Real-time compensation method for ACF attaching precision in liquid crystal screen assembly

Abstract

The present application relates to a real-time compensation method for ACF attachment precision in liquid crystal screen assembly, aiming to solve the problems of inaccurate process state recognition and difficult adaptive adjustment of compensation parameter boundary under multi-source heterogeneous signals. The core scheme includes high-frequency synchronous acquisition of multi-channel physical quantity signals, realization of time-frequency joint feature embedding of each channel by state encoder, construction of compact process state fingerprint space by unsupervised contrast learning, and generation of safety boundary fingerprint library with compensation amplitude label based on historical process safety window. In the real-time attachment process, the current working condition fingerprint is dynamically generated, the safety compensation boundary is obtained through library retrieval, and then the compensation parameters are subjected to bidirectional amplitude limiting check and continuity optimization. Combined with fingerprint space drift warning and small sample re-labeling, the compensation boundary is dynamically updated. The application improves the abnormal prevention and control capability of the process, realizes adaptive fine control of the compensation parameter boundary and improves the operation safety.

Description

Technical Field

[0001] This invention relates to the field of liquid crystal display manufacturing process control technology, and in particular to a real-time compensation method for ACF (Anchor Fluorescent Foam) attachment accuracy used in liquid crystal display assembly. Background Technology

[0002] In the assembly and manufacturing process of LCD screens, high-precision control of the ACF (Anisotropic Conductive Film) bonding process is a crucial step in ensuring product performance and yield. With the increasing resolution of display panels and the miniaturization of material structures, the real-time control requirements for multi-dimensional parameters such as bonding head pressure, thermal field distribution, conductive particle uniformity, and tape tension during the ACF bonding process have significantly increased. Current mainstream ACF bonding compensation boundary control technologies generally rely on physical modeling, numerical simulation, or pre-set multi-level rule thresholds to classify and limit safety boundaries for characteristic quantities such as pressure, temperature, tension, and vibration during the process. On the one hand, physical model-driven control methods require cumbersome prior process knowledge determination, have limited adaptability to factory environments, and struggle to cope with complex interactive process drift and nonlinear dynamic fluctuations such as material aging. On the other hand, relying on static rule trees and empirically determined thresholds makes it difficult to cover the "window drift" caused by minor disturbances, batch-to-batch differences, and equipment randomness that arise during actual mass production. This can easily lead to quality problems such as ACF overflow, broken conductive particle chains, or ITO circuit damage caused by exceeding compensation parameter limits, thus reducing bonding consistency and material integrity in the bonding area.

[0003] The limitations of existing technologies in predicting the ACF attachment compensation boundary under dynamic process fluctuations are mainly reflected in the following aspects: (1) It relies on static models, simulation parameters or hierarchical rules and lacks the ability to directly express and dynamically adapt to changes in multi-source signal coupling and process state evolution; (2) The real-time performance is difficult to meet the "millisecond-level" boundary perception and decision output of high-frequency production lines. The boundary judgment is often delayed due to the delay, which affects the accuracy correction of the attachment action. (3) The characterization of process safety windows relies on subjective experience and batch average statistics, making it difficult to capture the implicit state switching and window drift phenomena across batches and environmental conditions. (4) The incremental calibration process is cumbersome and lacks a convenient small-sample dynamic boundary update mechanism, which can easily cause the boundary label to become disconnected from the actual safety extreme value; (5) The lack of interpretability of parameter boundaries that are easy for process debugging and production line teams to understand hinders the actual implementation and large-scale promotion of intelligent boundary control. Summary of the Invention

[0004] This application provides a real-time compensation method for ACF attachment accuracy in LCD assembly, which aims to solve one of the problems or issues of the prior art mentioned in the background section.

[0005] The real-time compensation method for ACF bonding accuracy in LCD panel assembly provided in this application specifically includes: S1: Simultaneously acquire five types of multi-source heterogeneous signals during the ACF bonding process of the LCD screen: bonding head pressure fluctuation sequence, pressure head temperature gradient timing, ACF tape unwinding tension spectrum, platform vibration acceleration envelope, and sub-pixel displacement residual feedback from the optical positioning system, in order to form a raw data acquisition set containing complete process status information.

[0006] S2: Based on the original data acquisition set, perform time-frequency joint embedding processing independently on each type of signal, and map physical quantities of different dimensions into low-dimensional vectors of uniform length to generate independent channel embedding vectors that characterize the characteristics of a single signal channel.

[0007] S3: The independent channel embedding vectors are aligned by timestamp and then spliced ​​and fused. An unsupervised contrastive learning strategy is used to narrow the distance between adjacent time vectors within the same batch and widen the distance between similar working condition vectors between different batches in the latent space, so as to construct a process state fingerprint space with a compact cluster structure.

[0008] S4: Based on the time segments corresponding to the safety windows in historical yield data that do not cause ACF warping, glue overflow, conductive particle breakage or ITO circuit damage, extract the corresponding process state fingerprints and map them to each process state cluster in the process state fingerprint space to generate a safety boundary fingerprint library carrying the maximum allowable compensation amplitude label.

