Wafer test map file uniform format intelligent conversion system

Abstract

This invention relates to the field of semiconductor manufacturing technology, specifically to an intelligent conversion system for a unified format of wafer test mapping files. The system includes a file acquisition and reception module, an intelligent recognition and unified conversion engine, a standard mapping file generation and output module, and an access control and management module. The intelligent recognition and unified conversion engine automatically parses multi-format mapping files using a deep learning model and quickly adapts to new formats through a meta-learning mechanism; the access control and management module ensures operational security and data integrity; the system can also achieve cross-node collaborative evolution through federated learning. This invention achieves fully automatic, high-precision, and high-security conversion of wafer test mapping files to internal standard formats, overcoming the shortcomings of traditional manual processing, such as low efficiency, error-proneness, and inability to quickly adapt to new formats, significantly improving the intelligence, flexibility, and reliability of semiconductor packaging production.

Description

Technical Field

[0001] This invention relates to the field of semiconductor manufacturing technology, and more specifically to a unified format intelligent conversion system for wafer test mapping files. Background Technology

[0002] In the semiconductor packaging process, after the wafer completes circuit probe testing, a mapping file is generated that records the coordinates of each chip and the test results (usually represented by classification codes). This file guides the subsequent die-bonding equipment to accurately pick up qualified chips. However, due to the different brands, models, and software versions of testing equipment used by different chip design companies and testing plants, the formats of the wafer test mapping files generated vary greatly. It is estimated that there are hundreds of common non-standard formats. Currently, packaging plants generally handle these diverse file formats manually: process engineers manually identify the file structure, extract the effective information from the source file, and organize it into a single standard format that can be recognized by the die-bonding equipment inside the packaging plant through specific scripts or manual operations. This method has significant drawbacks: First, it relies heavily on the engineer's personal experience, resulting in low processing efficiency, difficulty in adapting to the fast-paced production demands, and the high risk of errors due to manual operation, leading to quality issues such as misplacement or omissions. Second, whenever a new customer or a change in testing procedures brings a new file format, engineers need to re-analyze, rewrite, and debug the parsing rules, resulting in a long import cycle and high costs. Furthermore, in manual processing, raw test data is at risk of being intentionally or unintentionally tampered with, making it difficult to guarantee data security and production traceability. Finally, the data and processing experience of each packaging plant are isolated, unable to be effectively shared and reused, forming "data silos" that hinder the collaborative improvement of processing capabilities across the entire industry. Therefore, there is an urgent need for an intelligent solution that can automatically, accurately, and securely process multiple wafer test mapping file formats, quickly adapt to new formats, and possess continuous learning capabilities.

[0003] Therefore, the existing technology still needs further development. Summary of the Invention

[0004] The purpose of this invention is to overcome the above-mentioned technical deficiencies and provide a unified format intelligent conversion system for wafer test mapping files to solve the problems existing in the prior art.

[0005] To achieve the above technical objectives, this invention provides a unified format intelligent conversion system for wafer test mapping files, comprising: The file acquisition and reception module is used to receive wafer test mapping files with different formats generated by different test equipment; The intelligent recognition and unified conversion engine is connected to the file acquisition and receiving module. It is used to automatically recognize the format and content structure of the wafer test mapping file, and convert mapping files of different formats into mapping files of internal standard format according to the preset conversion logic. The standard mapping file generation and output module is connected to the intelligent recognition and unified conversion engine and is used to generate and output target mapping files that conform to the internal standard format. The access control and management module is used to manage the access and operation permissions of system operators, ensuring that the content of the original mapping file is not modified by unauthorized personnel.

[0006] Specifically, the intelligent recognition and unified conversion engine includes a file parsing submodule and a format conversion submodule; The file parsing submodule is used to perform deep parsing on the received wafer test mapping file and extract the core data fields in the file. The core data fields include the chip coordinate matrix, the classification code characterizing the chip test results, the wafer batch number, the equipment identification number, and the flat edge direction information. The format conversion submodule is used to reorganize and map the parsed core data fields according to a preset internal standard format specification; The file parsing submodule includes a semantic and structural recognition model trained based on deep learning. This model learns the encoding rules, delimiter usage conventions, and data layout features of various known mapping file formats through training. It can perform structural inference and key field location for new format mapping files without predefined parsing rules, and output structured data objects to the format conversion submodule.

[0007] Specifically, the semantic and structural recognition model of the file parsing submodule is a multimodal fusion neural network, and its processing flow includes: The character stream encoding layer is used to transform the original character sequence of the mapped file into a high-dimensional vector representation, capturing the text-level sequential features; The visual feature extraction layer is used to input the mapped file into the convolutional neural network as a two-dimensional image to extract its spatial features such as layout, table structure and specific symbol arrangement. The feature fusion and attention layer is used to fuse the text features output by the character stream encoding layer with the spatial features output by the visual feature extraction layer, and to focus on key regions in the file through an attention mechanism. The structure prediction and field annotation header, based on the fused features, predicts the data structure tree of the file and annotates the boundaries and semantic labels of the chip coordinates, classification code, and batch number fields.

[0008] Specifically, the format conversion submodule includes an adaptive encoding conversion unit and a classification code standardization unit; The adaptive encoding conversion unit is used to automatically identify the encoding method of chip coordinates and row and column numbers in the source file based on the structured data object output by the file parsing submodule, and convert it into the unified coordinate representation system required by the internal standard format; The classification code standardization unit is used to process the classification codes in the core data field. It has a pre-built classification code mapping rule library. Based on the semantics and context of the parsed source file classification codes, the unit maps them to a set of standardized classification codes defined by internal standards to indicate qualified chips, unqualified chips and chips of specific categories, ensuring that the chip bonding device can pick up chips based on unified semantics.

[0009] Specifically, it also includes a work order and process management module; The work order and process management module is communicatively connected to the file acquisition and receiving module, the access control and management module, and the intelligent recognition and unified conversion engine, respectively. The work order and process management module is used to receive production task information, which includes at least customer identifier, product model and mapping file identifier to be converted; The work order and process management module automatically matches or triggers the corresponding conversion logic in the intelligent identification and unified conversion engine based on the production task information, and records the execution status, operator, and timestamp of the conversion task. The permission control and management module dynamically configures the operator's operation permissions in the work order and process management module according to the operator's role, including conversion task initiation permission, conversion program configuration permission, and system management permission.

[0010] Specifically, the access control and management module includes a dynamic access configuration submodule and an operation audit submodule; The dynamic permission configuration submodule is used to divide system roles into at least three categories: engineering configuration roles, production operation roles, and system management roles, and to assign differentiated function access and data operation permissions to different roles. The engineering configuration role has the authority to create, modify, and verify conversion logic for new customers or new formats, but has no right to directly execute production conversion tasks; the production operation role has the authority to execute mapping file conversion tasks based on work orders and view historical conversion records, but has no right to modify any configured conversion logic rules; the system management role has the highest authority for user management, role allocation, and viewing system logs; the operation audit submodule is used to record all users' key operation logs on the system, including but not limited to: user login and logout, addition or modification of conversion logic, execution of mapping file conversion, and changes in permission settings, forming an immutable operation chain for traceability.

[0011] Specifically, the work order and process management module also includes a content change awareness submodule; The content change awareness submodule is used to automatically compare the content of a new mapping file with the source mapping file used in the most recent successful conversion when it receives a new mapping file with the same customer and product model identifier as the historical task. The content change awareness submodule is pre-set with anomaly judgment logic. When the comparison finds that the key fields have changed, an alarm is triggered and the automatic conversion process is paused. The key fields include, but are not limited to: the semantic definition of the classification code, the dimension of the coordinate matrix, the horizontal direction identifier, and the file structure separator. After the alarm is triggered, the conversion task is marked as requiring manual review and the user with the engineering configuration role is notified to confirm and process it. After confirmation that there are no errors or the conversion logic is updated, the task continues to be executed.

[0012] Specifically, the intelligent recognition and unified conversion engine also includes a meta-learning adaptation submodule; The meta-learning adaptation submodule is coupled with the semantic and structural recognition model in the file parsing submodule; the meta-learning adaptation submodule is used to drive the semantic and structural recognition model to quickly adapt when receiving a small number of newly formatted mapped file samples; its workflow includes: Based on the general file parsing knowledge already learned by the semantic and structural recognition model as meta-knowledge, the model can perform gradient updates or rapid model parameter adjustments in a small number of steps using a small number of newly provided samples. This enables the model to obtain effective parsing capabilities for new formats without retraining and without significantly reducing the parsing performance of the original formats.

