Multi-process collaborative detection method and system in chip packaging process

Abstract

The invention discloses a multi-process collaborative detection method and system in a chip packaging process, and the method comprises the steps: constructing a whole-process material pedigree tree with a chip unique code as an index, achieving the time-space alignment and standardization of multi-source heterogeneous detection data, and employing a conditional variation auto-encoder to intelligently interpolate a missing value; cooperative feature coding of process track perception is carried out based on a Transform model with a process mask, and cooperative feature representation containing whole process dependency information is obtained; a specific inter-process early warning transmission mechanism is constructed, and multi-process risk fusion and dynamic threshold adjustment are carried out; and carrying out dynamic fusion and collaborative decision-making on the multi-process detection evidence based on the D-S evidence theory, and outputting a quality judgment result and a disposal instruction. According to the invention, information sharing and intelligent cooperation of the whole packaging process are realized, false detection accumulation between processes is effectively reduced, and the defect root cause tracing efficiency and the quality determination accuracy are improved.

Description

A multi-process collaborative inspection method and system for chip packaging Technical Field

[0001] This invention relates to the field of chip packaging inspection, and in particular to a multi-process collaborative inspection method and system for chip packaging. Background Technology

[0002] Chip packaging is a crucial back-end process in semiconductor manufacturing, encompassing multiple steps such as die preparation, die bonding, bonding, molding, and lead trimming. Each step can introduce defects, and these defects have a propagating and cumulative effect. For example, misalignment during die bonding can directly lead to wire breakage or poor soldering in subsequent bonding processes.

[0003] Traditional testing methods are mostly "single-point isolated" testing, with each process's automated optical inspection equipment or electrical test data operating independently. This leads to the following technical problems: information lag and gaps, making it difficult to trace the root cause of problems in subsequent processes, resulting in high troubleshooting costs; waste of testing resources, as using a uniform testing standard for all chips may lead to insufficient testing of high-risk chips and over-testing of low-risk chips; and biased decision-making, as quality judgments based on data from a single process have limited accuracy and are prone to misjudgments and omissions.

[0004] Therefore, there is an urgent need for an intelligent testing solution that can break down process barriers, integrate multi-source data, and achieve risk warning and collaborative decision-making. Summary of the Invention

[0005] To address the technical problems in the prior art, this invention provides a multi-process collaborative detection method and system for chip packaging.

[0006] This invention is achieved through the following technical solution: a multi-process collaborative detection method in chip packaging, comprising: spatiotemporal alignment and standardization of multi-source heterogeneous detection data, including collecting detection data from multiple processes in the entire chip packaging process, constructing a full-process material hierarchy tree to perform spatiotemporal alignment of the data, and standardizing the aligned data; collaborative feature encoding based on process trajectory perception, including obtaining collaborative feature representations containing full-process dependency information based on the Transformer model; anomaly perception prediction and cross-process early warning transmission, including constructing a specific inter-process early warning transmission mechanism, performing multi-process risk fusion, and dynamically adjusting the threshold based on the calculated composite risk value; dynamic fusion and collaborative decision-making of multi-process evidence, including comprehensively judging the chip packaging quality through a dynamic decision-making model and outputting collaborative handling instructions.

[0007] Furthermore, the spatiotemporal alignment and standardization of the multi-source heterogeneous detection data also includes data interpolation using a conditional variational autoencoder based on process context. The encoder optimization objectives include negative reconstruction log-likelihood and KL divergence.

[0008] Furthermore, the collaborative feature encoding based on process trajectory awareness also includes introducing process position encoding and process type encoding into the self-attention mechanism of Transformer.

[0009] Furthermore, the specific inter-process early warning transmission mechanism includes early warning transmission from the front-end die bonding process to the back-end bonding process, early warning transmission from the front-end bonding process to the back-end molding process, and early warning transmission from the front-end bonding process to the back-end molding process.

[0010] Furthermore, the early warning transmission of the front-end die bonding process to the back-end bonding process includes constructing a die bonding-bonding transfer function, extracting die bonding process features from the collaborative feature representation, and calculating the comprehensive risk index of the die bonding process to the bonding process based on the chip offset, angle offset, and adhesive layer thickness anomaly index.