[0009] S5: During the real-time attachment process, a new current operating condition fingerprint is generated every preset time interval, and a nearest neighbor search is performed in the security boundary fingerprint library to locate the target process state cluster to which the current operating condition fingerprint belongs, so as to obtain the dynamic compensation parameter boundary value bound to the target process state cluster.

[0010] S6: Input the boundary values ​​of the dynamic compensation parameters as hard upper limit constraints into the compensation strategy generation module, and truncate or limit the preliminary compensation parameters calculated in real time to generate the final execution compensation parameters that meet the safety requirements of the current process window.

[0011] S7: Based on the final execution compensation parameters, drive the attachment actuator to adjust the position or pressure of the attachment head, and at the same time monitor whether the continuously generated current working condition fingerprint falls into an unknown cluster without a safety label or triggers the compensation parameter to exceed the threshold frequency, so as to generate a boundary drift warning signal.

[0012] S8: If the boundary drift warning signal is detected, the security boundary fingerprint library label of the current target process state cluster is automatically frozen and the high-definition AOI image and equipment log of the most recent sample are retrieved for small sample re-labeling to update the maximum allowable compensation amplitude label and complete the incremental calibration of the dynamic compensation parameter boundary.

[0013] The real-time compensation method for ACF bonding accuracy in LCD screen assembly provided in this application has the following beneficial effects: (1) To address the technical bottlenecks in traditional LCD screen ACF attachment compensation control, which relies on complex physical modeling, simulation calculations, or rigid rule logic leading to poor real-time performance and weak adaptability, this solution proposes a dynamic safety boundary construction mechanism based on semantic fingerprinting. By extracting low-coupling and highly discriminative state representations from multi-source process signals, it achieves efficient online mapping and adaptive evolution of compensation parameter boundaries. Compared to existing technologies that require frequent calls to finite element analysis, causal reasoning, or multi-level condition judgments to determine the allowable compensation range, this solution abandons the reliance on explicit modeling and preset criteria. Instead, it utilizes a lightweight encoder to perform time-frequency joint embedding of heterogeneous signals such as pressure, temperature, tension, vibration, and displacement residuals to generate a unified format state fingerprint. In the latent space, it achieves refined clustering of process states through unsupervised comparative learning, so that each stable operating condition corresponds to a compact cluster with clear physical meaning. This design significantly improves the system's sensitivity to identifying minute but critical process deviations in actual production lines, so that the compensation boundary is no longer statically fixed, but dynamically anchored according to the actual production state. This effectively overcomes the control lag problem caused by equipment aging, environmental fluctuations, or material batch differences, and greatly improves the robustness and consistency of the bonding process.

[0014] (2) Furthermore, this solution innovatively associates the "safety window" in historical good product data with the clustered fingerprint space through labeling, establishing a direct mapping path from state perception to boundary output. This transforms the originally complex multivariate optimization problem into a high-speed retrieval and matching task, greatly reducing the computational load and latency of online inference. During the real-time operation phase, the new fingerprint generated every 20ms can be quickly located to its cluster through nearest neighbor search, and its pre-labeled "maximum allowable compensation range" is immediately called as a hard constraint to ensure that the compensation action is always within the verified safe range, avoiding defects such as edge warping, glue overflow, or damage to conductive structures. At the same time, a boundary drift warning and incremental calibration mechanism is introduced. When an unknown working condition cluster or an existing cluster's out-of-bounds frequency exceeds the standard, the system automatically triggers a small sample relabeling process. Combined with AOI images and equipment logs, the local labels are updated by expert feedback, realizing the gradual evolution of the model boundary without global retraining or rule reconstruction. This closed-loop mechanism not only significantly improves the system's adaptability to long-term process drift, but also retains interpretable paths of expert experience, facilitating quality traceability and engineering maintenance, and demonstrating excellent stability and maintainability in high-speed mass production environments.

[0015] The aforementioned technical approaches collectively construct a novel boundary control paradigm that does not rely on numerical simulation, rule tree determination, or reinforcement learning search, achieving a fundamental shift from "computationally intensive boundary generation" to "semantically driven boundary retrieval." Through semantic fingerprinting as the core medium, the system achieves millisecond-level response, continuous adaptive evolution, and expert knowledge fusion while ensuring security. This provides a compensated boundary protection system for high-precision electronic assembly scenarios, offering real-time performance, robustness, and interpretability, making it particularly suitable for automated production lines in the back-end processes of LCD screen manufacturing, where stability and efficiency are extremely critical. Attached Figure Description

[0016] Figure 1 This is the main flowchart of the real-time compensation method for ACF bonding accuracy in LCD screen assembly.

[0017] Figure 2 This is a sub-flowchart of a method for real-time compensation of ACF bonding accuracy in LCD screen assembly.

