Big data mining method and system based on implantation sub-packaging process development

Abstract

The invention provides a big data mining method and system based on implantation sub-packaging process development, and relates to the field of data processing, and the method comprises the steps: carrying out the process parameter detection based on a to-be-mined packaging process development data set to obtain first confidence distribution, carrying out the process parameter detection through a detection neural network to obtain second confidence distribution, in the detection process, in combination with the detection performance of the detection neural network, the problem of inaccurate detection caused by the inherent classification capability in the detection mining of the pre-trained neural network is relieved, and the detection precision is improved. Based on the packaging process development data set to be mined and the target process parameters, performing type mapping on the target process parameters to obtain third confidence degree distribution and fourth confidence degree distribution so as to obtain target process parameter types; in the type mapping process, the classification performance of the type mapping neural network is combined again, inaccurate detection in classification mining of the pre-training neural network is relieved, and the type mapping precision is improved.

Description

Technical Field

[0001] This application relates to the field of data processing, and more specifically, to a big data mining method and system developed based on implantation sub-encapsulation technology. Background Technology

[0002] In modern manufacturing, particularly in the medical device industry, the packaging process for implantable medical devices is crucial for ensuring product quality and safety. The packaging process not only affects the physical stability and durability of the device but also directly relates to its performance and long-term reliability within the human body. Therefore, developing and optimizing packaging processes is a key step in improving the overall performance of implantable medical devices.

[0003] Traditionally, packaging process development relies on trial and error and limited experimental verification. This approach is not only time-consuming and labor-intensive, but also struggles to comprehensively cover all possible combinations of process parameters and types. With the rapid development of big data and artificial intelligence technologies, it has become possible to use data mining techniques to assist in the development and optimization of packaging processes. However, existing data mining methods often suffer from inaccurate detection and imprecise type mapping when faced with complex and ever-changing packaging process data. This results in low reliability of the mined process parameters and type information, failing to effectively guide actual production. Especially in implant packaging processes, subtle differences in process parameters can significantly impact the performance of the final product, thus placing higher demands on the accuracy and robustness of data mining methods. Existing data mining methods typically rely on a single neural network model for process parameter detection and type mapping. The limitation of this approach is that a single model often struggles to simultaneously achieve both detection accuracy and generalization ability, particularly when facing new process parameters and types, easily leading to false positives and false negatives. Furthermore, traditional data mining methods often neglect the correlations and dependencies between data when processing packaging process data, resulting in a lack of globality and consistency in the mined process parameter and type information. This not only affects the accuracy and reliability of data mining, but also limits its application and promotion in actual production. Summary of the Invention

[0004] The purpose of this invention is to provide a big data mining method and system based on implantation sub-encapsulation technology.

[0005] In a first aspect, this application provides a big data mining method based on implantation sub-encapsulation technology, the method comprising:

[0006] Obtain the packaging process development dataset to be mined;

[0007] Based on the proposed packaging process development dataset, a pre-trained neural network is used to detect process parameters in the proposed packaging process development dataset to obtain a first confidence distribution for a first process parameter library. A detection neural network is then used to detect process parameters in the proposed packaging process development dataset to obtain a second confidence distribution for the first process parameter library, which includes multiple undetermined process parameters.

[0008] Based on the first confidence distribution and the second confidence distribution, the target process parameters in the packaging process development dataset to be mined are obtained;

[0009] Based on the proposed packaging process development dataset and target process parameters, the target process parameters are type-mapped using the pre-trained neural network to obtain a third confidence distribution for the second process parameter library. The target process parameters are then type-mapped using the type mapping neural network to obtain a fourth confidence distribution for the second process parameter library, which includes multiple undetermined process parameter types.

[0010] Based on the third and fourth confidence distributions, the target process parameter type is obtained.

[0011] Based on the target process parameters and the target process parameter type, the mining results of the packaging process development dataset to be mined are obtained.

[0012] Optionally, the detection neural network is obtained by adjusting the commonality measurement results between the first inference process parameters and the prior labels of process parameters in each first training dataset; the type mapping neural network is obtained by adjusting the commonality measurement results between the second inference process parameter types and the prior labels of types in each second training dataset.

[0013] The pre-trained neural network is obtained by debugging based on the commonality measurement results between the third mining results and the training prior labels of each third training dataset. The third mining results include third inference process parameters and third process parameter types, and the training prior labels include process parameter prior labels and type prior labels of the third training dataset.

[0014] Optionally, the method further includes:

[0015] Based on the proposed packaging process development dataset, process parameters are detected on the proposed packaging process development dataset using a type mapping neural network to obtain a fifth confidence distribution for the first process parameter library.

[0016] Based on the packaging process development dataset and target process parameters to be mined, the target process parameters are type-mapped through the detection neural network to obtain the sixth confidence distribution for the second process parameter library;

[0017] The step of obtaining the target process parameters in the packaging process development dataset to be mined based on the first confidence distribution and the second confidence distribution includes:

[0018] The target process parameters are obtained based on the first confidence distribution, the second confidence distribution, and the fifth confidence distribution.

[0019] The step of obtaining the target process parameter type based on the third confidence distribution and the fourth confidence distribution includes:

[0020] Based on the third confidence distribution, the fourth confidence distribution, and the sixth confidence distribution, the target process parameter type corresponding to the target process parameter is obtained.

[0021] Optionally, the method further includes:

[0022] Obtain parameter mining instructions; wherein, the parameter mining instructions include detection instructions and type mapping instructions, the detection instructions indicate that the process parameters of the target process flow in the packaging process development dataset to be mined are detected, and the type mapping instructions indicate that the detected process parameters are type-mapped;

[0023] The first confidence distribution is obtained by detecting process parameters in the proposed packaging process development dataset using a pre-trained neural network based on the proposed packaging process development dataset and detection instructions. The second confidence distribution is obtained by detecting process parameters in the proposed packaging process development dataset using a detection neural network based on the proposed packaging process development dataset and detection instructions. The third confidence distribution is obtained by performing type mapping on the target process parameters using a pre-trained neural network based on the proposed packaging process development dataset, target process parameters, and type mapping instructions. The fourth confidence distribution is obtained by performing type mapping on the target process parameters using a type mapping neural network based on the proposed packaging process development dataset, target process parameters, and type mapping instructions.

[0024] Optionally, obtaining the target process parameters in the packaging process development dataset to be mined based on the first confidence distribution and the second confidence distribution includes:

[0025] Based on the commonality measurement results between the first confidence distribution and the second confidence distribution, the first detection difference of the second confidence distribution is determined;

[0026] Based on the difference between the second confidence distribution and the first detection difference, a first corrected confidence level is determined for the first confidence distribution;

[0027] The first confidence distribution is corrected based on the first corrected confidence level to obtain the first target confidence distribution;

[0028] The target process parameters are obtained based on the first target confidence distribution.

[0029] Optionally, obtaining the target process parameters based on the first confidence distribution, the second confidence distribution, and the fifth confidence distribution includes:

[0030] Based on the commonality measurement results between the first confidence distribution and the second confidence distribution, the first detection difference of the second confidence distribution is determined;

[0031] Based on the commonality measurement results between the first confidence distribution and the fifth confidence distribution, the second detection difference of the fifth confidence distribution is determined;

[0032] Based on the difference between the second confidence distribution and the first detection difference, a first corrected confidence level for the first confidence distribution is determined;

[0033] The first confidence level distribution is corrected based on the first corrected confidence level;

[0034] The second target confidence distribution is determined based on the difference between the corrected first confidence distribution and the second detection difference.

[0035] The target process parameters are obtained based on the second target confidence distribution.

[0036] Optionally, the process of determining the first confidence distribution and the second confidence distribution includes:

[0037] Based on the proposed packaging process development dataset, a pre-trained neural network is used to detect process parameters in the dataset to obtain a first basic confidence distribution for a first process parameter library. Then, a detection neural network is used to detect process parameters in the dataset to obtain a second basic confidence distribution for the first process parameter library. The first basic confidence distribution includes the first basic confidence of each undetermined process parameter, and the second basic confidence distribution includes the second basic confidence of each undetermined process parameter.

[0038] Based on the first basic confidence distribution and the second basic confidence distribution, target process parameters with both the first basic confidence and the second basic confidence not less than a set critical value are determined among the various undetermined process parameters.

[0039] The confidence scores corresponding to the undetermined process parameters other than the target undetermined process parameters in the first basic confidence score distribution and the second basic confidence score distribution are cleaned to obtain the first confidence score distribution and the second confidence score distribution.

[0040] Optionally, the packaging process development dataset to be mined includes multiple target process parameters, and the mining results include multiple target process parameters and the target process parameter type of the target process parameters;

[0041] The mining results are obtained based on completing x rounds of mining. Each round of mining is either a detection mining or a type mapping mining, and the result of each round of mining is the target process parameter or the target process parameter type.

[0042] The e-th round of mining includes:

[0043] Based on the results of the (e-1)th round of mining, the result category of the eth round is determined. The result category is either a detection category or a type mapping category. The result category corresponding to the first round of mining is the detection category.

[0044] Based on the packaging process dataset to be mined and the results of the first e-1 rounds of mining, a shared confidence distribution is obtained through a pre-trained neural network, a detection confidence distribution is obtained through a detection neural network, and a mapping confidence distribution is obtained through a type mapping neural network.

[0045] Wherein, the shared confidence distribution is a first confidence distribution or a third confidence distribution, the detection confidence distribution is a second confidence distribution or a sixth confidence distribution, and the mapping confidence distribution is a fourth confidence distribution or a fifth confidence distribution;

[0046] Based on the shared confidence distribution, detection confidence distribution, and mapping confidence distribution, the results of the e-th round of mining are obtained.

[0047] Optionally, the debugging process of the detection neural network includes:

[0048] Obtain a first sample library, which includes multiple first training datasets carrying training prior labels. The training prior labels of each first training dataset include prior labels of process parameters for that first training dataset.

[0049] Based on each first training dataset, process parameters are detected in each first training dataset using a detection neural network to be tuned, thereby obtaining the first inference process parameters for each first training dataset.

[0050] Based on each first training dataset and the first inference process parameter, the first inference process parameter is type-mapped through the detection neural network to be tuned to obtain the first inference process parameter type. Based on the first inference process parameter and the first inference process parameter type of each first training dataset, the mining result of each first training dataset is obtained.

[0051] Based on the commonality measurement results between the first inference process parameters and the prior labels of process parameters in each first training dataset, a first calibration error value is determined, and based on the first calibration error value, the network parameters of the detection neural network to be calibrated are optimized in detail.

[0052] The training process of the type mapping neural network includes:

[0053] Obtain a second sample library, which includes multiple second training datasets carrying training prior labels. The training prior labels of each second training dataset include type prior labels of each process parameter prior label of the second training dataset.

[0054] Based on each second training dataset, process parameters are detected for each second training dataset through a type mapping neural network to be tuned, thereby obtaining the second inference process parameters for each second training dataset.

[0055] Based on each second training dataset and the second inference process parameter, the second inference process parameter is type-mapped through a type mapping neural network to be tuned, thereby obtaining the second inference process parameter type of the second inference process parameter. Based on the second inference process parameter and the second inference process parameter type of each second training dataset, the mining result of each second training dataset is obtained.

[0056] Based on the common measurement results between the second inference process parameter type and the type prior label of each second training dataset, the second calibration error value is determined, and the network parameters of the type mapping neural network to be calibrated are optimized in detail based on the second calibration error value.

[0057] The training process of the pre-trained neural network includes:

[0058] Obtain a third sample library, which includes multiple third training datasets carrying training prior labels. The training prior labels of each third training dataset include the process parameter prior labels of the third training dataset and the type prior labels of each process parameter prior label.

[0059] Based on each third training dataset, process parameters are detected in each third training dataset through a pre-trained neural network to be tuned, thereby obtaining the third inference process parameters for each third training dataset.

[0060] Based on each third training dataset and third inference process parameter, the third inference process parameter is type-mapped through a pre-trained neural network to be tuned, thereby obtaining the third inference process parameter type. Based on the third inference process parameter and third inference process parameter type of each third training dataset, the mining result of each third training dataset is obtained.

[0061] Based on the commonality measurement results between the third mining results of each third training dataset and the training prior labels, the third tuning error value is determined, and the network parameters of the pre-trained neural network to be tuned are optimized in detail based on the third tuning error value.

[0062] In a second aspect, this application provides a computer system comprising: one or more processors; a memory; and one or more computer programs; wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, they implement the method described above.

[0063] The beneficial effects of this application include: This application provides a big data mining method and system based on implantable packaging technology development. Based on the packaging technology development dataset to be mined, a first confidence distribution is obtained through process parameter detection using a pre-trained neural network, and a second confidence distribution is obtained through process parameter detection using a detection neural network. Based on the first and second confidence distributions, the target process parameters in the packaging technology development dataset to be mined are obtained. In this way, during the detection process, the detection performance of the detection neural network can be combined to alleviate the inaccurate detection caused by the inherent classification ability of the pre-trained neural network, increasing the detection accuracy. Furthermore, based on the packaging technology development dataset to be mined and the target process parameters, a third confidence distribution is obtained by type mapping of the target process parameters using a pre-trained neural network, and a fourth confidence distribution is obtained by type mapping using a type mapping neural network. The type of the target process parameter is obtained through the third and fourth confidence distributions. In this way, during the type mapping process, the classification performance of the type mapping neural network is combined again to alleviate the inaccurate detection problem caused by the detection ability in the classification mining of the pre-trained neural network, increasing the accuracy of type mapping. Based on this, this application divides the mining process into detection and type mapping, using different networks for detection and type mapping tasks at each stage. This makes the mining results more accurate, facilitating the extraction of precise process parameters and helping to predict potential quality issues. This not only improves the stability and efficiency of the implant packaging process but also provides data support for continuous improvement of the manufacturing process, ensuring the safety and reliability of implantable medical devices. Attached Figure Description

[0064] Figure 1 This is a flowchart of a big data mining method based on implantation sub-encapsulation technology provided in an embodiment of this application.

[0065] Figure 2 This is a schematic diagram of the composition of a computer system provided in an embodiment of this application. Detailed Implementation

[0066] In this application embodiment, the big data mining method developed based on implantation sub-encapsulation technology is executed by a computer system, including but not limited to servers, personal computers, laptops, tablets, and smartphones. Figure 1 As shown, the method includes:

[0067] Step S100: Obtain the packaging process development dataset to be mined.

[0068] In step S100, the computer system acquires the packaging process development dataset to be mined. The packaging process development dataset is a collection of data related to the implantation sub-packaging process, containing various data information involved in the development of the implantation sub-packaging process.

[0069] For example, in the encapsulation process, data may be involved, including material property data (such as electrical conductivity, thermal conductivity, and hardness), process equipment parameter data (such as operating temperature, pressure, and processing speed), process flow sequence data (such as the order of encapsulation operations), and product quality-related data (such as the sealing performance test results and electrical performance test results of the encapsulated product). These data collectively constitute the encapsulation process development dataset. Computer systems can acquire this dataset using various techniques. A common approach is extraction from a database. Assuming there exists a dedicated database storing data related to the encapsulation process, accumulated over a long period of process development and production, the computer system can retrieve the required data using specific database query statements. For instance, if the database uses SQL (Structured Query Language) to manage data, the computer system can use a statement like "SELECT * FROM encapsulation_process_development_data WHERE [condition]" to retrieve the data. Here, "*" indicates that all columns (i.e., all attributes of the data) are selected, "encapsulation_process_development_data" is the table name where the data is stored, and "[conditions]" can be filtering conditions set according to specific needs, such as only obtaining process data within a specific time period, or only obtaining process data related to a specific material, etc.