[0011] Furthermore, the early warning transmission from the front-end bonding process to the subsequent molding process includes constructing a bonding-molding transfer function, extracting bonding process-related features from the collaborative feature representation, and calculating the risk index of the die bonding process to the molding process based on the gold wire arc height, gold ball offset, and wire collapse index.

[0012] Furthermore, the early warning transmission of the front-end bonding process to the subsequent molding process is constructed as a die bonding and bonding-bead cutting transfer function, which integrates die bonding offset and bonding anomaly index to calculate the mechanical stress sensitivity risk of the bead cutting process.

[0013] Furthermore, the dynamic fusion and collaborative decision-making of multi-process evidence is based on DS evidence theory to perform multi-process evidence fusion. Based on the comprehensive anomaly index and historical statistical distribution, the evidence of each process generates a basic probability allocation function. The comprehensive anomaly index is obtained based on its own detection anomaly score, composite risk value and key process parameters.

[0014] This invention also provides a multi-process collaborative detection system for chip packaging, based on the multi-process collaborative detection method for chip packaging described above. It includes: a detection data spatiotemporal alignment and standardization module, used to collect detection data from multiple processes throughout the chip packaging process, construct a full-process material hierarchy tree to perform spatiotemporal alignment of the data, and standardize the aligned data; a collaborative feature encoding module, used to obtain collaborative feature representations containing full-process dependency information based on a Transformer model; and an anomaly perception prediction and cross-process early warning transmission module, used to construct a specific inter-process early warning transmission mechanism, perform multi-process risk fusion, and dynamically adjust thresholds.

[0015] The dynamic fusion and collaborative decision-making module for multi-process evidence is used to comprehensively determine the chip packaging quality through a dynamic decision-making model and output collaborative processing instructions.

[0016] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing program instructions for a multi-step collaborative detection method in the chip packaging process. The program instructions for the multi-step collaborative detection method in the chip packaging process can be executed by one or more processors to implement the steps of the multi-step collaborative detection method in the chip packaging process as described above.

[0017] Compared with the prior art, the beneficial effects of the present invention are: (1) The present invention constructs a full-process material spectrum tree with the chip unique code as the index, realizes the spatiotemporal alignment of heterogeneous data of multiple processes; introduces a conditional variational autoencoder based on process context to intelligently interpolate missing values, ensures the physical consistency of data, and lays a unified and reliable data foundation for collaborative analysis.

[0018] (2) This invention utilizes a Transformer model with process masking to strictly learn the defect evolution law according to the physical process and constructs an equal quantitative transfer function between processes. When an anomaly is detected in the previous process, the system automatically calculates the risk index of the next process and sends an early warning command in advance, dynamically adjusts the detection sensitivity, effectively blocks the propagation of defects, and improves the yield.

[0019] (3) Based on the DS evidence theory, this invention integrates multi-dimensional evidence such as the self-inspection results of each process, the risk of upstream transmission, and the deviation of process parameters to achieve uncertainty. This method effectively reduces misjudgment and missed judgment caused by single-point data noise, and can also provide specific handling suggestions for suspicious chips, significantly improving the accuracy of quality judgment and the level of intelligent decision-making. Attached Figure Description

[0020] The accompanying drawings, which are included to provide a further understanding of this application and constitute a part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 is a schematic flowchart of a multi-step collaborative detection method in a chip packaging process according to an embodiment of this application. Detailed Implementation

[0021] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0022] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0023] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. The illustrations only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the shape, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0024] Referring to Figure 1, a multi-process collaborative detection method in chip packaging includes the following steps: S1: Spatiotemporal alignment and standardization of multi-source heterogeneous detection data;

[0025] This step aims to break down data silos between different processes and establish a unified data foundation for subsequent collaborative analysis.

[0026] First, test data are collected from multiple processes throughout the chip packaging process. These multiple processes include at least the die preparation process, the mounting process, the bonding process, and the molding process. The test data includes visual inspection data, electrical inspection data, process parameter data, and equipment status data for each process.

