Omr bonding process bubble detection and traceability method based on sensory data

Abstract

The application relates to the field of high-end display manufacturing and sensing technology, and discloses an OMR attaching process bubble detection and tracing method based on sensing data; the application aims to solve the industry problems of low efficiency and poor accuracy caused by the fact that traditional bubble detection relies on manual work and problem tracing relies on experience. The method comprehensively integrates multi-source sensing data in the production process, and realizes two core functions by using intelligent algorithms: one is high-precision and automatic real-time detection of small bubbles; the other is to build a correlation model of defect phenomena and upstream process parameters, and to open a precise data tracing path from the result to the physical root cause; the application can quickly lock the root cause of the defect and provide engineers with accurate quantitative diagnosis reports, completely replacing the traditional trial-and-error troubleshooting. It significantly improves the accuracy of defect detection and the efficiency of problem tracing, and provides key technical support for yield improvement and process stability of high-end display manufacturing.

Description

Technical Field

[0001] This invention relates to the fields of high-end display manufacturing and sensing technology, and in particular to a method for detecting and tracing bubbles in the OMR bonding process based on sensor data. Background Technology

[0002] In high-precision manufacturing scenarios such as flexible electronics, power battery cell winding or stacking, and composite material lamination, real-time, in-situ quality monitoring of process interfaces (such as bonding and lamination interfaces) is crucial to ensuring the performance and yield of the final product. Currently, mainstream online monitoring technologies mainly rely on machine vision systems or discrete point sensors (such as thermocouples and pressure gauges) deployed at key locations. Machine vision technology identifies macroscopic defects such as bubbles, wrinkles, or foreign objects by analyzing images of the product surface, while point sensors are used to monitor whether process parameters such as temperature or pressure at specific points exceed limits.

[0003] However, while existing monitoring technologies can identify defects or single-point anomalies in process parameters to some extent, they typically only provide information about "what" or "where" the defect is, but struggle to provide a direct and reliable judgment on "why" the defect occurred. When a monitoring system issues an alarm, its output is often a simple "non-conforming" signal or the coordinates of the defect, creating a critical information gap between defect tracing and process optimization. Production line operators still need to rely on personal experience or time-consuming offline analysis methods, repeatedly experimenting to deduce and correct the underlying process parameter deviations causing the defect (e.g., inappropriate roller pressure, excessively fast bonding speed, or localized temperature anomalies). Therefore, establishing a real-time, automated correlation between the observed physical characterization of defects and their potential process root causes during production is a pressing technical challenge in this field. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] How to establish a connection between the observed physical characterization of defects and their potential process roots during the production process, so as to achieve a comprehensive, real-time, and high-resolution insight into the state of the entire process interface, thereby enabling timely detection and location of early defects occurring in sensor blind spots.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a method for detecting and tracing air bubbles in the OMR bonding process based on sensor data, which includes the following steps:

[0008] S1. Sensor data acquisition: Dynamic physical field distribution data during the bonding process is acquired in real time through a sensor array that integrates multiple physical sensing units and is deployed in the bonding interface area.

[0009] S2. Dynamic physical field reconstruction: Based on the dynamic physical field distribution data, a dynamic physical field characterizing the physical state of the bonding interface is reconstructed;

[0010] S3. Defect signature recognition: In the dynamic physical field, the formation event of bubble defects is detected by recognizing the defect physical signatures preset in the knowledge base;

[0011] S4. Root Cause Diagnosis: After detecting the bubble defect, based on the quantitative characteristics of the defect's physical signature, matching and reasoning are performed in a preset process diagnosis model to determine the root cause of the bubble defect.

[0012] As a preferred embodiment of the OMR bonding process bubble detection and tracing method based on sensor data described in this invention, the specific implementation of step S1, sensor data acquisition, includes: deploying a flexible sensor array that integrates at least two types of distributed sensing units on its flexible substrate, enabling its sensing surface to dynamically conform to the contour of the bonding interface area; and synchronously acquiring the response data of all sensing units through a time synchronization acquisition mechanism, wherein the time synchronization acquisition mechanism ensures strict consistency of the data on the time reference, thereby generating a data stream composed of continuous synchronous multidimensional sensing snapshots.

