Method and system for detecting surface defects of a release film

Abstract

The application provides a release film surface defect detection method and system, relates to the technical field of defect detection, and the method comprises the following steps: through the non-contact multi-physical field information collection on the same detection area of the continuously running functional film to be detected on a film production line, holographic quality field original data are acquired, space-time correlation fusion is performed on corresponding real-time process parameters, the quality field data cube is input into a pre-trained defect diagnosis model, double-channel defect traceability diagnosis and performance attenuation prediction are executed in parallel, and a release film surface defect diagnosis result is generated. The technical problem that the prior art only performs simple detection on the surface defects of the release film, cannot distinguish whether the performance has defects, and leads to the reduction of release film defect detection precision is solved. Through multi-physical field synchronous detection and space-time fusion analysis, accurate defect traceability and performance attenuation prediction are realized, and the technical effects of improving the release film detection precision and performance reliability are achieved.

Description

Technical Field

[0001] This invention relates to the field of defect detection technology, and specifically to a method and system for detecting defects on the surface of release films. Background Technology

[0002] For functional materials such as quantum dot films and OLED encapsulation release films, surface defects may not be obvious in appearance, but they can significantly affect the film's key functional properties, such as electroluminescence efficiency, barrier properties, and mechanical stability, thereby reducing the reliability of the end device. However, existing visual inspection methods for release films typically only perform simple surface defect detection, lacking quantitative analysis of the impact on material functionality. This makes it impossible to determine the true impact of defects on film performance, reducing the accuracy and reliability of release film defect detection, and thus increasing the risk of release film functional failure.

[0003] Existing technologies only perform simple surface defect detection on release films, and cannot distinguish whether there are performance defects. This leads to reduced accuracy in release film defect detection and affects the reliability of release film functions. Summary of the Invention

[0004] The purpose of this application is to provide a method and system for detecting surface defects in release films, in order to solve the technical problem that the existing technology only performs simple surface defect detection on release films, which cannot distinguish whether there are performance defects, resulting in reduced detection accuracy of release film defects and affecting the functional reliability of release films.

[0005] In view of the above problems, this application provides a method and system for detecting defects on the surface of release film.

[0006] The first aspect of this application provides a method for detecting defects on the surface of a release film. The method includes: simultaneously performing non-contact multiphysics information acquisition on the same detection area of ​​a continuously moving functional film under test on a thin film production line to obtain holographic mass field raw data including three-dimensional morphological field information, spectral feature field information, and electromagnetic functional field information; spatiotemporally fusing the holographic mass field raw data with the real-time process parameters of the production line corresponding to the acquisition of the holographic mass field raw data to form a mass field data cube with spatiotemporal coordinates; inputting the mass field data cube into a pre-trained defect diagnosis model, and based on the defect diagnosis model, performing parallel dual-channel defect tracing diagnosis and performance degradation prediction to generate a release film surface defect diagnosis result.

[0007] Optionally, three-dimensional topographic field information is obtained through confocal white light interferometry scanning, which includes the depth, width, three-dimensional contour, and edge steepness information of the micro-defect; spectral feature field information is obtained through pulsed laser excitation and microscopic spectral acquisition, which includes the fluorescence lifetime decay curve and transient absorption spectrum information of the defect region at multiple characteristic excitation wavelengths; and electromagnetic functional field information is obtained through polarized terahertz time-domain spectral scanning, which includes the complex permittivity tensor information of the defect region under different polarization states.

[0008] Optionally, a known phase modulation is introduced into the reference optical path of the confocal white light interferometric scanning, and the excitation timing of the pulsed laser and the emission pulse of the polarized terahertz wave are phase-locked by the same ultrafast optical clock source.

[0009] Optionally, when acquiring the original holographic mass field data, a high-precision spatiotemporal reference signal is injected synchronously to mark the corresponding absolute position coordinates of the production line and the precise acquisition time for the original holographic mass field data of each pixel unit; real-time process parameters of the production line corresponding to the precise acquisition time and absolute position coordinates are captured in real time from the distributed process sensor network, the real-time process parameters of the production line including environmental parameters, equipment status parameters and material attribute parameters; the original holographic mass field data with high-precision spatiotemporal reference signal and the real-time process parameters of the production line are aligned and fused in multiple dimensions according to the absolute position coordinates and precise acquisition time to generate a mass field data cube aligned in four dimensions: space, time, physical field, and process parameters.

[0010] Optionally, based on the absolute position coordinates and precise acquisition time, a spatiotemporal trajectory of each pixel unit in the original holographic mass field data is established in the production line coordinate system. According to the spatiotemporal trajectory and combined with the production line process transfer model, spatiotemporal interpolation and resampling of the real-time process parameters of the production line are performed on the time series of the original holographic mass field data to obtain a spatiotemporally aligned real-time process parameter sequence of the production line. The original holographic mass field data and the spatiotemporally aligned real-time process parameter sequence of the production line are spliced ​​and channel superimposed according to the corresponding timestamps and spatial coordinates to generate the mass field data cube aligned in four dimensions: space, time, physical field, and process parameters.

[0011] Optionally, a feature decoupling encoder is used to encode the mass field data cube into a shared feature vector, a process-sensitive feature vector, and a physical field feature vector. The shared feature vector contains common information related to defects, the process-sensitive feature vector characterizes process disturbance patterns, and the physical field feature vector characterizes anomalous patterns in specific physical fields. A causal inference defect tracing module is used to receive the process-sensitive feature vector, infer based on a structural causal model, and output a causal relationship diagram and corresponding responsibility weights characterizing the causal relationship between defects and specific process deviations. A physical constraint performance prediction module is used to receive the shared feature vector and the physical field feature vector, and under the constraints of a preset material degradation physical equation, simulate future performance degradation in the region corresponding to the mass field data cube, and output a performance degradation prediction curve.

[0012] Optionally, the mass field data cube is input into the feature decoupling encoder to generate the shared feature vector, the process-sensitive feature vector, and the physical field feature vector; the process-sensitive feature vector is input into the causal inference defect tracing module to calculate and output the defect tracing diagnosis result; the shared feature vector and the physical field feature vector are input into the physical constraint performance prediction module to calculate and output the performance degradation prediction result.

[0013] Optionally, the shared feature vector and the physical field feature vector are used as the initial and boundary conditions of the physical equations; during the forward propagation of the physical constraint performance prediction module, a physical calculation unit containing the material constitutive equation and the preset service load spectrum is used to iteratively solve the performance degradation process of the local region corresponding to the mass field data cube; the output of the physical calculation unit is fused with the hidden layer features of the neural network to output the quantified performance degradation prediction result.

