, electroplating solution composition, temperature and deposition time. Typical fault prediction scenarios are uneven plating, open circuit and short circuit. Cutting equipment includes wafer dicing machines and laser dicing machines. The characteristics related to fault prediction include cutting speed, blade wear and cooling system parameters. Typical fault prediction scenarios are chip breakage, cutting deviation and thermal damage. The testing equipment includes probe stations and automated testing equipment. Its fault prediction-related features include test voltage / current, contact resistance, and signal integrity. Typical fault prediction scenarios are open circuit / short circuit detection failure and parameter drift misjudgment. Cleaning equipment includes plasma cleaners and wet cleaning equipment. Its fault prediction features include cleaning time, gas flow rate, temperature and chemical solution concentration. Typical fault prediction scenarios are contaminant residue and surface damage.

[0014] A method for predicting faults in integrated circuit packaging includes the following steps: 1) Acquire multi-source data from equipment at each stage of integrated circuit packaging during operation, and preprocess the multi-source data; 2) Filter features related to fault prediction within the preprocessed multi-source data; 3) Input the features related to fault prediction into the fault prediction model to obtain the current fault status of any device in each stage of integrated circuit packaging.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: The integrated circuit packaging fault prediction model establishment method and fault prediction method provided by this invention improve the quality of input data for the fault prediction model by preprocessing multi-source data generated by various devices in the integrated circuit packaging process and converting it into a unified scale. Then, by selecting fault prediction-related features within the standardized dataset and using them as training datasets, the machine learning algorithm is repeatedly trained using k-fold cross-validation. By calculating the mean of precision, recall, and F1 score of all initial fault prediction models obtained from repeated training and comparing it with a set threshold, it is determined whether further training of the initial fault prediction models is necessary, thereby enhancing the generalization ability of the fault prediction model. By establishing a fault prediction model and applying it to actual production, fault prediction can be performed using real-time data, eliminating the limitations of subjective judgment and meeting the needs of complex production environments. Detailed Implementation

[0016] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0017] A method for establishing an integrated circuit packaging fault prediction model includes the following steps: 1) Obtain multi-source data generated by various devices in the integrated circuit packaging process during operation, filter the multi-source data, and remove invalid, incomplete, outlier or noise data; Calculate the mean and standard deviation of all multi-source data; Divide the difference between each multi-source data point and the mean by the standard deviation, and remove the multi-source data points that do not fall within the (0,1) interval to obtain the standardized dataset; The standardized dataset includes historical operating data, historical process parameters, historical operating environment parameters, historical fault reports, real-time operating data, real-time process parameters, real-time operating environment parameters, and real-time fault reports. Filter out features related to fault prediction from the standardized dataset and use them as the training dataset; In this step, any one of the following methods can be used to screen features within the standardized dataset that are relevant to fault prediction (strong correlation and high information contribution). 2) Divide the training dataset into k equal training subsets, and use random sampling to select any one of the training subsets as the test set, and the remaining k-1 training subsets as the training set; 3) Use historical operating data, historical process parameters, historical operating environment parameters, real-time operating data, real-time process parameters, and real-time operating environment parameters in the training set as input features, and historical fault reports and real-time fault reports as target outputs; Input features and target outputs are fed into a machine learning algorithm, which learns the mapping relationship between the input features and the target output. In other words, the machine learning algorithm learns the correspondence between input features (historical operating data, historical process parameters, historical operating environment parameters, real-time operating data, real-time process parameters, and real-time operating environment parameters) and target outputs (historical fault reports and real-time fault reports). Simply put, the machine learning algorithm learns the correspondence between features and faults. After training, an initial fault prediction model is obtained. The initial fault prediction model can be denoted as Y=M(X); Where X is the input feature, M is the machine learning algorithm, and Y is the target output; The process of model training involves continuously optimizing the function M to make the output of M(X) as close as possible to the real fault label Y.