[0013] As a preferred embodiment of the OMR bonding process bubble detection and tracing method based on sensor data described in this invention, the flexible sensor array further includes an integrated composite structure comprising multiple composite sensing units, each of which integrates at least two different types of sensing functions at a single location point, thereby ensuring the inherent colocation of multidimensional sensing data in spatial sampling.

[0014] As a preferred embodiment of the OMR bonding process bubble detection and tracing method based on sensor data described in this invention, the specific implementation of step S2 dynamic physical field reconstruction includes: preprocessing: preprocessing the discrete distribution data collected by the sensor array, which includes performing at least one operation in filtering and calibration to suppress noise in the original sensor signal and improve data quality; continuity: processing the preprocessed discrete distribution data using at least one spatial interpolation algorithm to generate a physical field distribution that is continuous in time and space for each sensed physical quantity; fusion: aligning and fusing the generated multiple continuous physical field distributions at the data level to construct the dynamic physical field.

[0015] As a preferred embodiment of the OMR bonding process bubble detection and tracing method based on sensor data described in this invention, the fusion further includes: specifically, it is achieved by integrating multiple continuous physical field distribution data into a unified mathematical model, wherein the mathematical model aims to characterize the inherent coupling relationship between various physical quantities on the bonding interface as they evolve with time and space.

[0016] As a preferred embodiment of the OMR bonding process bubble detection and traceability method based on sensor data described in this invention, step S3, which involves identifying defect physical signatures preset in a knowledge base, further includes: retrieving a set of defect physical signatures from the knowledge base, wherein each signature is defined as a standard data pattern characterizing the dynamic characteristics of a specific type of process defect; continuously scanning and comparing the real-time generated dynamic physical field using a pattern recognition algorithm to calculate the matching degree between a local region of the dynamic physical field and each standard data pattern in the defect physical signature; and when any calculated matching degree exceeds a preset threshold, the defect type associated with that standard data pattern is identified as the defect information of the current process.

[0017] As a preferred embodiment of the OMR bonding process bubble detection and tracing method based on sensor data described in this invention, the specific implementation of step S3, defect signature recognition, includes: analyzing the dynamic physical field through a data-driven feature extraction model to extract a set of quantitative feature vectors that characterize the bonding interface process state from the spatiotemporal evolution process; inputting the quantitative feature vectors into a pre-trained classification model trained based on historical process data, and having the pre-trained classification model output a judgment result regarding the bonding interface process defect; and, when a process defect is determined to exist, further inversely associating the dynamic physical field based on the judgment result to locate the spatial region in the dynamic physical field that causes the defect to occur.

[0018] As a preferred embodiment of the OMR bonding process bubble detection and tracing method based on sensor data described in this invention, the quantized feature vector includes at least one feature selected from the following group: static spatial features describing the distribution state, uniformity, and local anomalies of a single physical field in the spatial dimension at any given time; dynamic temporal features describing the evolution of the dynamic physical field over time at single and multiple spatial locations; and coupling features describing the spatiotemporal correlation and synchronicity between two or more different physical fields.

[0019] As a preferred embodiment of the OMR bonding process bubble detection and tracing method based on sensor data described in this invention, the specific implementation of step S4 root cause tracing diagnosis includes: taking the quantified features of the defect physical signature as an input vector and providing it to a preset process diagnosis model trained based on historical data, and the process diagnosis model outputting a specific diagnosis result identifying the root cause through the internally established mapping relationship from features to causes.

[0020] As a preferred embodiment of the OMR bonding process bubble detection and tracing method based on sensor data described in this invention, the process of establishing the process diagnostic model includes: collecting historical sensor data of the process containing known root causes, and for each known root cause, extracting the quantitative features of its defect physical signature from the corresponding historical sensor data, thereby constructing a training dataset composed of quantitative features and root cause labels; and based on the training dataset, training the process diagnostic model through a supervised learning algorithm so that it learns and establishes the mapping relationship from features to causes.