[0014] Optionally, the defect source diagnosis results in the release film surface defect diagnosis results are mapped to specific executable process compensation instructions. The executable process compensation instructions include at least the adjustment amount of the set point of a specific process chamber or the maintenance warning of a specific actuator. The process compensation instructions are bound to the performance degradation prediction results and pushed to the manufacturing execution unit in real time to drive the adaptive closed-loop adjustment of the corresponding production line links, and to dynamically sort and plan film products with different performance degradation prediction levels.

[0015] A second aspect of this application provides a release film surface defect detection system, comprising: a data acquisition module for simultaneously performing non-contact multi-physics field information acquisition on the same detection area of ​​a continuously moving functional film under test on a film production line, acquiring holographic mass field raw data including three-dimensional morphological field information, spectral feature field information, and electromagnetic functional field information; a data association module for spatiotemporally associating and fusing the holographic mass field raw data with the real-time process parameters of the production line corresponding to the acquisition of the holographic mass field raw data, forming a mass field data cube with spatiotemporal coordinates; and a defect diagnosis module for inputting the mass field data cube into a pre-trained defect diagnosis model, and based on the defect diagnosis model, performing dual-channel defect tracing diagnosis and performance degradation prediction in parallel to generate release film surface defect diagnosis results.

[0016] One or more technical solutions provided in this application have at least the following technical effects or advantages: The method provided in this application involves simultaneously performing non-contact multi-physics field information acquisition on the same detection area of ​​a continuously moving functional film under test on a thin film production line. This acquires holographic mass field raw data, including three-dimensional morphological field information, spectral feature field information, and electromagnetic functional field information. The holographic mass field raw data is then spatiotemporally correlated and fused with the real-time process parameters of the production line corresponding to the acquisition of the holographic mass field raw data, forming a mass field data cube with spatiotemporal coordinates. This mass field data cube is input into a pre-trained defect diagnosis model. Based on the defect diagnosis model, dual-channel defect tracing diagnosis and performance degradation prediction are performed in parallel to generate defect diagnosis results for the release film surface. This achieves the technical effect of improving the accuracy and reliability of release film detection by realizing accurate defect tracing and performance degradation prediction through synchronous multi-physics field detection and spatiotemporal fusion analysis.

[0017] The above description is merely an overview of the technical solution of this application. To better understand the technical means of this application and to facilitate its implementation according to the description, and to make the above and other objects, features, and advantages of this application more apparent, specific embodiments of this application are described below. It should be understood that the content described in this section is not intended to identify key or important features of the embodiments of this application, nor is it intended to limit the scope of this application. Other features of this application will become readily apparent through the following description. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the method for detecting surface defects in release film provided in this application.

[0020] Figure 2 This is a schematic diagram of the release film surface defect detection system provided in this application.

[0021] Figure labeling: Data acquisition module 11, data association module 12, defect diagnosis module 13. Detailed Implementation

[0022] This application provides a method and system for detecting surface defects in release films, addressing the technical problem that existing technologies only perform simple surface defect detection on release films, failing to distinguish whether performance defects exist, leading to reduced detection accuracy and affecting the functional reliability of release films. The method achieves the technical effect of accurately tracing the source of defects and predicting performance degradation through simultaneous multi-physics field detection and spatiotemporal fusion analysis, thereby improving the detection accuracy and performance reliability of release films.

[0023] The technical solutions of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. It should be understood that the present invention is not limited to the exemplary embodiments described herein. 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. It should also be noted that, for ease of description, only the parts related to the present invention are shown in the accompanying drawings, not all of them.

[0024] Example 1, as Figure 1 As shown, this application provides a method for detecting defects on the surface of a release film, the method comprising: On the thin film production line, non-contact multi-physics information acquisition is performed simultaneously on the same testing area of ​​the continuously moving functional thin films to be tested, to obtain holographic mass field raw data including three-dimensional morphology field information, spectral feature field information, and electromagnetic functional field information.

[0025] Furthermore, for the same detection area of ​​the continuously moving functional thin film under test, non-contact multiphysics information acquisition is performed simultaneously, including: acquiring three-dimensional morphological field information through confocal white light interferometry scanning, wherein the three-dimensional morphological field information includes the depth information, width information, three-dimensional contour information, and edge steepness information of micro-defects; acquiring spectral feature field information through pulsed laser excitation and microspectral acquisition, wherein the spectral feature field information includes the fluorescence lifetime decay curve and transient absorption spectrum information of the defect region at multiple characteristic excitation wavelengths; and acquiring electromagnetic functional field information through polarized terahertz time-domain spectral scanning, wherein the electromagnetic functional field information includes the complex permittivity tensor information of the defect region under different polarization states.

[0026] Furthermore, a known phase modulation is introduced into the reference optical path of the confocal white light interferometric scanning, and the excitation timing of the pulsed laser and the emission pulse of the polarized terahertz wave are phase-locked by the same ultrafast optical clock source.

[0027] Specifically, on the thin-film production line, the functional film under test, such as the release film used for OLED packaging, is continuously moved at a uniform speed, and data is simultaneously collected from the same area of ​​the functional film under test using multiple non-contact physical field detection methods. Among them, confocal white light interferometry is used to obtain the three-dimensional height data of the surface of the functional film under test by scanning the laser beam and measuring the coherent interference fringes of the reflected light. Simultaneously, a known phase modulation is introduced into the reference optical path of the confocal white light interferometry scanning. The known phase modulation is a controllable phase modulator, such as a piezoelectric ceramic driven mirror or a liquid crystal spatial light modulator. By periodically changing the optical path difference of the reference arm, a known phase shift is generated, and then a predictable phase change is superimposed on the interference signal. By demodulating the phase of the interference fringes, high-precision three-dimensional topographic information is obtained, including the depth information, width information, three-dimensional contour information, and edge steepness information of the micro-defect. Among them, the micro-defect depth information is the vertical distance between the lowest point of the scratch and the surrounding reference surface; the width information is the horizontal distance between the two edges of the scratch; the three-dimensional contour information is the complete topographic surface of the defect along the thin film and in the lateral direction; and the edge steepness information is the slope of the defect sidewall, which is obtained by calculating the ratio of depth to half width.

[0028] Simultaneously, nanosecond or picosecond pulsed lasers are used to excite the same detection area of ​​the thin film under test. A high-speed microspectrometry instrument is used to acquire the transient fluorescence response and absorption spectrum of this area, obtaining fluorescence lifetime decay curves and transient absorption spectra at multiple characteristic excitation wavelengths. The fluorescence lifetime decay curve reflects the local energy transfer efficiency of molecules or quantum dots, and can reflect micro-scratches or structural anomalies within the thin film. The transient absorption spectrum is a graph showing the change in light transmittance as a function of wavelength and time, used to reflect the deep-level density of states caused by defects. Transient fluorescence response and absorption spectra can be used to identify release films that appear intact but whose internal electronic structure has changed.