[0027] Secondly, using the chip's unique identifier as an index, a full-process material hierarchy tree is constructed, linking and aligning multi-source heterogeneous data from each process according to processing time sequence. For time-series data, time misalignment caused by inter-process transmission delays and buffering is eliminated.

[0028] Specifically, the full-process material hierarchy tree uses a unique identifier as the root node and loads the detection data, process parameters and equipment status data of each process in chronological order to form a multi-level tree structure containing a complete processing history, supporting subsequent process collaborative analysis and root cause tracing.

[0029] Finally, the aligned data is standardized to construct a process co-feature matrix X ∈ R^(N×D), where N is the number of chip samples and D is the sum of the feature dimensions of all processes. For missing detection items, a conditional variational autoencoder based on process context is used for imputation. Its optimization objectives include negative reconstruction log-likelihood and KL divergence, expressed as: Where x represents the detection data to be interpolated, c represents the process context condition variable, and z represents the latent variable. For encoder, Let E be the decoder, E be the expectation, and KL be the KL divergence. This represents the KL divergence from distribution Q to distribution P.

[0030] The first term, negative reconstruction log-likelihood, measures the accuracy with which the decoder reconstructs the original data x from the latent variable z and the condition c.

[0031] Specifically, It is a probability distribution (usually Gaussian or Bernoulli) defined by parameter θ, representing the likelihood of generating data x given the latent variable z and condition c.

[0032] encoder Map the true data x and condition c to a probability distribution of the latent variable z. Take the expectation of all possible z from this distribution, hoping that the reconstructed x will be as close as possible to the true x.

[0033] When some features are missing, this term forces the model to learn how to reconstruct the complete data using conditional c and latent variable z. After training, for missing samples, the latent variable z can be inferred first based on the observed parts and conditional c, and then the decoder generates appropriate imputation for the missing values.

[0034] The second term, KL divergence, is a regularization term that measures the posterior distribution of the encoder output. With prior distribution The differences between them. Prior knowledge. Typically, a standard Gaussian distribution is assumed, representing the prior belief in the latent variable z without observing specific data x. A smaller KL divergence indicates that the posterior distribution is closer to the prior distribution, thus ensuring a good structure in the latent space (e.g., continuous and smooth) and preventing overfitting. This term ensures that even for unseen missing patterns, the sampling of the latent variable z still falls within a reasonable region, avoiding the generation of unrealistic imputation values. Simultaneously, conditional priors... This distinguishes the hidden space sub-regions corresponding to different process contexts, thereby improving interpolation accuracy.

[0035] In one embodiment, when image acquisition fails during the "bonding" process of chip A, resulting in a decrease in the "gold ball thrust" indicator... The missing term is not specified. However, we know the offset x of the previous die-bonding process, as well as its batch and model information, i.e., condition c. The encoding process model, based on c (including the known x of die-bonding) and all known characteristics of the current chip, infers the most probable "overall quality state" distribution, i.e., the latent variable z. This latent variable z may represent "the internal stress distribution state of the chip after undergoing die-bonding offset." The decoding process generates the missing term. The model uses this inferred "stress distribution state z" and the known context condition c to calculate, through the decoder, the most probable value of "gold ball thrust" in this situation (i.e., generating the missing term). If the die-bonding offset is large (an anomaly in c), the model's learned knowledge will tell it that this offset often leads to stress concentration (a certain shape in z), which in turn leads to lower gold ball thrust (the generated...). This gives a lower interpolation value for the gold ball thrust that conforms to physical laws.

[0036] S2: Collaborative feature encoding based on process trajectory perception; targeting the sequential dependency characteristics of semiconductor manufacturing processes, a feature encoding model capable of perceiving the sequential relationship between processes is constructed to capture the propagation and evolution of defects between processes.

[0037] Specifically, a Transformer-based process trajectory encoder is constructed, and the process collaborative feature matrix is ​​organized into a sequence X=[ according to the process sequence. , ,..., ], where T is the total number of processes, Let be the detection feature vector of the t-th process.