[0021] The beneficial effects of this invention are as follows: Compared with the prior art, the core advantage of this invention is that by reconstructing sparse sensor data into spatial continuity, this invention transforms the invisible monitoring blind zone into a global, continuous, and visualized physical field map, providing a foundation for a comprehensive understanding of the process status.

[0022] Secondly, this invention innovatively introduces a knowledge-based defect signature recognition mechanism, enabling the system not only to "see" but also to "understand" abnormal patterns in the physical field. This allows for accurate early warning and location of defects at their nascent stage, elevating quality control from post-event remediation to pre-event prevention. Furthermore, by modeling the spatiotemporal evolution of the physical field using a deep learning model, this invention significantly improves the intelligence and robustness of defect recognition, effectively reducing reliance on human experience and the misjudgment rate under complex conditions. Attached Figure Description

[0023] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart of a bubble detection and traceability method for the OMR bonding process based on sensor data.

[0025] Figure 2 This is a flowchart of the sensor data acquisition process.

[0026] Figure 3 This is a flowchart for reconstructing dynamic physical fields.

[0027] Figure 4 This is a flowchart for defect signature recognition.

[0028] Figure 5 This is a flowchart for root cause diagnosis. Detailed Implementation

[0029] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0030] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0031] Secondly, the term "one embodiment" or "example" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the invention. The appearance of an embodiment in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that mutually excludes other embodiments.

[0032] Reference Figures 1-5 This is one embodiment of the present invention, which provides a method for bubble detection and tracing in an OMR bonding process based on sensor data, comprising the following steps:

[0033] S1: Sensor Data Acquisition: Dynamic physical field distribution data during the bonding process is acquired in real time through a sensor array integrating multiple physical sensing units deployed in the bonding interface area. The specific steps are as follows:

[0034] Deploy a flexible sensor array that integrates at least two types of distributed sensing units on its flexible substrate, enabling its sensing surface to dynamically conform to the contour of the interface area; and through a time-synchronized acquisition mechanism, synchronously acquire response data of all sensing units, which ensures strict consistency of data on a time base, thereby generating a data stream consisting of continuous synchronous multidimensional sensing snapshots.

[0035] The flexible sensor array adopts an integrated composite structure, which contains multiple composite sensing units. Each composite sensing unit integrates at least two different types of sensing functions at a single location point, thereby ensuring the inherent colocation of multidimensional sensing data in spatial sampling.

[0036] This invention first deploys a flexible sensor array in a critical area of ​​the OMR bonding equipment, near the bonding line between the roller and the substrate. The sensor array is constructed using polyimide as a flexible substrate. Through microelectromechanical systems (MEMS) technology, 64 composite sensing units (8x8) are integrated onto it. Each unit integrates both a pressure sensing unit and a temperature sensing unit at the same location. This integrated composite structure ensures inherent co-location of pressure and temperature data in spatial sampling, avoiding spatial alignment errors during subsequent data fusion.

[0037] A composite sensing unit refers to a basic sensor building block that achieves integrated design at the microscopic level. It is not simply a matter of juxtaposing or gluing together multiple discrete, single-function sensor components (such as a separate pressure sensor and a separate temperature sensor) on a macroscopic level. Rather, it refers to a unit that, at the same physical location, possesses the ability to simultaneously respond to and measure at least two or more different physical quantities (such as normal pressure, shear force, temperature, etc.) through a high degree of integration of materials, structure, or processes.

[0038] Inherent colocation is a key technical attribute used to describe the composite sensing unit employed in this invention. This attribute refers to the integration of multiple different physical sensing functions at a single, indivisible physical point, thereby enabling the measurement of different physical quantities (such as normal pressure, temperature, shear force, etc.) to have an inherent, hardware-structure-determined uniformity in spatial coordinates.

[0039] Regardless of how the surface shape of the bonding interface changes in real time due to the process, this flexible sensor array can always cover it tightly and seamlessly, like a smart plastic wrap, ensuring that the collected physical field data (such as pressure and temperature) truly, accurately, and with high fidelity reflect the real state of the interface.