[0018] Figure 3 This is another sub-flowchart of the ACF bonding accuracy real-time compensation method used for LCD screen assembly. Detailed Implementation

[0019] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0020] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0021] like Figure 1 As shown, this application provides a real-time compensation method for ACF (Anchor Fluorescent Foliar Fabric) attachment accuracy in LCD panel assembly, specifically including: S1: Simultaneously acquire five types of multi-source heterogeneous signals during the ACF bonding process of the LCD screen: bonding head pressure fluctuation sequence, pressure head temperature gradient timing, ACF tape unwinding tension spectrum, platform vibration acceleration envelope, and sub-pixel displacement residual feedback from the optical positioning system, in order to form a raw data acquisition set containing complete process status information.

[0022] S2: Based on the original data acquisition set, perform time-frequency joint embedding processing independently on each type of signal, and map physical quantities of different dimensions into low-dimensional vectors of uniform length to generate independent channel embedding vectors that characterize the characteristics of a single signal channel.

[0023] S3: The independent channel embedding vectors are aligned by timestamp and then spliced ​​and fused. An unsupervised contrastive learning strategy is used to narrow the distance between adjacent time vectors within the same batch and widen the distance between similar working condition vectors between different batches in the latent space, so as to construct a process state fingerprint space with a compact cluster structure.

[0024] S4: Based on the time segments corresponding to the safety windows in historical yield data that do not cause ACF warping, glue overflow, conductive particle breakage or ITO circuit damage, extract the corresponding process state fingerprints and map them to each process state cluster in the process state fingerprint space to generate a safety boundary fingerprint library carrying the maximum allowable compensation amplitude label.

[0025] S5: During the real-time attachment process, a new current operating condition fingerprint is generated every preset time interval, and a nearest neighbor search is performed in the security boundary fingerprint library to locate the target process state cluster to which the current operating condition fingerprint belongs, so as to obtain the dynamic compensation parameter boundary value bound to the target process state cluster.

[0026] S6: Input the boundary values ​​of the dynamic compensation parameters as hard upper limit constraints into the compensation strategy generation module, and truncate or limit the preliminary compensation parameters calculated in real time to generate the final execution compensation parameters that meet the safety requirements of the current process window.

[0027] S7: Based on the final execution compensation parameters, drive the attachment actuator to adjust the position or pressure of the attachment head, and at the same time monitor whether the continuously generated current working condition fingerprint falls into an unknown cluster without a safety label or triggers the compensation parameter to exceed the threshold frequency, so as to generate a boundary drift warning signal.

[0028] S8: If the boundary drift warning signal is detected, the security boundary fingerprint library label of the current target process state cluster is automatically frozen and the high-definition AOI image and equipment log of the most recent sample are retrieved for small sample re-labeling to update the maximum allowable compensation amplitude label and complete the incremental calibration of the dynamic compensation parameter boundary.

[0029] Step S1: Simultaneously acquire five types of multi-source heterogeneous signals during the ACF (Anti-Fluorescent Film) bonding process of the LCD screen: the bonding head pressure fluctuation sequence, the pressure head temperature gradient timing sequence, the ACF tape unwinding tension spectrum, the platform vibration acceleration envelope, and the sub-pixel-level displacement residual feedback from the optical positioning system, to form a raw data acquisition set containing complete process status information. Specifically, this includes: S1.1: Perform high-frequency synchronous sampling processing on the pressure fluctuation sequence and temperature gradient time sequence of the attachment head to obtain the original mechanical-thermal coupled signal stream containing dynamic mechanical load characteristics and transient thermal field distribution characteristics.

[0030] The acquisition of the pressure fluctuation sequence and temperature gradient timing of the attachment head is performed on the high-precision pressure sensor installed inside the attachment head and the multi-point temperature sensor array in the preheating zone of the attachment head during the attachment operation. The signal output interface is uniformly a digital acquisition bus, and the initial sampling condition is that the synchronous trigger signal arrives at the control unit of the two types of sensors.

[0031] The synchronous trigger signal is input to the sampling controller of the pressure sensor acquisition module and the temperature sensor acquisition module, and a unified high-frequency sampling rate and sampling time reference are set in the sampling controller to ensure that the time domains of the two types of signals are completely matched.

[0032] High-precision analog-to-digital conversion is performed on the original sampled values ​​of the pressure fluctuation sequence to convert the continuous voltage response curve into a time series array of pressure values, and the floating-point pressure values ​​are marked with nanosecond-level timestamps to ensure accurate mapping with the temperature time series.

[0033] Multi-channel synchronous acquisition is performed on the raw sampled values ​​of the temperature gradient time series. The resistance change curve of each temperature sensor node is converted into Celsius temperature value and merged into a multi-point temperature field distribution matrix. The same nanosecond-level timestamp as the pressure fluctuation sequence is added to each data point in the matrix.

[0034] Bandwidth matched filters are used to perform bandpass filtering on the pressure fluctuation sequence and temperature gradient time sequence, respectively, to retain the frequency bands that are significantly related to the dynamic mechanical load and transient thermal field distribution during the attachment process, while suppressing low-frequency baseline drift and high-frequency noise interference.