[0018] Among them, the machine learning algorithm is a combination of one or more algorithms such as decision tree, support vector machine or neural network; 4) Use historical operating data, historical process parameters, historical operating environment parameters, real-time operating data, real-time process parameters, and real-time operating environment parameters within the test set as test input features, and historical fault reports and real-time fault reports as test target outputs; The test input features in the test set are input into the initial fault prediction model. The fault prediction model outputs the corresponding prediction target output based on the test input features. At the same time, the fault prediction model compares the prediction target output with the test target output and constructs a confusion matrix. The precision, recall and F1 score of the fault prediction model are calculated through the basic statistics in the confusion matrix. Precision P = TP / (TP + FP); Recall rate R = TP / (TP + FN); F1 value = 2 * P * R / (P + R); In the formula, TP, FP, TN, and FN are the basic statistics of the confusion matrix, where TP represents true positive, FP represents false positive, TN represents true negative, and FN represents false negative. 5) Repeat steps 2) to 4) k times to complete the training and testing of the initial fault prediction model; 6) Calculate the mean of precision, recall, and F1 score for all initial fault prediction models, and determine whether the mean of precision, recall, and F1 score all exceed the threshold. The thresholds for precision, recall, and mean F1 score were 70%, 90%, and 0.85, respectively. If any of the mean values ​​of precision, recall, or F1 score do not exceed the threshold, then increase the training data or re-filter the features related to fault prediction, and repeat step S1. If the mean of precision, recall, and F1 score all exceed the threshold, the fault prediction model is established.

[0019] This embodiment mainly includes the following stages in establishing an integrated circuit packaging fault prediction model: preliminary data acquisition and processing, establishment of the fault prediction model, training of the fault prediction model, experimental verification and result analysis; each stage is briefly described below: I. Preliminary Data Collection and Processing (a) Historical data collection: Historical data can be obtained from the company's internal database or other relevant systems. Historical data may include historical operating data, historical process parameters, historical operating environment parameters, historical fault reports, maintenance records, etc.

[0020] When collecting historical data, it is necessary to ensure the integrity and accuracy of the data; the data should include information such as the time, location, type of equipment failure, and corresponding maintenance measures; select an appropriate time period to collect data to ensure that the data can reflect the failure situation in different time periods; and it should also cover data of various equipment and processes in the integrated circuit packaging process as much as possible to obtain more comprehensive and accurate analysis results.

[0021] After historical data collection is completed, the data also needs to be processed and cleaned to remove duplicate data, handle missing values ​​and outliers, etc.

[0022] (II) Real-time monitoring data collection: Real-time monitoring data includes real-time operating data, real-time process parameters, real-time operating environment parameters, and real-time fault reports, which can be collected using the following methods: 1) Sensor technology: Using sensors to monitor the operating status and process parameters of equipment. Sensors can be installed on the equipment to obtain real-time data by measuring indicators such as temperature, pressure, and vibration.

[0023] 2) Remote monitoring system: A remote monitoring system allows for real-time acquisition of equipment operating status and process parameters. This method typically uses a network connection between the equipment and the monitoring center to obtain data promptly.

[0024] 3) Wireless communication technology: Real-time data transmission between devices and monitoring systems can be achieved using wireless communication technology. Wireless sensor networks and Internet of Things (IoT) technologies can provide flexible data acquisition solutions.

[0025] 4) Data acquisition software: Specialized data acquisition software is used to collect real-time monitoring data. This software can connect to devices or sensors and automatically record and store data.

[0026] The following aspects should be noted when collecting real-time monitoring data: 1) Data Quality: Ensure the accuracy and reliability of the collected real-time monitoring data. Data quality can be improved by calibrating sensors and eliminating interference and noise.

[0027] 2) Data Storage: Choose an appropriate data storage method to facilitate subsequent processing and analysis of real-time monitoring data. Technologies such as databases, cloud storage, or distributed file systems can be used to store the data.

[0028] 3) Data Security: Protect the security and confidentiality of real-time monitoring data. Take appropriate security measures, such as encrypted communication, access control, and data backup, to prevent data leakage or loss.

[0029] (III) Data Cleaning and Denoising: During data acquisition, noise or outliers may appear in the data due to various reasons, which can adversely affect the accuracy and reliability of the model. Therefore, data cleaning and noise reduction are necessary before establishing a fault prediction model.