[0040] Data Acquisition System: The array connects to a 64-channel time-synchronized data acquisition card, synchronously acquiring all 128 data channels (64 pressure + 64 temperature) at a sampling rate of 1 kHz. This ensures strict consistency of all data on a time base, thereby generating a data stream consisting of continuous, synchronous multidimensional sensor snapshots.

[0041] By collecting dynamic physical field distribution data in real time during the bonding process, an unprecedented dimension of information is provided for defect detection and tracing:

[0042] For detection: a pressure field distribution map of just one frame is sufficient for the present invention to locate a bubble through a local low-pressure pit.

[0043] For tracing the source: Observing how this low-pressure crater forms in dynamic video—for example, whether it appears instantaneously or forms slowly, and whether its center is accompanied by a momentary increase in temperature field—these dynamic evolution processes constitute defect physical labels that distinguish bubbles of different origins (such as gas residue vs. particulate foreign matter), thereby achieving precise physical tracing.

[0044] S2. Dynamic Physical Field Reconstruction: Based on the dynamic physical field distribution data, the dynamic physical field representing the physical state of the bonding interface is reconstructed. The specific steps are as follows:

[0045] Preprocessing: The discrete distributed data acquired by the sensor array is preprocessed, including at least one operation of filtering and calibration to suppress noise in the original sensing signal and improve data quality; Continuation: The preprocessed discrete distributed data is processed using at least one spatial interpolation algorithm to generate a physical field distribution that is continuous in time and space for each sensed physical quantity; Fusion: The multiple generated continuous physical field distributions are aligned and fused at the data level to construct a dynamic physical field.

[0046] Fusion is specifically achieved by integrating multiple continuous physical field distribution data into a unified mathematical model, which aims to characterize the intrinsic coupling relationship between various physical quantities on the bonding interface as they evolve over time and space.

[0047] The original data collected is discrete. In order to obtain a continuous physical field, the present invention performs the following operations:

[0048] Preprocessing: A Kalman filter is applied to the raw data stream of each sensing channel to effectively suppress electromagnetic interference and thermal noise, thereby improving the signal-to-noise ratio. Simultaneously, the data is calibrated according to a pre-calibrated lookup table to compensate for minor differences in sensitivity between the sensing units.

[0049] Continuity: This invention uses the Kriging interpolation algorithm to spatially interpolate 64 discrete pressure data points and 64 discrete temperature data points after preprocessing, generating continuous pressure field distribution maps and temperature field distribution maps with a resolution of 256x256, respectively.

[0050] estimated value The measurement points x of its n neighboring sensors i The actual measured value Z(x) of (i = 1, 2, ..., n) i Calculated using a weighted linear combination of )

[0051]

[0052] in, This is the predicted value at the unknown point x0 that we want to know in this solution. For example, the pressure value at a point that fits the center of the interface.

[0053] Z(x1), Z(x2), ..., Z(x) n ): These are the actual measured values ​​of the n sensors that are already known in this scheme.

[0054] λ1,λ2,...,λ n These are the weights assigned to each sensor. These weights are the core and most intelligent part of the entire Kriging method.

[0055] Fusion: The continuous pressure and temperature fields are fused at the data level to construct a unified dynamic physical field. In this embodiment, the state of each spatial point (x, y) at time t is represented as a two-dimensional state vector.

[0056] V(x,y,t)=[P(x,y,t),T(x,y,t)].

[0057] Where P represents pressure and T represents temperature. This vectorized model characterizes the intrinsic coupling relationship between pressure and temperature at the bonding interface as they evolve over time and space.

[0058] This unified mathematical model can be characterized by a generalized system of transport equations, in vector form as follows:

[0059]

[0060] The rate of state evolution over time.

[0061] The convection term represents the transport of physical quantities during the bonding process (velocity U).

[0062] The diffusion / conduction term is the core of the model, describing the propagation and interaction of physical quantities in the medium.