[0035] The pressure sequence and temperature matrix are mapped one-to-one in the time domain using a timestamp alignment method. The temperature matrix is ​​expanded into a time-series vector of temperature gradient corresponding to the pressure sampling points and fused to form a coupled original signal stream containing dynamic mechanical load characteristics and transient thermal field distribution characteristics.

[0036] Through the above processing methods, the original sensor signal from the previous step is transformed into a structured, synchronized, and bandwidth-optimized mechanical-thermal coupled original signal stream, thereby achieving a precise joint characterization of force and heat during the attachment process.

[0037] For example, in the ACF mounting station for the LCD screen, the pressure sensor sampling rate is configured to 50,000 points per second, the temperature sensor array contains 8 distributed nodes, and the sampling rate is the same as that of the pressure sensor. The sampling controller is uniformly set to a sampling resolution of 0.1 nanoseconds. The frequency band range of the pressure signal processed by the bandpass filter is set to... Hz to remove mechanical baseline drift and high-frequency resonance; the bandpass filter range of the temperature signal is set to The Hz frequency is set to preserve transient thermal distribution characteristics. The filter order is uniformly set to a 4th-order Butterworth filter, with a damping coefficient of [missing value]. Pick To ensure a flat passband response, the timestamp alignment method employs a nanosecond-level interpolation alignment strategy, calculating the maximum permissible deviation between the pressure sequence and the temperature matrix during the alignment process. s serves as a sampling accuracy constraint. The processed coupled signal stream significantly improves the ability to capture mechanical and thermal synchronization features during verification and exhibits greatly enhanced boundary prediction stability in subsequent compensation strategy analysis.

[0038] S1.2: Based on the original signal stream of the mechanical-thermal coupling, the ACF tape unwinding tension spectrum and the platform vibration acceleration envelope are subjected to wideband parallel capture processing to generate an electromechanical multidimensional sensing dataset that integrates material transport stability information and mechanical structure micro-motion disturbance information.

[0039] Based on the mechanical and thermal coupling original signal stream obtained in the preceding step S1.1, the high-speed acquisition interface corresponding to the ACF tape unwinding tension sensor and the platform vibration acceleration sensor is called to synchronously enable the dual-channel parallel acquisition mode within the specified wideband operating parameter range.

[0040] During the capture process, the tension signal and vibration signal are input to independent broadband signal conditioning modules. The effective frequency band of the tension signal is set to 0.5Hz to 500Hz and the effective frequency band of the vibration signal is set to 1Hz to 1000Hz through hardware bandpass filters to ensure that the frequency components related to material transport stability and mechanical structure micro-motion interference are effectively preserved.

[0041] High-speed sampling operations are performed on the tension signal and vibration signal after wideband filtering. The sampling rates are set to 1024Hz and 2048Hz respectively according to the upper limit of the frequency band. The sampling start point of the two data streams is aligned with the time reference of the mechanical and thermal coupling original signal stream through the timestamp synchronization module.

[0042] Using a multi-channel time-frequency analysis method, continuous wavelet transform is performed on the tension signal to extract the transient energy value and phase information of each sub-band. Short-time Fourier transform is performed on the vibration signal to extract the high-frequency disturbance spectrum of the mechanical structure.

[0043] The time-frequency characteristics of the tension and vibration signals mentioned above are fused according to the time series, and a synchronization index related to the mechanical and thermal signal flow is introduced into the fusion matrix to construct an electromechanical multidimensional sensing dataset covering the stability of material transport and the micro-motion disturbance information of mechanical structure.

[0044] By using wideband parallel acquisition and synchronous time-frequency feature extraction processing, the original mechanical-thermal coupled signal stream from the previous step is expanded into electromechanical multidimensional sensing data that includes material transport stability indicators and mechanical structure micro-motion perturbation spectrum, thereby achieving comprehensive and highly reliable input data for the subsequent optical positioning system drive link.

[0045] For example, in a high-precision LCD screen mounting production line, the time reference accuracy of the mechanical-thermal coupling original signal stream is set to the nanosecond level, and the range of the ACF tape unwinding tension sensor is configured to 0~5N with a resolution of [missing information]. N, the range configuration of the platform vibration acceleration sensor is ± g, resolution is g. The broadband filtering parameters are set to 0.5Hz~500Hz for tension signals and 1Hz~1000Hz for vibration signals, with high-speed sampling rates set to 1024Hz and 2048Hz, respectively. The Morlet wavelet basis is selected in the continuous wavelet transform, with a scale range of [missing information]. The energy and phase of the tension signal at each scale are extracted; a Hanning window is selected in the short-time Fourier transform, with a window length of [missing information]. Point, frame shift is The high-frequency disturbance spectrum of the vibration signal within the range of 0~500Hz was extracted. After the fusion matrix was constructed, the output included quantitative data such as tension stability indicators (mean, standard deviation, maximum transient fluctuation value) and mechanical micro-motion disturbance spectrum (peak frequency, peak amplitude, energy concentration). The verification results show that this electromechanical multidimensional sensing dataset can significantly improve the robustness and positioning accuracy of the optical positioning system in the sub-pixel displacement residual extraction process.