[0029] Furthermore, by scanning the same detection area of ​​the functional thin film under test using a polarized terahertz time-domain spectrometer, and adjusting the propagation characteristics of terahertz electromagnetic waves under different polarization states, the complex permittivity tensor information of the defect region is obtained, reflecting the anisotropy of the local electromagnetic response and changes in the internal structure of the material. The pulsed laser and polarized terahertz emission pulses are triggered by the same ultrafast optical clock source. The reference pulse signal output by the ultrafast optical clock simultaneously drives the laser emitter and the terahertz emitter through a delay line or electronic trigger module, aligning their pulse times to avoid data misalignment caused by thin film motion and vibration, and ensuring the accuracy of multi-physics field information time synchronization. The acquired information is then integrated to form holographic mass field raw data including three-dimensional morphological field information, spectral feature field information, and electromagnetic functional field information.

[0030] For example, on an OLED encapsulation release film production line, a continuously moving section of the film surface was inspected. Using confocal white light interferometry, a micro-scratch approximately 2.1 μm wide, 85 nm deep, and with an edge steepness of 42 nm / μm was detected at a lateral position (X = 152.3 mm) and a vertical position (Y = 880.6 mm). Pulsed laser excitation and microspectroscopy were used to collect data from the same area. The fluorescence lifetime decay curve showed that the fluorescence lifetime at 625 nm wavelength decreased from 12.3 ns in the normal region to 6.7 ns, with a transient absorption peak shift of 5 nm. Polarized terahertz time-domain spectroscopy revealed that the real part of the complex permittivity decreased from 3.12 in the normal region to 2.98, while the imaginary part increased from 0.08 to 0.17. By integrating the collected data, complete raw data of the holographic mass field were formed.

[0031] By synchronously acquiring multi-physics information and performing precise spatiotemporal alignment, a more reliable and accurate original high-dimensional data foundation is provided for subsequent processing methods. This makes the defects on the release film surface not only visible but also allows for the quantification of their potential impact on the film's performance energy, thereby improving the comprehensiveness and accuracy of release film defect detection.

[0032] The original holographic mass field data is spatiotemporally correlated and fused with the real-time process parameters of the production line corresponding to the acquisition of the original holographic mass field data to form a mass field data cube with spatiotemporal coordinates.

[0033] Furthermore, the original holographic mass field data is spatiotemporally correlated and fused with the real-time production line process parameters corresponding to the acquisition of the original holographic mass field data to form a mass field data cube with spatiotemporal coordinates. This includes: simultaneously injecting a high-precision spatiotemporal reference signal when acquiring the original holographic mass field data to mark the corresponding absolute position coordinates and precise acquisition time of the original holographic mass field data for each pixel unit; capturing the real-time production line process parameters corresponding to the precise acquisition time and absolute position coordinates from the distributed process sensor network in real time, wherein the real-time production line process parameters include environmental parameters, equipment status parameters, and material attribute parameters; and aligning and fusing the original holographic mass field data with the high-precision spatiotemporal reference signal and the real-time production line process parameters according to the absolute position coordinates and precise acquisition time to generate a mass field data cube aligned in four dimensions: space, time, physical field, and process parameters.

[0034] Specifically, while acquiring the raw holographic mass field data of the functional film under test, a high-precision spatiotemporal reference signal is synchronously injected into the raw holographic mass field data of each pixel unit through a spatiotemporal reference unit. This signal is used to mark the corresponding absolute position coordinates of the production line and the precise acquisition time. The spatiotemporal reference unit can be based on the IEEE1588 precision time protocol synchronization clock of the PCIe time card, combined with an incremental encoder or grating ruler installed on the main drive roller of the production line. The incremental encoder or grating ruler provides the absolute position coordinates of the production line, and the time protocol synchronization clock ensures the accuracy of the acquisition time.

[0035] A distributed process sensor network pre-deployed on the production line collects real-time process parameters corresponding to spatiotemporal reference signals. This distributed process sensor network includes environmental sensors, equipment detection sensors, and material detection equipment. Environmental sensors include, but are not limited to, online temperature and humidity sensors, air pressure sensors, and cleanliness particle counters, used to collect environmental parameters of the production workshop, such as temperature, humidity, air pressure, and air cleanliness. Equipment detection sensors include, but are not limited to, encoders and torque sensors, thermocouples, and tension sensors, used to collect equipment-level status parameters such as the speed and torque of the coating roller servo motor, and the real-time temperature and tension of the drying chamber. Material detection equipment includes, but is not limited to, online viscometers, near-infrared online analyzers, and RFID or barcode scanners, used to collect material-level attribute parameters such as the batch number of the base film, solution concentration, and coating liquid-solid content.

[0036] Based on the absolute position coordinates of each pixel unit and the precise acquisition time, the original holographic mass field data with high-precision spatiotemporal reference signals is aligned and fused with the real-time process parameters of the production line in multiple dimensions to generate a mass field data cube aligned in four dimensions: space, time, physical field, and process parameters. This provides a comprehensive data foundation for defect analysis and performance prediction of release films.

[0037] By accurately calibrating the spatiotemporal data and aligning the process data, the original data of the holographic mass field is fused with the real-time process parameters of the production line to form a traceable high-dimensional mass field data cube. This not only preserves the spatial and temporal information of defects but also associates them with process parameters, providing a reliable foundation for subsequent causal analysis, performance degradation prediction, and intelligent process closed-loop control.

[0038] Furthermore, the holographic mass field raw data with high-precision spatiotemporal reference signals and the real-time process parameters of the production line are aligned and fused using multi-dimensional heterogeneous data based on the absolute position coordinates and precise acquisition time. This includes: establishing the spatiotemporal trajectory of each pixel unit in the holographic mass field raw data in the production line coordinate system based on the absolute position coordinates and precise acquisition time; performing spatiotemporal interpolation and resampling of the real-time process parameters of the production line on the time series of the holographic mass field raw data according to the spatiotemporal trajectory and the production line process transfer model, to obtain a spatiotemporally aligned real-time process parameter sequence; and splicing and channel overlaying of the holographic mass field raw data and the spatiotemporally aligned real-time process parameter sequence according to the corresponding timestamps and spatial coordinates to generate a mass field data cube aligned in four dimensions: space, time, physical field, and process parameters.