[0013] Specifically, the implementation of the meta-learning adaptation submodule includes a few-shot learning support set construction unit and a model fast tuning unit; The few-shot learning support set construction unit is used to receive support set samples provided by users for a new format. The support set samples contain a small number of mapping file instances that have been manually annotated or have been confirmed to be correctly parsed. The fast model tuning unit employs an optimization-based meta-learning algorithm to initialize the basic network parameters of the semantic and structural recognition model as meta-parameters sensitive to multi-format parsing tasks. When processing a new format, the fast model tuning unit targets the parsing loss on the support set samples and performs a finite number of gradient descent iterations near the basic network meta-parameters to quickly generate model parameters adapted to the new format. The adapted model parameters are then dynamically loaded into the file parsing submodule for processing subsequent files in the new format.

[0014] Specifically, it also includes a federated learning model update module; The federated learning model update module is connected to the AI ​​model in the intelligent recognition and unified conversion engine to collaboratively optimize model performance while protecting the privacy of each customer's data; its operation is as follows: On multiple production nodes where this system is deployed, the local AI model is trained using the historical conversion data of their respective wafer test mapping files, and only the update of the model parameters is encrypted and uploaded to the central server. The central server aggregates the update of the model parameters from multiple nodes, generates a global model update, and then distributes the updated global model parameters to each node.

[0015] Beneficial effects: Compared with existing technologies, the intelligent conversion system for unified format of wafer test mapping files provided by this invention has the following advantages: First, by integrating semantic and structural recognition models trained based on deep learning, the system achieves automated intelligent parsing of the contents of wafer test mapping files. It can accurately extract core fields such as chip coordinates and classification codes, greatly reducing reliance on human experience and significantly improving the speed and accuracy of file processing, thereby reducing production quality risks caused by human error from the source.

[0016] Secondly, the system innovatively introduces a meta-learning adaptation mechanism, which can drive the parsing model to quickly adapt when faced with a small number of samples in a completely new format. It can obtain effective parsing capabilities for the new format in a very short time, overcoming the bottleneck of traditional methods that require rewriting a large number of parsing rules for the new format. This greatly shortens the cycle of introducing new customers and new products and enhances the agility and flexibility of the production line.

[0017] Third, the system achieves a strict separation between the right to configure conversion rules and the right to execute production through built-in access control and work order process management. Combined with the content change perception function, it automatically controls abnormal changes in key fields, building a complete data security protection and operation traceability chain, effectively preventing data tampering and ensuring the reliability and compliance of the production process.

[0018] Fourth, through the federated learning model update module, the system can achieve collaborative evolution of model knowledge across factories and organizations while strictly protecting the local data privacy of each production node. This enables systems deployed in various locations to continuously learn from a wider range of data distributions, jointly improve their ability to process complex and rare file formats, break down "data silos," and achieve collective intelligence enhancement under privacy computing.

[0019] In summary, this invention not only solves the problem of automating the unified conversion of multi-format mapping files, but also endows the system with self-adaptive, self-evolving, and highly secure capabilities through the deep integration of artificial intelligence and advanced learning mechanisms, providing key technical support for the intelligent upgrading of semiconductor packaging. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the system composition of the intelligent conversion system for unified format of wafer test mapping files provided in a specific embodiment of the present invention. Detailed Implementation

[0021] To enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Based on the embodiments in this application, other similar embodiments obtained by those skilled in the art without creative effort should all fall within the scope of protection of this application. Furthermore, directional terms mentioned in the following embodiments, such as "up," "down," "left," and "right," are only for reference to the directions in the accompanying drawings; therefore, the directional terms used are for illustrative purposes and not for limiting the invention.

[0022] The present invention will be further described below with reference to the accompanying drawings and preferred embodiments.

[0023] Please see Figure 1 This invention provides a unified format intelligent conversion system for wafer test mapping files, comprising: (1) File acquisition and reception module, used to receive wafer test mapping files with different formats generated by different test equipment.

[0024] It should be further explained that the file acquisition and reception module is responsible for receiving raw wafer test mapping files from network shared directories, FTP servers, or through a manual upload interface. These files may come from test equipment from different manufacturers and have different formats, such as G85 format, specific ASCII text format, XML format, etc., covering a total of about 120 common formats.

[0025] (2) Intelligent recognition and unified conversion engine, connected to the file acquisition and receiving module, is used to automatically recognize the format and content structure of the wafer test mapping file, and convert mapping files of different formats into mapping files of internal standard format according to the preset conversion logic.

[0026] It should be further explained that the intelligent recognition and unified conversion engine is the core processing unit of the system. It has built-in conversion programs (such as rule sets based on VB scripts) pre-configured for different customers and product models, which can parse and convert the received files.

[0027] (3) Standard mapping file generation and output module, which is connected to the intelligent recognition and unified conversion engine, is used to generate and output target mapping files that conform to the internal standard format.

[0028] It should be further explained that the standard mapping file generation and output module encapsulates the converted data according to the company's unified standard format (such as the format named "ZKT") and outputs it to a specified directory for downstream die bonding equipment to read and use directly.

[0029] (4) Access control and management module, used to manage the access and operation permissions of system operators, and to ensure that the contents of the original mapping file are not modified by unauthorized personnel.

[0030] It should be further explained that the access control and management module provides security for the system. Users log in with an account and password and are divided into different roles. For example, the engineering role only has the permission to configure the conversion program but not to modify the original files, while the operation role only has the permission to execute the conversion. This prevents the original test data from being intentionally or unintentionally tampered with during the conversion process, ensuring the originality of the data and the reliability of production.

[0031] Furthermore, the file acquisition and reception module can be configured with monitoring functions to periodically scan specified directories and automatically trigger subsequent processes upon discovering a new file. The intelligent identification and unified conversion engine's preset conversion logic includes key steps: first, identifying the file source (e.g., through file naming rules or specific header identifiers); then, calling the corresponding parsing rules to extract valid information, such as chip coordinates on the wafer (DieX, DieY), Bin Code representing chip quality levels, wafer batch number (LotID), device number, etc.; finally, mapping and reorganizing this information into the fields and arrangement order required by the internal standard format. Before generating the final file, the standard mapping file generation and output module can attach a conversion log, recording information such as conversion time, operator, source file name, and target file name. The access control and management module can adopt a role-based access control model, where the system administrator creates users and assigns roles, with each role bound to a series of specific operation permissions.

[0032] Understandably, through the aforementioned modular design, the system achieves fully automated and standardized conversion from multi-format mapped files to a single internal standard format, completely replacing the inefficient and error-prone traditional method of relying on manual identification and copy-pasting by engineers. This not only significantly improves production efficiency and reduces production interruptions or material errors caused by human error, but also eliminates the risk of production data being tampered with during the preparation stage through strict access control, meeting the high standards of data integrity and traceability required in semiconductor manufacturing. The unified output format also simplifies the configuration of downstream die-bonding equipment and improves the flexibility of the production line.

[0033] Specifically, the intelligent recognition and unified conversion engine includes a file parsing submodule and a format conversion submodule; The file parsing submodule is used to perform deep parsing on the received wafer test mapping file and extract the core data fields in the file. The core data fields include the chip coordinate matrix, the classification code characterizing the chip test results, the wafer batch number, the equipment identification number, and the flat edge direction information. The format conversion submodule is used to reorganize and map the parsed core data fields according to a preset internal standard format specification; The file parsing submodule includes a semantic and structural recognition model trained based on deep learning. This model learns the encoding rules, delimiter usage conventions, and data layout features of various known mapping file formats through training. It can perform structural inference and key field location for new format mapping files without predefined parsing rules, and output structured data objects to the format conversion submodule.

[0034] Furthermore, the intelligent recognition and unified conversion engine consists of a file parsing submodule and a format conversion submodule working together. The file parsing submodule is the "understanding" part of the engine, and its primary task is to accurately extract all the core data required for subsequent conversion from the original file. This core data includes: a coordinate matrix defining the position of each chip on the wafer (usually represented by row Y and column X); a classification code (Bin Code, such as 1 for good, 2 for power short circuit failure, etc.) identifying whether the chip test result is pass, specific type of failure, or complete failure; a wafer batch number (LotNumber) for production traceability; a tester ID that generated the file; and orientation information indicating the wafer notch or flat direction, which is crucial for the subsequent chip placement and removal orientation calibration during die bonding.

[0035] Furthermore, the format conversion submodule is the engine's "output" part. It receives structured data from the parsing submodule and reassembles the data according to a set of strict internal standard format specifications (such as ZKT format). For example, it converts the parsed coordinates represented in the form of "(X,Y)" into the standard format of "Row=Y, Column=X"; and it converts the classification codes that may exist in the source file in the form of numbers, letters, or strings into standard format-defined numerical codes with clear process meanings by querying the built-in mapping table.