[0070] Alternatively, data may be imported from external files. If the data is stored in CSV (Comma-Separated Values) file format, the computer system can use appropriate file reading functions to read the file content and convert it into a data structure that it can process internally. For example, in Python, the `read_csv` function from the `pandas` library can be used to read CSV files, as shown in the following code:

[0071] Python

[0072] import pandas as pd

[0073] data=pd.read_csv('encapsulation_process_data.csv')

[0074] ```

[0075] The "encapsulation_process_data.csv" file here is an external file storing packaging process development data. Using the code described above, the computer system can read the data from this file into a variable named "data". This variable represents a portion of the packaging process development dataset to be mined. By combining or using various similar techniques individually, the computer system can obtain the complete packaging process development dataset to be mined, laying the foundation for subsequent process parameter mining.

[0076] Step S200: Based on the packaging process development dataset to be mined, process parameters are detected on the packaging process development dataset to be mined through a pre-trained neural network to obtain a first confidence distribution for the first process parameter library. Process parameters are detected on the packaging process development dataset to be mined through a detection neural network to obtain a second confidence distribution for the first process parameter library. The first process parameter library includes multiple undetermined process parameters.

[0077] In step S200, the computer system develops a dataset based on the packaging process to be mined, and uses a pre-trained neural network to detect process parameters in the dataset to obtain a first confidence distribution for the first process parameter library. At the same time, it uses a detection neural network to detect process parameters in the same dataset to obtain a second confidence distribution for the first process parameter library. Here, the first process parameter library contains multiple undetermined process parameters.

[0078] A pre-trained neural network is a neural network model that has been pre-trained on a large amount of data. For example, in the field of implant packaging technology, this pre-trained neural network may be trained on a large amount of previously collected data on different packaging processes. This data contains various combinations of process parameters and their corresponding results. Through this large-scale data training, the network learns the relationships between different process parameters and their impact on the final result.

[0079] For process parameter detection, let's take temperature as an example in implanted sub-packages. Temperature is a critical process parameter during packaging; different temperatures can affect the performance of packaging materials, the strength of the package, and so on. The computer system inputs the packaging process development dataset to be mined into a pre-trained neural network. The network analyzes the temperature-related features in the dataset based on its internal neuron connections and weights, thereby evaluating the probability that each undetermined process parameter (temperature in this case) is in the first process parameter library. This evaluation result is presented in the form of a confidence distribution. The confidence distribution represents the probability distribution of each undetermined process parameter belonging to the correct process parameter. Assuming there are three undetermined temperature values ​​in the first process parameter library, T1, T2, and T3, the first confidence distribution that the pre-trained neural network might derive is P1(T1) = 0.3, P1(T2) = 0.4, and P1(T3) = 0.3, indicating that T2 has a relatively high probability of being the correct process parameter, but this is only a preliminary evaluation.

[0080] Detection neural networks are also specifically designed for process parameter detection, although they may differ from pre-trained neural networks in structure, training data, or training methods. The computer system also inputs the packaging process development dataset to be mined into the detection neural network for process parameter detection. Taking temperature parameters as an example, the detection neural network might derive a second confidence distribution of P2(T1) = 0.25, P2(T2) = 0.5, and P2(T3) = 0.25 for these three undetermined temperature values.

[0081] In terms of technical means, deep learning frameworks such as TensorFlow or PyTorch can be used to implement neural networks. Taking TensorFlow as an example, when building a pre-trained neural network, the network structure must first be defined. Assuming this is a simple multilayer perceptron (MLP) network, the number of neurons in the input layer depends on the number of features in the packaging process development dataset to be mined. There can be several hidden layers, and the number of neurons in each layer can be determined based on experience or experimentation. The number of neurons in the output layer is the same as the number of undetermined process parameters in the first process parameter library. The code for defining the network structure is as follows:

[0082] Python

[0083] import tensorflow as tf

[0084] # Define the input layer

[0085] input_layer=tf.keras.layers.Input(shape=(input_feature_size,))

[0086] # Define hidden layer

[0087] hidden_layer1=tf.keras.layers.Dense(units=hidden_size1,activation='relu')(input_layer)

[0088] hidden_layer2=tf.keras.layers.Dense(units=hidden_size2,activation='relu')(hidden_layer1)

[0089] # Define the output layer

[0090] output_layer=tf.keras.layers.Dense(units=num_pending_parameters,activation='softmax')(hidden_layer2)

[0091] #Building a Model

[0092] model=tf.keras.Model(inputs=input_layer, outputs=output_layer)

[0093] ```

[0094] Where `input_feature_size` is the number of input features, `hidden_size1` and `hidden_size2` are the number of hidden layer neurons, and `num_pending_parameters` is the number of pending process parameters in the first process parameter library.

[0095] When performing process parameter detection, the packaging process development dataset to be mined is converted into a format suitable for network input (e.g., normalization), and then the `model.predict` function is used for prediction to obtain the confidence distribution.

[0096] Similar network construction and detection methods can be used for detection neural networks, but the network parameters (such as weight initialization and learning rate) or structure (such as the number of hidden layers and neurons) may differ. The first and second confidence distributions obtained from the pre-trained and detection neural networks, respectively, provide important basis for subsequently determining the target process parameters in the proposed packaging process development dataset. These two confidence distributions evaluate the process parameters from different perspectives; their combination can improve the accuracy of process parameter detection, avoid the limitations of a single neural network, and thus better uncover accurate process parameters for implantation sub-packaging process development.

[0097] Step S300: Based on the first confidence distribution and the second confidence distribution, obtain the target process parameters in the packaging process development dataset to be mined.

[0098] In step S300, the computer system obtains the target process parameters in the packaging process development dataset to be mined based on the first confidence distribution and the second confidence distribution obtained in step S200.

[0099] The first confidence distribution is the confidence level of each undetermined process parameter in the first process parameter library, obtained by the computer system through pre-trained neural networks detecting process parameters in the packaging process development dataset to be mined. For example, if the undetermined process parameter in the first process parameter library is the soldering temperature in the implantation sub-packaging process, and assuming there are three undetermined soldering temperature values ​​T1, T2, and T3, the first confidence distribution might be P1(T1) = 0.3, P1(T2) = 0.4, and P1(T3) = 0.3. This means that according to the detection of the pre-trained neural network, T2 is relatively likely to be the correct soldering temperature parameter, but it cannot be determined yet.

[0100] The second confidence distribution is the confidence level of each undetermined process parameter in the first process parameter library, obtained by the computer system through a detection neural network detecting process parameters on the same dataset. Continuing with the welding temperature example above, the second confidence distribution might be P2(T1) = 0.25, P2(T2) = 0.5, and P2(T3) = 0.25, indicating that the detection neural network considers T2 to be the most likely correct parameter.

[0101] The computer system combines these two confidence distributions to determine the target process parameter. One possible technique is to compare their consistency and difference. If the confidence level of a certain process parameter is relatively high in both confidence distributions, then this parameter is more likely to be the target process parameter. For example, a comprehensive evaluation index, such as the weighted average method, can be set to calculate the comprehensive confidence level of each process parameter. Let the weight of the first confidence distribution be α, and the weight of the second confidence distribution be β (α+β=1). For welding temperature T1, the comprehensive confidence level P(T1)=α×P1(T1)+β×P2(T1); for T2, P(T2)=α×P1(T2)+β×P2(T2); for T3, P(T3)=α×P1(T3)+β×P2(T3).

[0102] Assuming α = 0.4 and β = 0.6, then for T1, P(T1) = 0.4 × 0.3 + 0.6 × 0.25 = 0.27; for T2, P(T2) = 0.4 × 0.4 + 0.6 × 0.5 = 0.46; and for T3, P(T3) = 0.4 × 0.3 + 0.6 × 0.25 = 0.27. By comparing the magnitudes of P(T1), P(T2), and P(T3), the computer system can determine T2 as the target process parameter because it has the highest overall confidence level.

[0103] Another approach is based on confidence intervals. According to the law of large numbers in probability theory, when the sample size is sufficiently large, confidence intervals can be used to measure the reliability of a process parameter to be determined. The computer system calculates the confidence interval for each process parameter under a first confidence distribution and a second confidence distribution. If the confidence intervals for a process parameter under the two confidence distributions have a large overlap, and the confidence level corresponding to this overlap is high, then this process parameter is more likely to be the target process parameter.

[0104] For example, for welding temperature T2, the 95% confidence interval under the first confidence distribution is [39.5, 40.5], and the 95% confidence interval under the second confidence distribution is [39.8, 40.2]. These two intervals have a large overlap, which indicates that T2 has high confidence in both detection results. Therefore, the computer system can determine T2 as the target process parameter.

[0105] In this way, the computer system, based on the first and second confidence distributions and taking into account the detection results of the pre-trained neural network and the detection neural network, can more accurately obtain the target process parameters in the packaging process development dataset to be mined, providing key process parameter support for subsequent packaging process development.

[0106] Step S400: Based on the packaging process development dataset and target process parameters to be mined, the target process parameters are type-mapped through a pre-trained neural network to obtain a third confidence distribution for the second process parameter library. The target process parameters are then type-mapped through a type mapping neural network to obtain a fourth confidence distribution for the second process parameter library, which includes multiple undetermined process parameter types.

[0107] In step S400, the computer system, based on the packaging process development dataset to be mined and the target process parameters obtained in step S300, performs type mapping on the target process parameters through a pre-trained neural network to obtain a third confidence distribution for the second process parameter library. Furthermore, it performs type mapping on the target process parameters through a type mapping neural network to obtain a fourth confidence distribution for the second process parameter library, where the second process parameter library contains multiple undetermined process parameter types.

[0108] First, let's clarify the concept of target process parameters. Target process parameters are those that are significant to the packaging process and are determined through comprehensive analysis in step S300. For example, in the implantation sub-packaging process, if the determined target process parameter is the soldering temperature, this temperature value has a crucial impact on the quality and effectiveness of the entire packaging process.

[0109] Next, we will explain type mapping. Type mapping is the process of classifying target process parameters into specific process parameter types. There are multiple ways to classify the undetermined process parameter types in the second process parameter library. For example, based on the nature of their impact on the packaging process, they can be divided into critical impact parameter types and minor impact parameter types; based on the physical properties of the parameters, they can be divided into temperature-related types, pressure-related types, electrical characteristic-related types, etc.

[0110] When a computer system uses a pre-trained neural network for type mapping, taking soldering temperature as an example, the pre-trained neural network has already learned from a large amount of packaging process data, which includes different temperature values ​​and their corresponding process parameter types. When a target process parameter (soldering temperature) is input, the network evaluates the probability of each undetermined process parameter type in a second process parameter library based on its internal neuron connection weights and learned patterns, obtaining a third confidence distribution. Assuming there are three undetermined process parameter types in the second process parameter library, namely Temperature Critical Influence Type (TKI), Temperature Secondary Influence Type (TSI), and Temperature-Independent Type (NT), the third confidence distribution that the pre-trained neural network might obtain is P3(TKI) = 0.6, P3(TSI) = 0.3, and P3(NT) = 0.1, indicating that the target process parameter soldering temperature is highly likely to belong to the Temperature Critical Influence Type. However, this is only a preliminary evaluation based on the pre-trained neural network.

[0111] Similarly, type mapping neural networks perform a similar task. This network is specifically built and trained for mapping process parameters. When the computer system inputs the target process parameter, welding temperature, into the type mapping neural network, it evaluates each potential process parameter type based on its structure and training results, deriving a fourth confidence distribution. For example, the fourth confidence distribution derived by the type mapping neural network might be P4(TKI) = 0.55, P4(TSI) = 0.35, and P4(NT) = 0.1.

[0112] In terms of technical means, type mapping for pre-trained neural networks can be implemented using methods based on deep learning frameworks (such as TensorFlow or PyTorch). Taking TensorFlow as an example, the first step is to ensure that the pre-trained neural network has an output layer structure suitable for the type mapping task. If the output layer uses a Softmax function to represent the probability distribution, for the three undetermined process parameter types mentioned above, the output layer will have three neurons.

[0113] Assuming the input layer of the pre-trained neural network receives target process parameters (represented as a vector, containing various features related to soldering temperature, such as the specific stage in the packaging process and its correlation with other process parameters), the hidden layers can consist of multiple fully connected layers, each using an activation function such as ReLU. During training, the pre-trained neural network adjusts the connection weights between neurons by minimizing a loss function. The loss function can be the cross-entropy loss function, with the following formula:

[0114]

[0115] Where n is the number of undetermined process parameter types (here n = 3), y i This is the actual label of the process parameter type to which the target process parameter belongs (1 for temperature-critical influence type, 0 for others), p i It is the probability of the i-th undetermined process parameter type predicted by the pre-trained neural network (i.e., the value in the third confidence distribution).

[0116] During type mapping, the computer system inputs the target process parameters into the pre-trained neural network after preprocessing (such as normalization), and calculates the output layer result, i.e., the third confidence distribution, through the forward propagation algorithm.

[0117] The construction and training process for type mapping neural networks is similar, but they may differ in parameters such as network depth, number of neurons, or learning rate. For example, the hidden layers of a type mapping neural network may use different numbers of neurons or different inter-layer connection methods to adapt to its specific type mapping task. In this way, the computer system uses the third and fourth confidence distributions obtained from the pre-trained neural network and the type mapping neural network, respectively, to provide important basis for accurately determining the type of target process parameters. This helps to further understand the properties and functions of various process parameters in the packaging process, thereby providing more accurate information for optimizing the packaging process.

[0118] Step S500: Based on the third confidence distribution and the fourth confidence distribution, obtain the target process parameter type of the target process parameter.

[0119] In step S500, the computer system obtains the target process parameter type based on the third confidence distribution and the fourth confidence distribution obtained in step S400.

[0120] The third confidence distribution is obtained by the computer system after mapping the target process parameters to types using a pre-trained neural network. It represents the confidence level of each undetermined process parameter type in the second process parameter library. For example, the undetermined process parameter types in the second process parameter library might be material property types (such as electrical property types, mechanical property types, etc.) or process operation types (such as welding types, packaging types, etc.) related to the implantation and packaging process. Assuming there are three undetermined process parameter types: electrical property type (E), mechanical property type (M), and process operation type (P), the third confidence distribution obtained by the pre-trained neural network might be P3(E) = 0.3, P3(M) = 0.4, and P3(P) = 0.3. This indicates that, according to the evaluation of the pre-trained neural network, the target process parameter is relatively likely to belong to the mechanical property type, but it cannot be determined yet.

[0121] The fourth confidence distribution is the confidence level of each undetermined process parameter type in the second process parameter library, obtained by the computer system through type mapping neural network to map the target process parameter to its type. For the same three undetermined process parameter types, the fourth confidence distribution obtained by the type mapping neural network may be P4(E) = 0.25, P4(M) = 0.5, and P4(P) = 0.25, indicating that the type mapping neural network considers the target process parameter to be most likely to belong to the mechanical characteristic type.