[0038] In the Transformer's self-attention mechanism, process position encoding and process type encoding are introduced to distinguish the temporal importance differences between different processes: Where Q, K, and V are the query, key, and value matrices, respectively. Let K be the transpose of K, and M be the process relation mask matrix, used to constrain the attention mechanism to only focus on physically allowed process propagation paths, avoiding learning non-existent reverse dependencies. The scaling factor is used. This encoder outputs a collaborative feature representation containing information about all process dependencies. .

[0039] S3: Anomaly detection and prediction, and cross-process early warning transmission; collaborative feature representation based on S2. To address the physical causal relationships and defect propagation patterns between processes in the chip packaging process, a specific inter-process early warning transmission mechanism is constructed. This includes early warning transmission from the front-end die bonding process to the back-end bonding process, early warning transmission from the front-end bonding process to the back-end molding process, and early warning transmission from the front-end die bonding and bonding processes to the back-end lead trimming and forming process. Ultimately, the risks of multiple processes are integrated and dynamically transmitted. The specific steps include the following.

[0040] S31: Early warning transmission from the front-end die bonding process to the subsequent bonding process; quality defects in the die bonding process directly affect the bonding quality of subsequent bonding processes. According to packaging process practice, chip corner misalignment, microcracks, and adhesive abnormalities in the die bonding process will directly lead to problems such as gold wire collapse, gold ball misalignment, and insufficient bonding pull force in the bonding process. In particular, chip tilting or uneven adhesive layer thickness will cause differences in bonding surface height, resulting in uneven ultrasonic energy transfer in the bonding machine and leading to poor solder joints.

[0041] The specific early warning mechanism is as follows: Constructing a die-bonding transfer function. Extract die bonding process-related features from the collaborative feature representation. When the offset of the i-th chip is detected... (X-direction offset) (Y-direction offset) (Angle offset) and adhesive layer thickness anomaly index At that time, calculate its comprehensive risk index for the bonding process: ;in, The Sigmoid function maps risk to the [0,1] interval. Preset weighting coefficients, optional, for example .

[0042] when threshold At this time, the system sends an early warning command to the bonding process. Optional early warning commands include: Imaging strategy adjustment: The AOI system of the bonding process performs multi-angle high-resolution imaging of the chip, focusing on detecting the consistency of the gold wire arc height on the die bonding offset side; Test parameter adjustment: The online monitoring system of the bonding process doubles the frequency of online spot checks on the gold ball thrust of the chip; Bonding parameter compensation: The ultrasonic power or pressure parameters of the bonding head are automatically adjusted according to the offset direction to compensate for the impact of die bonding unevenness.

[0043] S32: Early warning transmission from the preceding bonding process to the subsequent molding process: The condition of the gold wire in the bonding process directly affects the molding quality in the molding process. If the bonded gold wire has collapsed wires, excessive wire arc, gold ball misalignment, or gold wire damage, the flow impact of epoxy resin during the high-speed injection molding process will cause the gold wire to further shift, bend, or even break, potentially resulting in exposed gold wires or short circuits after molding. In addition, residual foreign matter in the bonding area may damage the mold under the high pressure of molding.

[0044] The specific early warning mechanism is as follows: Constructing a bond-encapsulation transfer function In the collaborative feature representation, bonding process-related features are extracted. When the bonding process detects abnormal features of the gold wire, such as the gold wire arc height... Gold ball offset Collapse Index Calculate its risk index for the molding process: ;in, For indicator functions, The maximum arc height allowed by the design rules. For the gold ball offset tolerance These are the weighting coefficients.

[0045] when threshold At this time, the system sends an early warning command to the molding process. Optional early warning commands include: Molding process parameter adjustment: Adjust the injection pressure curve and mold temperature of the molding machine, reduce the initial injection speed, and reduce the impact force on high-risk gold wires; Incoming material sorting: Divert high-risk chips to a low-speed, low-pressure molding chamber or use a gentler injection process; Foreign object re-inspection: If foreign objects are detected in the bonding process at the same time, a secondary foreign object detection is triggered in the molding loading area. Before the lead frame enters the mold, a high-resolution camera performs a focused scan of the non-molding area to ensure that no foreign objects are brought into the mold and cause damage.