[0036] It should be further explained that the core innovation of the file parsing submodule lies in the integration of a semantic and structural recognition model trained based on deep learning. This model is trained and deployed using an architecture that integrates convolutional neural networks and a Transformer encoder. The specific design includes: (1) Model training phase: A large number (e.g., more than 100,000) of labeled wafer test mapping files covering various known formats (such as G85, ASCII variants, XML, etc.) are used as training samples. The annotation information of each training sample includes: the start and end positions of the chip coordinate region in the file, the position of the column where the classification code is located, the text or symbol of the flat edge direction identifier, etc.

[0037] (2) Model Input and Processing: For a new file to be parsed, the model processes it in two ways simultaneously. First, the file content is treated as a character sequence, and the contextual semantic information of the text is captured through the embedding layer and the Transformer encoder, such as recognizing keywords such as "LotID:", "Notch", and "Bin" and their relationship with subsequent data. Second, the file content is rendered as a two-dimensional image (or its character position matrix is ​​directly used), and visual structural features are extracted through CNN, such as recognizing the border lines of tables, the alignment of data, and the arrangement pattern of specific symbols (such as "X" representing bad images).

[0038] (3) Model output: The model ultimately outputs a structured data object (e.g., a JSON or an instance of a specific class), which explicitly contains the values ​​and data types of the core data fields located and extracted from the file. For new format files not found in the training set, the model can perform analogical reasoning and structural inference based on its "pattern knowledge" learned from a large number of formats, thereby attempting to locate key fields, greatly reducing the reliance on manual rule programming for entirely new formats.

[0039] Understandably, the introduction of deep learning models significantly enhances the system's generalization ability and intelligence. Traditional parsing methods based on fixed rules or regular expressions often fail when faced with minor format adjustments or entirely new, unknown formats, requiring manual intervention. In contrast, the model in this solution can automatically learn the deep-seated features of file formats, possessing a certain level of understanding and parsing ability for unseen formats. This improves the system's robustness and adaptability, reduces reliance on senior engineers, and accelerates the import of new formats.

[0040] Specifically, the semantic and structural recognition model of the file parsing submodule is a multimodal fusion neural network, and its processing flow includes: The character stream encoding layer is used to transform the original character sequence of the mapped file into a high-dimensional vector representation, capturing the text-level sequential features; The visual feature extraction layer is used to input the mapped file into the convolutional neural network as a two-dimensional image to extract its spatial features such as layout, table structure and specific symbol arrangement. The feature fusion and attention layer is used to fuse the text features output by the character stream encoding layer with the spatial features output by the visual feature extraction layer, and to focus on key regions in the file through an attention mechanism. The structure prediction and field annotation header, based on the fused features, predicts the data structure tree of the file and annotates the boundaries and semantic labels of the chip coordinates, classification code, and batch number fields.

[0041] Specifically, this multimodal fusion neural network is designed to synergistically utilize the textual semantics and visual layout information of the mapped files. Its processing flow is an end-to-end deep network containing the following key components.

[0042] It should be further explained that the detailed implementation methods of each level are as follows: (1) Character Stream Encoding Layer: This layer receives the raw string of the mapped file. First, an embedding layer maps each character (including letters, numbers, punctuation, spaces, etc.) to a dense vector of a fixed dimension (e.g., 128 dimensions). Then, a multi-layered bidirectional Transformer encoder (e.g., containing 6 layers, each with 8 attention heads) processes this vector sequence. The encoder's output is a context-aware vector representation of each character position. ,in , It is the sequence length. This is the model dimension (e.g., 512). This process enables the model to understand the meaning of "Bin1" as a whole unit, rather than just isolated characters "B", "i", and "n".

[0043] (2) Visual Feature Extraction Layer: In parallel, this layer processes the mapped file content. A preferred approach is to render the file as a grayscale image using a fixed-width font (such as Courier New). Image size can be standardized to (For example The image is fed into a convolutional neural network backbone, such as the first few layers of a ResNet-34 pre-trained on ImageNet (excluding the last fully connected layers). The CNN backbone outputs a spatial feature map. ,in and It is the height and width of the feature map. This is the number of channels (e.g., 512). This feature map encodes the visual structure of the file, such as the alignment of text lines, the position of table lines, and the division of data blocks into regions.

[0044] (3) Feature fusion and attention layer: The goal of this layer is to fuse text features and visual features Effective fusion is achieved. First, the two features need to be aligned to a common spatial reference frame. One approach is to map the position of each character in the text sequence back to its approximate two-dimensional coordinates in the original image, thereby providing... Each vector in the matrix is ​​assigned a spatial location. Then, a cross-attention mechanism is used to allow text features to "query" visual features. Specifically, an attention weight matrix is ​​computed, whose elements... Indicates the first The text feature vector and the first The correlation strength of each visual spatial location feature vector is determined. Finally, this attention weight is used to perform a weighted summation of the visual features to obtain an enhanced text feature with the same dimension as the text features, but infused with visual context. This process enables the model to recognize that the word "Bin" typically appears in the column header position of a table, while the classification code number appears in the data area below it.

[0045] (4) Structure prediction and field annotation header: based on fused features The network completes the final task through several parallel prediction heads. A segmentation head (e.g., using a fully convolutional network) predicts whether each pixel in the image belongs to the chip coordinate matrix region, classification code region, title information region, etc. A sequence labeling head (e.g., connected to a conditional random field layer) labels each character in the text sequence (using the BIOES labeling system), for example, identifying that "Lot-" is the beginning of the batch number field (B-LOT) and "123456" is inside the batch number field (I-LOT). Finally, these prediction results are integrated into a structured data representation, clearly indicating which parts of the file belong to the chip coordinates (usually a two-dimensional array), which columns are classification codes, and the specific values ​​and positions of each metadata field (such as Lot, Wafer ID, Notch Direction) in the title row.

[0046] Understandably, this multimodal fusion method overcomes the limitations of single text or visual analysis. For example, text analysis alone makes it difficult to distinguish whether a number is a coordinate value or a classification code; visual analysis alone makes it difficult to understand the specific semantics of "NotchDown". This solution deeply integrates the two through an attention mechanism, enabling the model to accurately understand the structure of a document by combining its "appearance" (layout) and "content" (text), just like a human engineer. This allows for high-precision automated parsing of complex and varied mapping files, laying a solid foundation for subsequent accurate conversion.

[0047] Specifically, the format conversion submodule includes an adaptive encoding conversion unit and a classification code standardization unit; The adaptive encoding conversion unit is used to automatically identify the encoding method of chip coordinates and row and column numbers in the source file based on the structured data object output by the file parsing submodule, and convert it into the unified coordinate representation system required by the internal standard format; The classification code standardization unit is used to process the classification codes in the core data field. It has a pre-built classification code mapping rule library. Based on the semantics and context of the parsed source file classification codes, the unit maps them to a set of standardized classification codes defined by internal standards to indicate qualified chips, unqualified chips and chips of specific categories, ensuring that the chip bonding device can pick up chips based on unified semantics.

[0048] Specifically, after the format conversion submodule receives the structured data object from the file parsing submodule, two specialized units handle the most critical and error-prone data conversion task.

[0049] It should be further explained that the adaptive encoding conversion unit is responsible for handling the standardization of chip coordinates. The coordinate system and origin definition may differ in mapping files generated by different test equipment or software. Specific design considerations include: (1) Coordinate starting value identification: Some file chip coordinates start counting from (0,0), while others start from (1,1). This unit determines this by analyzing the minimum value of the parsed coordinate data. For example, if both the minimum row coordinate and the minimum column coordinate are 0, it is determined to start from 0; if both are 1, it is determined to start from 1. Then, a translation transformation is performed according to internal standards (such as a unified rule to start from (0,0)).

[0050] (2) Row and Column Direction and Order Identification: Some documents may define X as a row and Y as a column, which is the opposite of the usual definition. This unit can make logical inferences by analyzing the description fields in the document (if any) or by combining the shape (number of rows and columns) and flat edge direction of the wafer map. For example, in a typical 8-inch wafer map, the number of rows and columns of the effective chip area roughly conforms to a certain physical ratio, and the row and column definitions can be confirmed by comparison. The conversion unit will then perform the necessary transpose operation.

[0051] (3) Coordinate representation transformation: The identified source coordinates are uniformly converted into the representation specified by the internal standard format, such as the string "row index: column index" or a linearized list of a two-dimensional array.

[0052] The classification code standardization unit is responsible for handling the semantic uniformity of classification codes. This is crucial for the conversion because different test programs define Bin Codes very differently.