[0030] Data cleaning refers to the process of filtering and processing data to remove invalid or incomplete data.

[0031] Data denoising refers to the processing of outliers or noise in data.

[0032] When cleaning and denoising data, it is necessary to select appropriate methods and techniques based on the specific characteristics of the data and the research objectives. At the same time, it is also necessary to pay attention to preserving the integrity and reliability of the data and avoid over-processing or losing important information.

[0033] (iv) Feature selection and extraction: To facilitate the training of the fault prediction model, the most representative and relevant features need to be selected and extracted from the historical data and real-time monitoring data collected in the early stage, so as to reduce the dimensionality of the feature space and improve the efficiency and accuracy of the model.

[0034] In the process of feature selection, statistical methods such as correlation coefficient analysis, information gain analysis, and analysis of variance can be used to evaluate the importance and contribution of features, respectively. The importance of a feature is determined by calculating the correlation between it and the target variable. Additionally, statistical methods such as information gain analysis and analysis of variance can be used to assess the contribution of a feature.

[0035] Feature extraction transforms collected historical and real-time monitoring data into more representative and interpretable features. Commonly used feature extraction methods include Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA); these methods can improve model performance by reducing data dimensionality or extracting the most discriminative information from the data.

[0036] (v) Data standardization: To transform data with different features into a uniform scale, thereby eliminating dimensional differences between features and enabling historical and real-time monitoring data to be better applied to the training of machine learning algorithms, the original data is scaled according to a certain ratio to conform to a standard normal distribution. This is typically done using Z-score normalization and MinMax normalization.

[0037] Z-score standardization is achieved by calculating the difference between each sample and the overall sample mean, and then dividing by the overall sample standard deviation. Specifically, it involves filtering multi-source data to remove invalid, incomplete, outlier, or noisy data. Calculate the mean and standard deviation of all multi-source data; Divide the difference between each multi-source data point and the mean by the standard deviation, and remove multi-source data points that do not fall within the (0,1) interval to obtain a standardized dataset.

[0038] MinMax standardization maps raw data to a specified range, typically between 0 and 1. Specifically, after acquiring multi-source data generated by various devices in the integrated circuit packaging process, the minimum and maximum values ​​of all multi-source data are extracted. Divide the difference between each multi-source data point and the minimum value by the difference between the maximum value and the minimum value, and remove the data that does not fall within the (0,1) interval to obtain the standardized dataset.

[0039] II. Establishment of Fault Prediction Model To effectively evaluate model performance, the dataset needs to be divided into training and testing sets. The majority of the data is used to train the fault prediction model, while the remaining data is used to test the model's accuracy.

[0040] Fault prediction models typically employ machine learning algorithms such as decision trees, support vector machines, or a combination of one or more neural networks.

[0041] In this embodiment, the network structure of the fault prediction model (described using Mermaid syntax) is as follows: graphTD A [Input Layer] --> B [Feature Data] B-->C [Model Selection and Structure] C-->C1 [Decision Tree] C1-->C11["Node Splitting (Information Gain / Gini Coefficient)"] C1-->C12["Pruning strategies (pre-pruning / post-pruning)"] C --> C2 [Support Vector Machine (SVM)] C2-->C21["Kernel function (linear / Gaussian / RBF)"] C2-->C22["Hyperplane Optimization (Margin Maximization)"] C --> C3 [Neural Network] C3-->C31["Input Layer (Feature Dimension)"] C3 --> C32 ["Hidden Layer (ReLU / Sigmoid Activation)"] C3 --> C33 ["Output Layer (Softmax / Sigmoid)"] C --> C4 [Integration Method] C4 --> C41 ["Bagging (Random Forest)"] C4-->C42["Boosting(XGBoost / LightGBM)"] C4 --> C43 ["Stacking (Model Stacking)"] C --> D [Output Layer] D-->D1["Fault Probability Prediction"] D-->D2["Fault Classification (Normal / Abnormal)"] D-->D3["Maintenance Strategy Recommendations"] III. Model Training Dataset partitioning and cross-validation: The dataset is typically divided into training and test sets using random sampling.