[0046] S1.3: Using the electromechanical multidimensional sensing dataset, drive the optical positioning system to perform subpixel-level displacement residual extraction operation to obtain an optical measurement feedback vector characterizing the micro-deviation of the attachment position.

[0047] Based on the electromechanical multidimensional sensing dataset generated in the previous steps, the sub-pixel-level imaging module of the optical positioning system is called to capture multiple high-resolution images of the LCD screen attachment area, forming an image sequence data containing micro-displacement information of the current working condition.

[0048] The image sequence is input to the subpixel matching module. By performing phase correlation calculation on each frame of the image and the reference image, a two-dimensional displacement vector in the pixel coordinate system is obtained. The pixel-level displacement vector is then improved to subpixel accuracy using a subpixel interpolation operator.

[0049] Based on the platform vibration acceleration envelope signal in the electromechanical multidimensional sensing dataset, a vibration compensation model is constructed. This model is used to perform dynamic vibration de-bias calculation on the displacement vector after sub-pixel interpolation to eliminate spurious displacement components introduced by mechanical vibration.

[0050] A temperature-coupled correction operator is used to incorporate the thermal expansion coefficient parameter in the pressure head temperature gradient time-series signal into the displacement correction calculation, and the displacement residual after temperature compensation is calculated.

[0051] The temperature-compensated displacement residual vector is input into the coordinate system normalization module to perform numerical scaling and orientation unification processing on the displacement residuals of different measurement channels, generating an optical measurement feedback vector that can be directly called by the downstream timestamp alignment logic.

[0052] Through the above processing method, the electromechanical multidimensional sensing dataset from the previous step is transformed into sub-pixel displacement residual data after vibration and temperature compensation, realizing high-precision quantification of micro-deviations in the attachment position, and providing a unique optical input index for the subsequent construction of a synchronous collection of multi-source heterogeneous signals.

[0053] For example, in a certain LCD screen ACF bonding process, the electromechanical multidimensional sensing dataset includes the amplitude of the platform vibration acceleration envelope signal (0.02g) and the time-series signal of the pressure head temperature gradient (temperature difference). Celsius, coefficient of thermal expansion / ℃. The phase correlation displacement between the reference image captured by the optical positioning system and the current image is... The pixel value was improved to 0.0005 pixels through subpixel interpolation. The vibration correction displacement calculated using the vibration compensation model was 0.00045 pixels. In the temperature compensation calculation, the displacement correction amount was... = The displacement residual value after temperature compensation is [pixel value]. =0.00044655 pixels. After normalization, the numerical range of the optical measurement feedback vector and other physical quantities is unified within [...]. The interval [1,1] is used to maintain directional consistency constraints, which are then used in the subsequent multi-source signal alignment process, effectively improving the stability and repeatability of the process state fingerprint construction.

[0054] S1.4: Based on the optical measurement feedback vector, perform nanosecond-level timestamp alignment and format standardization on the obtained mechanical and thermal coupling original signal stream, electromechanical multidimensional sensing dataset and optical measurement feedback vector to construct a multi-source heterogeneous signal synchronization set with strict spatiotemporal consistency.

[0055] S1.5: Perform integrity verification and noise floor assessment on the multi-source heterogeneous signal synchronization set to finally form a raw data acquisition set containing complete process status information, which serves as the sole input source for subsequent time-frequency joint embedding processing of the lightweight status encoder.

[0056] Step S2: Based on the original data acquisition set, perform time-frequency joint embedding processing independently on each type of signal, mapping physical quantities of different dimensions into low-dimensional vectors of uniform length to generate independent channel embedding vectors characterizing the properties of a single signal channel. Specifically, this includes: S2.1: Based on the five types of multi-source heterogeneous signals in the original data acquisition set, namely the attachment head pressure fluctuation sequence, the pressure head temperature gradient time series, the ACF tape unwinding tension spectrum, the platform vibration acceleration envelope, and the sub-pixel displacement residual, sliding window truncation and zero-mean normalization preprocessing are performed on each type of signal to eliminate dimensional differences and generate standardized time series signal segments.

[0057] Based on five types of multi-source heterogeneous signals from the original data acquisition set—the attachment head pressure fluctuation sequence, the pressure head temperature gradient time series, the ACF tape unwinding tension spectrum, the platform vibration acceleration envelope, and the sub-pixel displacement residual—an independent sliding sampling window structure was established for each type of signal. The window length and step size parameters were set to values ​​that matched the highest effective frequency of the signal and the process synchronization cycle to ensure that all dynamic characteristics were retained after signal truncation. The signal segments truncated by the sliding window were input into a zero-mean normalization operator. The arithmetic mean of each signal segment was calculated, and a mean subtraction operation was performed at each sampling point of the entire signal segment to eliminate the reference bias between different physical quantities. The normalized signal segments were normalized using the standard deviation normalization method, divided by the square root of the segment's variance to eliminate differences in the amplitude of different signals and ensure the stability of the numerical distribution under the same dimension. The mean and standard deviation calculations during the signal normalization process were completed within the independent segments after window truncation to ensure that the normalization parameters were not affected by external segment data. By using the sliding window truncation and zero-mean normalization methods described above, the multi-source heterogeneous signals from the previous step are transformed into standardized time-series signal segments that eliminate dimensional differences and possess uniform statistical characteristics, thus achieving the expected technical effect of providing high-quality input for subsequent frequency domain decomposition and embedding mapping.