[0053] (a) Mapping Rule Base Construction: The core of this unit is a configurable mapping rule base. Each record in the rule base is associated with a specific "customer-product model-test procedure" combination and defines the mapping relationship from the source file Bin Code to the internal standard Bin Code. For example, for product X of customer A, the rules might be defined as: "0" in the source file -> internal standard "1" (qualified chip); "1-9" -> internal standard "2" (unqualified chip); specific code "10" -> internal standard "3" (chips requiring special handling, such as repairable memory). The rule base supports wildcards and range matching.

[0054] (b) Context-Aware Mapping: When applying mapping rules, the unit incorporates the parsed context information. For example, some files may define a separate table of meanings for Bin Codes in the file header. The unit will preferentially use such explicit definitions. If not, it will use the preset rules matched for the file source. If a source Bin Code not defined in the rules is encountered, the unit can be configured to map it to a specific "unknown" code and trigger an alert to prompt the engineer to investigate.

[0055] (c) Standardized Output: Ultimately, all chip classification codes are converted to the same standardized semantic space (e.g., 1, 2, 3, ...). This ensures that regardless of how the source file is defined, the downstream die-attaching equipment receives instructions with a unified meaning: pick up the chip with code 1, skip the chip with code 2, and place the chip with code 3 into the specific hopper.

[0056] Understandably, the adaptive encoding conversion unit solves the normalization problem of physical location representation, avoiding chip picking errors caused by inconsistent coordinate systems. The classification code standardization unit solves the semantic normalization problem of test results, which is the core value of achieving "unified format" conversion. It ensures the clarity of production instructions, enabling the die bonder to accurately execute the "picking" operation, fundamentally preventing major quality incidents such as discarding good products or mounting defective products due to semantic confusion in the code. The combination of these two units achieves comprehensive standardization from data form to data meaning.

[0057] Specifically, it also includes a work order and process management module; The work order and process management module is communicatively connected to the file acquisition and receiving module, the access control and management module, and the intelligent recognition and unified conversion engine, respectively. The work order and process management module is used to receive production task information, which includes at least customer identifier, product model and mapping file identifier to be converted; The work order and process management module automatically matches or triggers the corresponding conversion logic in the intelligent identification and unified conversion engine based on the production task information, and records the execution status, operator, and timestamp of the conversion task. The permission control and management module dynamically configures the operator's operation permissions in the work order and process management module according to the operator's role, including conversion task initiation permission, conversion program configuration permission, and system management permission.

[0058] Specifically, the work order and process management module serves as the system's process scheduling and control center, incorporating discrete file conversion tasks into standardized production process management. This module typically provides a user interface (UI) or application programming interface (API) for receiving or creating production tasks.

[0059] It should be further explained that the workflow and internal logic of the work order and process management module are as follows: (1) Work Order Creation and Information Binding: Production task information, i.e., work orders, can be created in various ways. For example, by integrating with the Manufacturing Execution System (MES), production instructions containing "Customer ID," "Device Model," and "Corresponding Wafer Map File Name" can be automatically received. Alternatively, authorized operators can manually create work orders within the system by inputting the aforementioned key information. Once created, a work order is associated with a specific physical mapping file.

[0060] (2) Automatic matching and triggering: The work order and process management module automatically searches for and matches the pre-set conversion program (rule or model configuration) in the configuration library of the intelligent recognition and unified conversion engine based on the "customer identifier" and "product model" in the work order. After successful matching, the module automatically triggers the conversion process: the notification file acquisition and receiving module locates the corresponding physical file and sends it to the intelligent recognition and unified conversion engine, while transmitting the identifier of the matched conversion program.

[0061] (3) Status Tracking and Recording: The module monitors the execution status of the conversion task in real time. The status includes "Waiting", "Conversion in progress", "Conversion successful", and "Conversion failed (and the error reason is recorded)". Every status change and every operation step (such as "Start conversion", "Complete parsing", "Complete output") will be recorded in a detailed log, including the timestamp of the occurrence and the operator ID of the operator who performed the operation. This forms a complete and tamper-proof traceability chain of the conversion process.

[0062] (4) Dynamic control of permissions: The permission control and management module is deeply integrated with the work order module. After a user logs into the system, the function buttons and the range of data they can operate in the work order and process management module interface are entirely determined by their role. For example, only users with the "Production Operator" role can see the "Execute Conversion" button and can only operate on work orders assigned to them that are in the "Waiting" state. Users with the "Process Engineer" role can see advanced function buttons such as "Configure Conversion Rules" and "Create Work Order Template," but may not be able to directly execute production conversion tasks. "System Administrator" has access to all functions.

[0063] Understandably, the introduction of work order and process management modules transforms what might have been arbitrary and scattered file conversion operations into a standardized production process that is planned, monitorable, and traceable. This ensures that every conversion task is traceable (which file, for which customer / product, by whom, when, which conversion rule was used, and what the result was), greatly enhancing the standardization and transparency of production management. Combined with the access control system, it achieves a clear division of responsibilities, guaranteeing both production efficiency and the security of system configurations and the compliance of data operations, meeting the stringent process control requirements of semiconductor manufacturing.

[0064] Specifically, the access control and management module includes a dynamic access configuration submodule and an operation audit submodule; The dynamic permission configuration submodule is used to divide system roles into at least three categories: engineering configuration roles, production operation roles, and system management roles, and to assign differentiated function access and data operation permissions to different roles. The engineering configuration role has the authority to create, modify, and verify conversion logic for new customers or new formats, but has no right to directly execute production conversion tasks; the production operation role has the authority to execute mapping file conversion tasks based on work orders and view historical conversion records, but has no right to modify any configured conversion logic rules; the system management role has the highest authority for user management, role allocation, and viewing system logs; the operation audit submodule is used to record all users' key operation logs on the system, including but not limited to: user login and logout, addition or modification of conversion logic, execution of mapping file conversion, and changes in permission settings, forming an immutable operation chain for traceability.

[0065] Specifically, the access control and management module achieves refined access control and complete operation traceability through its dynamic access configuration submodule and operation audit submodule.

[0066] It should be further explained that the dynamic permission configuration submodule implements a role-based permission model. The system predefines at least three core roles, each associated with a specific set of permission policies: (1) Engineering Configuration Role: This role is typically held by a process engineer or product engineer. Their permissions include: accessing the "Conversion Logic Configuration" interface to create new parsing and conversion rules for new customer-product model combinations; modifying existing rules to optimize conversion performance; and using test files to simulate and verify the configured rules, ensuring correct conversion. However, this role is explicitly prohibited from operating actual production-related file conversion tasks in the "Production Work Order Execution" interface. This separation of "configuration rights" and "execution rights" is a crucial mechanism to prevent core conversion rules from being accidentally or maliciously compromised.

[0067] (2) Production Operation Role: This role is held by production line operators. Their permissions include: after logging into the system, viewing the list of tasks assigned to them in the "Production Work Order" interface; performing operations such as "Start Conversion" and "Confirm Result" on tasks; and viewing the history of tasks completed by themselves or their shift and conversion logs. This role cannot see or access the "Conversion Logic Configuration" interface, and therefore cannot access or modify any conversion rules. They can only use the rules pre-configured and verified by engineers to execute tasks, ensuring the standardization and consistency of production operations.

[0068] (3) System Management Role: This role is held by an IT administrator or system maintenance personnel. They have the highest privileges, including: creating, disabling, and deleting system user accounts; assigning or changing roles for users (e.g., changing a user from "Production Operation Role" to "Engineering Configuration Role"); and viewing all system operation logs and audit trails. However, the system management role typically does not have the authority to directly modify specific product conversion rules, which further isolates the responsibilities of system management and process configuration.

[0069] The operation auditing submodule silently records all key user actions in the background. The recorded data items include at least: operation timestamp, user ID, user IP address, operation type (e.g., "user login", "add conversion rule A001", "execute work order #20240520001 conversion", "modify user X's role"), operation object (e.g., rule ID, work order number, username), operation result (success / failure), and more detailed contextual information (e.g., comparison of differences before and after rule modification). These logs are stored in a secure, append-only database or file, ensuring that once recorded, they cannot be modified or deleted by ordinary users, forming a reliable audit trail.

[0070] Understandably, this finely defined role division and access control achieve a balance between convenience and security. Engineering roles focus on the accuracy and optimization of technical rules, production roles focus on the efficiency and reliability of task execution, and management roles focus on the stable operation of the system and user management. The comprehensive traceability capabilities provided by the operations audit submodule not only enable rapid identification of the cause and responsible party when problems occur (e.g., which user modified the rules at which time due to a batch file conversion error), but also effectively deter user behavior, standardize operational processes, and fully comply with the semiconductor industry's stringent standards for data integrity and production traceability.