[0042] The training set is used to train the model and tune its parameters, while the test set is used independently to evaluate the model's performance and ensure its effectiveness in real-world applications.

[0043] Cross-validation, as a powerful model evaluation technique, can further enhance the reliability of evaluation results. Among them, k-fold cross-validation and leave-one-out cross-validation are two commonly used methods.

[0044] k-fold cross-validation: The dataset is divided into k equal subsets. Each time, one subset is selected as the validation set, and the remaining k-1 subsets are used for training. This process is repeated k times, each time selecting a different subset as the validation set. Finally, the average of the k evaluation results is used as the model's performance metric.

[0045] Leave-one-out cross-validation: Only one sample is left as validation data each time, and the rest are used for training. This process is repeated once for each sample. Although computationally expensive, it provides an unbiased estimate of the model's performance.

[0046] IV. Experimental Verification and Result Analysis The test set was fed into the trained and optimized fault prediction model. Subsequently, by comparing the model with actual fault records, we statistically analyzed the following key metrics: true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN). These metrics provide detailed classification information about the model's prediction results.

[0047] Based on these classification results, we further calculated the following important evaluation metrics: Precision: Reflects the proportion of samples that the model predicts as faults but are actually faults. It is calculated as TP / (TP+FP).

[0048] Recall: Also known as true probability, it represents the proportion of actual faulty samples that are correctly predicted by the model. The formula is TP / (TP+FN).

[0049] F1 score: The harmonic mean of precision and recall, providing a comprehensive metric for evaluating model performance. The formula is 2*(Precision*Recall) / (Precision+Recall).

[0050] By calculating these evaluation metrics, a comprehensive and in-depth understanding of the performance of the fault prediction model on different datasets can be obtained. Based on the evaluation results, necessary adjustments can be made to the model. This may involve strategies such as increasing training data and re-selecting features relevant to fault prediction. For example, collecting more raw data: legally collecting data according to the model scenario, such as images under different conditions; transforming existing data to generate new samples, such as image flipping and text synonym replacement; using statistics or algorithms to select important features, such as variance screening and tree models to assess importance. Extracting new features from original features: SIFT for images, and bag-of-words model for text; feature transformation: mathematical transformation of features, such as normalization, logarithmic transformation, and optimization of distribution correlation.

[0051] This embodiment establishes a fault prediction model and applies it to a real production environment for fault prediction. This facilitates the development of corresponding maintenance strategies based on the prediction results, thereby reducing fault risks and improving the overall reliability and stability of the equipment. The specific aspects of maintenance strategy development include: 1. Fault Risk Assessment and Inspection Cycle Determination: In practical applications, suitable assessment indicators can be selected based on specific circumstances, and multiple indicators can be combined for comprehensive evaluation. Classifying fault types and selecting appropriate assessment indicators provides a foundation and basis for subsequent fault prediction model establishment and maintenance strategy formulation. The determination of the inspection cycle should comprehensively consider factors such as equipment characteristics, operating environment, data analysis results, maintenance costs, and production loss costs. By reasonably adjusting the inspection cycle, the accuracy of fault prediction and the effectiveness of maintenance strategies can be improved, thereby achieving timely prevention of production faults and stable equipment operation.

[0052] 2. Equipment Maintenance Plan Development: By analyzing historical fault records, monitoring data, and maintenance records, the operating status of the equipment can be assessed. Simultaneously, considering factors such as equipment type, quality, and working environment, the equipment's service life can be determined. Based on this information, reasonable maintenance strategies can be developed to extend the equipment's lifespan and reduce the risk of failure.

[0053] When developing a maintenance plan, the following aspects need to be considered comprehensively. First, the maintenance cycle for different components and systems should be determined based on the equipment type and working environment. Generally, critical components and systems require more frequent maintenance. Next, the specific content and methods of maintenance should be determined based on the equipment's usage and workload. For example, for equipment operating under high load, maintenance can be performed by periodically replacing parts or lubricating oil. As can be seen from the above discussion, it is necessary to establish an execution and monitoring mechanism for the maintenance plan to ensure its effective implementation.