[0039] Specifically, based on the absolute position coordinates and precise acquisition time of each pixel unit in the original holographic mass field data, a corresponding spatiotemporal trajectory is established in the production line coordinate system. First, the marked absolute position coordinates and precise acquisition time of each pixel unit are used as anchor points for the trajectory. Then, a mapping relationship is established between the film conveying speed and the production line direction coordinate system, converting the position of each pixel unit on the film belt into global coordinates (x, y) of the production line. Finally, combined with the corresponding precise acquisition time, a time series is formed, generating the trajectory of each pixel point in three-dimensional space changing with time, that is, the spatiotemporal trajectory of each pixel unit in the original holographic mass field data. This represents the change of the spatial position of each pixel during the continuous movement of the film belt, ensuring precise matching between the physical field data and the production line state.

[0040] A process transfer model for the production line is determined. This model is a pre-calibrated physical or empirical model for each production line, describing the dynamic impact and time delay of parameters at each stage on the local properties of the thin film. For example, empirical modeling is used, and the response relationship and time lag between changes in process parameters and local performance indicators are extracted through statistical analysis of historical production data. For instance, regression analysis is used to establish hysteresis functions for temperature, humidity, and tension fluctuations on thin film thickness and surface defects, resulting in a process transfer model for each production line. This model can predict the dynamic response of the local physical properties of the thin film under given changes in production line parameters.

[0041] Based on the spatiotemporal trajectory of each pixel unit, its spatial position and acquisition time are determined in the production line coordinate system. Then, according to the production line process transfer model, the actual impact of each process parameter on the pixel unit at the acquisition time is calculated. Interpolation and resampling are then performed on the time series of the original holographic mass field data. Real-time production line process parameters such as temperature, tension, and humidity, which are only measured at the sampling points, are extrapolated to the acquisition time point corresponding to each pixel unit using linear interpolation. This results in a sequence of real-time production line process parameters that are spatiotemporally aligned with the original holographic mass field data of each pixel unit. This ensures that the data of each pixel unit not only includes the original holographic mass field data but also accurately reflects its real-time process parameters during production, guaranteeing the precise consistency of the multidimensional mass field data cube in spatial, temporal, and process parameter dimensions. Each data point in the obtained original holographic mass field data and real-time production line process parameters is normalized using the min-max normalization method to eliminate dimensional differences and improve data fusion accuracy.

[0042] The normalized holographic mass field raw data and the spatiotemporally aligned real-time process parameter sequence of the production line are fused together. First, the data is stitched together according to the corresponding timestamps and spatial coordinates, and the holographic mass field raw data and the real-time process parameters belonging to the same spatiotemporal point, i.e., the same pixel unit and the same acquisition time, are stitched together and integrated. Then, the stitched multidimensional data is channel-overlayed, and the original physical field feature channels and the newly added process parameter channels are merged in the feature dimensions to form a four-dimensional aligned mass field data cube. The four dimensions of the mass field data cube are space, time, physical field, and process parameters. The spatial dimension represents the position of the thin film surface, and each pixel unit corresponds to a specific physical coordinate. The temporal dimension represents the acquisition time, ensuring that the data of each pixel unit is traceable during the continuous movement of the thin film. The physical field includes multi-physical field data such as three-dimensional morphology, spectral characteristics, and electromagnetic functions. The process parameters include environmental, equipment, and material level parameters, reflecting the production line status and the thin film formation process.

[0043] For example, during the data acquisition process, each pixel is marked with its absolute position coordinates and acquisition time. For instance, at x=152.3mm, y=880.6mm, t=12.345s, the corresponding production line process parameters are: ambient temperature 23.5℃, humidity 44.8%, coating machine speed 1.21m / min, tension 2.48N, and raw material concentration 5.15wt%. Through spatiotemporal interpolation and resampling, each pixel obtains accurate process parameter values, which are then stitched together with 3D morphology, spectral characteristics, and electromagnetic function data to generate a comprehensive and reliable mass field data cube.

[0044] By strictly aligning physical field information with process parameters in space and time, a mass field data cube is constructed to achieve the fusion of multi-source heterogeneous data. This provides a reliable and comprehensive data foundation for the analysis of surface defects, performance prediction, and closed-loop control of release films, thereby improving the accuracy of release film defect diagnosis and tracing, as well as the reliability and accuracy of release film performance prediction.

[0045] The mass field data cube is input into a pre-trained defect diagnosis model. Based on the defect diagnosis model, dual-channel defect tracing diagnosis and performance degradation prediction are performed in parallel to generate defect diagnosis results for the release film surface.

[0046] Furthermore, the pre-trained defect diagnosis model includes: a feature decoupling encoder, used to encode the mass field data cube into a shared feature vector, a process-sensitive feature vector, and a physical field feature vector, wherein the shared feature vector contains common information related to defects, the process-sensitive feature vector is used to characterize process disturbance patterns, and the physical field feature vector is used to characterize abnormal patterns of specific physical fields; a causal inference defect tracing module, used to receive the process-sensitive feature vector, infer based on a structural causal model, and output a causal relationship graph characterizing the causal relationship between defects and specific process deviations, along with corresponding responsibility weights; and a physical constraint performance prediction module, used to receive the shared feature vector and the physical field feature vector, and under the constraints of a preset material degradation physical equation, simulate future performance degradation in the region corresponding to the mass field data cube, and output a performance degradation prediction curve.

[0047] Specifically, the defect diagnosis model is a multi-task deep learning architecture that integrates causal inference and physical constraints. It consists of three working sub-modules, including a feature decoupling encoder, a causal inference defect tracing module, and a physical constraint performance prediction module.

[0048] The feature decoupling encoder is used to encode a four-dimensional mass field data cube into a shared feature vector, a process-sensitive feature vector, and a physical field feature vector. The feature decoupling encoder is a hybrid structure of multi-branch convolutional and fully connected layers, specifically as follows: The physical field branch uses a 5-layer convolutional neural network to extract spatial and local physical anomaly features. The convolutional layers are configured with 32, 64, 128, 128, and 256 kernels in sequence, with each kernel size of 3×3 and a stride of 1. Every two layers, a 2×2 max-pooling layer is added to reduce spatial dimensionality and increase the receptive field. The activation function of the convolutional layers uses ReLU to enhance nonlinear representation capabilities. The convolutional features are flattened and input into the fully connected layer for further mapping to form a 64-dimensional physical field feature vector.