[0071] Specifically, the work order and process management module also includes a content change awareness submodule; The content change awareness submodule is used to automatically compare the content of a new mapping file with the source mapping file used in the most recent successful conversion when it receives a new mapping file with the same customer and product model identifier as the historical task. The content change awareness submodule is pre-set with anomaly judgment logic. When the comparison finds that the key fields have changed, an alarm is triggered and the automatic conversion process is paused. The key fields include, but are not limited to: the semantic definition of the classification code, the dimension of the coordinate matrix, the horizontal direction identifier, and the file structure separator. After the alarm is triggered, the conversion task is marked as requiring manual review and the user with the engineering configuration role is notified to confirm and process it. After confirmation that there are no errors or the conversion logic is updated, the task continues to be executed.

[0072] Specifically, the content change awareness submodule is a crucial quality control and safety checkpoint within the work order and process management module. Its design aims to capture situations where files appear to have the same format but whose critical internal definitions have changed. If such changes are automatically converted directly, they could lead to systemic errors in downstream production.

[0073] It should be further explained that the working mechanism of this submodule includes the following steps: (1) Historical baseline file establishment: When the system successfully completes the first conversion of the mapping file for a specific customer and product model (e.g., product P of customer C), the content change awareness submodule will automatically store the original source file (or its main feature fingerprint) used for the successful conversion as a "historical baseline file" and associate it with the customer-product model combination.

[0074] (2) New File Arrival and Comparison Trigger: When the work order and process management module receives a new conversion task whose customer and product model identifiers match a combination associated with an existing "historical baseline file", the content change awareness submodule is automatically triggered. It will obtain the source file of this task and perform a deep comparison with the stored "historical baseline file".

[0075] (3) Key Field Comparison and Anomaly Detection: The comparison is not a simple binary comparison, but focuses on key fields that may affect the conversion results and downstream production. The submodule has pre-defined anomaly detection logic, including but not limited to: a) Classification Code Semantic Definition: Compare whether the descriptive text about the meaning of the Bin Code in the header area (or specific comment section) of two files is consistent. For example, if the historical baseline file defines "Bin1=PASS" while the new file defines "Bin1=FAIL", this is a fatal change and must be alerted.

[0076] b) Coordinate Matrix Dimensions: Compare the total number of rows and columns of the wafer described in the two files to see if they are the same. If the dimensions change, it means that the wafer size or test range may have changed, and this needs to be confirmed.

[0077] c) Flat edge direction marking: Compare whether the strings or symbols marking the wafer's flat edge or notch direction are consistent, such as "Notch Down" and "Flat Left".

[0078] d) File Structure Separators: Analyze whether the separators (commas, tabs, spaces, etc.) used in the data areas (such as chip Bin Code lists) are consistent. Changes in separators may lead to parsing errors. The submodule extracts and compares these key fields; if any inconsistency is found, it is considered a "key field change".

[0079] (4) Alarms and Process Control: Once an alarm is triggered, the content change awareness submodule will immediately perform the following actions: First, change the status of the current conversion work order from "Waiting" or "Conversion in progress" to "Requires manual review - Content change", and suspend any automatic conversion process. Second, send an alarm notification to the pre-configured notification list (such as the email address or instant messaging tool of the relevant process engineer), attaching the change details (which field, from what value to what value). Finally, wait for the intervention of the user with the engineering configuration role.

[0080] (5) Manual Review and Processing: After receiving the alarm, the engineer logs into the system to view the change details and determine whether the change is a normal engineering change (such as a test program update causing a change in the Bin definition) or an abnormal file error. If it is a normal change, the engineer needs to review and update the conversion logic configuration for that customer-product model (e.g., modify the BinCode mapping rules) and confirm the change is valid in the system. If it is a file error, the engineer needs to contact the file provider for correction. Only after the engineer completes the review and takes appropriate action will the work order be released, allowing the conversion to continue.

[0081] Understandably, the content change awareness submodule acts as an intelligent "security checkpoint" before production. It addresses the risks associated with the "blind trust" of input files in traditional automated systems. Even if the file format and name remain unchanged, minor modifications to the internal definitions can lead to catastrophic batch production errors. This module, through automated comparison and control mechanisms, mandates manual review when critical information changes, significantly improving system robustness and production safety. It prevents batch misoperations caused by unsynchronized notifications of upstream test data definition changes, making it a crucial link in achieving highly reliable intelligent manufacturing.

[0082] Specifically, the intelligent recognition and unified conversion engine also includes a meta-learning adaptation submodule; The meta-learning adaptation submodule is coupled with the semantic and structural recognition model in the file parsing submodule; the meta-learning adaptation submodule is used to drive the semantic and structural recognition model to quickly adapt when receiving a small number of newly formatted mapped file samples; its workflow includes: Based on the general file parsing knowledge already learned by the semantic and structural recognition model as meta-knowledge, the model can perform gradient updates or rapid model parameter adjustments in a small number of steps using a small number of newly provided samples. This enables the model to obtain effective parsing capabilities for new formats without retraining and without significantly reducing the parsing performance of the original formats.

[0083] Furthermore, the meta-learning adaptation submodule is designed to further enhance the system's ability to quickly adapt to new and unseen mapping file formats. Traditional deep learning models typically require collecting a large number of new samples for retraining or fine-tuning when faced with entirely new data distributions, which is often time-consuming and costly in production environments. Meta-learning aims to enable models to "learn," allowing them to leverage knowledge gained from previous multi-task learning to quickly adapt to new tasks.

[0084] It should be further explained that the working principle of the meta-learning adaptation submodule is based on the idea of ​​model-independent meta-learning algorithms. Assume the backbone network parameters of the semantic and structural recognition model in the file parsing submodule are... In the initial meta-training phase, the model is not trained on a single format, but rather on a task distribution containing a large number of mapping files in different formats. Training is conducted on [the platform]. Each task [is performed on this platform]. Learning to analyze a specific file format involves: (1) Meta-training process: In meta-training, each training iteration (episode) simulates a process of "rapidly adapting" to a new task. The specific steps are as follows: a) From the task distribution A task of sampling a batch .

[0085] b) For each task A support set is sampled from the dataset corresponding to this task. And a query set. The support set is usually small, for example, K=5 samples (5-shot learning), simulating a scenario where a new task only provides a small number of samples.

[0086] c) Model parameters As an initial value, in the task Support set Calculate loss Then, perform gradient descent in a small number of steps (e.g., 1-5 steps) to obtain task-specific adapted parameters. This process is called the inner loop.

[0087] d) Use the adapted parameters In the mission query set Calculate loss This loss is used to evaluate the model's performance after rapid adaptation.

[0088] e) For all sampling tasks, accumulate the loss on the query set, and adjust the original model parameters based on this accumulated loss. Gradient updates are performed. This process is called the outer loop. Its optimization objective is to find a set of initial parameters. This makes it possible for those from For any new task sampled in the middle, the model can achieve good performance after a few gradient updates using only a small amount of support set data for that task.

[0089] (2) Meta-adaptation (inference) process: When the system encounters a completely new mapping file format (new task) in production, the model has not been trained on. When this occurs, the meta-learning adaptation submodule is activated, and its specific design includes: a) Project configuration users need to provide a small number (e.g., 3-5) of correctly annotated sample files for this new format to form a support set. Annotations include indicating coordinate regions and category code columns in the file.

[0090] b) The meta-learning adaptation submodule loads the parameters of the meta-trained master model. .by Starting from the new task support set Perform a few steps (e.g., 3 steps) of gradient descent to quickly adjust the model parameters and obtain model parameters adapted to the new format. .

[0091] c) The data is dynamically loaded into the file parsing submodule to parse subsequent files of the same format. This process takes only a few minutes, and because of the small number of update steps and the small amount of data, the forgetting effect on the model's original knowledge is minimized, ensuring that its ability to parse other existing formats is largely unaffected.

[0092] Understandably, the introduction of the meta-learning adaptation submodule upgrades the system's intelligent parsing capability from "recognizing the known" to "quickly learning to recognize the unknown." It reduces the time required for creating parsing rules for new formats from the traditional hours or even days spent by engineers analyzing file structures, writing and debugging parsing scripts, to simply annotating a few sample files and initiating a rapid adaptation process. This significantly lowers the technical barrier and time cost of importing new formats, and dramatically improves the system's agility in handling diverse customer needs and rapid iteration of test programs, representing a key competitive advantage for intelligent systems.