[0063] F(V,x,y,t): Source term, representing external forces (such as equipment vibration) or internal generation (such as instantaneous stress sources caused by foreign particles).

[0064] 1. Explanation of each part of the model

[0065] K: Coupling coefficient matrix (key to this model). This is a D x D matrix (D is the physical field dimension), whose off-diagonal elements directly represent the intrinsic coupling relationships between physical fields.

[0066] K = [[K p p,K p t],[K t p,Kt t]).

[0067] Diagonal element: K p p and K t t represents the diffusion / conduction coefficients of pressure and temperature, respectively.

[0068] Off-diagonal elements (coupling terms):

[0069] K p t: Characterizes the effect of the temperature field on the pressure field. For example, a local increase in temperature causes material expansion or changes in the curing rate of adhesives, thereby altering the pressure distribution.

[0070] K t p: Characterizes the effect of the pressure field on the temperature field. For example, the temperature change caused by adiabatic compression / expansion due to abrupt pressure changes (thermoelastic effect).

[0071] When K p t or K t When p is not zero, the model mathematically represents the "intrinsic coupling relationship".

[0072] Specific examples of coupling relationships:

[0073]

[0074] S3. Defect Signature Recognition: In a dynamic physical field, the formation event of bubble defects is detected by recognizing the defect physical signatures preset in the knowledge base. The specific steps are as follows:

[0075] A set of defect physical signatures is retrieved from the knowledge base, where each signature is defined as a standard data pattern representing the dynamic characteristics of a specific type of process defect. Through a pattern recognition algorithm, continuous scanning and comparison are performed in a dynamically generated physical field in real time to calculate the matching degree between a local region of the dynamic physical field and each standard data pattern in the defect physical signature. Furthermore, when any calculated matching degree exceeds a preset threshold, the defect type associated with that standard data pattern is identified as the defect information of the current process.

[0076] By analyzing the dynamic physical field through a data-driven feature extraction model, a set of quantitative feature vectors that can characterize the process state of the bonding interface can be extracted from the spatiotemporal evolution process. The quantitative feature vectors are then input into a pre-trained classification model trained on historical process data. The pre-trained classification model outputs a judgment result on the bonding interface process defect. Furthermore, when a process defect is determined to exist, the dynamic physical field is further correlated with the judgment result to locate the spatial region in the dynamic physical field that caused the defect.

[0077] The quantization feature vector is selected from at least one feature from the following group: static spatial features used to describe the distribution, uniformity and local anomalies of a single physical field in the spatial dimension at any time; dynamic temporal features used to describe the evolution of a dynamic physical field over time at a single or multiple spatial locations; and coupling features used to describe the spatiotemporal correlation and synchronicity between two or more different physical fields.

[0078] This invention predefines physical signatures for various defects in a knowledge base. For example, a typical bubble defect signature is defined as follows: within a circular region with a radius of less than 2 mm, the pressure value rapidly drops to below 120% of the ambient pressure within 50 milliseconds, while the temperature of this region steadily increases by 0.5 °C relative to the surrounding area.

[0079] This invention employs a normalized cross-correlation algorithm to continuously scan and match real-time dynamic physical field local regions with bubble signature templates in the knowledge base.

[0080] In this embodiment, the core of the knowledge base-based defect signature recognition is to perform template matching through a normalized cross-correlation algorithm.

[0081] Calculated using the following formula:

[0082]

[0083] Sliding window and region definition: The algorithm uses the defect template t as a sliding window, moving pixel by pixel across the physical field map f to be inspected. At each stopping position (u,v), the algorithm defines a sub-region in the map f that is exactly the same size as the template t for comparison.

[0084] Brightness independence processing (numerator operations): To eliminate interference from absolute physical quantities (or image brightness), the algorithm first performs "zero-mean normalization" on the template and the current sub-region. It calculates the average value of the template t. and the local average value of the current sub-region Then, when calculating the cross-correlation, the average value of the region within each pixel is subtracted from its original value. This step... and The operation shifts the focus of comparison from absolute values ​​to internal structure and texture undulations. Regardless of whether a region is generally brighter or darker, as long as its internal relative change trend is consistent with the template, the molecule will generate a large positive value.