[0058] For example, in an LCD screen ACF bonding production line, the window length for the bonding head pressure fluctuation signal is set to 1024 sampling points, the step size to 256 sampling points, and the sampling frequency to 2048 points per second, to meet the Nyquist sampling requirement for the highest frequency component of 1kHz. The mean of the truncated signal segment is calculated as follows. =2.35N, standard deviation calculated as follows =0.52N, the signal is normalized to a zero-mean unit variance distribution. For the pressure head temperature gradient signal, the window length is set to 512 sampling points, the step size to 128 sampling points, and the sampling frequency to 1024 points per second. The mean of the normalized signal is... ℃, standard deviation is After processing, temperature reference differences were eliminated while maintaining amplitude consistency. Similarly, window and step size parameters were set for the ACF tape tension spectrum, platform vibration acceleration envelope, and displacement residual signal according to their respective frequency characteristics, and normalization was performed to ensure that all signal segments maintained statistical consistency. This embodiment was verified in the subsequent short-time Fourier transform decomposition stage. The spectral energy distribution of the normalized pressure, temperature, tension, vibration, and displacement residual signals was free from distortion caused by amplitude differences, significantly improving the discriminative power and stability of the state encoder embedding vector.

[0059] S2.2: Based on the standardized time-series signal segments, the short-time Fourier transform method is used to perform frequency domain decomposition on the signal within each time window to extract the time-frequency matrix containing instantaneous energy distribution characteristics, thereby converting the one-dimensional time-series data into two-dimensional time-frequency spectrum data.

[0060] S2.3: Based on the two-dimensional time-frequency spectrum data, a lightweight state encoder network architecture including a convolutional feature extraction layer and a temporal attention mechanism layer is constructed. The network weight parameters are optimized through backpropagation training to generate a pre-trained lightweight state encoder model with nonlinear mapping capabilities.

[0061] Based on the aforementioned two-dimensional time-frequency spectrum data, the input shape of the convolutional feature extraction layer is selected to match the resolution of the time-frequency matrix, and the input tensor is mapped to a three-dimensional structure. This allows the subsequent convolutional kernels to slide simultaneously along both the time and frequency dimensions to form local feature perception. The kernel size of the convolutional feature extraction layer is set to a composite parameter set that includes both frequency and time axis coverage. Multiple sets of convolutional kernels with different kernel sizes are determined based on the main energy concentration range of the signal to achieve multi-scale spectral feature capture. The feature map output by the convolution is input to the batch normalization layer to perform feature scale unification and numerical stabilization processing, thereby mitigating the internal covariate shift effect and improving the network's generalization performance across different batches of data. The normalized feature map is input to the temporal attention mechanism layer, where attention weights are generated by calculating the similarity distribution between the feature vectors at each time step and the global context features. Based on these attention weights, the feature sequences are weighted and combined to highlight time segment features containing key frequency components. The weighted feature vector output by the temporal attention mechanism layer is input to the global average pooling layer to perform spatial dimension compression, thereby reducing the number of model parameters while retaining the main dynamic information of the overall process state. The network training objective function is established based on the initial weight randomization setting and the selection of the loss function, where the loss function adopts the mean squared error form: in For the desired output vector, To predict the output after combining convolution and attention, The sample size is specified. The two-dimensional time-frequency spectrum data from the training set is sequentially input into the network. The gradient of the loss function with respect to the weights of each layer is calculated using the backpropagation algorithm, and a learning rate decay strategy is employed to optimize the weight parameters until the validation set error converges. Through this processing, the two-dimensional time-frequency spectrum data from the previous step is transformed into a pre-trained lightweight state encoder model with nonlinear mapping capabilities that highlights key frequency components, achieving the technical effect of mapping the original multi-source signals to a unified low-dimensional embedding space.

[0062] For example, in the time-frequency spectra of the pressure fluctuation sequence and temperature gradient sequence of the LCD screen attachment head, the kernel size of the convolutional feature extraction layer is set to two sets: 3×3 and 5×5, with a stride of 1, and the padding method is set to "same" to maintain the feature map size. Batch normalization uses a momentum coefficient of 0.9, and the query vector dimension in the temporal attention mechanism is 64, with attention weights calculated using the softmax function. The global average pooling layer compresses the 64-dimensional temporal features into 1-dimensional global features. The loss function is as described in the aforementioned formula, the initial learning rate is set to 0.001 during training, and it decays by 50% every 10 iterations. The optimizer used is Adam. After 100 iterations, the validation set loss value stabilizes in the low error range, and the model can resolve the key frequency components in pressure and temperature changes in milliseconds, significantly improving the discriminability and robustness of subsequent process state fingerprints.