[0093] Specifically, the implementation of the meta-learning adaptation submodule includes a few-shot learning support set construction unit and a model fast tuning unit; The few-shot learning support set construction unit is used to receive support set samples provided by users for a new format. The support set samples contain a small number of mapping file instances that have been manually annotated or have been confirmed to be correctly parsed. The fast model tuning unit employs an optimization-based meta-learning algorithm to initialize the basic network parameters of the semantic and structural recognition model as meta-parameters sensitive to multi-format parsing tasks. When processing a new format, the fast model tuning unit targets the parsing loss on the support set samples and performs a finite number of gradient descent iterations near the basic network meta-parameters to quickly generate model parameters adapted to the new format. The adapted model parameters are then dynamically loaded into the file parsing submodule for processing subsequent files in the new format.

[0094] Specifically, the meta-learning adaptation submodule achieves rapid adaptation to small samples through two collaboratively working units.

[0095] It should be further explained that the few-shot learning support set construction unit provides a user interface that guides users in creating effective support sets for new formats. When a user registers a new format in the system, the interface prompts the user to upload a small number (preferably 3-5) of representative sample files in that format. The system then launches a simplified annotation wizard.

[0096] (1) Annotation process: Users are required to identify several key areas in each sample file by selecting or clicking, such as: the start and end lines of the chip coordinate data block; the column (or region) containing the classification code; and the position of the header line containing metadata such as Lot ID and Wafer ID. For files with simple structures, only 1-2 samples may need to be annotated. The system will automatically infer and generate structured parsing annotations for the file (such as BIOES tags for each character) based on these sparse annotations. Once the user confirms that the automatic inference results are correct, the support set construction is complete. Support set ,in It is the original content of the i-th sample file. It is its corresponding label. It supports set size, usually This range was chosen because, within the meta-learning framework, 3-5 samples are usually sufficient to provide enough information for the model to understand the basic structure of a new format, while more than 10 samples may increase the user's annotation burden, with no significant incremental gains.

[0097] (2) The core of the fast model tuning unit is the use of optimization-based meta-learning algorithms, such as model-independent meta-learning. During the meta-training phase, this algorithm is used to find a set of excellent initial model parameters. (i.e., meta-parameters). This set of parameters is not optimal for a particular format, but rather located in a "sensitive" position, allowing for rapid achievement of good performance for each new task (format) with only a few gradient steps on the support set. The specific algorithm steps are as follows: a) Initialization: Randomly initialize model parameters .in This represents the set of all trainable parameters in a semantic and structural recognition model.

[0098] b) Iterative loop (outer loop): Repeat the following steps until the model performance converges.

[0099] i. From the task distribution A task that randomly samples a batch .in Representing the The task of parsing various file formats. It is a task probability distribution that covers multiple known formats.

[0100] ii. For each sampled task Execute the inner loop adaptation process: - From the task In the corresponding dataset, a small support set is sampled. and a query set for evaluation .in and Each file contains several sample files and their annotations for this task. - The computational model is set to the current parameters. Below, in support of the set Losses: .in The parameter is The model, This is a loss function, such as cross-entropy loss. - Calculate the above loss with respect to the model parameters. gradient: .in Indicates the parameter Find the gradient. It is the gradient vector. - Use this gradient to perform one or more gradient descent updates on the model parameters, and calculate the task-specific adapted parameters: .in The inner loop learning rate is a hyperparameter, and its optimal value is [value missing]. .choose This is to ensure that the inner loop update step size is small, avoiding changes to the meta parameters. This can cause excessive single-step perturbation, while also allowing the model parameters to effectively adapt to the new task.

[0101] iii. After completing the inner loop adaptation for all tasks in the current batch, calculate the performance of this batch of tasks within their respective query sets. The sum of losses. Note that the task-specific parameters updated via the inner loop are used when calculating the loss. : .in The summation represents the total query loss of the current batch of tasks, and is achieved by iterating through all tasks in the batch. .

[0102] iv. Calculate the total query loss Regarding initial meta-parameters The gradient is calculated, and the meta-parameters are updated accordingly. : .in The outer loop learning rate (also known as the meta-learning rate) is a hyperparameter, and its optimal value is [value missing]. .choose It is to make the meta-parameters The updates are more stable, allowing for slow and continuous optimization, eventually converging to a set of initial points that have good adaptability to all tasks.

[0103] c) Repeat step b) for multiple iterations until the model's performance on the validation task set stabilizes, which is considered convergence. The model parameters obtained at this point are the trained meta-parameters. .in These are the initial parameters of the model obtained after meta-learning training, which have the ability to adapt quickly.

[0104] (3) Quickly adapt to new formats: When faced with a completely new format and its support set At this time, the model fast tuning unit performs the following operations: a) Loading meta parameters As initial parameters.

[0105] b) New format support set Calculate loss .

[0106] c) Execution Step gradient descent (inner loop) to generate adaptive parameters: Steps It is a hyperparameter, with an optimal value of 5. Choosing 5 steps is based on an empirical trade-off; too few steps may not be adequately adapted, while too many may lead to overfitting to a small support set and increase computation time. 5 steps usually achieve a good balance between speed and efficiency.

[0107] d) will As a dedicated parsing model parameter for this new format, it is dynamically loaded into the file parsing submodule.

[0108] Understandably, by learning support set building blocks from few samples, the system significantly reduces the difficulty for users to provide knowledge for new formats, transforming the writing of complex rules into simple example annotation. Through rapid model tuning of the building blocks and its core MAML algorithm, the system can leverage the "meta-knowledge" previously learned on large-scale, multi-format data. This enables rapid and effective adaptation on a very small number of new samples. This addresses the pain points of deep learning models being prone to overfitting and difficult to generalize in small data scenarios, giving the intelligent parsing system a rapid learning ability similar to humans' "learning by analogy," significantly enhancing the system's practicality and scalability.

[0109] Specifically, it also includes a federated learning model update module; The federated learning model update module is connected to the AI ​​model in the intelligent recognition and unified conversion engine to collaboratively optimize model performance while protecting the privacy of each customer's data; its operation is as follows: On multiple production nodes where this system is deployed, the local AI model is trained using the historical conversion data of their respective wafer test mapping files, and only the update of the model parameters is encrypted and uploaded to the central server. The central server aggregates the update of the model parameters from multiple nodes, generates a global model update, and then distributes the updated global model parameters to each node.

[0110] Specifically, the federated learning model update module provides a secure and compliant collaborative evolution mechanism for system instances deployed across different factories and even different companies. Because wafer test mapping files contain sensitive chip test results and product information, directly sharing raw data between different customers or factories poses a serious risk of privacy and trade secret leaks. Federated learning perfectly resolves this contradiction by training locally on the data and exchanging only model updates, not the original data.

[0111] It should be further explained that the workflow of the federated learning model update module is a periodic, multi-round iterative process: (1) Initialization: The central server initializes a global semantic and structural recognition model, the parameters of which are: It is then distributed to all client nodes participating in federated learning (i.e., each factory that has deployed this system).

[0112] (2) Local training (the first) wheel): a) Each client node The system stores its own historical transformation data (anonymized mapping files and their parsing results) locally, forming a local dataset. At the start of each round of federated learning, the nodes... Download the latest global model parameters from the central server. .

[0113] b) Nodes Use local data Train the model. The training objective is to minimize the local loss function. Training will be performed in multiple batches (epochs), for example... Five local epochs were chosen to allow for sufficient local learning without overfitting the local data. Local learning rate. The optimal value is 0.001 to ensure training stability.

[0114] c) After local training is completed, the node Calculate the update amount (i.e., gradient, or parameter change) of the local model parameters. ,in These are the model parameters after local training. Node For update volume The data is encrypted (e.g., using homomorphic encryption or differential privacy techniques) and then uploaded to a central server. (Original data) It always remains local and never leaves the node.

[0115] (3) Global aggregation (the first) wheel): a) After the central server collects a sufficient number (e.g., exceeding a set proportion) of client encrypted updates, it decrypts (if applicable) and aggregates them. The most commonly used aggregation algorithm is FedAvg. The specific formula is: in, This represents the total number of clients participating in this round of training. It is a client Local data volume This is the total amount of data used in this round of training. It is a client The parameters are the ones trained locally. This weighted average ensures that clients with larger datasets contribute more to the global model.

[0116] b) The central server will aggregate the new global model parameters. The data is encrypted and then distributed to all client nodes.

[0117] (4) Iteration and Deployment: After receiving new global model parameters, the client node updates its local AI model using them. Processes (2) and (3) above are repeated continuously. As the number of federated learning rounds increases, the global model... By learning from the data distribution of all participating nodes, its generalization ability and ability to handle rare formats will continuously improve. Ultimately, the validated and stable new global model can be deployed to all nodes, enabling the intelligent recognition and unified transformation engine of each node to be enhanced synchronously.