[0049] The process parameter branch processes multidimensional parameter sequence data related to the thin film production process, such as coating machine temperature, tension fluctuations, speed, and humidity. The process parameter channel input is a spatiotemporally aligned sequence of process parameters. The channel structure is a 3-layer Temporal Convolutional Network (TCN), with each layer having a kernel size of 3 and dilation coefficients increasing by 1, 2, and 4, respectively. The number of convolutional channels is 64, 128, and 128, respectively, used to capture the local and long-range dependencies of process disturbances over time. The channel output generates a 64-dimensional process-sensitive feature vector through linear mapping, used to characterize the response pattern of local defects to process disturbances, and simultaneously provides input to the causal inference module.

[0050] A shared feature branch is used to fuse feature information from the physical field channel and the process parameter channel to extract common patterns related to defects. The physical field convolutional output features and the process parameter channel output features are concatenated by channel and then input into a two-layer fully connected network (256 nodes per layer, ReLU activation) for feature fusion and compression. The fused output is linearly mapped to generate a 128-dimensional shared feature vector. This shared feature vector characterizes the general patterns of thin film surface defects, such as the spatial, temporal, and physical field characteristics of micro-scratches, pits, or bubbles. This provides a stable, high-dimensional input to the physical constraint performance prediction module while preserving the interpretability of defect information.

[0051] During the training of the feature-decoupled encoder, a multi-task loss function is adopted, including defect recognition cross-entropy loss for defect classification, mean squared error of process-sensitive feature regression for predicting process disturbance patterns, and mean squared error of physical field anomaly regression for characterizing physical anomaly patterns. The overall loss function is a weighted sum, and the initial weights can be set based on actual needs and experience. The Adam optimizer is selected as the optimization algorithm, and the initial learning rate is set to 1×10⁻⁶. -3The cosine annealing strategy is employed to improve convergence stability, with batch sizes set to 16-32. The convergence criteria are a validation set defect identification accuracy exceeding 95% or no significant decrease in loss over 10 consecutive rounds, and a performance prediction mean square error below 5%. Based on this feature decoupling encoder, effective separation and fusion of physical fields, process parameters, and common defect features can be achieved, providing high-dimensional, interpretable, and spatiotemporally aligned feature inputs for causal inference defect tracing and physical constraint performance prediction, thereby improving defect diagnosis accuracy and performance prediction reliability.

[0052] The causal inference defect tracing module employs a structural causal model, receiving process-sensitive feature vectors from the feature decoupling encoder for causal source analysis of defects. First, for the release film production line and product type, domain knowledge combined with offline causal analysis algorithms, such as the PC algorithm, is used to learn causal hypotheses between process parameters from historical data, constructing a structural causal graph. The nodes of this graph include process parameters such as tension, temperature, coating speed, humidity, solution concentration, and coating liquid-solid content. The directed edges in the graph represent prior causal hypotheses; for example, coating speed affects actual dynamic tension, and tension affects the probability of micro-scratches forming.

[0053] A causal inference network is trained based on a structural causal graph. This network employs a graph neural network, with process-sensitive feature vectors as inputs. Each feature vector corresponds to a process parameter node in the structural causal graph. First, the input feature vectors are mapped to a 128-dimensional high-dimensional space through a node embedding layer, providing sufficient representation for subsequent information transmission. The node features output from the embedding layer serve as input to graph convolutional layers. Each graph convolutional layer achieves information transmission by aggregating features from neighboring nodes and its own features, updating the representation of each node. The causal inference network consists of three graph convolutional layers, with output dimensions of 128, 128, and 64 respectively. ReLU is used as the activation function for all layers, and edge weights are weighted based on prior causal assumptions in the causal graph. The node features after each convolutional layer are fused with the features from the previous layer through skip connections to avoid gradient vanishing and enhance information retention. After the graph convolutional layers, the node features are mapped to predicted causal association weights (i.e., corresponding responsibility weights) through two fully connected layers. Simultaneously, a causal association graph is calculated based on the predicted edge weights and node features.

[0054] The physical constraint performance prediction module receives shared feature vectors and physical field feature vectors from the feature decoupling encoder. These are used to simulate future performance degradation in the current detection area, i.e., the microscopic thin film region corresponding to the mass field data cube, and output a performance degradation prediction curve. For the functional failure modes of the release membrane, the material constitutive equation and degradation kinetic equation are pre-selected or derived as constraints on the material degradation physical equation. For example, for barrier performance degradation, Fick's second law of diffusion combined with the Arrhenius equation is used to describe the accelerated permeation of water vapor around the defect: ∂C / ∂t=D(T)∇2C, D(T)=D0exp(−Ea / kT), where D(T) is the temperature-dependent diffusion coefficient, C is the water vapor concentration, t is time, D0 is the membrane material reference diffusion coefficient, Ea is the activation energy (i.e., the preset membrane material parameter), k is the Boltzmann constant, and T is the temperature.

[0055] The shared feature vector and the physical field feature vector are concatenated and mapped to the initial and boundary conditions of the physical equation through a fully connected network. The preset material degradation physical equation is input, and the performance degradation process of the micro-region corresponding to the mass field data cube is iteratively solved through finite difference. The performance degradation prediction curve is output, which reflects the possible future decline trend of optical, electrical and mechanical properties of the defect region.

[0056] By performing feature decoupling, causal tracing, and physical constraint prediction on the mass field data cube, we can improve the performance degradation from the visibility of defects to the explanation of defect causes and the predictability of performance degradation, thereby enhancing the accuracy of release film defect diagnosis, process optimization capabilities, and the predictability of thin film product quality.

[0057] Furthermore, the mass field data cube is input into a pre-trained defect diagnosis model. Based on the defect diagnosis model, dual-channel defect tracing diagnosis and performance degradation prediction are performed in parallel, including: inputting the mass field data cube into the feature decoupling encoder to generate the shared feature vector, process-sensitive feature vector, and physical field feature vector; inputting the process-sensitive feature vector into the causal inference defect tracing module to calculate and output the defect tracing diagnosis result; and inputting the shared feature vector and physical field feature vector into the physical constraint performance prediction module to calculate and output the performance degradation prediction result.

[0058] Furthermore, the shared feature vector and the physical field feature vector are input into the physical constraint performance prediction module to calculate and output the performance degradation prediction result, including: using the shared feature vector and the physical field feature vector as the initial and boundary conditions of the physical equation; during the forward propagation process of the physical constraint performance prediction module, using a physical calculation unit containing the material constitutive equation and the preset service load spectrum to iteratively solve the performance degradation process of the local region corresponding to the mass field data cube; and fusing the output result of the physical calculation unit with the hidden layer features of the neural network to output the quantified performance degradation prediction result.