[0046] S33: Warning transmission from the front-end die bonding and bonding process to the subsequent lead trimming and molding process; the lead trimming and molding process involves cutting and molding the lead frame after molding. If there is chip position misalignment or gold wire abnormality in the front-end die bonding or bonding process, after molding and curing, the mechanical stress of the lead trimming process may cause internal crack propagation or gold wire breakage, resulting in electrical test failure after lead trimming.

[0047] The specific early warning mechanism is as follows: Construct the die-bonding and bonding-beam shearing transfer function. Integrate die-bonding offset. and bonding anomaly index Calculate the mechanical stress sensitivity risk of the rebar cutting process: ;when threshold At that time, the system sends an early warning command to the lead cutting and forming process. Optional early warning commands include: mechanical parameter adjustment: adjust the closing speed and stamping pressure of the lead cutting die to reduce mechanical impact on high-risk chips; enhanced online testing: after lead cutting, perform rapid online electrical parameter testing on the chip to monitor whether open circuits or short circuits are caused by mechanical stress; stricter offline sampling inspection: list the batch of chips as key targets for reliability sampling inspection and conduct destructive tests such as shear force and gold wire tension to verify the integrity of the internal structure.

[0048] In one embodiment, the die bonding offset is obtained by an automated optical inspection device after die bonding; the bonding anomaly index is obtained by a weighted fusion of solder joint offset, gold wire arc height, and gold ball thrust.

[0049] S34: Multi-process risk fusion and dynamic threshold adjustment: Feedforward collaboration mechanism: Establish a multi-process risk fusion model, weighted fusion of early warning information from multiple upstream processes to form a composite risk heat map of downstream processes.

[0050] Let the risk vector of the k-th process be... Where M is the number of risk types. For the subsequent process j, the composite risk it receives is: ;in, The risks arising from the j-th process itself, The transfer weight matrix from process i to process j can be obtained through training on historical data. The risk of the i-th process being transferred to the j-th process can optionally be obtained from steps S31-S33.

[0051] Furthermore, based on the composite risk value, the detection thresholds for subsequent processes are dynamically adjusted. : ;in, As the baseline detection threshold, λ represents the sensitivity coefficient (0 < λ < 1). For high-risk chips, the detection threshold is lowered, increasing detection sensitivity; for low-risk chips, the detection threshold is raised, reducing over-detection.

[0052] Furthermore, the warning information is transmitted to downstream process operators or AOI systems in the form of heat map annotations. For example, on the display screen of the bonding process, chips with high risk of die bonding are marked with a red box, indicating "This chip has a large die bonding offset, it is recommended to focus on checking the gold wire on the corresponding side"; on the MES system interface of the molding process, it indicates "The risk of bonding wire collapse in this batch is high, and the injection pressure has been automatically reduced".

[0053] Through the above-described process early warning transmission mechanism, this invention achieves a leap from single-point detection to process awareness. Minor deviations in the die bonding process become key detection areas in the bonding process, and gold wire anomalies in the bonding process become the basis for process adjustment in the molding process, truly constructing an intelligent collaborative detection network for the entire chip packaging process.

[0054] S4: Dynamic fusion and collaborative decision-making of evidence from multiple processes;

[0055] This step further integrates the inspection evidence and risk information from each process, uses a dynamic decision-making model to comprehensively assess the chip packaging quality, and outputs collaborative handling instructions. Specifically, it includes the following sub-steps:

[0056] S41: Construct a multi-stage evidence matrix;

[0057] The collaborative feature representation H, early warning information, and process parameters of the detection results of each chip sample at each process stage are organized into a multi-dimensional evidence matrix. Where T is the total number of processes, and K is the number of evidence types extracted for each process. Evidence types may include: process-specific anomaly scores, forward-transfer composite risk values, and critical process parameter offsets. For the i-th chip, the evidence vector for the t-th process can be represented as: ;in, The self-detection anomaly score for process t is obtained by comparing the data in the collaborative feature representation H with the set quantity; The composite risk value of process t obtained in step S34; This refers to the deviation of the actual value of the key process parameter t from the ideal design value.