[0122] The computer system combines these two confidence distributions to determine the target process parameter type. A feasible technique is to use a weighted summation method. The weight of the third confidence distribution is set as α, and the weight of the fourth confidence distribution is set as β (α + β = 1). The overall confidence level for each type of process parameter to be determined is calculated. For the electrical characteristic type (E), the overall confidence level P(E) = α × P3(E) + β × P4(E); for the mechanical characteristic type (M), P(M) = α × P3(M) + β × P4(M); for the process operation type (P), P(P) = α × P3(P) + β × P4(P).

[0123] Assuming α = 0.4 and β = 0.6, then for E, P(E) = 0.4 × 0.3 + 0.6 × 0.25 = 0.27; for M, P(M) = 0.4 × 0.4 + 0.6 × 0.5 = 0.46; and for P, P(P) = 0.4 × 0.3 + 0.6 × 0.25 = 0.27. By comparing the magnitudes of P(E), P(M), and P(P), the computer system can determine that the mechanical characteristic type (M) is the target process parameter type because it has the highest overall confidence level.

[0124] Another technical approach can be based on Bayesian decision theory. According to Bayes' formula... The third and fourth confidence distributions are considered as prior probabilities. The posterior probability of the target process parameter belonging to each of the various undetermined process parameter types is calculated given these two distributions. Here, A represents the target process parameter belonging to a certain undetermined process parameter type, and B represents the combination of the known third and fourth confidence distributions. The undetermined process parameter type with the highest posterior probability calculated is the target process parameter type.

[0125] In this way, the computer system, based on the third and fourth confidence distributions and considering the type mapping results of the pre-trained neural network and the type mapping neural network, can more accurately obtain the target process parameter type. This helps to deeply understand the specific properties and categories of the target process parameters in the implantation sub-packaging process, providing an important basis for process optimization and quality control.

[0126] Step S600: Based on the target process parameters and the target process parameter type, obtain the mining results of the packaging process development dataset to be mined.

[0127] In step S600, the computer system obtains the mining results of the packaging process development dataset to be mined based on the target process parameters and the target process parameter type.

[0128] The target process parameter is the parameter that is critical to the development of the packaging process, as determined in step S300. For example, in the implantation sub-packaging process, the target process parameter might be a specific soldering temperature value. The target process parameter type is the type to which the process parameter belongs, as determined in step S500. For example, this soldering temperature value belongs to the critical temperature parameter type.

[0129] The mining results represent a comprehensive analysis of the packaging process development dataset to be mined, encompassing the target process parameters extracted from the dataset and their corresponding types. This result is significant for a deeper understanding of the various factors and their interrelationships in the packaging process development process.

[0130] Taking the implantation sub-packaging process as an example, suppose the packaging process development dataset to be mined contains various data from multiple batches of implantation sub-packaging processes, such as temperature, pressure, material properties, and operation time at different stages. In the previous steps, the target process parameter was determined to be a certain soldering temperature (e.g., 200℃), and its target process parameter type is the critical temperature parameter type.

[0131] The computer system integrates the target process parameter and its type into the overall dataset analysis framework to derive mining results. One possible technique is to construct a data structure to store and display the mining results. For example, a table structure with multiple fields can be used, where one column stores the target process parameter, another column stores the corresponding target process parameter type, and other columns can be used to record other data related to the target process parameter, such as the frequency of the process parameter in the dataset and its correlation with other process parameters.

[0132] From a correlation perspective, the relationship between the target process parameter and other process parameters can be quantified by calculating the correlation coefficient. Assuming another process parameter is welding pressure, the Pearson correlation coefficient between the two can be calculated. x i This represents a sample value indicating the welding temperature. It is the average welding temperature, y i The sample value represents the welding pressure. Here, is the average welding pressure, and n is the sample size. This coefficient determines the degree of linear correlation between welding temperature and welding pressure; this correlation value can be recorded in a table as part of the data mining results. Furthermore, the computer system can classify or filter the entire dataset based on target process parameters and their types. For example, if the target process parameter type is a critical temperature parameter, the system can filter out all data records related to critical temperatures and further analyze the patterns and characteristics in these records, such as the success rate of implanted sub-packages and product quality characteristics at different critical temperatures. These analytical results will also be integrated into the data mining results.

[0133] In this way, the computer system uses the target process parameters and target process parameter types to comprehensively and deeply mine the packaging process development dataset to be mined. The mining results can provide valuable reference for the optimization of implanted sub-packaging processes, quality control, cost reduction and other aspects, which helps to improve the overall efficiency of packaging processes and product quality.

[0134] In one implementation, the detection neural network is obtained by tuning based on the commonality measurement results between the first inference process parameters and the prior labels of the process parameters in each first training dataset; the type mapping neural network is obtained by tuning based on the commonality measurement results between the second inference process parameter types and the prior labels of the types in each second training dataset.

[0135] The pre-trained neural network is obtained by debugging based on the commonality measurement results between the third mining results and the training prior labels of each third training dataset. The third mining results include the third inference process parameters and the third process parameter types. The training prior labels include the process parameter prior labels and type prior labels of the third training dataset.

[0136] In this implementation, the computer system debugs the detection neural network, the type mapping neural network, and the pre-trained neural network to meet the corresponding functional requirements.

[0137] The detection neural network is obtained by tuning based on the commonality measurement results between the first inference process parameters and the prior labels of process parameters in each first training dataset (i.e., the packaging process development dataset samples). Firstly, the first training dataset is the sample data used to train the detection neural network. This data is a collection of data from the packaging process development process, containing information related to various process parameters. For example, in implant packaging process development, the first training dataset may contain data on temperature, pressure, and material composition ratios of different batches of implants during packaging.

[0138] The first inferred process parameter is the result obtained by the computer system through the detection neural network to be calibrated, which detects process parameters for each first training dataset. Suppose a first training dataset involves the welding process during sub-package implantation. After the computer system inputs this dataset into the detection neural network to be calibrated, the network, based on its initial weights and structure, infers key parameters (such as welding temperature) in the welding process, obtaining an inferred welding temperature value. This value is the first inferred process parameter.

[0139] The prior label of the process parameter is the correct process parameter label in each first training dataset that is known in advance. For example, for the welding process mentioned above, the actual accurate welding temperature value is known, and this known accurate value is the prior label of the process parameter.

[0140] Commonality metrics are used to measure the similarity or consistency between the first inference process parameters and the prior labels of those parameters. Computer systems use these commonality metrics to evaluate the performance of the detection neural network being tuned. For example, mean squared error (MSE) can be used to calculate the commonality metric, as shown in the formula: Where n is the number of samples, y i These are priori markers (true values) of process parameters. This is the first inference process parameter (predicted value). If the MSE value is large, it indicates a significant difference between the first inference process parameter and the prior label of the process parameter, suggesting that the network performance needs improvement. Based on this commonality metric result, the computer system adjusts the network parameters of the detection neural network, such as adjusting the connection weights between neurons, to reduce the value of the commonality metric result and improve the accuracy of the network.

[0141] For the type mapping neural network, it is obtained by debugging based on the commonality measurement results between the second inference process parameter type and the type prior label of each second training dataset. The second training dataset is also sample data used for training, similar to the first training dataset but possibly different in data content or structure.

[0142] The second inference process parameter type is the result obtained by the computer system after detecting process parameters and further mapping them using a type mapping neural network to be tuned for each second training dataset. For example, for a certain process parameter in the implantation sub-packaging process, after processing by the type mapping neural network, it is inferred that this process parameter belongs to the "critical temperature influence type", which is the second inference process parameter type.

[0143] Type prior labels are the correct type labels for process parameters in each second training dataset, which are predetermined in advance. For example, if a process parameter is known to belong to the "critical temperature influence type", this is a type prior label.

[0144] The computer system calculates a commonality measure between the second inference process parameter type and the type prior label, which can be measured, for example, by accuracy, as shown in the formula: If the accuracy is low, it indicates that the type mapping neural network is underperforming. Based on this commonality metric, the computer system adjusts the network parameters of the type mapping neural network to improve the accuracy of the network in mapping process parameters to types.

[0145] For pre-trained neural networks, adjustments are made based on the commonality metrics between the third mining results of various third training datasets and the training prior labels. The third training datasets contain more comprehensive information, and their training prior labels include process parameter prior labels and type prior labels.

[0146] The third mining results include third inferred process parameters and third process parameter types. These are obtained by the computer system through process parameter detection and type mapping on the third training dataset using a pre-trained neural network to be tuned. For example, for a process parameter in the implantation sub-packaging process, the pre-trained neural network infers the value of this process parameter (third inferred process parameter) and determines the type to which this process parameter belongs (third process parameter type).

[0147] The computer system calculates a commonality metric between the third-party mining results and the training prior labels. This metric comprehensively considers the accuracy of both process parameters and the type of process parameters. For example, an F1-score can be used for comprehensive evaluation. Precision and recall are calculated by comparing the results of the third mining operation with the training prior labels. If the F1-score is low, the computer system adjusts the network parameters of the pre-trained neural network based on this commonality metric, improving the accuracy of the pre-trained neural network in both process parameter detection and type mapping. Through this debugging process, the three neural networks can better adapt to data mining tasks in packaging process development, improving the accuracy and reliability of the overall big data mining method.

[0148] In one implementation, the method may further include:

[0149] Step S101: Based on the packaging process development dataset to be mined, process parameters are detected on the packaging process development dataset to be mined through a type mapping neural network to obtain the fifth confidence distribution for the first process parameter library.

[0150] In step S101, the computer system uses a type mapping neural network to detect process parameters in the packaging process development dataset to be mined, based on the packaging process development dataset to be mined, and obtains the fifth confidence distribution for the first process parameter library.

[0151] First, the proposed packaging process development dataset contains multifaceted data related to the implantation sub-packaging process. For example, during the implantation sub-packaging process, the dataset may cover material properties (such as the electrical and thermal conductivity of materials), equipment parameters (such as the temperature and pressure settings of the packaging equipment), and operational procedures (such as the execution sequence and time of each step) for different production batches.

[0152] Type mapping neural networks are specifically designed for type mapping tasks, and here they are used for process parameter detection. They may differ from other neural networks (such as the pre-trained neural networks and detection neural networks mentioned earlier) in structure, function, and training methods. When constructing a type mapping neural network, the structural parameters such as the connection method of its neurons, the number of layers, and the number of neurons per layer are set according to the requirements of processing the type mapping task.

[0153] The first process parameter library contains multiple undetermined process parameters. Taking the welding process in the implantation sub-packaging process as an example, the undetermined process parameters may include welding temperature, welding time, welding current, etc. After the computer system inputs the packaging process development dataset to be mined into the type mapping neural network, the network evaluates the probability of each undetermined process parameter in the first process parameter library according to its internal algorithm and weights, thereby obtaining the fifth confidence distribution.

[0154] Assuming the undetermined process parameters in the first process parameter library are welding temperature (T1), welding time (T2), and welding current (T3), the fifth confidence distribution that the type mapping neural network might derive is P5(T1) = 0.2, P5(T2) = 0.3, and P5(T3) = 0.5. This means that, according to the detection of the type mapping neural network, welding current (T3) is relatively likely to be the correct process parameter, but this is only a preliminary evaluation result.

[0155] In terms of technical means, type-mapping neural networks can be implemented based on deep learning frameworks such as TensorFlow or PyTorch. Taking TensorFlow as an example, when building a type-mapping neural network, the network structure needs to be defined first. Assuming the network is a multilayer perceptron (MLP), the number of neurons in the input layer depends on the number of features in the packaging process development dataset to be mined. For example, if the dataset contains 10 features related to the welding process, then the input layer will have 10 neurons. Several hidden layers can be set, with the number of neurons in each layer determined empirically or experimentally. For example, with two hidden layers, the first layer has 20 neurons, and the second layer has 15 neurons. The number of neurons in the output layer is the same as the number of undetermined process parameters in the first process parameter library, which is 3 in this case. The code defining the network structure is as follows:

[0156] import tensorflow as tf

[0157] # Define the input layer

[0158] input_layer=tf.keras.layers.Input(shape=(10,))

[0159] # Define hidden layer

[0160] hidden_layer1=tf.keras.layers.Dense(units=20,activation='relu')(input_layer)

[0161] hidden_layer2=tf.keras.layers.Dense(units=15,activation='relu')(hidden_layer1)

[0162] # Define the output layer

[0163] output_layer=tf.keras.layers.Dense(units=3,

[0164] activation='softmax')(hidden_layer2)

[0165] # Construct the model: model = tf.keras.Model(inputs = input_layer, outputs = output_layer)

[0166] When performing process parameter detection, the dataset for developing the packaging process to be mined needs to be preprocessed to make it suitable for network input. For example, the data can be normalized, mapping each feature value to the [0,1] interval. Then, the model.predict function is used for prediction to obtain the fifth confidence distribution.

[0167] Step S102: Based on the packaging process development dataset and target process parameters to be mined, the target process parameters are type-mapped by a detection neural network to obtain the sixth confidence distribution for the second process parameter library.

[0168] In step S102, the computer system, based on the packaging process development dataset and target process parameters to be mined, performs type mapping on the target process parameters through a detection neural network to obtain the sixth confidence distribution for the second process parameter library.

[0169] The target process parameters are parameters that are important to the packaging process and are determined in the preceding steps (such as certain operations in step S300). For example, in the implantation sub-packaging process, if the soldering temperature has been determined as the target process parameter, the value of this parameter has a critical impact on the quality and efficiency of the packaging process.

[0170] The detection neural network is a neural network used for process parameter detection; here, it is used to perform type mapping of target process parameters. The second process parameter library contains multiple types of pending process parameters. Taking the implantation sub-packaging process as an example, the types of pending process parameters may include temperature-related types (such as critical temperature influence types and minor temperature influence types), pressure-related types, and material property-related types, etc.

[0171] The computer system provides the target packaging process development dataset and the target process parameters as input to the detection neural network. Based on its internal weights and algorithm, the detection neural network evaluates the probability that the target process parameter belongs to each type of undetermined process parameter in the second process parameter library, thus obtaining the sixth confidence distribution.

[0172] Assuming the types of undetermined process parameters in the second process parameter library are critical temperature influence (KTI), minor temperature influence (STI), and non-temperature-dependent (NTI), the sixth confidence distribution that the detection neural network might derive for the target process parameter, welding temperature, is P6(KTI) = 0.4, P6(STI) = 0.3, and P6(NTI) = 0.3. This indicates that, according to the assessment of the detection neural network, the target process parameter welding temperature is relatively likely to belong to the critical temperature influence (KTI) type, but this is only a preliminary assessment result.

[0173] The implementation of detection neural networks can also be based on deep learning frameworks. For example, a detection neural network can be built using the PyTorch framework. First, the network structure is defined, assuming a Convolutional Neural Network (CNN) structure (which may be more effective in some cases for processing data with certain structural features). The input layer of the network needs to be designed based on the representation of the dataset and target process parameters of the packaging process to be mined. If the dataset and target process parameters are represented as a two-dimensional matrix (e.g., the data is converted into a matrix through some encoding method), the size of the input layer is determined by the size of this matrix.

[0174] Assuming the input layer size is [3,3] (this is just an example), multiple convolutional layers can be set, such as one convolutional layer with a kernel size of [2,2], a stride of 1, and padding of 0. The pooling layer can use max pooling with a kernel size of [2,2]. Then, after several fully connected layers, the data is mapped to a dimension equal to the number of undetermined process parameter types in the second process parameter library, which is assumed to be 3. When training and using this network, the input data needs to be appropriately processed, such as encoding the dataset and target process parameters according to a predefined encoding method, and then inputting the encoded data into the network for forward propagation calculation to obtain the sixth confidence distribution.