[0059] Specifically, the mass field data cube is input into a pre-trained defect diagnosis model, and dual-channel defect tracing diagnosis and performance degradation prediction are executed in parallel. First, the feature decoupling encoder in the defect diagnosis model receives the mass field data cube and generates a shared feature vector, a process-sensitive feature vector, and a physical field feature vector. Then, the process-sensitive feature vector is input into the causal inference defect tracing module, and the shared feature vector and physical field feature vector are input into the physical constraint performance prediction module to execute dual-channel defect tracing diagnosis and performance degradation prediction in parallel.

[0060] The process-sensitive feature vector is input into the causal inference defect tracing module. Then, the graph neural network is used for inference to output the causal weight of each process parameter on the formation of defects and the causal relationship graph that characterizes the causal relationship between defects and specific process deviations, thus forming the defect tracing diagnosis result.

[0061] The shared feature vector and the physical field feature vector are concatenated and input into the physical constraint performance prediction module. This is transformed into initial and boundary conditions for physical equations through a fully connected network, describing the initial state of the defect region and external influences, such as film thickness distribution, micro-scratch depth, local stress state, and water vapor or oxygen concentration gradients. During the forward propagation of the physical constraint performance prediction module, the built-in physical calculation unit iteratively solves the local region. This physical calculation unit includes the material constitutive equation and a preset service load spectrum. For example, the barrier properties of the thin film can be described by Fick's second law of diffusion and the temperature-dependent diffusion coefficient, while optical or electrical degradation can be expressed by a material refractive index evolution function with respect to defects or a dynamic model of the dielectric constant with respect to defects.

[0062] During the iterative solution process, the physical computing unit discretizes the degradation process of the local thin film in both spatial and temporal steps, outputting a predicted performance state sequence. Then, the output of the physical computing unit, i.e., the predicted performance state sequence, is fused with the hidden layer features of a neural network through a nonlinear mapping of shared feature vectors. Specifically, the shared feature vectors are mapped to a time dimension consistent with the time steps of the performance state sequence via a forward fully connected network, resulting in time-series-aligned hidden layer features. Then, through weighted concatenation, the predicted value of the physical computing unit at each time step and the corresponding hidden layer feature are fused element-wise or channel-wise to form an enhanced feature. The enhanced feature is calculated as: Enhanced Feature = α × Predicted value of physical computing unit + (1 − α) × Corresponding hidden layer feature, where α is the fusion weight, adaptively learned from the training data, with a value ranging from 0 to 1. The fused enhanced feature is further mapped through a fully connected network to the final performance degradation prediction result, reflecting the trend of declining optical, electrical, and mechanical properties of the defective region during its future service life.

[0063] By processing these two channels in parallel, the results of defect cause analysis and performance prediction can be obtained simultaneously. This enables the prediction of the impact of defects on the performance of the release film while detecting defects, and finally outputs the defect diagnosis results on the release film surface, including defect location, cause, responsibility weight, and future performance degradation curve.

[0064] By using parallel dual-channel analysis, we can not only quickly and accurately detect and trace the causes of defects on the release film surface, but also quantitatively predict the impact of defects on film performance. This provides an interpretable and operable scientific basis for real-time process adjustment, product grading, and quality control in the production line, thereby improving the intelligence level of the release film production process and the reliability of functional materials.

[0065] Furthermore, the method also includes: mapping the defect source diagnosis results in the defect diagnosis results of the release film surface to specific executable process compensation instructions, wherein the executable process compensation instructions include at least the set point adjustment amount for a specific process chamber or the maintenance warning for a specific actuator; binding the process compensation instructions with the performance degradation prediction results, pushing them to the manufacturing execution unit in real time, driving the adaptive closed-loop adjustment of the corresponding production line links, and dynamically sorting and planning film products with different performance degradation prediction levels.

[0066] Specifically, defect source diagnosis is extracted from the defect diagnosis results of the release film surface, and the process responsibility weight and abnormal parameter characteristics are analyzed. For example, the formation of a micro-scratch is mainly attributed to coating tension fluctuations, insufficient drying temperature, or high material solid content. Based on a pre-established process mapping rule base, abnormal process parameter characteristics are mapped to specific operating parameters or maintenance actions, generating executable process compensation instructions. The process mapping rule base is established through a combination of offline data analysis and expert experience. First, historical production data is collected, including process parameters and film defect records, as well as the performance degradation corresponding to defects. Then, causal analysis and statistical regression methods are used to identify the degree of influence of key process parameters on defect formation, and a mapping relationship is established between different parameter variation ranges and defect types. Combined with process expert experience and equipment operating specifications, rules are supplemented for special abnormal situations and maintenance needs, forming the final process mapping rule base. This base is used to automatically map abnormal process parameter characteristics to executable operations based on real-time detected process anomalies and defect source tracing results, including adjusting the setpoint of a specific process chamber or providing maintenance warnings for the actuator. For example, an executable process compensation instruction could increase the coating machine tension by 0.5 N / m. 2 The dryer temperature is increased by 2°C, or maintenance warnings are issued to specific actuators, such as reminders for coating roller cleaning or pump and valve maintenance, to ensure that the causal analysis results can be directly converted into actionable measures for the production line, thus achieving a closed-loop connection from diagnosis to control.

[0067] Based on performance degradation prediction results, film products are divided into high-risk, medium-risk, and low-risk zones. For example, the high-risk zone corresponds to films with a barrier performance degradation exceeding 10%, the medium-risk zone corresponds to films with a barrier performance degradation between 5% and 10%, and the low-risk zone corresponds to films with a barrier performance degradation of less than 5%. Process compensation instructions are linked to performance degradation prediction results, ensuring that each instruction includes the corresponding performance prediction level information. This information is then sent in real-time to the Manufacturing Execution Unit (MES) via a high-speed data interface. The MES automatically adjusts production line parameters, such as tension, temperature, and humidity control, or triggers maintenance actions based on the performance level and process compensation instructions, driving adaptive closed-loop adjustments in the corresponding production line stages. Simultaneously, based on the performance prediction level, film products with different performance degradation prediction levels are dynamically sorted and planned. Low-performance films are directly sent to the mainstream production line or packaged for shipment, medium-performance films are sent to reprocessing or process adjustment stages, and low-performance films are sent to rework, downgraded use, or substandard product channels.

[0068] By transforming defect detection, cause analysis, and performance prediction results into executable intelligent control strategies and product sorting schemes, closed-loop intelligent manufacturing is achieved, moving from passive detection to proactive regulation and optimization. Simultaneously, through dynamic sorting strategies, not only is production line optimization realized, allowing various release films to flow to different processing paths rationally based on predicted performance, but also product hierarchical management is achieved, ensuring the functional reliability of high-value release films, improving production efficiency, and reducing scrap losses caused by functional failures.