[0058] S42: Intra-process multi-feature fusion generates basic probability assignments; to handle the uncertainty and potential conflicts of multi-process evidence, an improved Dempster-Shafer (DS) evidence theory is used for fusion. First, the identification framework is defined. These correspond to the chip quality status: qualified, questionable, and unqualified. Based on the comprehensive anomaly index... Based on historical statistical distributions and evidence for each process, a basic probability allocation (BPA) function is generated. ,satisfy .

[0059] A three-part membership function is used: ; ; ;in , The comprehensive anomaly index threshold for process t can be set based on the quantiles of the historical qualified product statistical distribution (e.g., the 95th and 99th quantiles).

[0060] In one embodiment, the method for calculating the comprehensive anomaly index is as follows: ; where , , are weights respectively, , , are the self-detection anomaly score of process t, the composite risk value, and the deviation degree of the actual value of the key process parameter relative to the designed ideal value respectively.

[0061] S43: Collaborative decision-making based on the fusion result: According to the fused BPA, calculate the belief function Bel(A) and the plausibility function Pl(A) of each state: ; ; where A is the target proposition and B is the evidence set.

[0062] Define the decision rule: If , the chip is judged as qualified and normally flows into the next process; if , the chip is judged as unqualified, triggering the rejection or rework process; otherwise, it is judged as suspicious and manual rejudgment or additional offline sampling inspection is required.

[0063] where is the passing judgment threshold, is the non-passing judgment threshold, which can be dynamically adjusted according to the production quality requirements. For suspicious chips, specific disposal suggestions are generated in combination with the warning information in step three, such as "It is recommended to复测 the gold wire tension in the bonding process" or "Reduce the injection speed in the plastic encapsulation process".

[0064] Through the above steps, the present invention realizes the leap from single-process independent judgment to multi-process collaborative decision-making, effectively integrates the in-process detection information and the risk transfer law between processes, significantly improves the credibility of chip packaging quality judgment, reduces misjudgment and missed judgment, and provides strong support for intelligent quality control.

[0065] In this embodiment, a full-process material pedigree tree is constructed with the chip unique code as the index to achieve the spatio-temporal alignment of multi-process heterogeneous data; a conditional variational autoencoder based on process context is introduced to intelligently interpolate missing values to ensure data physical consistency, laying a unified and reliable data foundation for collaborative analysis.

[0066] Use the Transformer model with process masks to strictly learn the defect evolution law according to the physical process and construct an equal-quantity transfer function between processes. When an anomaly in the previous process is detected, the system automatically calculates the risk index of the subsequent process and sends a warning instruction in advance, dynamically adjusting the detection sensitivity, effectively blocking the defect propagation, and improving the yield.

[0067] Based on the DS evidence theory, this method integrates multi-dimensional evidence, including self-inspection results from each process, risks in the upstream process, and deviations in process parameters, to address uncertainties. This approach effectively reduces misjudgments and omissions caused by single-point data noise and provides specific handling suggestions for suspicious chips, significantly improving the accuracy of quality assessment and the level of intelligent decision-making.

[0068] This invention also proposes a multi-process collaborative detection system for chip packaging, based on the multi-process collaborative detection method for chip packaging described above. The system includes: a detection data spatiotemporal alignment and standardization module, used to collect detection data from multiple processes throughout the chip packaging process, construct a full-process material hierarchy tree to perform spatiotemporal alignment of the data, and standardize the aligned data; a collaborative feature encoding module, used to obtain collaborative feature representations containing full-process dependency information based on a Transformer model; an anomaly perception prediction and cross-process early warning transmission module, used to construct a specific inter-process early warning transmission mechanism, perform multi-process risk fusion, and dynamically adjust thresholds; and a multi-process evidence dynamic fusion and collaborative decision-making module, used to comprehensively determine the chip packaging quality through a dynamic decision-making model and output collaborative handling instructions.

[0069] Furthermore, this invention also proposes a computer-readable storage medium storing program instructions for a multi-step collaborative detection method in the chip packaging process. These program instructions can be executed by one or more processors to implement the steps of the multi-step collaborative detection method in the chip packaging process as described above.