[0175] Based on this, step S300, according to the first confidence distribution and the second confidence distribution, obtains the target process parameters in the packaging process development dataset to be mined, which may include:

[0176] Step S310: Obtain the target process parameters based on the first confidence distribution, the second confidence distribution, and the fifth confidence distribution.

[0177] Step S500, which obtains the target process parameter type based on the third confidence distribution and the fourth confidence distribution, may include:

[0178] Step S510: Based on the third confidence distribution, the fourth confidence distribution and the sixth confidence distribution, obtain the target process parameter type corresponding to the target process parameter.

[0179] As one implementation method, step S310, obtaining the target process parameters based on the first confidence distribution, the second confidence distribution, and the fifth confidence distribution, may include:

[0180] Step S311: Based on the commonality measurement results between the first confidence distribution and the second confidence distribution, determine the first detection difference of the second confidence distribution;

[0181] Step S312: Based on the commonality measurement results between the first confidence distribution and the fifth confidence distribution, determine the second detection difference of the fifth confidence distribution;

[0182] Step S313: Based on the difference between the second confidence distribution and the first detection difference, determine the first corrected confidence level of the first confidence distribution;

[0183] Step S314: Correct the first confidence level distribution based on the first corrected confidence level;

[0184] Step S315: Determine the second target confidence distribution based on the difference between the corrected first confidence distribution and the second detection difference;

[0185] Step S316: Obtain the target process parameters based on the second target confidence distribution.

[0186] In step S101, the computer system uses a type mapping neural network to detect process parameters in the packaging process development dataset to be mined, based on the packaging process development dataset to be mined, and obtains the fifth confidence distribution for the first process parameter library.

[0187] First, the proposed packaging process development dataset contains multifaceted data related to the implantation sub-packaging process. For example, during the implantation sub-packaging process, the dataset may cover material properties (such as the electrical and thermal conductivity of materials), equipment parameters (such as the temperature and pressure settings of the packaging equipment), and operational procedures (such as the execution sequence and time of each step) for different production batches.

[0188] Type mapping neural networks are specifically designed for type mapping tasks, and here they are used for process parameter detection. They may differ from other neural networks (such as the pre-trained neural networks and detection neural networks mentioned earlier) in structure, function, and training methods. When constructing a type mapping neural network, the structural parameters such as the connection method of its neurons, the number of layers, and the number of neurons per layer are set according to the requirements of processing the type mapping task.

[0189] The first process parameter library contains multiple undetermined process parameters. Taking the welding process in the implantation sub-packaging process as an example, the undetermined process parameters may include welding temperature, welding time, welding current, etc. After the computer system inputs the packaging process development dataset to be mined into the type mapping neural network, the network evaluates the probability of each undetermined process parameter in the first process parameter library according to its internal algorithm and weights, thereby obtaining the fifth confidence distribution.

[0190] Assuming the undetermined process parameters in the first process parameter library are welding temperature (T1), welding time (T2), and welding current (T3), the fifth confidence distribution that the type mapping neural network might derive is P5(T1) = 0.2, P5(T2) = 0.3, and P5(T3) = 0.5. This means that, according to the detection of the type mapping neural network, welding current (T3) is relatively likely to be the correct process parameter, but this is only a preliminary evaluation result.

[0191] In terms of technical means, type-mapping neural networks can be implemented based on deep learning frameworks such as TensorFlow or PyTorch. Taking TensorFlow as an example, when building a type-mapping neural network, the network structure needs to be defined first. Assuming the network is a multilayer perceptron (MLP), the number of neurons in the input layer depends on the number of features in the packaging process development dataset to be mined. For example, if the dataset contains 10 features related to the welding process, then the input layer will have 10 neurons. Several hidden layers can be set, with the number of neurons in each layer determined empirically or experimentally. For example, with two hidden layers, the first layer has 20 neurons, and the second layer has 15 neurons. The number of neurons in the output layer is the same as the number of undetermined process parameters in the first process parameter library, which is 3 in this case. The code defining the network structure is as follows:

[0192] Python

[0193] import tensorflow as tf

[0194] # Define the input layer

[0195] input_layer=tf.keras.layers.Input(shape=(10,))

[0196] # Define hidden layer

[0197] hidden_layer1=tf.keras.layers.Dense(units=20,activation='relu')(input_layer)

[0198] hidden_layer2=tf.keras.layers.Dense(units=15,activation='relu')(hidden_layer1)

[0199] # Define the output layer

[0200] output_layer=tf.keras.layers.Dense(units=3,activation='softmax')(hidden_layer2)

[0201] #Building a Model

[0202] model=tf.keras.Model(inputs=input_layer, outputs=output_layer)

[0203] ```

[0204] When performing process parameter detection, the dataset for developing the packaging process to be mined needs to be preprocessed to make it suitable for network input. For example, the data can be normalized, mapping each feature value to the [0,1] interval. Then, the `model.predict` function is used for prediction to obtain the fifth confidence distribution.

[0205] In step S102, the computer system, based on the packaging process development dataset and target process parameters to be mined, performs type mapping on the target process parameters through a detection neural network to obtain the sixth confidence distribution for the second process parameter library.

[0206] The target process parameters are parameters that are important to the packaging process and are determined in the preceding steps (such as certain operations in step S300). For example, in the implantation sub-packaging process, if the soldering temperature has been determined as the target process parameter, the value of this parameter has a critical impact on the quality and efficiency of the packaging process.

[0207] The detection neural network is a neural network used for process parameter detection; here, it is used to perform type mapping of target process parameters. The second process parameter library contains multiple types of pending process parameters. Taking the implantation sub-packaging process as an example, the types of pending process parameters may include temperature-related types (such as critical temperature influence types and minor temperature influence types), pressure-related types, and material property-related types, etc.

[0208] The computer system provides the target packaging process development dataset and the target process parameters as input to the detection neural network. Based on its internal weights and algorithm, the detection neural network evaluates the probability that the target process parameter belongs to each type of undetermined process parameter in the second process parameter library, thus obtaining the sixth confidence distribution.

[0209] Assuming the types of undetermined process parameters in the second process parameter library are critical temperature influence (KTI), minor temperature influence (STI), and non-temperature-dependent (NTI), the sixth confidence distribution that the detection neural network might derive for the target process parameter, welding temperature, is P6(KTI) = 0.4, P6(STI) = 0.3, and P6(NTI) = 0.3. This indicates that, according to the assessment of the detection neural network, the target process parameter welding temperature is relatively likely to belong to the critical temperature influence (KTI) type, but this is only a preliminary assessment result.

[0210] In terms of technical means, the implementation of detection neural networks can also be based on deep learning frameworks. For example, a detection neural network can be built using the PyTorch framework. First, the network structure is defined, assuming a convolutional neural network (CNN) structure is used (which may be more effective in some cases for processing data with certain structural features). The input layer of the network needs to be designed according to the representation of the dataset and target process parameters of the packaging process to be mined. If the dataset and target process parameters are represented as a two-dimensional matrix (e.g., the data is converted into a matrix through some encoding method), the size of the input layer is determined according to the size of this matrix.

[0211] Assuming the input layer size is [3,3] (this is just an example), multiple convolutional layers can be set, such as one convolutional layer with a kernel size of [2,2], a stride of 1, and padding of 0. The pooling layer can use max pooling with a kernel size of [2,2]. Then, after several fully connected layers, the data is mapped to a dimension equal to the number of undetermined process parameter types in the second process parameter library, which is assumed to be 3. When training and using this network, the input data needs to be appropriately processed, such as encoding the dataset and target process parameters according to a predefined encoding method, and then inputting the encoded data into the network for forward propagation calculation to obtain the sixth confidence distribution.

[0212] In step S310, the computer system obtains the target process parameters based on the first confidence distribution, the second confidence distribution, and the fifth confidence distribution.

[0213] The first confidence distribution is the confidence distribution of a first process parameter library obtained by a computer system through pre-trained neural networks to detect process parameters in the packaging process development dataset to be mined. For example, if the undetermined process parameters in the first process parameter library are welding temperature (T1), welding time (T2), and welding current (T3), the first confidence distribution obtained by the pre-trained neural network may be P1(T1) = 0.3, P1(T2) = 0.4, and P1(T3) = 0.3.

[0214] The second confidence distribution is the confidence distribution against the first process parameter library obtained by the computer system through the detection neural network to detect process parameters of the packaging process development dataset to be mined. It is assumed that the second confidence distribution obtained by the detection neural network is P2(T1) = 0.25, P2(T2) = 0.5, and P2(T3) = 0.25.

[0215] The fifth confidence distribution is the confidence distribution against the first process parameter library obtained by the computer system through the type mapping neural network to detect process parameters of the packaging process development dataset to be mined. As mentioned above, it is assumed that P5(T1) = 0.2, P5(T2) = 0.3, and P5(T3) = 0.5.

[0216] In a specific implementation of step S310, the computer system obtains the target process parameters based on the first confidence distribution, the second confidence distribution, and the fifth confidence distribution through steps S311-S316.

[0217] In step S311, the computer system first calculates the commonality measurement result between the first confidence distribution and the second confidence distribution. Here, the first confidence distribution is the confidence distribution of the first process parameter library obtained by the computer system through process parameter detection on the packaging process development dataset to be mined using a pre-trained neural network, and the second confidence distribution is the confidence distribution of the first process parameter library obtained by detecting process parameters on the same dataset using a detection neural network.

[0218] For example, the undetermined process parameters in the first process parameter library are welding temperature (T1), welding time (T2), and welding current (T3). Assume the first confidence distribution obtained from the pre-trained neural network is P1(T1) = 0.3, P1(T2) = 0.4, and P1(T3) = 0.3, and the second confidence distribution obtained from the detection neural network is P2(T1) = 0.25, P2(T2) = 0.5, and P2(T3) = 0.25.

[0219] Computer systems can use various methods to calculate commonality metrics, a common method being the calculation of Euclidean distance. Treating the first and second confidence distributions as three-dimensional vectors (each undetermined process parameter corresponds to one dimension), i.e., P1 = [0.3, 0.4, 0.3], P2 = [0.25, 0.5, 0.25], the Euclidean distance formula is... Here n = 3, x i It is an element in P1, y i It is an element in P2. Substituting it into the calculation yields... The Euclidean distance d is a measure of the commonality between the first and second confidence distributions, reflecting the degree of difference between them.

[0220] The first detection difference of the second confidence distribution is determined based on the calculated commonality metric. Assuming a threshold T1 is set (this threshold can be determined empirically or experimentally, for example, T1 = 0.1), if d > T1, it indicates a large difference between the first and second confidence distributions.

[0221] At this point, the second confidence distribution can be adjusted according to certain rules to determine the first detection difference. For example, a method can be used to scale each element in the second confidence distribution according to a ratio of distance d, with a scaling factor f = 1 / (1+d). Then, the elements in the adjusted second confidence distribution... This adjusted result is the second confidence distribution after taking into account the differences in the first detection.

[0222] In step S312, the computer system then calculates the commonality measure between the first confidence distribution and the fifth confidence distribution. The fifth confidence distribution is the confidence distribution against the first process parameter library obtained by the computer system through process parameter detection of the packaging process development dataset to be mined using a type mapping neural network.

[0223] Assuming the first confidence distribution mentioned above is P1(T1) = 0.3, P1(T2) = 0.4, P1(T3) = 0.3, the fifth confidence distribution obtained by the type mapping neural network is P5(T1) = 0.2, P5(T2) = 0.3, P5(T3) = 0.5.

[0224] Similarly, Euclidean distance is used to calculate the commonality metric results. P1 and P5 are treated as vectors, and the Euclidean distance is calculated. This d' is the commonality measure between the first confidence distribution and the fifth confidence distribution.

[0225] 2. Determination of the difference in the second detection

[0226] The second detection difference of the fifth confidence distribution is determined based on this distance d'. Assuming the preset threshold is T2 (e.g., T2 = 0.15), if d' > T2, the fifth confidence distribution is adjusted according to a certain adjustment rule.

[0227] For example, using scaling factors The fifth confidence level distribution is adjusted. The elements in the adjusted fifth confidence level distribution... This is the fifth confidence distribution after taking into account the differences in the second detection.

[0228] In step S313, the computer system determines the first corrected confidence level of the first confidence distribution based on the difference between the second confidence distribution and the first detection difference. After obtaining the second confidence distribution P2' after considering the first detection difference, P2'-P2 is calculated (the subtraction here is the subtraction of corresponding elements).

[0229] For example, the difference obtained by P2'-P2 is ΔP2 as [P2′(T1)-P2(T1), P2′(T2)-P2(T2), P2′(T3)-P2(T3)], that is...

[0230] The first revised confidence level of the first confidence distribution is determined based on this difference vector ΔP2. One method is to add each element of the first confidence distribution to the corresponding element of the difference vector ΔP2, obtaining the first revised confidence distribution P1' = P1 + ΔP2P2. This P1' is the result after the first correction to the first confidence distribution.

[0231] In step S314, the computer system corrects the first confidence distribution based on the first corrected confidence level P1'. That is, the original first confidence distribution is replaced with P1'. This step ensures that the first confidence distribution takes into account the relationship between the second confidence distribution and the first detection difference, thereby more accurately reflecting the probability of the process parameters to be determined.

[0232] In step S315, the computer system determines the second target confidence distribution based on the difference between the corrected first confidence distribution P1' and the second detection difference. First, P1'-P5' (where P5' is the fifth confidence distribution after considering the second detection difference) is calculated to obtain a new difference vector ΔP15.

[0233] For example, ΔP15 is [P1′(T1)-P5′(T1),P1′(T2)-P5′(T2),P1′(T3)-P5′(T3)]. Substituting the previously calculated values ​​of P1′ and P5′ into this value will give you the specific difference vector.

[0234] The confidence distribution of the second target is determined based on this difference vector ΔP15. Assuming a weighted average method is used, with weights w1 = 0.6 and w2 = 0.4, for each undetermined process parameter T... i The element P(T) in the second target confidence distribution i )=w1×P1′(T i )+w2×P5′(T i )-ΔP15(T iFor example, for welding temperature T1, P(T1) = 0.6 × P1′(T1) + 0.4 × P5′(T1) - ΔP15(T1). Similarly, P(T2) and P(T3) can be calculated, thus obtaining a second target confidence distribution that comprehensively considers the first, second, and fifth confidence distributions. In step S316, the computer system determines the target process parameters based on the obtained second target confidence distribution. For example, in the second target confidence distribution, the confidence values ​​corresponding to each undetermined process parameter are compared. Assuming that for welding temperature (T1), welding time (T2), and welding current (T3), the second target confidence distribution is P(T1) = 0.25, P(T2) = 0.35, and P(T3) = 0.4, since the confidence value corresponding to welding current (T3) is the highest, the computer system determines welding current (T3) as the target process parameter.

[0235] Through steps S311-S316, the computer system uses the first confidence distribution, the second confidence distribution, and the fifth confidence distribution to perform a series of calculations and adjustments to finally determine the target process parameters. This method comprehensively considers the confidence distribution of multiple neural network detection results, which helps to improve the accuracy of the determination of target process parameters.

[0236] In step S510, the computer system obtains the target process parameter type corresponding to the target process parameter based on the third confidence distribution, the fourth confidence distribution, and the sixth confidence distribution.