[0069] Example 2, based on the same inventive concept as the release film surface defect detection method in the foregoing examples, such as... Figure 2 As shown, this application provides a release film surface defect detection system, wherein the release film surface defect detection system includes: The data acquisition module 11 is used to simultaneously perform non-contact multi-physics field information acquisition on the same detection area of ​​a continuously moving functional film under test on the thin film production line, and acquire holographic mass field raw data including three-dimensional morphology field information, spectral feature field information, and electromagnetic functional field information; the data association module 12 is used to perform spatiotemporal association and fusion of the holographic mass field raw data with the real-time process parameters of the production line corresponding to the acquisition of the holographic mass field raw data, and form a mass field data cube with spatiotemporal coordinates; the defect diagnosis module 13 is used to input the mass field data cube into a pre-trained defect diagnosis model, and based on the defect diagnosis model, to perform dual-channel defect tracing diagnosis and performance degradation prediction in parallel, and generate defect diagnosis results for the release film surface.

[0070] Furthermore, the data acquisition module 11 is also used to: acquire three-dimensional topographic field information through confocal white light interferometry scanning, wherein the three-dimensional topographic field information includes the depth information, width information, three-dimensional contour information and edge steepness information of the micro-defect; acquire spectral feature field information through pulsed laser excitation and microscopic spectral acquisition, wherein the spectral feature field information includes the fluorescence lifetime decay curve and transient absorption spectrum information of the defect region under multiple characteristic excitation wavelengths; and acquire electromagnetic functional field information through polarized terahertz time-domain spectral scanning, wherein the electromagnetic functional field information includes the complex permittivity tensor information of the defect region under different polarization states.

[0071] Furthermore, the data acquisition module 11 is also used to: introduce a known phase modulation into the reference optical path of the confocal white light interferometric scanning, and the excitation timing of the pulsed laser and the emission pulse of the polarized terahertz wave are phase-locked by the same ultrafast optical clock source.

[0072] Furthermore, the data association module 12 is also used to: synchronously inject a high-precision spatiotemporal reference signal when acquiring the original holographic mass field data, and mark the corresponding production line absolute position coordinates and precise acquisition time for the original holographic mass field data of each pixel unit; capture the real-time process parameters of the production line corresponding to the precise acquisition time and absolute position coordinates from the distributed process sensor network in real time, the real-time process parameters of the production line including environmental parameters, equipment status parameters and material attribute parameters; and perform multi-dimensional heterogeneous data alignment and fusion of the original holographic mass field data with the high-precision spatiotemporal reference signal and the real-time process parameters of the production line according to the absolute position coordinates and precise acquisition time to generate a mass field data cube aligned in four dimensions: space, time, physical field, and process parameters.

[0073] Furthermore, the data association module 12 is also used to: establish the spatiotemporal trajectory of each pixel unit in the original holographic mass field data in the production line coordinate system based on the absolute position coordinates and the precise acquisition time; according to the spatiotemporal trajectory, combined with the production line process transfer model, perform spatiotemporal interpolation and resampling on the time series of the original holographic mass field data to obtain a spatiotemporally aligned real-time production line process parameter sequence; and perform data splicing and channel superposition on the original holographic mass field data and the spatiotemporally aligned real-time production line process parameter sequence according to the corresponding timestamps and spatial coordinates to generate the mass field data cube aligned in four dimensions: space, time, physical field, and process parameters.

[0074] Furthermore, the defect diagnosis module 13 is also used for: a feature decoupling encoder, used to encode the mass field data cube into a shared feature vector, a process-sensitive feature vector, and a physical field feature vector, wherein the shared feature vector contains common information related to defects, the process-sensitive feature vector is used to characterize process disturbance patterns, and the physical field feature vector is used to characterize abnormal patterns of specific physical fields; a causal inference defect tracing module, used to receive the process-sensitive feature vector, and based on structural causal model reasoning, output a causal relationship diagram characterizing the causal relationship between defects and specific process deviations and corresponding responsibility weights; and a physical constraint performance prediction module, used to receive the shared feature vector and the physical field feature vector, and under the constraints of a preset material degradation physical equation, perform future performance degradation simulation on the region corresponding to the mass field data cube, and output a performance degradation prediction curve.

[0075] Furthermore, the defect diagnosis module 13 is also used to: input the mass field data cube into the feature decoupling encoder to generate the shared feature vector, the process-sensitive feature vector, and the physical field feature vector; input the process-sensitive feature vector into the causal inference defect tracing module to calculate and output the defect tracing diagnosis result; and input the shared feature vector and the physical field feature vector into the physical constraint performance prediction module to calculate and output the performance degradation prediction result.

[0076] Furthermore, the defect diagnosis module 13 is also used to: use the shared feature vector and the physical field feature vector as the initial and boundary conditions of the physical equation; during the forward propagation process of the physical constraint performance prediction module, use a physical calculation unit containing the material constitutive equation and the preset service load spectrum to iteratively solve the performance degradation process of the local region corresponding to the mass field data cube; fuse the output of the physical calculation unit with the hidden layer features of the neural network to output the quantified performance degradation prediction result.

[0077] Furthermore, the system is also used to: map the defect tracing diagnosis results in the defect diagnosis results of the release film surface to specific executable process compensation instructions, wherein the executable process compensation instructions include at least the adjustment amount of the set point of a specific process chamber or the maintenance warning of a specific actuator; bind the process compensation instructions with the performance degradation prediction results, push them to the manufacturing execution unit in real time, drive the adaptive closed-loop adjustment of the corresponding production line links, and dynamically sort and plan film products with different performance degradation prediction levels.

[0078] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The release film surface defect detection method and specific examples in the aforementioned embodiment one are also applicable to the release film surface defect detection system of this embodiment. Through the foregoing detailed description of the release film surface defect detection method, those skilled in the art can clearly understand the release film surface defect detection system of this embodiment. Therefore, for the sake of brevity, it will not be described in detail here.

[0079] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0080] Obviously, those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A method for detecting defects on the surface of release film, characterized in that, The method includes: On the thin film production line, non-contact multi-physics field information is simultaneously collected in the same testing area of ​​the continuously moving functional thin film to be tested, and holographic mass field raw data including three-dimensional morphology field information, spectral feature field information, and electromagnetic functional field information are obtained. The original holographic mass field data is spatiotemporally correlated and fused with the real-time process parameters of the production line corresponding to the time of acquisition of the original holographic mass field data to form a mass field data cube with spatiotemporal coordinates. The mass field data cube is input into a pre-trained defect diagnosis model. Based on the defect diagnosis model, dual-channel defect tracing diagnosis and performance degradation prediction are performed in parallel to generate defect diagnosis results for the release film surface.