[0085] Contrast Independence Processing (Denominator Operations): To further eliminate the influence of contrast (i.e., the dynamic range of the signal), the algorithm calculates the denominator as a normalization factor. The two terms in the denominator... and These are measures of the energy (or standard deviation) of the current sub-region and the template, respectively. Dividing the molecule by this normalization factor is equivalent to canceling out their respective "energy intensities." This ensures that even if a defect feature appears very weak (low contrast) on the physical field map, while the template feature is very clear (high contrast), the algorithm can still give a high similarity score as long as their basic morphological structures are consistent.

[0086] Response Map Generation and Defect Location: The above calculation is performed once at every possible location (u,v) in the physics map f, generating a final similarity score R(u,v). This score ranges from -1 to +1, where +1 represents a perfect match, 0 represents no correlation, and -1 represents a perfect negative correlation. The scores R(u,v) at all locations together constitute a correlation response map. Locations with significant peaks (values ​​close to +1) in this response map correspond to the best-matching locations of the defect template in the original physics map. By setting a reasonable threshold (e.g., 0.85), the system can automatically identify these peak points, thereby achieving accurate and robust defect location.

[0087] In a preferred embodiment, in order to extract quantized feature vectors from a dynamic physical field (essentially a video or time-series image sequence), a hybrid neural network model can be constructed:

[0088] Convolutional neural networks as spatial feature extractors:

[0089] For each snapshot of the dynamic physics field (i.e., a two-dimensional or three-dimensional image of the physics field), a pre-trained convolutional neural network is used to extract its spatial features. The convolutional layers of the convolutional neural network can effectively capture information such as local patterns, gradients, and textures in the physics field, and its output is a compact feature vector.

[0090] Long Short-Term Memory Networks as Time Series Analyzers:

[0091] A series of time-ordered feature vectors extracted by a convolutional neural network are input into a long short-term memory (LSM) network. The LSM network can learn the evolution and dependencies of these features over time, thereby capturing the dynamic process of defect formation. The final hidden state of the LSM network is a quantized feature vector that integrates spatiotemporal information.

[0092] S4. Root Cause Diagnosis: After detecting bubble defects, based on the quantitative characteristics of the defect's physical signature, matching and reasoning are performed in a pre-set process diagnosis model to determine the root cause of the bubble defects. The specific steps are as follows:

[0093] The quantified features of the physical signature of the defect are used as input vectors and fed into a pre-set process diagnostic model trained on historical data. The process diagnostic model then outputs a specific diagnostic result that identifies the root cause through an internally established mapping relationship from features to causes.

[0094] Collect historical sensor data of the process containing known root causes, and for each known root cause, extract the quantitative features of its defect physical signature from the corresponding historical sensor data to construct a training dataset consisting of quantitative features and root cause labels; and based on the training dataset, train the process diagnostic model through a one-supervised learning algorithm so that it learns and establishes the mapping relationship from features to causes.

[0095] Establishment of the process diagnostic model:

[0096] Step 1: Building a training knowledge base

[0097] Collecting raw data: First, collect a large amount of historical sensor data from past production lines during the bonding process. This data constitutes dynamic physical field data.

[0098] Identify physical defects: Find all cases of defects that have occurred in this historical data. Then, a person skilled in the art manually analyzes and annotates each defect case, clearly pointing out the root cause of the defect.

[0099] For example:

[0100] The data in Case A was labeled by the engineer as: "Roller pressure too high".

[0101] The data in Case B was labeled by the engineer as: "Adhesion speed mismatch".

[0102] The data in Case C was labeled by the engineer as: "Microparticles exist on the substrate."

[0103] Forming a knowledge base: After the above two steps, this invention obtains a defect case training dataset. Each record in this dataset contains two parts:

[0104] Problem: A quantitative feature vector describing the changes in the physical field when a defect occurs.

[0105] Answer: The root cause label that caused the defect.