[0070] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A multi-process collaborative detection method in chip packaging, characterized in that, include: The spatiotemporal alignment and standardization of multi-source heterogeneous detection data includes collecting detection data from multiple processes in the entire chip packaging process, constructing a full-process material hierarchy tree to perform spatiotemporal alignment of the data, and standardizing the aligned data; collaborative feature encoding based on process trajectory perception includes obtaining collaborative feature representations containing full-process dependency information based on the Transformer model; anomaly detection prediction and cross-process early warning transmission includes constructing specific inter-process early warning transmission mechanisms, performing multi-process risk fusion, and dynamically adjusting thresholds based on the calculated composite risk value; dynamic fusion and collaborative decision-making of multi-process evidence includes comprehensively judging chip packaging quality through a dynamic decision model and outputting collaborative handling instructions.

2. The multi-process collaborative detection method in the chip packaging process according to claim 1, characterized in that, The spatiotemporal alignment and standardization of the multi-source heterogeneous detection data also includes data interpolation using a conditional variational autoencoder based on process context. The encoder optimization objectives include negative reconstruction log-likelihood and KL divergence.

3. The multi-process collaborative detection method in the chip packaging process according to claim 1, characterized in that, The collaborative feature encoding based on process trajectory awareness also includes introducing process position encoding and process type encoding into the self-attention mechanism of Transformer.

4. The multi-process collaborative detection method in the chip packaging process according to claim 1, characterized in that, The specific inter-process early warning transmission mechanism includes early warning transmission from the previous die bonding process to the subsequent bonding process, early warning transmission from the previous bonding process to the subsequent molding process, and early warning transmission from the previous die bonding and bonding processes to the subsequent bead cutting and forming process.

5. The multi-process collaborative detection method in the chip packaging process according to claim 4, characterized in that, The early warning transmission from the front-end die bonding process to the back-end bonding process includes constructing a die bonding-bonding transfer function, extracting die bonding process features from the collaborative feature representation, and calculating the comprehensive risk index of the die bonding process to the bonding process based on the chip offset, angle offset, and adhesive layer thickness anomaly index.

6. The multi-process collaborative detection method in the chip packaging process according to claim 4, characterized in that, The early warning transmission from the front-end bonding process to the subsequent molding process includes constructing a bonding-molding transfer function, extracting bonding process-related features from the collaborative feature representation, and calculating the risk index of the die bonding process to the molding process based on gold wire arc height, gold ball offset, and wire collapse index.

7. The multi-process collaborative detection method in the chip packaging process according to claim 4, characterized in that, The early warning transmission of the preceding bonding process to the subsequent molding process is constructed using a die bonding and bonding-bead cutting transfer function. This function integrates die bonding offset and bonding anomaly index to calculate the mechanical stress sensitivity risk of the bead cutting process.

8. The multi-process collaborative detection method in the chip packaging process according to claim 1, characterized in that, The dynamic fusion and collaborative decision-making of multi-process evidence is based on DS evidence theory to fuse multi-process evidence. Based on the comprehensive anomaly index and historical statistical distribution, the evidence of each process generates a basic probability allocation function. The comprehensive anomaly index is obtained based on its own anomaly detection score, composite risk value, and key process parameters.

9. A multi-process collaborative inspection system for chip packaging, based on the multi-process collaborative inspection method for chip packaging as described in any one of claims 1 to 8, characterized in that, include: The module for spatiotemporal alignment and standardization of inspection data is used to collect inspection data from multiple processes in the entire chip packaging process, construct a full-process material hierarchy tree to align the data spatiotemporally, and standardize the aligned data. The module for collaborative feature encoding is used to obtain collaborative feature representations containing the dependency information of the entire process based on the Transformer model. The module for anomaly detection prediction and cross-process early warning transmission is used to construct a specific inter-process early warning transmission mechanism, perform multi-process risk fusion, and dynamically adjust the threshold. The module for dynamic fusion and collaborative decision-making of multi-process evidence is used to comprehensively judge the chip packaging quality through a dynamic decision-making model and output collaborative handling instructions.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program instructions for a multi-step collaborative detection method in the chip packaging process. The program instructions for the multi-step collaborative detection method in the chip packaging process can be executed by one or more processors to implement the steps of the multi-step collaborative detection method in the chip packaging process as described in any one of claims 1 to 8.

Images