[0118] Understandably, the federated learning model update module creates a collaborative intelligence paradigm where "the data remains still while the model moves." It allows production facilities geographically dispersed and belonging to different stakeholders to collaboratively contribute knowledge to train a more powerful and general-purpose AI parsing model, while strictly adhering to data privacy regulations and business confidentiality requirements. Each participant benefits from a richer collective experience, resulting in a model capable of handling more complex document variations and emerging testbed formats without any party disclosing its sensitive raw data. This breaks down data silos, enabling privacy-preserving cross-organizational knowledge sharing and improved model performance—a key technology for building industry-level intelligent solutions.

[0119] To better understand this invention, a specific application example is provided below to illustrate the workflow, data processing, and beneficial effects of the unified format intelligent conversion system for wafer test mapping files.

[0120] Specifically, suppose a packaging factory receives a wafer with product model "IC2024" from customer "Alpha," along with a mapping file generated after CP testing, named "Alpha_IC2024_Wafer25.xml". This file is in XML format, but its specific tag structure and internal Bin definitions are entirely new to ZKT's die-attaching station, lacking pre-configured rules. This example will demonstrate how the system collaborates with multiple modules to safely, accurately, and efficiently complete the conversion of this file, as follows: (1) Work Order Creation and File Receiving: Production planner "Zhang San" (with a production operation role) logs into the system. In the interface of the work order and process management module, he creates a new production work order, fills in the customer identifier as "Alpha", the product model as "IC2024", and associates the mapping file to be converted as "Alpha_IC2024_Wafer25.xml". The work order and process management module sets the status of the work order to "waiting" and notifies the file acquisition and receiving module to retrieve the physical file from the specified server path.

[0121] (2) New format recognition and rapid adaptation through meta-learning: Since this was the first production run of the customer "Alpha's" product "IC2024", the system could not find a matching preset conversion program in the configuration library of the intelligent recognition and unified conversion engine. At this time, the system automatically triggered the meta-learning adaptation process. Engineer "Li Si" (with the engineering configuration role) received a system notification. He uploaded the mapping files ("Wafer23.xml", "Wafer24.xml", "Wafer26.xml") of three other wafers of the same model as support set samples through the interface of the small sample learning support set building unit, and performed rapid annotation: In "Wafer23.xml", he selected the XML node area where the chip coordinate data was located. <diemap>The system annotated the attributes representing the X and Y coordinates; clicked on the BinCode attribute containing the Bin code; and highlighted the comment section in the file header defining the meaning of Bin, specifying that BinCode="1" represents PassDie and BinCode="8" represents FailDie. Based on the sparse annotations of these three samples, the system automatically generated complete structured annotations. Subsequently, the model fast tuning unit was activated, loading the pre-trained semantic and structural recognition model meta-parameters. .by Starting with 3 support set samples Calculate loss and execute Step gradient descent for fast adaptation: The inner loop learning rate Approximately 2 minutes later, the dedicated parsing model parameters adapted to the "Alpha_IC2024" XML format were obtained. And dynamically load it into the file parsing submodule.

[0122] (3) Intelligent file parsing and data extraction: The file parsing submodule uses the newly adapted model (parameters are...). The file "Alpha_IC2024_Wafer25.xml" is parsed. The model's multimodal fusion neural network then begins operation, with specific design elements including: a) The character stream encoding layer reads the file text and converts the character sequence into a vector. When reading... <die x=""10"Y="20"Bin" code=""1" / ">At that time, the Transformer encoder can understand the semantics of tags such as "Die", "X", "Y", and "Bin Code" and their attribute value relationships, and output a context vector.

[0123] b) Visual Feature Extraction Layer (Although the source file is text, it is converted into regular text image processing) uses CNN to identify the tree-like indentation structure of XML and determine...<Die Map> The content under a node consists of well-organized data blocks.

[0124] c) Feature fusion and attention layer calculation of the association between text features and visual features. For example, it will focus the attention weight of the text feature "BinCode" on the spatial features of the XML attribute value region. The model finally outputs a structured data object, accurately extracting: the wafer has a total of 300 rows and 300 columns; the flat edge direction is Notch Down; the wafer batch number is LotAlpha2024Q1; and parsing the entire chip coordinate matrix and the corresponding Bin Code, for example, the chip Bin Code at coordinate (10,20) is "1", and the chip Bin Code at coordinate (10,21) is "8".

[0125] (4) Format Conversion and Standardization: The format conversion submodule receives the above structured data objects, and its specific design includes: a) The adaptive encoding conversion unit recognizes that the source file coordinates start from (0,0), which is consistent with the internal standard. No conversion is required, and the coordinates (10,20) are used directly.

[0126] b) The classification code standardization unit performs a conversion based on the Bin definition parsed from the file ("1"=Pass,"8"=Fail), combined with the mapping rules set by engineer "Li Si" for this customer-product in the configuration rule base ("1"->internal standard code "1"[PassDie], "8"->internal standard code "2"[FailDie]). Therefore, the chip at coordinates (10,20) has its internal standard Bin Code set to "1"; the chip at coordinates (10,21) has its internal standard Bin Code set to "2".

[0127] (5) Content Change Awareness and Security Checkpoint: Upon successful conversion, the content change awareness submodule in the work order and process management module automatically activates. It compares the source file "Alpha_IC2024_Wafer25.xml" with the support set sample "Wafer23.xml" (used as a historical baseline) stored in the system for meta-learning adaptation. The comparison reveals that the two files have different notch direction identifiers: the new file is Notch Down, while the historical baseline file is Notch Up. This difference triggers a critical field change alarm. The submodule immediately sets the current work order status to "Requires Manual Review - Content Change" and suspends any subsequent automatic processing (such as publishing the converted file to the production line). The system automatically sends an alarm notification to engineer "Li Si".

[0128] (6) Manual review and process continuation: Engineer "Li Si" checked the alarm and found that the change was due to the physical orientation of the wafer provided by the customer being different from the previous wafers, which is normal. He confirmed the change in the system and updated the configuration information of the customer-product model, noting "Wafer25 orientation is Notch Down". After confirmation, the work order status was restored and the process continued.

[0129] (7) Standard File Generation and Output: The standard mapping file generation and output module receives all converted and verified data and encapsulates it according to the ZKT internal standard format. An example of the generated target file content fragment is as follows: [HEADER] Lot ID=LotAlpha2024Q1 Wafer ID=25 Rows=300 Columns=300 Notch=Down [DIEMAP] 10,20,1 10,21,2 ...(other chip data) ".

[0130] This file is automatically output to a network shared directory accessible by the die bonding station equipment. The die bonding machine automatically retrieves this file through the EAP system, and can accurately pick up qualified chips based on the unified coordinates (10,20) and the standard Bin Code "1" without any manual settings by engineers.

[0131] (8) Federated Learning Knowledge Accumulation: After this successful conversion, the relevant anonymized parsing data (excluding sensitive customer information) is stored on the local node. In the next federated learning cycle, the federated learning model update module will use the local data containing the parsing experience of this "Alpha_IC2024" format to train the local model and calculate the model update amount. The updates are then encrypted and uploaded to the central server. The central server aggregates updates from multiple factory nodes and updates the global model parameters. This gives all participating intelligent parsing models the potential to be enhanced to handle such new formats.

[0132] As can be seen from the above specific operational embodiments, the system of the present invention exhibits the following significant effects: 1. High efficiency in handling new formats: When faced with new customers and new formats (XML variants) that have never been seen before, the system, with the help of meta-learning capabilities, only requires engineers to provide 3 labeled samples and completes the rapid adaptation of the parsing model in about 2 minutes, breaking the bottleneck of traditional manual analysis and programming that requires several days.

[0133] 2. Accuracy of data conversion: The intelligent parsing model accurately extracted all key fields, and the classification code standardization unit correctly performed the mapping. The calculation process shows that the chip at coordinates (10,21) was correctly converted from the source file Bin Code "8" to the internal standard code "2" (Fail Die), ensuring that the die bonder would not pick up the chip incorrectly.

[0134] 3. Reliability of safety checkpoints: The content change perception submodule successfully captured the critical change of the flat edge direction (NotchDown VS NotchUp) and forcibly interrupted the process for manual review, effectively preventing a large-scale accident caused by the incorrect die bonding direction of the entire wafer due to the change in the physical direction of the incoming materials.