[0106] Step 2: Extract quantized feature vectors

[0107] Simply feeding the massive amount of raw video data directly to the model is inefficient and fails to capture the key information. In this embodiment, the key information of each defect case needs to be extracted and translated into a language that the model can understand.

[0108] This process involves extracting a quantitative feature vector from the dynamic physical field that accurately describes the defect's condition and behavior. This vector may contain:

[0109] Static spatial characteristics:

[0110] What is the area of ​​the defective low-pressure zone?

[0111] Is the low-pressure area circular or elongated?

[0112] How high is the peak value of the pressure spike?

[0113] Dynamic time-domain characteristics:

[0114] Does the low-pressure area appear instantaneously or form slowly? (For example, what is the pressure drop rate in Pa / ms?)

[0115] Does the pressure fluctuation exhibit periodicity? What is the frequency?

[0116] Coupling characteristics:

[0117] When the pressure decreases, does the temperature rise synchronously or remain unchanged?

[0118] How closely related are the pressure field and the temperature field?

[0119] Through feature vector extraction, each complex defect case is transformed into a set of vectorized numbers.

[0120] Step 3: Training the diagnostic model

[0121] Algorithm selection: Random Forest algorithm is used. Training: In this invention, the vectorized numbers extracted in step two are input into the Random Forest algorithm in large quantities.

[0122] Learning the mapping relationship: During the learning process, the Random Forest algorithm searches for the non-linear, intrinsic pattern between features and causes. By continuously adjusting its internal parameters, it establishes a precise mapping relationship: given a specific combination of feature vectors, it can output the corresponding root cause with extremely high accuracy.

[0123] Step 4: Model Building and Solution

[0124] After detecting bubble defects, the quantized feature Φ of the defect physical signature extracted in S3 is used. defect In the preset process diagnostic model Θ diag The reasoning is used to determine the root cause of the defect.

[0125] 1. Diagnostic Inference: Input the feature vector into a pre-trained diagnostic model (e.g., a random forest model), and output the predicted root cause.

[0126] hatC=M diag (Φ defect ;Θ diag ).

[0127] 2. Establishment of the diagnostic model: This model is trained using a supervised learning algorithm on a historical dataset containing labels of known root causes. The goal of training is to find a set of optimal model parameters Θ*. diag This minimizes the loss function L:

[0128] Θ* diag =argmin Θ [(1 / N)*ΣL(M diag (Φ i ;Θ),C i )+Ω(Θ)].

[0129] Where (Φ i C i ) represents the training samples, and Ω(Θ) represents the regularization term.

[0130] Step 5: Model Validation and Deployment

[0131] In this invention, after the model training is completed, it needs to be validated using a portion of data that was never used in the training to verify its diagnostic accuracy. If the accuracy reaches a preset standard, the model is considered successfully established and can be deployed to the production line for real-time online diagnosis, enabling the detection of defects in steam drums, identification of when they occur, and the cause of their occurrence.

[0132] In summary, compared with existing technologies, the core advantage of this invention lies in its complete breakthrough of the technical bottleneck of traditional discrete point monitoring, achieving a fundamental leap from data silos to global understanding. This invention provides a new method for global, accurate, and intelligent process monitoring, which can significantly improve defect detection rate and final product yield, providing strong technical support for quality control and process optimization in precision manufacturing.