[0237] The third confidence distribution is the confidence distribution of the target process parameter to the second process parameter library obtained by the computer system through type mapping of the target process parameter using a pre-trained neural network. For example, if the types of the undetermined process parameters in the second process parameter library are critical temperature influence type (KTI), minor temperature influence type (STI), and non-temperature related type (NTI), the third confidence distribution obtained by the pre-trained neural network may be P3(KTI) = 0.6, P3(STI) = 0.3, and P3(NTI) = 0.1.

[0238] The fourth confidence distribution is the confidence distribution of the second process parameter library obtained by the computer system through type mapping neural network to perform type mapping on the target process parameters. Assume that the fourth confidence distribution obtained by the type mapping neural network is P4(KTI) = 0.55, P4(STI) = 0.35, and P4(NTI) = 0.1.

[0239] The sixth confidence distribution is the confidence distribution of the second process parameter library obtained by the computer system through the detection neural network to perform type mapping on the target process parameters. As mentioned above, it is assumed that P6(KTI) = 0.4, P6(STI) = 0.3, and P6(NTI) = 0.3.

[0240] The computer system calculates the commonality measure between the third and fourth confidence distributions. For example, using cosine similarity as the commonality measure, for vectors P3 = [0.6, 0.3, 0.1] and P4 = [0.55, 0.35, 0.1], the cosine similarity formula is: Here n = 3, x i It is an element in P3, y i It is an element in P4. The calculated cosine similarity value reflects the commonality measure between the third and fourth confidence distributions.

[0241] The first type of mapping difference is determined based on this commonality measurement result. Assuming a threshold T3 is set, if the calculated cosine similarity is less than T3, it indicates a significant difference between the two. In this case, the fourth confidence distribution can be adjusted according to certain rules. For example, the adjustment rule is to scale each element in the fourth confidence distribution according to the difference in cosine similarity. Let the scaling factor f1 = 1 - (1 - cosine_similarity) × k (k is a coefficient determined empirically, such as k = 0.5). The elements in the adjusted fourth confidence distribution are P4′(KTI) = P4(KTI) × f1, P4′(STI) = P4(STI) × f1, and P4′(NTI) = P4(NTI) × f1. This adjusted result is the fourth confidence distribution after considering the first type of mapping difference.

[0242] Following a similar method, the commonality measure between the third and sixth confidence distributions is calculated. For example, for vectors P3 = [0.6, 0.3, 0.1] and P6 = [0.4, 0.3, 0.3], the cosine similarity formula is applied again. (Here n = 3, x i It is an element in P3, y i (The element in P6) is used to calculate the cosine similarity value between the two, which reflects the result of the commonality measurement between them.

[0243] The second type of mapping difference is determined based on this commonality measurement result. A threshold T4 is set; if the cosine similarity value is less than T4, it indicates a significant difference between the two. The sixth confidence distribution is then adjusted according to certain rules. For example, let the adjustment factor f2 = 1 - (1 - cosine_similarity′) × m (where m is a coefficient determined empirically, such as m = 0.3). Then, the elements in the adjusted sixth confidence distribution are P6′(KTI) = P6(KTI) × f2, P6′(STI) = P6(STI) × f2, and P6′(NTI) = P6(NTI) × f2. This is the sixth confidence distribution after considering the second type of mapping difference.

[0244] After obtaining the fourth confidence distribution considering the first type of mapping difference, the computer system calculates the difference between the fourth confidence distribution and the first type of mapping difference. For example, for each type of undetermined process parameter, P4'-P4 (element-wise subtraction) is calculated to obtain a difference vector ΔP4. Based on this difference vector ΔP4, the first corrected type confidence of the third confidence distribution is determined. For example, using addition, the first corrected type confidence distribution P3' = P3 + ΔP4 is obtained. This P3' is the result after the first correction of the third confidence distribution, taking into account the relationship between the fourth confidence distribution and the first type of mapping difference.

[0245] The computer system directly uses the first corrected type confidence level P3' obtained in the previous steps to correct the third confidence level distribution. That is, the original third confidence level distribution is replaced with P3', making the third confidence level distribution more accurately reflect the probability that the target process parameter belongs to each of the various undetermined process parameter types. After correcting the third confidence level distribution to obtain P3', the computer system calculates the difference between P3' and the sixth confidence level distribution (i.e., P6') after considering the difference in the second type mapping. For example, for each undetermined process parameter type, P3'-P6' is calculated to obtain a new difference vector ΔP36.

[0246] The target confidence distribution for the target process parameter type is determined based on this difference vector ΔP36. Assuming a weighted average method is used, with weights w3 = 0.7 and w4 = 0.3, for each undetermined process parameter type Tt... i The element P(Tt) in the target confidence distribution i )=w3×P3′(Tt i )+w4×P6′(Tt i )-ΔP36(Tt i This yields a target confidence distribution that comprehensively considers the third, fourth, and sixth confidence distributions.

[0247] The computer system determines the target process parameter type corresponding to the target process parameter based on the obtained target confidence distribution. For example, in the target confidence distribution, the confidence values ​​corresponding to each undetermined process parameter type are compared. Assuming that for the critical temperature influence type (KTI), minor temperature influence type (STI), and non-temperature-dependent type (NTI), the target confidence distribution is P(KTI) = 0.5, P(STI) = 0.3, and P(NTI) = 0.2, since the critical temperature influence type (KTI) has the highest confidence value, the computer system determines that the critical temperature influence type (KTI) is the target process parameter type corresponding to the target process parameter (such as the welding temperature mentioned earlier).

[0248] Through these steps S101-S102, S310 and S510, the computer system can more comprehensively and accurately obtain the target process parameters and the corresponding target process parameter types from the packaging process development dataset to be mined. This helps to deeply understand the various relationships in the packaging process and provides more reliable data support for process optimization, quality control and other tasks.

[0249] In one implementation, the method may further include:

[0250] Step S202: Obtain parameter mining instructions; wherein, the parameter mining instructions include detection instructions and type mapping instructions. The detection instructions indicate that the process parameters of the target process flow in the packaging process development dataset to be mined are detected, and the type mapping instructions indicate that the detected process parameters are type-mapped.

[0251] In step S202, the computer system obtains parameter mining instructions, which include detection instructions and type mapping instructions.

[0252] A detection instruction indicates the need to detect process parameters of a target process flow in a dataset for developing packaging processes. For example, in a dataset for developing implantable packaging processes, the target process flow might be a soldering process. The detection instruction is an instruction to detect process parameters in this soldering process, such as soldering temperature, soldering time, and soldering current. These process parameters have a direct or indirect impact on soldering quality and the final packaging effect. Technically, a computer system can store detection instructions in memory in a specific data format, such as a structure containing the instruction type (detection instruction), the target process flow name (soldering process), and relevant process parameter identifiers (temperature, time, current, etc.).

[0253] Type mapping instructions represent the type mapping of detected process parameters. Continuing with the welding process as an example, after detecting process parameters such as welding temperature and welding time, the type mapping instruction determines the type of these parameters. For example, welding temperature might be mapped to a critical influence temperature type or a secondary influence temperature type; welding time might be mapped to an operation time type or a process delay type, etc. This type mapping helps to gain a deeper understanding of the nature and role of process parameters in the entire packaging process. In terms of technical implementation, computer systems can process type mapping instructions by defining a type mapping rule table. For example, a two-dimensional table can be created, where rows represent different process parameters, columns represent possible types, and each cell in the table can store an association weight or matching degree value, indicating the probability or degree of association of a certain process parameter belonging to a certain type.

[0254] In the context of big data mining throughout the development of implantable packaging processes, acquiring parameter mining instructions is a crucial step. Computer systems can obtain these instructions from external devices (such as user input terminals) or internal instruction queues (if an instruction pre-generation and queuing mechanism exists). Assuming a scenario where a user inputs instructions through a specific software interface, the computer system listens for events generated by the interaction between the software interface and the user. When it detects user input and confirms it as a parameter mining instruction, it acquires and parses it. This allows for subsequent operations on the packaging process development dataset to be mined, such as calling a pre-trained neural network and a detection neural network to detect process parameters based on detection instructions, or calling a pre-trained neural network and a type mapping neural network to perform type mapping operations based on type mapping instructions.

[0255] In one implementation, the first confidence distribution is obtained by detecting process parameters in the proposed packaging process development dataset using a pre-trained neural network based on the proposed packaging process development dataset and detection instructions; the second confidence distribution is obtained by detecting process parameters in the proposed packaging process development dataset using a detection neural network based on the proposed packaging process development dataset and detection instructions; the third confidence distribution is obtained by performing type mapping on the target process parameters using a pre-trained neural network based on the proposed packaging process development dataset, target process parameters, and type mapping instructions; and the fourth confidence distribution is obtained by performing type mapping on the target process parameters using a type mapping neural network based on the proposed packaging process development dataset, target process parameters, and type mapping instructions.

[0256] In this implementation, the computer system obtains a first confidence distribution, a second confidence distribution, a third confidence distribution, and a fourth confidence distribution through different operations. These distributions are of great significance in the big data mining process of the entire implant sub-packaging process development.

[0257] First, there's the first confidence distribution, obtained by the computer system using a pre-trained neural network to detect process parameters within the target packaging process development dataset, based on the dataset to be mined and the detection instructions. The target packaging process development dataset contains various data related to the implantation sub-packaging process. For example, in the implantation sub-packaging process, it may involve performance data of different materials at different process stages, such as the hardness and flexibility of a certain polymer material changing with temperature and pressure during the packaging process; it may also include equipment operating parameters, such as the temperature settings, pressure settings, and running time of the packaging equipment under different processes. The detection instructions specify that the computer system needs to detect the process parameter-related parts of this dataset.

[0258] A pre-trained neural network is a neural network model pre-trained on a large amount of data. For example, this network may have been trained on numerous previously collected implantation packaging process data, learning some data patterns and feature relationships. When the computer system receives a detection command and inputs the packaging process development dataset to be mined into the pre-trained neural network, the network detects the process parameters in the dataset based on its internal neuron connection weights and activation functions. Assuming there are several undetermined process parameters in the first process parameter library, such as welding temperature, welding time, and packaging pressure, the pre-trained neural network will assign a confidence level to each undetermined process parameter. The set of these confidence levels constitutes the first confidence level distribution. For example, a confidence level of 0.3 might be assigned to welding temperature, indicating that the probability of this parameter being considered a correct process parameter in the current dataset is 0.3; a confidence level of 0.4 might be assigned to welding time, and so on.

[0259] The second confidence distribution is obtained by the computer system using a detection neural network to detect process parameters in the proposed packaging process development dataset based on the dataset and detection instructions. This detection neural network is specifically designed for process parameter detection and may differ from pre-trained neural networks in structure, training data, or training level. Again, taking implanted sub-packaging process data as an example, after the computer system inputs the dataset and detection instructions into the detection neural network, it will also provide confidence scores for the undetermined process parameters in the first process parameter library. For example, a confidence score of 0.25 for soldering temperature and 0.5 for soldering time, etc., constitute the second confidence distribution.

[0260] Next is the third confidence distribution, which is obtained by the computer system using a pre-trained neural network to type-map the target process parameters based on the packaging process development dataset to be mined, the target process parameters, and the type mapping instructions. The target process parameters are those determined in previous steps that are important to the packaging process, such as the previously mentioned soldering temperature. The type mapping instructions instruct the computer system to perform type mapping operations on this target process parameter. The pre-trained neural network has already learned some relationships between process parameters and their types during its previous training. After the computer system inputs the packaging process development dataset to be mined, the target process parameter (soldering temperature), and the type mapping instructions into the pre-trained neural network, the network will give a confidence score for each undetermined process parameter type in the second process parameter library, thus forming the third confidence distribution. For example, the types of undetermined process parameters in the second process parameter library may include temperature-critical influence types, temperature-secondary influence types, and non-temperature-related types. The pre-trained neural network may give a confidence level of 0.6 for temperature-critical influence types, 0.3 for temperature-secondary influence types, and 0.1 for non-temperature-related types.

[0261] Finally, there is the fourth confidence distribution. This is obtained by the computer system using a type mapping neural network to map the target process parameters to the desired packaging process development dataset, target process parameters, and type mapping instructions. The type mapping neural network is specifically designed for type mapping tasks. Taking soldering temperature as an example, after the computer system inputs the relevant data into the type mapping neural network, it assigns a confidence level to the type of the target process parameter in the second process parameter library. Assuming a confidence level of 0.55 for the temperature-critical influence type, 0.35 for the temperature-secondary influence type, and 0.1 for the non-temperature-related type, these confidence levels constitute the fourth confidence distribution.

[0262] Operations on neural networks can be implemented using deep learning frameworks. Taking the TensorFlow framework as an example, when building pre-trained and detection neural networks, the network structure must first be determined. For process parameter detection tasks, the number of neurons in the input layer depends on the number of features in the packaging process development dataset to be mined. Assuming the dataset has 100 feature-related data points, the input layer can have 100 neurons. Hidden layers can be set to several layers according to actual needs, and the number of neurons in each layer can be determined through experimentation and optimization, such as setting 3 hidden layers with 200, 150, and 100 neurons respectively. The number of neurons in the output layer is related to the number of process parameters to be determined or the number of types of process parameters to be determined. When pre-trained neural networks perform process parameter detection (obtaining the first confidence distribution), the mean squared error (MSE) can be used as the loss function to train the network. By continuously adjusting the network weights to minimize the loss function, the accuracy of the network in process parameter detection and type mapping is improved. A similar structure and training method can be used for type mapping neural networks, but the specific network parameters and training data may differ to adapt to the needs of the type mapping task.

[0263] In another implementation of step S300, namely, obtaining the target process parameters in the packaging process development dataset to be mined based on the first confidence distribution and the second confidence distribution, may include:

[0264] Step S301: Based on the commonality measurement results between the first confidence distribution and the second confidence distribution, determine the first detection difference of the second confidence distribution;

[0265] Step S302: Based on the difference between the second confidence distribution and the first detection difference, determine the first corrected confidence level for the first confidence distribution;

[0266] Step S303: Correct the first confidence distribution based on the first corrected confidence level to obtain the first target confidence distribution;

[0267] Step S304: Obtain the target process parameters based on the first target confidence distribution.

[0268] In this embodiment of step S300, the computer system obtains the target process parameters in the packaging process development dataset to be mined based on the first confidence distribution and the second confidence distribution through steps S301-S304.

[0269] In step S301, the computer system calculates the commonality measurement result between the first confidence distribution and the second confidence distribution. The first confidence distribution is the confidence distribution of the first process parameter library obtained by the computer system through process parameter detection on the packaging process development dataset to be mined using a pre-trained neural network. The second confidence distribution is the confidence distribution of the first process parameter library obtained by detecting process parameters on the same dataset using a detection neural network.

[0270] For example, suppose the undetermined process parameters in the first process parameter library are the soldering temperature (T1), soldering time (T2), and soldering current (T3) in the implantation sub-packaging process. The first confidence distribution obtained by the pre-trained neural network is P1(T1) = 0.3, P1(T2) = 0.4, and P1(T3) = 0.3, and the second confidence distribution obtained by the detection neural network is P2(T1) = 0.25, P2(T2) = 0.5, and P2(T3) = 0.25.