[0054] 3. Application of Intelligent Systems: In research on production fault prediction and maintenance strategies based on machine learning, real-time monitoring and early warning of abnormal situations play a crucial role. This relies on the precise coordination of sensors and data acquisition systems, which continuously track core parameters in production equipment and processes, such as equipment operating status, temperature, and pressure. This real-time data is transmitted to the central control system for in-depth processing and analysis.

[0055] The central control system, through the application of advanced algorithms and models, can quickly identify abnormal patterns or parameter values ​​deviating from normal ranges in the data. Once an anomaly is detected, the system automatically triggers an alarm mechanism, promptly notifying relevant personnel through multiple means to ensure they can respond quickly and take necessary measures.

[0056] The integrated circuit packaging fault prediction model established in this embodiment can be applied to the following equipment and processes: Based on the above-mentioned method for establishing an integrated circuit packaging fault prediction model, this invention also provides an integrated circuit packaging fault prediction method, specifically: 1) Acquire multi-source data from equipment at each stage of integrated circuit packaging during operation, and preprocess the multi-source data; 2) Filter features related to fault prediction within the preprocessed multi-source data; 3) Input the features related to fault prediction into the fault prediction model to obtain the current fault status of any device in each stage of integrated circuit packaging.

[0057] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions within the technical scope disclosed in the present invention should be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for establishing a fault prediction model for integrated circuit packaging, characterized in that, Includes the following steps: S1. Obtain multi-source data generated by various devices in the integrated circuit packaging process during operation, and preprocess the multi-source data to obtain a standardized dataset; Filter out features related to fault prediction from the standardized dataset and use them as the training dataset; S2. Divide the training dataset into k equal training subsets, and use random sampling to select any one of the training subsets as the test set, and the remaining k-1 training subsets as the training set. S3. Divide the data in the training set into input features and target output; Input features and target outputs are fed into a machine learning algorithm, which learns the mapping relationship between the input features and the target output to obtain an initial fault prediction model. S4. Divide the data in the test set into test input features and test target output; The test input features are input into the initial fault prediction model, so that the initial fault prediction model outputs the corresponding prediction target output based on the test input features. The output of the predicted target is compared with the output of the test target by the initial fault prediction model, and the precision, recall and F1 score of the initial fault prediction model are calculated. S5. Repeat steps S2 to S4k times to complete the training and testing of the initial fault prediction model; S6. Calculate the mean of precision, recall and F1 score of all initial fault prediction models, and determine whether the mean of precision, recall and F1 score all exceed the threshold. If any of the mean values ​​of precision, recall, or F1 score do not exceed the threshold, then increase the training data or re-filter the features related to fault prediction, and repeat step S1. If the mean of precision, recall, and F1 score all exceed the threshold, the fault prediction model is established.

2. The method for establishing an integrated circuit packaging fault prediction model according to claim 1, characterized in that, In step S1, the specific method for preprocessing the multi-source data to obtain standardized data is as follows: Filter multi-source data to remove invalid, incomplete, outlier, or noisy data; Calculate the mean and standard deviation of all multi-source data; Divide the difference between each multi-source data point and the mean by the standard deviation, and remove multi-source data points that do not fall within the (0,1) interval to obtain a standardized dataset.

3. The method for establishing an integrated circuit packaging fault prediction model according to claim 1, characterized in that, In step S1, the standardized dataset includes historical operating data, historical process parameters, historical operating environment parameters, historical fault reports, real-time operating data, real-time process parameters, real-time operating environment parameters, and real-time fault reports.

4. The method for establishing an integrated circuit packaging fault prediction model according to claim 2, characterized in that, In step S1, the minimum and maximum values ​​are extracted from all multi-source data; Divide the difference between each multi-source data point and the minimum value by the difference between the maximum value and the minimum value, and remove the data that does not fall within the (0,1) interval to obtain the standardized dataset.