[0133] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for bubble detection and traceability in OMR bonding process based on sensor data, characterized in that, Includes the following steps: S1. Sensor data acquisition: Dynamic physical field distribution data during the bonding process is acquired in real time through a sensor array that integrates multiple physical sensing units and is deployed in the bonding interface area. Deploy a flexible sensor array that integrates at least two types of distributed sensing units on its flexible substrate, so that its sensing surface can dynamically conform to the contour of the bonding interface area; Furthermore, a time synchronization acquisition mechanism is used to synchronously acquire response data from all sensing units. This time synchronization acquisition mechanism ensures strict consistency of data on a time base, thereby generating a data stream composed of continuous synchronous multidimensional sensing snapshots. The flexible sensor array further includes: an integrated composite structure comprising multiple composite sensing units, each of which integrates at least two different types of sensing functions at a single location point, thereby ensuring the inherent colocation of multidimensional sensing data in spatial sampling; S2. Dynamic Physical Field Reconstruction: Based on the dynamic physical field distribution data, a dynamic physical field characterizing the physical state of the bonding interface is reconstructed; Preprocessing: The discrete distribution data collected by the sensor array is preprocessed, including at least one operation of filtering and calibration to suppress noise in the original sensing signal and improve data quality; Continuity: The preprocessed discrete distribution data is processed using at least one spatial interpolation algorithm to generate a physical field distribution that is continuous in time and space for each sensed physical quantity; Fusion: The multiple continuous physical field distributions generated are aligned and fused at the data level to construct the dynamic physical field; Fusion further includes: integrating multiple continuous physical field distribution data into a unified mathematical model, the mathematical model being designed to characterize the intrinsic coupling relationship between the physical quantities on the bonding interface as they evolve with time and space; S3. Defect signature recognition: In the dynamic physical field, the formation event of bubble defects is detected by recognizing the defect physical signatures preset in the knowledge base; S4. Root cause diagnosis: After the bubble defect is detected, based on the quantitative characteristics of the physical signature of the defect, matching and reasoning are performed in a preset process diagnosis model to determine the root cause of the bubble defect.

2. The method for bubble detection and traceability in the OMR bonding process based on sensor data according to claim 1, characterized in that, Step S3, which involves identifying the defect physical signatures preset in the knowledge base, further includes: retrieving a set of defect physical signatures from the knowledge base, wherein each signature is defined as a standard data pattern characterizing the dynamic characteristics of a specific type of process defect; continuously scanning and comparing the real-time generated dynamic physical field using a pattern recognition algorithm to calculate the matching degree between a local region of the dynamic physical field and each standard data pattern in the defect physical signature; and when any calculated matching degree exceeds a preset threshold, the defect type associated with the standard data pattern is identified as the defect information of the current process.

3. The method for bubble detection and traceability in the OMR bonding process based on sensor data according to claim 1, characterized in that, The specific implementation of step S3, defect signature recognition, includes: analyzing the dynamic physical field using a data-driven feature extraction model to extract a set of quantized feature vectors that characterize the bonding interface process state from the spatiotemporal evolution process; inputting the quantized feature vectors into a pre-trained classification model trained based on historical process data, and having the pre-trained classification model output a judgment result regarding the bonding interface process defect; and, when a process defect is determined to exist, further inversely associating the dynamic physical field based on the judgment result to locate the spatial region in the dynamic physical field that caused the defect.

4. The method for bubble detection and traceability in the OMR bonding process based on sensor data according to claim 3, characterized in that, The quantized feature vector includes at least one feature selected from the following group: Static spatial characteristics used to describe the distribution, uniformity, and local anomalies of a single physical field in the spatial dimension at any given time; Dynamic time-domain features used to describe the evolution of the dynamic physical field over time at single and multiple spatial locations; And coupling characteristics used to describe the spatiotemporal correlation and synchronicity between two or more different physical fields.

5. The method for bubble detection and traceability in the OMR bonding process based on sensor data according to claim 1, characterized in that, The specific implementation of step S4, root cause diagnosis, includes: using the quantified features of the defect physical signature as an input vector, providing them to a preset process diagnosis model trained based on historical data, and having the process diagnosis model output a specific diagnosis result identifying the root cause through an internally established mapping relationship from features to causes.

6. The method for bubble detection and traceability in the OMR bonding process based on sensor data according to claim 5, characterized in that, The process of establishing the process diagnostic model includes: collecting historical sensor data of the process containing known root causes, and extracting the quantitative features of the physical signature of the defect from the corresponding historical sensor data for each known root cause, thereby constructing a training dataset composed of quantitative features and root cause labels; and training the process diagnostic model based on the training dataset using a supervised learning algorithm so that it learns and establishes the mapping relationship from features to causes.

Images