[0271] Computer systems can use various methods to calculate commonality metrics; here, we take cosine similarity calculation as an example. Treating the first and second confidence distributions as vectors, for vectors P1 = [0.3, 0.4, 0.3] and P2 = [0.25, 0.5, 0.25], the cosine similarity formula is: Here n = 3, x i It is an element in P1, y i It is an element in P2. Substituting it into the calculation yields... The cosine similarity value is a measure of the commonality between the first confidence distribution and the second confidence distribution, reflecting the degree of similarity between them.

[0272] The first detection difference of the second confidence distribution is determined based on the calculated commonality measure. Assuming a threshold T1 is set (this threshold can be determined empirically or experimentally, for example, T1 = 0.8), if the calculated cosine similarity is less than T1, it indicates that the similarity between the first and second confidence distributions has not met expectations, and there is a certain difference.

[0273] At this point, the second confidence distribution can be adjusted according to certain rules to determine the first detection difference. For example, a method can be used to scale each element in the second confidence distribution based on the ratio of the difference between the cosine similarity and the threshold. Let the scaling factor f = 1 - (1 - cosine_similarity) × k (k is a coefficient determined empirically, such as k = 0.5), then the elements in the adjusted second confidence distribution are P2′(T1) = P2(T1) × f = 0.25 × (1 - (1 - cosine_similarity) × 0.5), P2′(T2) = P2(T2) × f = 0.5 × (1 - (1 - cosine_similarity) × 0.5), P2′(T3) = P2(T3) × f = 0.25 × (1 - (1 - cosine_similarity) × 0.5). This adjusted result is the second confidence distribution after considering the first detection difference.

[0274] In step S302, the computer system determines the first corrected confidence level for the first confidence level distribution based on the difference between the second confidence level distribution and the first detection difference. After obtaining the second confidence level distribution P2' after considering the first detection difference, P2'-P2 is calculated (the subtraction here is the subtraction of corresponding elements). For example, the difference vector ΔP2 obtained by P2'-P2 is [P2′(T1)-P2(T1), P2′(T2)-P2(T2), P2′(T3)-P2(T3)], that is, [0.25×(1-(1-cosine_similarity)×0.5)-0.25, 0.5×(1-(1-cosine_similarity)×0.5)-0.5, 0.25×(1-(1-cosine_similarity)×0.5)-0.25].

[0275] The first revised confidence level of the first confidence distribution is determined based on this difference vector ΔP2. One method is to add each element in the first confidence distribution to the corresponding element in the difference vector ΔP2 to obtain the first revised confidence distribution P1' = P1 + ΔP2. That is, P1'(T1) = P1(T1) + ΔP2(T1) = 0.3 + (0.25 × (1 - (1 - cosine_similarity) × 0.5) - 0.25), P1'(T2) = P1(T2) + ΔP2(T2) = 0.4 + (0.5 × (1 - (1 - cosine_similarity) × 0.5) - 0.5), P1'(T3) = P1(T3) + ΔP2(T3) = 0.3 + (0.25 × (1 - (1 - cosine_similarity) × 0.5) - 0.25). This P1' is the result after the first correction to the first confidence distribution.

[0276] In step S303, the computer system corrects the first confidence distribution based on the first corrected confidence level P1'. That is, the original first confidence distribution is replaced with P1'. This step ensures that the first confidence distribution takes into account the relationship between the second confidence distribution and the difference in the first detection, thereby more accurately reflecting the probability of the process parameters to be determined.

[0277] In step S304, the computer system obtains the target process parameter based on the corrected first confidence distribution (i.e., P1'). For example, in the corrected first confidence distribution, the confidence value corresponding to each process parameter to be determined is compared. Assuming that for welding temperature (T1), welding time (T2), and welding current (T3), the corrected first confidence distribution is P1'(T1) = 0.28, P1'(T2) = 0.42, and P1'(T3) = 0.3, since the confidence value corresponding to welding time (T2) is the highest, the computer system determines welding time (T2) as the target process parameter.

[0278] Through steps S301-S304 of this implementation, the computer system utilizes the first and second confidence distributions to calculate commonality measurement results, determine detection differences, and correct confidence levels, ultimately determining the target process parameters. This method helps improve the accuracy of target process parameter determination, thereby providing more reliable data support for the development of implantable packaging processes. Technically, these computational operations can be performed using programming languages ​​and related mathematical libraries. For example, in Python, the NumPy library can be used for vector operations, such as calculating cosine similarity. For neural network operations, deep learning frameworks such as TensorFlow or PyTorch can be used to implement the construction, training, and prediction functions of pre-trained and detection neural networks.

[0279] In one implementation, the process of determining the first confidence distribution and the second confidence distribution includes the following steps:

[0280] Step S210: Based on the packaging process development dataset to be mined, process parameters are detected on the packaging process development dataset to be mined using a pre-trained neural network to obtain a first basic confidence distribution for the first process parameter library. Then, process parameters are detected on the packaging process development dataset to be mined using a detection neural network to obtain a second basic confidence distribution for the first process parameter library. The first basic confidence distribution includes the first basic confidence of each undetermined process parameter, and the second basic confidence distribution includes the second basic confidence of each undetermined process parameter.

[0281] Step S220: Based on the first basic confidence distribution and the second basic confidence distribution, determine the target process parameters among the undetermined process parameters whose first basic confidence and second basic confidence are not less than the set critical value;

[0282] Step S230: Clean the confidence scores of the undetermined process parameters other than the target undetermined process parameters in the first basic confidence score distribution and the second basic confidence score distribution respectively to obtain the first confidence score distribution and the second confidence score distribution.

[0283] In the implementation of the process for determining the first confidence distribution and the second confidence distribution, the computer system completes the relevant operations through steps S210-S230.

[0284] In step S210, based on the packaging process development dataset to be mined, a pre-trained neural network is used to detect process parameters in the packaging process development dataset to be mined, so as to obtain a first basic confidence distribution for the first process parameter library. At the same time, a detection neural network is used to detect process parameters in the dataset to obtain a second basic confidence distribution for the first process parameter library.

[0285] The proposed packaging process development dataset contains a wealth of data related to implant packaging processes. For example, in implant packaging processes, the dataset may contain various information about different batches of implants during the packaging process, such as the physical properties of the packaging materials (e.g., hardness, thermal conductivity), the operating parameters of the process equipment (e.g., temperature, pressure, operating speed), and the sequence and timing of operations in the packaging process.

[0286] The first process parameter library contains multiple undetermined process parameters. Taking the welding process in the implantation sub-packaging process as an example, the undetermined process parameters may include welding temperature, welding current, welding time, etc. The pre-trained neural network is a neural network model pre-trained with a large amount of data; it has learned some patterns and feature relationships related to process parameters. When the computer system inputs the packaging process development dataset to be mined into the pre-trained neural network, the network, based on its internal neuron connection weights, activation functions, and other structures, detects each undetermined process parameter in the first process parameter library, thereby obtaining the first basic confidence distribution.

[0287] Assuming the undetermined process parameters in the first process parameter database are welding temperature (T1), welding current (T3), and welding time (T2), the first base confidence distribution that the pre-trained neural network might derive is as follows: for welding temperature T1, the first base confidence is P1_base(T1) = 0.2; for welding current T3, P1_base(T3) = 0.3; and for welding time T2, P1_base(T2) = 0.5. Here, the confidence represents the probability that the pre-trained neural network considers the undetermined process parameter to be a correct process parameter in the current dataset.

[0288] The detection neural network is also used for process parameter detection. It may differ from the pre-trained neural network in terms of structure, training data, or training level. The computer system inputs the same dataset of the packaging process development to be mined into the detection neural network. The detection neural network detects the undetermined process parameters in the first process parameter library and obtains a second base confidence distribution. For example, for welding temperature T1, its second base confidence is P2_base(T1) = 0.15; for welding current T3, P2_base(T3) = 0.35; and for welding time T2, P2_base(T2) = 0.5.

[0289] In step S220, the computer system determines, based on the first basic confidence distribution and the second basic confidence distribution, the target process parameters whose first basic confidence and the second basic confidence are not less than the set critical value among the various process parameters to be determined.

[0290] Setting a critical value is a threshold determined based on actual needs and experience, used to screen out relatively reliable process parameters to be determined. For example, a critical value of 0.3 is set. For welding temperature T1, its first basic confidence level P1_base(T1) = 0.2 is less than 0.3, and its second basic confidence level P2_base(T1) = 0.15 is also less than 0.3, so welding temperature T1 does not meet the conditions for the target process parameter to be determined; for welding current T3, its first basic confidence level P1_base(T3) = 0.3 is not less than 0.3, and its second basic confidence level P2_base(T3) = 0.35 is not less than 0.3, so welding current T3 meets the conditions; for welding time T2, its first basic confidence level P1_base(T2) = 0.5 is not less than 0.3, and its second basic confidence level P2_base(T2) = 0.5 is not less than 0.3, so welding time T2 also meets the conditions. Therefore, in this example, welding current T3 and welding time T2 are the target process parameters to be determined.

[0291] In step S230, the computer system cleans the confidence levels of the undetermined process parameters other than the target undetermined process parameters in the first basic confidence distribution and the second basic confidence distribution, respectively, to obtain the first confidence distribution and the second confidence distribution.

[0292] Confidence cleaning is a data processing operation aimed at highlighting the confidence information of the target process parameter and reducing interference from other less reliable process parameters. A simple cleaning method is to set the confidence level of process parameters other than the target one to a low value (such as 0) or to directly delete these process parameters and their confidence information.

[0293] For the first base confidence distribution, the original confidence levels for welding temperature (T1), welding current (T3), and welding time (T2) were P1_base(T1) = 0.2, P1_base(T3) = 0.3, and P1_base(T2) = 0.5, respectively. After cleaning, since welding temperature T1 is not a target process parameter, its confidence level is set to 0, resulting in the following first confidence distribution: for welding current T3, P1(T3) = 0.3; for welding time T2, P1(T2) = 0.5.

[0294] For the second base confidence distribution, the original confidence levels for welding temperature (T1), welding current (T3), and welding time (T2) were P2_base(T1) = 0.15, P2_base(T3) = 0.35, and P2_base(T2) = 0.5, respectively. After cleaning, the confidence level for welding temperature T1 was set to 0, resulting in the second confidence distribution as follows: for welding current T3, P2(T3) = 0.35; for welding time T2, P2(T2) = 0.5.

[0295] Operations on neural networks can be implemented using deep learning frameworks. For example, in the TensorFlow framework, the construction, training, and prediction of pre-trained and detection neural networks can be easily performed. For the input data (the packaging process development dataset to be mined), data preprocessing operations can be performed, such as data normalization, mapping the feature values ​​of the data to a specific interval (e.g., [0,1]) to allow the neural network to better process the data. Determining the target process parameters and performing confidence level cleaning can be achieved using simple programming logic. For example, in Python, loop statements and conditional statements can be used to iterate through the target process parameters and their confidence levels, filter them according to set thresholds, and perform confidence level cleaning.

[0296] In one implementation, the packaging process development dataset to be mined includes multiple target process parameters, and the mining results include multiple target process parameters and target process parameter types. The mining results are obtained based on completing x rounds of mining, where each round of mining is either a detection mining round or a type mapping mining round, and the result of each round of mining is a target process parameter or a target process parameter type; wherein, the e-th round of mining includes:

[0297] Step S401: Based on the results of the (e-1)th round of mining, determine the result category of the eth round result. The result category is either the detection category or the type mapping category. The result category corresponding to the first round of mining is the detection category.

[0298] Step S402: Based on the packaging process development dataset to be mined and the results of the first e-1 rounds of mining, obtain the shared confidence distribution through a pre-trained neural network, obtain the detection confidence distribution through a detection neural network, and obtain the mapping confidence distribution through a type mapping neural network;

[0299] Among them, the shared confidence distribution is the first confidence distribution or the third confidence distribution, the detection confidence distribution is the second confidence distribution or the sixth confidence distribution, and the mapping confidence distribution is the fourth confidence distribution or the fifth confidence distribution;

[0300] Step S403: Based on the shared confidence distribution, the detection confidence distribution, and the mapping confidence distribution, obtain the results of the e-th round of mining.

[0301] In the implementation of the process for determining the mining results, the computer system determines the results of each round of mining through steps S401-S403, and then gradually obtains the final mining results.

[0302] In step S401, the computer system determines the result category of the e-th round of mining based on the results of the (e-1)th round of mining. The result category is either a detection category or a type mapping category, and the result category corresponding to the first round of mining is the detection category.

[0303] In the context of big data mining for implanted sub-packaging process development, each round of mining aims to extract specific information from the dataset related to the packaging process development. For example, suppose that in the first round of mining, since there are no results from the previous round as a reference, the result category is set as the detection category. This means that in the first round of mining, the computer system mainly focuses on directly detecting process parameters from the dataset.

[0304] When the system reaches the e-th round (e>1), it needs to determine the result category of the e-th round based on the results of the (e-1)-th round. For example, if the (e-1)-th round reveals a process parameter, such as welding temperature, then the e-th round might be a type mapping category, meaning it needs to determine which type this welding temperature process parameter belongs to, such as whether it's a critical temperature type or a secondary temperature type. This method of determining the result category of the next round based on the results of the previous round helps the system logically and progressively delve deeper into the information within the dataset.

[0305] In step S402, the computer system obtains the shared confidence distribution through a pre-trained neural network, the detection confidence distribution through a detection neural network, and the mapping confidence distribution through a type mapping neural network, based on the packaging process development dataset to be mined and the results of the previous e-1 rounds of mining.

[0306] The proposed dataset for developing packaging processes contains a wealth of information related to implanted sub-packaging processes, such as material properties, process equipment parameters, and operational procedures. The results of the first e-1 rounds of mining provide more background information or constraints for the current round of mining.

[0307] A pre-trained neural network is a neural network pre-trained on a large amount of data. It can derive a shared confidence distribution based on this input information. For example, suppose that in a certain round of data mining, the pre-trained neural network evaluates the types of undetermined process parameters in a second process parameter library based on the packaging process development dataset to be mined and some process parameters mined previously (results from the first e-1 rounds). If the undetermined process parameter types in the second process parameter library include critical temperature influence types (KTI), minor temperature influence types (STI), and non-temperature-dependent types (NTI), the pre-trained neural network might derive a shared confidence distribution of P_shared(KTI) = 0.4, P_shared(STI) = 0.3, and P_shared(NTI) = 0.3. This shared confidence distribution represents the probability of each undetermined process parameter type based on the current dataset and the mined results.

[0308] The detection neural network also derives the detection confidence distribution based on the input. For example, for the same set of undetermined process parameter types, the detection neural network may derive detection confidence distributions of P_detect(KTI) = 0.35, P_detect(STI) = 0.3, and P_detect(NTI) = 0.35.

[0309] Type mapping neural networks then derive the mapping confidence distribution. For example, the mapping confidence distribution derived by a type mapping neural network is P_map(KTI) = 0.45, P_map(STI) = 0.3, and P_map(NTI) = 0.25.

[0310] In terms of technical means, the operations of these neural networks can be implemented based on deep learning frameworks. Taking TensorFlow as an example, the first step is to construct the structure of these neural networks. Pre-trained neural networks, detection neural networks, and type mapping neural networks may have different structures, such as the number of layers and the number of neurons per layer. When inputting data, the dataset of the encapsulation process development to be mined and the results of the first e-1 rounds of mining need to be appropriately encoded and preprocessed so that they can be accepted by the neural network. For example, the numerical values ​​in the dataset and results can be normalized and mapped to the [0,1] interval so that the neural network can process them better. Then, through the forward propagation algorithm, the neural network calculates the corresponding confidence distribution based on its internal weights and activation functions.