2. The method for detecting defects on the surface of release film as described in claim 1, characterized in that, On the thin film production line, non-contact multiphysics information acquisition is performed simultaneously in the same testing area of ​​continuously moving functional thin films under test, including: Three-dimensional topographic field information is obtained by confocal white light interferometry scanning. The three-dimensional topographic field information includes the depth information, width information, three-dimensional contour information and edge steepness information of micro-defects. Spectral feature field information is obtained by pulsed laser excitation and microscopic spectral acquisition. The spectral feature field information includes fluorescence lifetime decay curves and transient absorption spectra of the defect region at multiple characteristic excitation wavelengths. Electromagnetic functional field information is obtained by polarization terahertz time-domain spectral scanning. The electromagnetic functional field information includes the complex permittivity tensor information of the defect region under different polarization states.

3. The method for detecting defects on the surface of release film as described in claim 2, characterized in that, A known phase modulation is introduced into the reference optical path of the confocal white light interferometry scan, and the excitation timing of the pulsed laser and the emission pulse of the polarized terahertz wave are phase-locked by the same ultrafast optical clock source.

4. The method for detecting surface defects of release film as described in claim 1, characterized in that, The original holographic mass field data is spatiotemporally correlated and fused with the real-time production line process parameters corresponding to the acquisition of the original holographic mass field data to form a mass field data cube with spatiotemporal coordinates, including: When acquiring the original data of the holographic mass field, a high-precision spatiotemporal reference signal is injected simultaneously to mark the corresponding absolute position coordinates of the production line and the precise acquisition time for the original data of the holographic mass field of each pixel unit; Real-time process parameters of the production line corresponding to the precise acquisition time and absolute position coordinates are captured from the distributed process sensor network. The real-time process parameters of the production line include environmental parameters, equipment status parameters and material attribute parameters. The holographic mass field raw data with high-precision spatiotemporal reference signal and the real-time process parameters of the production line are aligned and fused in a multidimensional heterogeneous manner based on the absolute position coordinates and the precise acquisition time to generate a mass field data cube aligned in four dimensions: space, time, physical field, and process parameters.

5. The method for detecting defects on the surface of release film as described in claim 4, characterized in that, The holographic mass field raw data with high-precision spatiotemporal reference signal is aligned and fused with the real-time process parameters of the production line based on the absolute position coordinates and precise acquisition time, including: Based on the absolute position coordinates and the precise acquisition time, the spatiotemporal trajectory of each pixel unit in the original holographic mass field data is established in the production line coordinate system. Based on the spatiotemporal trajectory and combined with the production line process transfer model, spatiotemporal interpolation and resampling are performed on the time series of the original data of the holographic mass field to obtain a spatiotemporally aligned sequence of real-time process parameters of the production line. The original holographic mass field data is spliced ​​and channel-overlaid with the spatiotemporally aligned real-time process parameter sequence of the production line according to the corresponding timestamps and spatial coordinates, generating the mass field data cube aligned in four dimensions: space, time, physical field, and process parameters.

6. The method for detecting surface defects of release film as described in claim 1, characterized in that, The pre-trained defect diagnosis model includes: A feature decoupling encoder is used to encode the mass field data cube into a shared feature vector, a process-sensitive feature vector, and a physical field feature vector. The shared feature vector contains common information related to defects, the process-sensitive feature vector is used to characterize process disturbance patterns, and the physical field feature vector is used to characterize abnormal patterns of specific physical fields. The causal inference defect tracing module is used to receive the process-sensitive feature vector, reason based on the structural causal model, and output a causal relationship diagram and corresponding responsibility weights that characterize the causal relationship between defects and specific process deviations. The physical constraint performance prediction module is used to receive the shared feature vector and the physical field feature vector, and under the constraints of the preset material degradation physical equation, to simulate the future performance degradation of the region corresponding to the mass field data cube and output the performance degradation prediction curve.

7. The method for detecting surface defects of release film as described in claim 6, characterized in that, The mass field data cube is input into a pre-trained defect diagnosis model. Based on the defect diagnosis model, dual-channel defect tracing diagnosis and performance degradation prediction are performed in parallel, including: The mass field data cube is input into the feature decoupling encoder to generate the shared feature vector, the process-sensitive feature vector, and the physical field feature vector. The process-sensitive feature vector is input into the causal inference defect tracing module to calculate and output the defect tracing diagnosis result; The shared feature vector and the physical field feature vector are input into the physical constraint performance prediction module to calculate and output the performance degradation prediction result.

8. The method for detecting defects on the surface of a release film as described in claim 7, characterized in that, The shared feature vector and the physical field feature vector are input into the physical constraint performance prediction module to calculate and output the performance degradation prediction result, including: The shared eigenvector and the physical field eigenvector are used as the initial and boundary conditions of the physical equations; During the forward propagation of the physical constraint performance prediction module, a physical calculation unit containing the material constitutive equation and the preset service load spectrum is used to iteratively solve the performance degradation process of the local region corresponding to the mass field data cube. The output of the physical computing unit is fused with the hidden layer features of the neural network to output the quantified performance degradation prediction result.

9. The method for detecting defects on the surface of a release film as described in claim 8, characterized in that, The method further includes: The defect source diagnosis results in the defect diagnosis results of the release film surface are mapped to specific executable process compensation instructions. The executable process compensation instructions include at least the set point adjustment amount of a specific process chamber or the maintenance warning of a specific actuator. The process compensation command is bound to the performance degradation prediction result and pushed to the manufacturing execution unit in real time to drive the adaptive closed-loop adjustment of the corresponding production line links, and to dynamically sort and plan thin film products with different performance degradation prediction levels.

10. A release film surface defect detection system, characterized in that, The step of implementing the release film surface defect detection method according to any one of claims 1 to 9, wherein the release film surface defect detection system comprises: The data acquisition module is used to simultaneously perform non-contact multi-physics field information acquisition on the same detection area of ​​continuously moving functional films under test on the thin film production line, and acquire holographic mass field raw data including three-dimensional morphology field information, spectral feature field information, and electromagnetic functional field information. The data association module is used to perform spatiotemporal association and fusion of the original holographic mass field data with the real-time process parameters of the production line corresponding to the time of collecting the original holographic mass field data, to form a mass field data cube with spatiotemporal coordinates. The defect diagnosis module is used to input the mass field data cube into a pre-trained defect diagnosis model, and based on the defect diagnosis model, to perform dual-channel defect tracing diagnosis and performance degradation prediction in parallel, and generate defect diagnosis results for the release film surface.

Images