[0063] S2.4: Based on the pre-trained lightweight state encoder model, the two-dimensional time-frequency spectrum data is input into the convolutional feature extraction layer for spatial feature compression, and key frequency components are focused through the temporal attention mechanism layer to output a high-dimensional feature tensor representing the local dynamic characteristics of the signal.

[0064] The two-dimensional time-frequency spectrum data generated in the previous steps is used as the input object. The weight parameters of the pre-trained lightweight state encoder model are loaded separately for each signal channel to ensure the consistency of feature extraction and the stability of network mapping during the processing.

[0065] Two-dimensional convolution operations are performed through a convolutional feature extraction layer. Local response values ​​are calculated by sliding within the time-frequency spectrum matrix using preset convolution kernel size and stride parameters. Spatially correlated local time-frequency energy patterns are extracted, and batch normalization is applied to balance the numerical distribution of different batches of data.

[0066] By combining multi-scale convolutional kernel groups to perform cascaded operations on the extracted feature maps, pattern information at different frequency and temporal resolutions can be captured in the same layer, thereby improving the ability to represent the non-stationary characteristics of signals.

[0067] The convolutional output feature map is fed into the temporal attention mechanism layer. Based on the self-attention weight calculation, each frequency component on the time axis is weighted. The query-key-value structure is used to automatically focus on the frequency components with significant energy changes and high correlation with the process state in the time series features.

[0068] We use a normalized attention weight vector to perform a weighted summation on the convolutional feature map to generate a temporally enhanced feature tensor that emphasizes key frequency components and suppresses redundant frequency components, while maintaining consistency with the spatial dimension of the convolution to facilitate subsequent fusion.

[0069] By employing the aforementioned convolutional compression and attention-focused sequential processing method, the time-frequency spectrum data is transformed into a signal local dynamic feature tensor with high discriminative power and high dimensionality. This achieves spatial compression and temporal emphasis enhancement of the features, providing optimized feature inputs for subsequent dimensionality reduction mapping and state fingerprint construction.

[0070] For example, in the ACF bonding process of a liquid crystal display, the two-dimensional time-frequency spectrum data of the bonding head pressure fluctuation sequence is input into the convolutional feature extraction layer of a lightweight state encoder, using a size of stride length The convolutional kernels are used for spatial local feature extraction, and the convolutional output is passed through a batch normalization layer to eliminate inter-batch data skew. The convolutional kernel group is configured with four different scales, the largest of which is... To capture low-frequency wide-time-window patterns, the minimum scale is This approach captures high-frequency, short-time-window patterns. The convolutional results enter a temporal attention mechanism layer, where attention is calculated using a scaled dot product. The calculated attention weight vectors are then summed across the convolutional feature maps step-by-step to generate a high-dimensional feature tensor that enhances key frequency components. In this scenario, the output tensor has a dimension of [missing information]. (Number of channels × time step × frequency step), after subsequent dimensionality reduction processing, is compressed into a 32-dimensional independent channel embedding vector, which significantly improves the feature discrimination and stability in the state fingerprint construction stage.

[0071] S2.5: Based on the high-dimensional feature tensor, a fully connected projection layer is used to perform linear dimensionality reduction mapping to forcibly compress the variable-length high-dimensional feature tensor into a fixed-length low-dimensional vector, so as to generate an independent channel embedding vector characterizing the characteristics of a single signal channel.

[0072] like Figure 2 As shown, step S3 involves aligning the independent channel embedding vectors by timestamp and then splicing and fusing them. An unsupervised contrastive learning strategy is used to narrow the distance between adjacent time-stamp vectors within the same batch and widen the distance between similar working condition vectors between different batches in the latent space, thereby constructing a process state fingerprint space with a compact cluster structure. Specifically, this includes: S3.1: Based on the independent channel embedding vectors generated in the previous steps, perform timestamp alignment processing on the independent channel embedding vectors to eliminate timing deviations in the multi-source heterogeneous signal acquisition process and generate a timing-aligned embedding vector set with strict time synchronization characteristics.

[0073] Based on the independent channel embedding vectors generated in the previous steps, the original timestamp information of each vector is extracted into a time sequence identifier value in a unified format, forming a time index set with comparability and retrieval. Using a high-precision global clock synchronization mechanism, the timestamps of all channel embedding vectors are uniformly converted, mapping the local time reference of the acquisition end to the global time axis of the process. Through nanosecond-level precision time difference calculation, a time deviation matrix between multiple channels is constructed, and elements with deviations exceeding a preset threshold are identified in the matrix to determine the channel vector indices that need correction. A time resampling method based on the fusion of linear interpolation and spline interpolation is used to perform numerical reconstruction on the channel embedding vectors with time deviations, interpolating the vectors to the target sampling point positions on the global time axis. After deviation correction, all channel embedding vectors are sequentially sorted according to the global time axis, generating a time-aligned embedding vector set strictly arranged according to time synchronization. Through the above alignment processing, the independent channel embedding vectors from the previous step are transformed into standard input data with strict time synchronization characteristics, achieving temporal consistency and feature correlation in the subsequent multimodal feature splicing and fusion process.