5. The method for establishing an integrated circuit packaging fault prediction model according to claim 1, characterized in that, Step S3 is as follows: The training set includes historical operating data, historical process parameters, historical operating environment parameters, real-time operating data, real-time process parameters, and real-time operating environment parameters as input features, and historical fault reports and real-time fault reports as target outputs. The input features and target output are fed into a machine learning algorithm, which learns the mapping relationship between the input features and the target output to obtain an initial fault prediction model; wherein, the machine learning algorithm is a combination of one or more algorithms such as decision tree, support vector machine or neural network.

6. The method for establishing an integrated circuit packaging fault prediction model according to claim 1, characterized in that, Step S4 is as follows: The test set includes historical operating data, historical process parameters, historical operating environment parameters, real-time operating data, real-time process parameters, and real-time operating environment parameters as test input features, and historical fault reports and real-time fault reports as test target outputs. The test input features in the test set are input into the initial fault prediction model. The fault prediction model outputs the corresponding prediction target output based on the test input features. At the same time, the fault prediction model compares the prediction target output with the test target output and constructs a confusion matrix. The precision, recall and F1 score of the fault prediction model are calculated through the basic statistics in the confusion matrix.

7. The method for establishing an integrated circuit packaging fault prediction model according to claim 6, characterized in that, In step S4, the accuracy P = TP / (TP + FP); Recall rate R = TP / (TP + FN); F1 value = 2 * P * R / (P + R); In the formula, TP, FP, TN, and FN are the basic statistics of the confusion matrix, where TP represents true positives, FP represents false positives, TN represents true negatives, and FN represents false negatives.

8. The method for establishing an integrated circuit packaging fault prediction model according to claim 1, characterized in that, In step S6, the thresholds for the precision, recall, and mean F1 score of all initial fault prediction models are 70%, 90%, and 0.85, respectively.

9. The method for establishing an integrated circuit packaging fault prediction model according to claim 1, characterized in that, In step S1, the equipment in the integrated circuit packaging process includes chip mounting equipment, bonding equipment, molding packaging equipment, electroplating equipment, dicing equipment, testing equipment, and cleaning equipment; among which: Chip placement equipment includes high-speed pick and place machines and flip-chip bonding machines. Its fault prediction-related features include placement pressure, positioning accuracy, temperature profile and solder paste amount. Typical fault prediction scenarios are chip misalignment, insufficient solder paste and cold solder joints. Bonding equipment includes wire bonding machines and carrier tape automatic bonding equipment. Its fault prediction-related features include bonding pressure, ultrasonic power, bonding time, and gold wire tension. Typical fault prediction scenarios are wire breakage, insufficient bonding strength, and poor solder joint. Molding and packaging equipment includes transfer molding machines and compression molding machines. Its failure prediction-related characteristics include molding pressure, temperature distribution, curing time and resin flowability. Typical failure prediction scenarios are bubble defects, incomplete filling and stress cracking. Electroplating equipment includes chemical plating equipment and electroplating production lines. Its fault prediction features include current density, electroplating solution composition, temperature and deposition time. Typical fault prediction scenarios are uneven plating, open circuit and short circuit. Cutting equipment includes wafer dicing machines and laser dicing machines. The characteristics related to fault prediction include cutting speed, blade wear and cooling system parameters. Typical fault prediction scenarios are chip breakage, cutting deviation and thermal damage. The testing equipment includes probe stations and automated testing equipment. Its fault prediction-related features include test voltage / current, contact resistance, and signal integrity. Typical fault prediction scenarios include open / short circuit detection failure and parameter drift misjudgment. Cleaning equipment includes plasma cleaners and wet cleaning equipment. Its fault prediction features include cleaning time, gas flow rate, temperature and chemical solution concentration. Typical fault prediction scenarios are contaminant residue and surface damage.

10. An integrated circuit packaging fault prediction method based on the integrated circuit packaging fault prediction model establishment method according to any one of claims 1-9, characterized in that, Includes the following steps: 1) Acquire multi-source data from equipment at each stage of integrated circuit packaging during operation, and preprocess the multi-source data; 2) Filter features related to fault prediction within the preprocessed multi-source data; 3) Input the features related to fault prediction into the fault prediction model to obtain the current fault status of any device in each stage of integrated circuit packaging.

Images