[0311] In step S403, the computer system obtains the results of the e-th round of mining based on the shared confidence distribution, the detection confidence distribution, and the mapping confidence distribution.

[0312] If the result category in round e is the detection category, the computer system may consider both the shared confidence distribution (here, the shared confidence distribution is the first confidence distribution, which is the confidence distribution regarding process parameters) and the detection confidence distribution to determine the process parameters. For example, suppose in the mining related to the soldering process in the implantation sub-packaging process, the shared confidence distribution gives confidence scores of P_shared(T1) = 0.3, P_shared(T3) = 0.4, and P_shared(T2) = 0.3 for soldering temperature (T1), soldering current (T3), and soldering time (T2), respectively, while the detection confidence distribution gives confidence scores of P_detect(T1) = 0.25, P_detect(T3) = 0.5, and P_detect(T2) = 0.25. The computer system can use a weighted average method to combine these two confidence distributions. Let the weight of the shared confidence distribution be w1 = 0.4, and the weight of the test confidence distribution be w2 = 0.6. Then, for welding temperature T1, the overall confidence P(T1) = w1 × P shared (T1)+w2×P detect (T1)=0.4×0.3+0.6×0.25=0.27; For welding current T3, P(T3)=w1×P shared (T3)+w2×P detect (T3)=0.4×0.4+0.6×0.5=0.46; For welding time T2, P(T2)=w1×P shared (T2)+w2×P detect (T2) = 0.4 × 0.3 + 0.6 × 0.25 = 0.27. Since the overall confidence level of welding current T3 is the highest, the process parameter mined in the e-th round is welding current T3.

[0313] If the result category in round e is a type mapping category, the computer system will comprehensively consider the shared confidence distribution (in this case, the third confidence distribution, which is a confidence distribution about the process parameter type), the detection confidence distribution (in this case, the sixth confidence distribution, which is a confidence distribution about the process parameter type), and the mapping confidence distribution to determine the process parameter type. For example, suppose the shared confidence distribution gives confidence scores of P_shared(KTI) = 0.4, P_shared(STI) = 0.3, and P_shared(NTI) = 0.3 for critical temperature influence type (KTI), minor temperature influence type (STI), and non-temperature related type (NTI), respectively; the detection confidence distribution gives confidence scores of P_detect(KTI) = 0.35, P_detect(STI) = 0.3, and P_detect(NTI) = 0.35, respectively; and the mapping confidence distribution gives confidence scores of P_map(KTI) = 0.45, P_map(STI) = 0.3, and P_map(NTI) = 0.25, respectively. Using the same weighted average method, let the weight of the shared confidence distribution be w3 = 0.3, the weight of the detection confidence distribution be w4 = 0.3, and the weight of the mapping confidence distribution be w5 = 0.4. For the critical temperature influence type KTI, the overall confidence P(KTI) = w3 × P shared (KTI)+w4×P detect (KTI)+w5×P map (KTI) = 0.3 × 0.4 + 0.3 × 0.35 + 0.4 × 0.45 = 0.405; For the minor temperature influence type STI, P(STI) = w3 × P shared (STI)+w4×P detect (STI)+w5×P map (STI) = 0.3 × 0.3 + 0.3 × 0.3 + 0.4 × 0.3 = 0.3; For non-temperature-dependent NTI, P(NTI) = w3 × P shared (NTI)+w4×P detect (NTI)+w5×P map (NTI) = 0.3 × 0.3 + 0.3 × 0.35 + 0.4 × 0.25 = 0.295. Since the critical temperature influence type KTI has the highest overall confidence level, the process parameter type mined in the e-th round is the critical temperature influence type KTI.

[0314] Through round after round of mining, the computer system gradually determines the mining results of each round. After x rounds of mining, the final mining results include multiple target process parameters and target process parameter types. These results help to deeply understand the various relationships in the development of implantation sub-packaging process and provide important basis for process optimization, quality control, etc.

[0315] In one implementation, the debugging process for the detection neural network includes the following steps:

[0316] Step S11: Obtain the first sample library, which includes multiple first training datasets carrying training prior labels. The training prior labels of each first training dataset include the process parameter prior labels of the first training dataset.

[0317] Step S12: Based on each first training dataset, perform process parameter detection on each first training dataset through the detection neural network to be tuned, and obtain the first inference process parameters for each first training dataset;

[0318] Step S13: Based on each first training dataset and the first inference process parameter, the first inference process parameter is type-mapped through the detection neural network to be tuned to obtain the first inference process parameter type. Based on the first inference process parameter and the first inference process parameter type of each first training dataset, the mining result of each first training dataset is obtained.

[0319] Step S14: Based on the commonality measurement results between the first inference process parameters and the prior labels of process parameters in each first training dataset, determine the first calibration error value, and based on the first calibration error value, perform detailed optimization on the network parameters of the detection neural network to be calibrated.

[0320] During the debugging process of the detection neural network, the computer system adjusts the network through steps S11-S14 to enable it to better detect process parameters.

[0321] In step S11, the computer system acquires a first sample library, which includes multiple first training datasets carrying training prior labels. The training prior labels for each first training dataset include prior labels for the process parameters of that first training dataset.

[0322] In the context of implantation sub-packaging process development, the first sample library is a crucial data source for training and debugging detection neural networks. For example, the first training dataset in the first sample library could be detailed process data records from different batches of implantation sub-packaging processes. These data records contain various process-related information, such as process parameters like temperature setpoints, pressure setpoints, and welding operation durations recorded during a particular batch of implantation sub-packaging.

[0323] Training prior labels are known, accurate information used to measure the difference between the output of the detection neural network and the correct result. For process parameter prior labels, these are the accurate process parameter values ​​corresponding to each first training dataset. For example, for a given first training dataset, which records the duration of a welding operation, the process parameter prior label explicitly states the accurate duration of this welding operation in that batch. This serves as the standard for judging the accuracy of the detection neural network's results.

[0324] In step S12, the computer system performs process parameter detection on each first training dataset using the detection neural network to be tuned, based on each first training dataset, to obtain the first inference process parameters for each first training dataset.

[0325] The detection neural network to be tuned is a not-yet-fully-tuned network whose purpose is to optimize its parameters by processing data from the first sample library. When the computer system inputs the first training dataset into this detection neural network to be tuned, the network analyzes and predicts the process parameters in the dataset based on its existing structure (including the connection method of neurons, the number of layers, the number of neurons per layer, etc.) and its initial weight values.

[0326] For example, given a first training dataset containing data related to the welding process in the implantation sub-packaging process, the detection neural network to be tuned might predict a certain process parameter value for the welding operation based on the input welding-related data features (such as current and voltage fluctuations in the welding equipment, and the characteristics of the welding materials). Assuming this process parameter is the welding temperature, the detection neural network might predict a welding temperature of 200°C; this 200°C would then be the first inference process parameter for this first training dataset.

[0327] In step S13, the computer system performs type mapping on the first inference process parameters through the detection neural network to be tuned, based on each first training dataset and the first inference process parameters, to obtain the first inference process parameter type of the first inference process parameters. Based on the first inference process parameters and the first inference process parameter type of each first training dataset, the system obtains the mining result of each first training dataset.

[0328] After obtaining the first inference process parameter (such as the previously mentioned soldering temperature of 200℃), the detection neural network to be calibrated further maps this process parameter to a different type. For example, in the implantation sub-packaging process, the soldering temperature can be mapped to different types, such as critical temperature type, secondary temperature type, or normal temperature range type. Assuming the detection neural network maps the 200℃ soldering temperature to a critical temperature type according to its internal mapping rules and algorithms, this critical temperature type is the first inference process parameter type.

[0329] Then, based on the first inference process parameter (200℃) and the first inference process parameter type (key temperature type) of each first training dataset, the computer system determines the mining result for each first training dataset. This mining result is a comprehensive result after process parameter detection and type mapping of the first training dataset, which includes the detected process parameter values ​​and the type information to which the process parameter belongs.

[0330] In step S14, the computer system determines the first calibration error value based on the commonality measurement results between the first inference process parameters and the prior labels of process parameters in each first training dataset, and performs detailed optimization of the network parameters of the detection neural network to be calibrated based on the first calibration error value.

[0331] Commonality measures are used to quantify the similarity or difference between the first inference process parameter and the prior label of the process parameter. For example, mean squared error (MSE) can be used to calculate the commonality measure. Suppose that for a certain first training dataset, the welding temperature in the prior label of the process parameter is 205℃ (accurate value), while the first inference process parameter (welding temperature) predicted by the detection neural network is 200℃. According to the mean squared error formula... (Here, n=1 because only one process parameter is being compared; y) i It is a priori marking of the process parameters, namely 205℃; The first inference process parameter, i.e., 200℃, was calculated. The value of 25 is the mean squared error for this first training dataset.

[0332] After calculating such mean squared error values ​​(or other common metrics) for all the first training datasets, the computer system combines these error values ​​to obtain the first calibration error value. For example, the mean squared error values ​​of all the first training datasets can be averaged, and the average value obtained is the first calibration error value.

[0333] Based on this initial calibration error value, the network parameters of the detection neural network to be calibrated are optimized in detail. Technically, gradient descent-based optimization algorithms can be used. For example, backpropagation is a commonly used method to calculate gradients and update network parameters (such as connection weights between neurons). After calculating the initial calibration error value, the weights in the network are adjusted according to the error, with the adjustment direction aimed at reducing the error. Assuming the weights of the detection neural network are w, according to the backpropagation algorithm, the weight update formula is: Where α is the learning rate (a pre-set small positive number used to control the step size of weight updates), and E is the first calibration error value (such as mean squared error). It is the derivative of the error E with respect to the weight w. By continuously updating all network parameters in this way, the detection neural network gradually adjusts its structure, improving the accuracy of process parameter detection and type mapping.

[0334] Through the operations in steps S11-S14, the detection neural network continuously learns and adjusts on the first training dataset in the first sample library, enabling it to detect process parameters more accurately, thus laying a good foundation for mining process parameters in the packaging process development dataset to be mined.

[0335] In one implementation, the training process of a type mapping neural network includes the following steps:

[0336] Step S21: Obtain the second sample library, which includes multiple second training datasets carrying training prior labels. The training prior labels of each second training dataset include the type prior labels of each process parameter prior label of the second training dataset.

[0337] Step S22: Based on each second training dataset, process parameters are detected for each second training dataset through the type mapping neural network to be tuned, and the second inference process parameters for each second training dataset are obtained.

[0338] Step S23: Based on each second training dataset and the second inference process parameter, perform type mapping on the second inference process parameter through the type mapping neural network to be tuned, and obtain the second inference process parameter type of the second inference process parameter. Based on the second inference process parameter and the second inference process parameter type of each second training dataset, obtain the mining result of each second training dataset.

[0339] Step S24: Based on the commonality measurement results between the second inference process parameter type and type prior label of each second training dataset, determine the second calibration error value, and perform detailed optimization of the network parameters of the type mapping neural network to be calibrated based on the second calibration error value.

[0340] During the training of the type mapping neural network, the computer system adjusts the network through steps S21-S24 to enable it to accurately map process parameter types.

[0341] In step S21, the computer system obtains a second sample library, which includes multiple second training datasets carrying training prior labels. The training prior labels of each second training dataset include type prior labels of each process parameter prior label of the second training dataset.

[0342] In the context of implantation sub-packaging process development, a second sample library is an important data source for training type-mapping neural networks. For example, the second training dataset in the second sample library could be a detailed record of process data during the implantation sub-packaging process under different conditions. These data records contain various process parameters, such as the conductivity of the materials used, the operating temperature of the packaging equipment, and the pressure, etc., recorded in a specific implantation sub-packaging operation.

[0343] For each second training dataset, the type prior label in its training prior labels is pre-determined type information for each process parameter. Taking temperature parameters in the implantation sub-packaging process as an example, suppose a second training dataset contains a temperature value from the soldering process, and the type prior label corresponding to this temperature value indicates that it belongs to a critical temperature type (because this temperature value has a very critical impact on soldering quality). These type prior labels provide an accurate standard for evaluating the output of the type mapping neural network.

[0344] In step S22, the computer system performs process parameter detection on each second training dataset using the type mapping neural network to be tuned, based on each second training dataset, to obtain the second inference process parameters for each second training dataset.

[0345] The type mapping neural network to be tuned is a neural network that has not yet been trained and needs to be optimized using data from a second sample library. When the computer system inputs the second training dataset into this network, the network analyzes and predicts the process parameters in the dataset based on its current structure (such as the connection method of neurons, the number of layers in the network, the number of neurons in each layer, etc.) and its internal initial weight values.

[0346] For example, given a second training dataset containing data related to the welding process in the implantation sub-packaging process, the type mapping neural network to be tuned might predict a certain process parameter in the welding operation based on the input welding-related data features (such as current and voltage fluctuations in the welding equipment, and the characteristics of the welding materials). Assuming this process parameter is the welding temperature, the type mapping neural network might predict a welding temperature of 200°C; this 200°C would then be the second inference process parameter for this second training dataset.

[0347] In step S23, the computer system performs type mapping on the second inference process parameters through a type mapping neural network to be tuned, based on each second training dataset and the second inference process parameter, to obtain the second inference process parameter type of the second inference process parameters. Based on the second inference process parameters and the second inference process parameter type of each second training dataset, the mining result of each second training dataset is obtained.

[0348] After obtaining the second inference process parameter (such as the previously mentioned soldering temperature of 200℃), the type mapping neural network to be tuned further performs type mapping on this process parameter. For example, in the implantation sub-packaging process, the soldering temperature can be mapped to different types, such as critical temperature type, secondary temperature type, or normal temperature range type. Assuming that the type mapping neural network maps the soldering temperature of 200℃ to a critical temperature type according to its internal mapping rules and algorithms, this critical temperature type is the type of the second inference process parameter.

[0349] Then, based on the second inference process parameter (200℃) and the second inference process parameter type (key temperature type) of each second training dataset, the computer system determines the mining result for each second training dataset. This mining result is a comprehensive result after process parameter detection and type mapping of the second training dataset, containing the detected process parameter values ​​and the type information to which the process parameter belongs.

[0350] In step S24, the computer system determines the second calibration error value based on the commonality measurement results between the second inference process parameter type and the type prior label of each second training dataset, and performs detailed optimization of the network parameters of the type mapping neural network to be calibrated based on the second calibration error value.

[0351] Commonality metrics are used to measure the similarity or difference between the type of the second inference process parameter and the type prior label. For example, accuracy can be used as a method to calculate the commonality metric. Suppose that for a certain second training dataset, the type prior label indicates that the process parameter (such as welding temperature) belongs to the critical temperature type, while the type mapping neural network predicts that the second inference process parameter type is a minor temperature type.

[0352] To calculate accuracy, first count the total number N of all second training datasets, then count the number n of datasets that correctly predicted the data (i.e., the second inference process parameter type is the same as the type prior label). The accuracy formula is Accuracy = n / N. A low accuracy indicates that the performance of the type mapping neural network needs improvement.