[0135] 4. Automation and Traceability of the Process: From work order creation, file parsing, format conversion, security checks to final file output, the entire process is automated and its status is traceable. All operations (who, when, what was done, what alarms were encountered, and how they were resolved) are recorded by the access control and management module, forming a complete audit chain.

[0136] 5. Co-evolution of knowledge: Through the federated learning mechanism, the processing experience can be contributed to the optimization of the global model while protecting the privacy of each individual's data, demonstrating the system's continuous learning and evolution capabilities.

[0137] This embodiment fully demonstrates the end-to-end process of the system from receiving files of unknown format to outputting secure and reliable standard files, verifying its comprehensive advantages in improving production efficiency, ensuring data security and product quality, and responding quickly to changes.

[0138] The technical features described above can be combined arbitrarily. Although not all possible combinations of these technical features are described, any combination of these technical features should be considered to be covered by this specification, provided that such combination does not contain contradictions.

[0139] The specific embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention. Any other corresponding changes and modifications made in accordance with the technical concept of the present invention should be included within the scope of protection of the claims of the present invention.< / die> < / diemap>

Claims

1. A unified format intelligent conversion system for wafer test mapping files, characterized in that, include: The file acquisition and reception module is used to receive wafer test mapping files with different formats generated by different test equipment; The intelligent recognition and unified conversion engine is connected to the file acquisition and receiving module. It is used to automatically recognize the format and content structure of the wafer test mapping file, and convert mapping files of different formats into mapping files of internal standard format according to the preset conversion logic. The standard mapping file generation and output module is connected to the intelligent recognition and unified conversion engine and is used to generate and output target mapping files that conform to the internal standard format. The access control and management module is used to manage the access and operation permissions of system operators, ensuring that the content of the original mapping file is not modified by unauthorized personnel.

2. The wafer test mapping file unified format intelligent conversion system according to claim 1, characterized in that, The intelligent recognition and unified conversion engine includes a file parsing submodule and a format conversion submodule; The file parsing submodule is used to perform deep parsing on the received wafer test mapping file and extract the core data fields in the file. The core data fields include the chip coordinate matrix, the classification code characterizing the chip test results, the wafer batch number, the equipment identification number, and the flat edge direction information. The format conversion submodule is used to reorganize and map the parsed core data fields according to a preset internal standard format specification; The file parsing submodule includes a semantic and structural recognition model trained based on deep learning. This model learns the encoding rules, delimiter usage conventions, and data layout features of various known mapping file formats through training. It can perform structural inference and key field location for new format mapping files without predefined parsing rules, and output structured data objects to the format conversion submodule.

3. The wafer test mapping file unified format intelligent conversion system according to claim 2, characterized in that, The semantic and structural recognition model of the file parsing submodule is a multimodal fusion neural network, and its processing flow includes: The character stream encoding layer is used to transform the original character sequence of the mapped file into a high-dimensional vector representation, capturing the text-level sequential features; The visual feature extraction layer is used to input the mapped file into the convolutional neural network as a two-dimensional image to extract its spatial features such as layout, table structure and specific symbol arrangement. The feature fusion and attention layer is used to fuse the text features output by the character stream encoding layer with the spatial features output by the visual feature extraction layer, and to focus on key regions in the file through an attention mechanism. The structure prediction and field annotation header, based on the fused features, predicts the data structure tree of the file and annotates the boundaries and semantic labels of the chip coordinates, classification code, and batch number fields.

4. The wafer test mapping file unified format intelligent conversion system according to claim 2 or 3, characterized in that, The format conversion submodule includes an adaptive encoding conversion unit and a classification code standardization unit; The adaptive encoding conversion unit is used to automatically identify the encoding method of chip coordinates and row and column numbers in the source file based on the structured data object output by the file parsing submodule, and convert it into the unified coordinate representation system required by the internal standard format; The classification code standardization unit is used to process the classification codes in the core data field. It has a pre-built classification code mapping rule library. Based on the semantics and context of the parsed source file classification codes, the unit maps them to a set of standardized classification codes defined by internal standards to indicate qualified chips, unqualified chips and chips of specific categories, ensuring that the chip bonding device can pick up chips based on unified semantics.

5. The wafer test mapping file unified format intelligent conversion system according to claim 1, characterized in that, It also includes a work order and process management module; The work order and process management module is communicatively connected to the file acquisition and receiving module, the access control and management module, and the intelligent recognition and unified conversion engine, respectively. The work order and process management module is used to receive production task information, which includes at least customer identifier, product model and mapping file identifier to be converted; The work order and process management module automatically matches or triggers the corresponding conversion logic in the intelligent identification and unified conversion engine based on the production task information, and records the execution status, operator, and timestamp of the conversion task. The permission control and management module dynamically configures the operator's operation permissions in the work order and process management module according to the operator's role, including conversion task initiation permission, conversion program configuration permission, and system management permission.

6. The wafer test mapping file unified format intelligent conversion system according to claim 5, characterized in that, The access control and management module includes a dynamic access configuration submodule and an operation audit submodule; The dynamic permission configuration submodule is used to divide system roles into at least three categories: engineering configuration roles, production operation roles, and system management roles, and to assign differentiated function access and data operation permissions to different roles. The engineering configuration role has the authority to create, modify, and verify conversion logic for new customers or new formats, but has no right to directly execute production conversion tasks; the production operation role has the authority to execute mapping file conversion tasks based on work orders and view historical conversion records, but has no right to modify any configured conversion logic rules; the system management role has the highest authority for user management, role allocation, and viewing system logs; the operation audit submodule is used to record all users' key operation logs on the system, including but not limited to: user login and logout, addition or modification of conversion logic, execution of mapping file conversion, and changes in permission settings, forming an immutable operation chain for traceability.

7. The wafer test mapping file unified format intelligent conversion system according to claim 5, characterized in that, The work order and process management module also includes a content change awareness sub-module; The content change awareness submodule is used to automatically compare the content of a new mapping file with the source mapping file used in the most recent successful conversion when it receives a new mapping file with the same customer and product model identifier as the historical task. The content change awareness submodule is pre-set with anomaly judgment logic. When the comparison finds that the key fields have changed, an alarm is triggered and the automatic conversion process is paused. The key fields include, but are not limited to: the semantic definition of the classification code, the dimension of the coordinate matrix, the horizontal direction identifier, and the file structure separator. After the alarm is triggered, the conversion task is marked as requiring manual review and the user with the engineering configuration role is notified to confirm and process it. After confirmation that there are no errors or the conversion logic is updated, the task continues to be executed.

8. The wafer test mapping file unified format intelligent conversion system according to claim 2, characterized in that, The intelligent recognition and unified conversion engine also includes a meta-learning adaptation submodule; The meta-learning adaptation submodule is coupled with the semantic and structural recognition model in the file parsing submodule; the meta-learning adaptation submodule is used to drive the semantic and structural recognition model to quickly adapt when a small number of newly formatted mapping file samples are received. Its workflow includes: Based on the general file parsing knowledge already learned by the semantic and structural recognition model as meta-knowledge, the model can perform gradient updates or rapid model parameter adjustments in a small number of steps using a small number of newly provided samples. This enables the model to obtain effective parsing capabilities for new formats without retraining and without significantly reducing the parsing performance of the original formats.

9. The wafer test mapping file unified format intelligent conversion system according to claim 8, characterized in that, The specific implementation of the meta-learning adaptation submodule includes a few-shot learning support set construction unit and a model fast tuning unit; The few-shot learning support set construction unit is used to receive support set samples provided by users for a new format. The support set samples contain a small number of mapping file instances that have been manually annotated or have been confirmed to be correctly parsed. The fast model tuning unit employs an optimization-based meta-learning algorithm to initialize the basic network parameters of the semantic and structural recognition model as meta-parameters sensitive to multi-format parsing tasks. When processing a new format, the fast model tuning unit targets the parsing loss on the support set samples and performs a finite number of gradient descent iterations near the basic network meta-parameters to quickly generate model parameters adapted to the new format. The adapted model parameters are then dynamically loaded into the file parsing submodule for processing subsequent files in the new format.

10. The wafer test mapping file unified format intelligent conversion system according to claim 1, characterized in that, It also includes a federated learning model update module; The federated learning model update module is connected to the AI ​​model in the intelligent recognition and unified conversion engine to collaboratively optimize model performance while protecting the privacy of each customer's data; its operation is as follows: On multiple production nodes where this system is deployed, the local AI model is trained using the historical conversion data of their respective wafer test mapping files, and only the update of the model parameters is encrypted and uploaded to the central server. The central server aggregates the update of the model parameters from multiple nodes, generates a global model update, and then distributes the updated global model parameters to each node.

Images