[0353] Besides accuracy, other metrics can be used, such as recall and F1-score. Recall is the ratio of the number of correctly predicted types to the number of actual correct types. Assuming the number of process parameters that are actually critical temperature types is M, and the number of parameters predicted as critical temperature types and actually being critical temperature types is m, then recall = m / M.

[0354] The F1-score takes into account both accuracy and recall, and the formula is as follows: Precision refers to accuracy.

[0355] The computer system determines the second calibration error value based on these common metrics (such as precision, recall, or F1-score). For example, if F1-score is used as the metric and the current F1-score is found to be low, this low F1-score value can be used as the second calibration error value.

[0356] Based on this second calibration error value, the network parameters of the type mapping neural network to be calibrated are optimized in detail. Technically, gradient descent-based optimization algorithms can be used. For example, stochastic gradient descent (SGD) is a commonly used method. Assuming the weights of the type mapping neural network are w, according to the SGD algorithm, the weight update formula is: Where α is the learning rate (a pre-set small positive number used to control the step size of weight updates), and E is the second tuning error value (such as the F1 score). It is the derivative of the error E with respect to the weight w. By continuously updating all network parameters in this way, the type mapping neural network gradually adjusts its structure, improving the accuracy of mapping process parameter types.

[0357] Through the operations in steps S21-S24, the type mapping neural network continuously learns and adjusts on the second training dataset in the second sample library, thereby enabling more accurate mapping of process parameters and providing a guarantee for type mapping tasks in the actual packaging process development dataset to be mined.

[0358] In one implementation, the training process of the pre-trained neural network includes the following steps:

[0359] Step S31: Obtain the third sample library. The third sample library includes multiple third training datasets carrying training prior labels. The training prior labels of each third training dataset include the process parameter prior labels of the third training dataset and the type prior labels of each process parameter prior label.

[0360] Step S32: Based on each third training dataset, process parameters are detected for each third training dataset using the pre-trained neural network to be tuned, and the third inference process parameters for each third training dataset are obtained.

[0361] Step S33: Based on each third training dataset and the third inference process parameters, the third inference process parameters are type-mapped through the pre-trained neural network to be tuned, and the third inference process parameter type is obtained. Based on the third inference process parameters and the third inference process parameter type of each third training dataset, the mining results of each third training dataset are obtained.

[0362] Step S34: Based on the commonality measurement results between the third mining results and the training prior labels of each third training dataset, determine the third tuning error value, and perform detailed optimization of the network parameters of the pre-trained neural network to be tuned based on the third tuning error value.

[0363] During the training process of the pre-trained neural network, the computer system adjusts the pre-trained neural network through steps S31-S34 to enable it to accurately detect process parameters and map types.

[0364] In step S31, the computer system obtains a third sample library, which includes multiple third training datasets carrying training prior labels. The training prior labels of each third training dataset include the process parameter prior labels of the third training dataset and the type prior labels of each process parameter prior label.

[0365] In the context of implant packaging technology development, third-party sample libraries are an important data source for training pre-trained neural networks. For example, a third training dataset in a third sample library may contain various combinations of data related to different implant packaging processes. Taking a specific third training dataset as an example, it may contain material property data (such as material hardness, conductivity, etc.), equipment operating parameters (such as temperature, pressure, etc. of the packaging equipment), and operation process information (such as the sequence and time of each process) for a specific implant packaging batch.

[0366] For each third training dataset, the process parameter prior labels specify the exact values ​​of each process parameter. For example, for the temperature parameters of the packaging equipment, the process parameter prior labels provide the accurate temperature setpoint for that specific packaging operation. The type prior labels of each process parameter further determine the type to which these process parameters belong. For example, for the aforementioned temperature parameter, the type prior label might indicate that the temperature belongs to the critical process parameter type because it has a crucial impact on packaging quality.

[0367] In step S32, the computer system performs process parameter detection on each third training dataset using the pre-trained neural network to be tuned, based on each third training dataset, to obtain the third inference process parameters for each third training dataset.

[0368] The pre-trained neural network to be tuned is a neural network that has not yet been fully trained. Its structure (such as the connection method of neurons, the number of layers, the number of neurons per layer, etc.) and initial weights are preset, but need to be optimized using data from a third sample library. When the computer system inputs the third training dataset into this pre-trained neural network to be tuned, the network analyzes and predicts the process parameters in the dataset based on its existing structure and internal initial weights.

[0369] For example, given a third training dataset containing data related to the welding process in the implantation sub-packaging process, the pre-trained neural network to be tuned might predict a certain process parameter in the welding operation based on the input welding-related data features (such as current and voltage fluctuations of the welding equipment, and the characteristics of the welding materials). Assuming this process parameter is the welding current, the pre-trained neural network might predict a welding current of 10A; this 10A would then be the third inference process parameter for this third training dataset.

[0370] In step S33, the computer system performs type mapping on the third inference process parameters through the pre-trained neural network to be tuned, based on each third training dataset and the third inference process parameters, to obtain the third inference process parameter type of the third inference process parameters. Based on the third inference process parameters and the third inference process parameter type of each third training dataset, the mining result of each third training dataset is obtained.

[0371] After obtaining the third inference process parameter (such as the welding current of 10A mentioned earlier), the pre-trained neural network to be tuned further performs type mapping on this process parameter. For example, in the implantation sub-packaging process, the welding current can be mapped to different types, such as critical current type, secondary current type, or normal current range type. Assuming that the pre-trained neural network maps the 10A welding current to the critical current type according to its internal mapping rules and algorithms, this critical current type is the third inference process parameter type. Then, based on the third inference process parameter (10A) and the third inference process parameter type (critical current type) of each third training dataset, the computer system determines the mining result for each third training dataset. This mining result is a comprehensive result after process parameter detection and type mapping of the third training dataset, which includes the detected process parameter value and the type information to which the process parameter belongs.

[0372] In step S34, the computer system determines the third tuning error value based on the commonality measurement results between the third mining results of each third training dataset and the training prior labels, and performs detailed optimization of the network parameters of the pre-trained neural network to be tuned based on the third tuning error value.

[0373] Commonality metrics are indicators used to measure the similarity or difference between third-party mining results and training prior labels. For example, commonality metrics can be calculated by combining mean squared error (MSE) with accuracy.

[0374] Regarding the mean squared error (MSE) component, suppose that for a certain third training dataset, the welding current in the prior label of the process parameters is 12A, while the third inference process parameter predicted by the pre-trained neural network is 10A. According to the mean squared error formula... (Here, n=1 because only one process parameter is being compared; y) i It is a priori marker for process parameters, namely 12A; The third inference process parameter, i.e., 10A, is calculated to obtain...

[0375] Regarding the accuracy component, it is assumed that for type mapping, if the type prior label indicates that the process parameter (such as welding current) belongs to the critical current type, and the third inference process parameter type predicted by the pre-trained neural network is also the critical current type, then this is considered a correct prediction. The total number N of all third training datasets and the number n of datasets with correct predictions (including those with correct process parameter values ​​and types) are counted, and the accuracy formula is Accuracy = n / N.

[0376] By combining mean squared error and accuracy, a comprehensive common metric can be obtained, which can be used as the third calibration error value. For example, a weighted summation method can be used, where the weight of mean squared error is w1 and the weight of accuracy is w2 (w1+w2=1), and the third calibration error value E=w1×MSE+w2×(1-Accuracy).

[0377] Based on the third calibration error value, the network parameters of the pre-trained neural network to be calibrated are optimized in detail. Technically, gradient descent-based optimization algorithms, such as stochastic gradient descent (SGD), can be used. Assuming the weights of the pre-trained neural network are w, according to the SGD algorithm, the weight update formula is as follows: Where α is the learning rate (a pre-set small positive number used to control the step size of weight updates), and E is the third calibration error value. This is the derivative of the error E with respect to the weight w. By continuously updating all network parameters in this way, the pre-trained neural network gradually adjusts its structure, improving its accuracy in process parameter detection and type mapping. Through steps S31-S34, the pre-trained neural network continuously learns and adjusts on the third training dataset in the third sample library, thereby enabling more accurate process parameter detection and type mapping, laying a good foundation for subsequent operations on the packaging process development dataset to be mined.

[0378] This application provides a computer system, such as... Figure 2 As shown, the computer system 100 includes a processor 101 and a memory 103. The processor 101 and the memory 103 are connected, for example, via a bus 102. Optionally, the computer system 100 may also include a transceiver 104. It should be noted that in practical applications, the transceiver 104 is not limited to one type, and the structure of this computer system 100 does not constitute a limitation on the embodiments of this application.

[0379] This application provides a computer system, which includes: one or more processors; a memory; and one or more computer programs, wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more programs are executed by the processors, they implement the above-described method.

Claims

1. A big data mining method based on implantation sub-encapsulation technology, characterized in that, The method includes: Obtain the packaging process development dataset to be mined; Based on the proposed packaging process development dataset, a pre-trained neural network is used to detect process parameters in the proposed packaging process development dataset to obtain a first confidence distribution for a first process parameter library. A detection neural network is then used to detect process parameters in the proposed packaging process development dataset to obtain a second confidence distribution for the first process parameter library, which includes multiple undetermined process parameters. Based on the first confidence distribution and the second confidence distribution, the target process parameters in the packaging process development dataset to be mined are obtained; Based on the proposed packaging process development dataset and target process parameters, the target process parameters are type-mapped using the pre-trained neural network to obtain a third confidence distribution for the second process parameter library. The target process parameters are then type-mapped using the type mapping neural network to obtain a fourth confidence distribution for the second process parameter library, which includes multiple undetermined process parameter types. Based on the third and fourth confidence distributions, the target process parameter type is obtained. Based on the target process parameters and the target process parameter type, the mining results of the packaging process development dataset to be mined are obtained.

2. The method according to claim 1, characterized in that, The detection neural network is obtained by adjusting the common measurement results between the first inference process parameters and the prior labels of process parameters in each first training dataset; The type mapping neural network is obtained by adjusting the common measurement results between the second inference process parameter type and the type prior label of each second training dataset; The pre-trained neural network is obtained by debugging based on the commonality measurement results between the third mining results and the training prior labels of each third training dataset. The third mining results include third inference process parameters and third process parameter types, and the training prior labels include process parameter prior labels and type prior labels of the third training dataset.

3. The method according to claim 2, characterized in that, The method further includes: Based on the proposed packaging process development dataset, process parameters are detected on the proposed packaging process development dataset using a type mapping neural network to obtain a fifth confidence distribution for the first process parameter library. Based on the packaging process development dataset and target process parameters to be mined, the target process parameters are type-mapped through the detection neural network to obtain the sixth confidence distribution for the second process parameter library; The step of obtaining the target process parameters in the packaging process development dataset to be mined based on the first confidence distribution and the second confidence distribution includes: The target process parameters are obtained based on the first confidence distribution, the second confidence distribution, and the fifth confidence distribution. The step of obtaining the target process parameter type based on the third confidence distribution and the fourth confidence distribution includes: Based on the third confidence distribution, the fourth confidence distribution, and the sixth confidence distribution, the target process parameter type corresponding to the target process parameter is obtained.

4. The method according to claim 1, characterized in that, The method further includes: Obtain parameter mining instructions; wherein, the parameter mining instructions include detection instructions and type mapping instructions, the detection instructions indicate that the process parameters of the target process flow in the packaging process development dataset to be mined are detected, and the type mapping instructions indicate that the detected process parameters are type-mapped; The first confidence distribution is obtained by detecting process parameters in the proposed packaging process development dataset using a pre-trained neural network based on the proposed packaging process development dataset and detection instructions. The second confidence distribution is obtained by detecting process parameters in the proposed packaging process development dataset using a detection neural network based on the proposed packaging process development dataset and detection instructions. The third confidence distribution is obtained by performing type mapping on the target process parameters using a pre-trained neural network based on the proposed packaging process development dataset, target process parameters, and type mapping instructions. The fourth confidence distribution is obtained by performing type mapping on the target process parameters using a type mapping neural network based on the proposed packaging process development dataset, target process parameters, and type mapping instructions.

5. The method according to claim 1, characterized in that, The step of obtaining the target process parameters in the packaging process development dataset to be mined based on the first confidence distribution and the second confidence distribution includes: Based on the commonality measurement results between the first confidence distribution and the second confidence distribution, the first detection difference of the second confidence distribution is determined; Based on the difference between the second confidence distribution and the first detection difference, a first corrected confidence level is determined for the first confidence distribution; The first confidence distribution is corrected based on the first corrected confidence level to obtain the first target confidence distribution; The target process parameters are obtained based on the first target confidence distribution.

6. The method according to claim 3, characterized in that, The step of obtaining the target process parameters based on the first confidence distribution, the second confidence distribution, and the fifth confidence distribution includes: Based on the commonality measurement results between the first confidence distribution and the second confidence distribution, the first detection difference of the second confidence distribution is determined; Based on the commonality measurement results between the first confidence distribution and the fifth confidence distribution, the second detection difference of the fifth confidence distribution is determined; Based on the difference between the second confidence distribution and the first detection difference, a first corrected confidence level for the first confidence distribution is determined; The first confidence level distribution is corrected based on the first corrected confidence level; The second target confidence distribution is determined based on the difference between the corrected first confidence distribution and the second detection difference. The target process parameters are obtained based on the second target confidence distribution.

7. The method according to any one of claims 1 to 5, characterized in that, The process of determining the first confidence distribution and the second confidence distribution includes: Based on the proposed packaging process development dataset, a pre-trained neural network is used to detect process parameters in the dataset to obtain a first basic confidence distribution for a first process parameter library. Then, a detection neural network is used to detect process parameters in the dataset to obtain a second basic confidence distribution for the first process parameter library. The first basic confidence distribution includes the first basic confidence of each undetermined process parameter, and the second basic confidence distribution includes the second basic confidence of each undetermined process parameter. Based on the first basic confidence distribution and the second basic confidence distribution, target process parameters with both the first basic confidence and the second basic confidence not less than a set critical value are determined among the various undetermined process parameters. The confidence scores corresponding to the undetermined process parameters other than the target undetermined process parameters in the first basic confidence score distribution and the second basic confidence score distribution are cleaned to obtain the first confidence score distribution and the second confidence score distribution.

8. The method according to claim 3, characterized in that, The packaging process development dataset to be mined includes multiple target process parameters, and the mining results include multiple target process parameters and the target process parameter type of the target process parameters. The mining results are obtained based on completing x rounds of mining. Each round of mining is either a detection mining or a type mapping mining, and the result of each round of mining is the target process parameter or the target process parameter type. The e-th round of mining includes: Based on the results of the (e-1)th round of mining, the result category of the eth round is determined. The result category is either a detection category or a type mapping category. The result category corresponding to the first round of mining is the detection category. Based on the packaging process dataset to be mined and the results of the first e-1 rounds of mining, a shared confidence distribution is obtained through a pre-trained neural network, a detection confidence distribution is obtained through a detection neural network, and a mapping confidence distribution is obtained through a type mapping neural network. Wherein, the shared confidence distribution is a first confidence distribution or a third confidence distribution, the detection confidence distribution is a second confidence distribution or a sixth confidence distribution, and the mapping confidence distribution is a fourth confidence distribution or a fifth confidence distribution; Based on the shared confidence distribution, detection confidence distribution, and mapping confidence distribution, the results of the e-th round of mining are obtained.

9. A computer system, characterized in that, include: One or more processors; Memory; One or more computer programs; The one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, they implement the method as described in any one of claims 1 to 8.

Images