[0074] For example, for the five independent channel embedding vectors collected during the ACF bonding process of the LCD screen, assuming that the sampling frequency of each vector is 2000Hz and there is a maximum trigger delay of 3 nanoseconds between the acquisition modules, a global clock synchronization mechanism is used to unify all timestamps to a millisecond-level reference, and a time deviation matrix is ​​constructed, where the deviation threshold is set to... Seconds. For channel vectors with deviations exceeding the limit, a hybrid interpolation method is applied for resampling, where linear interpolation is used for deviations less than the limit. In the case of seconds, spline interpolation is used for deviations greater than this value. After resampling, all channel vectors are fully aligned with the global time axis, and after sorting, a strictly temporally consistent set of embedding vectors is obtained. This processing shows a significant improvement in feature alignment accuracy in multi-batch process data validation, ensuring that the synchronous correlation between physical quantities can be maintained during subsequent process state fingerprint splicing operations, reducing feature drift caused by temporal misalignment, thereby enhancing the stability and reliability of process state fingerprint space construction.

[0075] S3.2: Based on the time-aligned embedded vector set, a vector splicing operator is used to perform dimension concatenation and fusion processing on the low-dimensional vectors of each channel to integrate the multi-modal feature information of pressure, temperature, tension, vibration and displacement residuals, and generate an original process state fingerprint vector that represents the complete instantaneous process state.

[0076] S3.3: Based on the original process state fingerprint vector, construct positive and negative sample pairs using an unsupervised contrastive learning method. The original process state fingerprint vectors at adjacent times within the same batch are defined as positive sample pairs, and the original process state fingerprint vectors of similar working conditions between different batches are defined as negative sample pairs, so as to establish a set of sample constraint relationships for latent space optimization.

[0077] Based on the original process state fingerprint vector, correlation analysis is performed on adjacent state fingerprints between different time steps within the same batch, and state pairs with high temporal continuity are extracted as a candidate set of positive samples to ensure that the positive samples contain stable evolution characteristics of the same type of working conditions.

[0078] The system performs differential discrimination on state fingerprints that are similar in physical quantity values ​​but lack temporal continuity between different batches, and selects easily confused cross-batch state pairs as a negative sample candidate set to construct a contrastive learning input with high discriminative requirements.

[0079] The Euclidean distance and cosine similarity between the embedding vectors of the positive and negative candidate sets are calculated using similarity measurement functions. Based on a preset distance threshold, adjacent samples in the same batch with a distance lower than the threshold are identified as positive sample pairs, and similar samples across batches with a distance higher than the threshold are identified as negative sample pairs.

[0080] An unsupervised contrastive learning objective function is used to map the labels of the positive and negative sample pairs, so that the system has the constraints of forward approach and backward separation in the latent space optimization step, providing a complete set of sample constraint relationships for subsequent latent space mapping operations.

[0081] By using a sample filtering and construction mechanism, the original state fingerprint vector from the previous step is transformed into a data structure with positive and negative labels, thereby achieving the expected technical effect of compressing similar samples and separating dissimilar samples in the latent space.

[0082] For example, on a high-precision LCD screen ACF bonding production line, 500 frames of original process status fingerprint vectors were collected for batch A and batch B respectively. The Euclidean distance between any two adjacent frames in batch A was within... Within this range, a positive sample distance threshold is set. These vector pairs are identified as positive sample pairs; the Euclidean distance between some state vectors in batch B and batch A is approximately And the cosine similarity is higher than However, due to the lack of temporal continuity, they were confirmed as negative sample pairs. In this scenario, through... The contrastive loss function of the form compresses the distance between positive sample pairs to... Below, the distance between negative sample pairs is expanded to... In summary, after constructing the sample constraint relationship set, the compactness of state clusters in the same batch is significantly improved during latent space clustering, the cross-batch confusion rate is significantly reduced, and the stability of the system compensation parameter safety boundary prediction is enhanced.

[0083] S3.4: Based on the set of sample constraint relationships, perform latent space mapping transformation operation. By minimizing the Euclidean distance between positive sample pairs in the latent space and maximizing the Euclidean distance between negative sample pairs in the latent space, the distribution range of similar process states is compressed and the distribution regions of dissimilar process states are separated, generating an optimized process state fingerprint latent vector with compact cluster structure characteristics.

[0084] Based on the set of sample constraint relationships, the latent space coordinates of each positive sample pair are read and their Euclidean distance is calculated to quantify the proximity of the process state fingerprints at adjacent times in the same batch in the latent space.

[0085] When calculating the Euclidean distance, the distance operation function is called, and the two latent vectors are used as component inputs respectively.

[0086] For negative sample pairs, repeat the above distance calculation steps and call the reverse optimization module to perform the negative sample pair distance maximization iteration. Input the distance value of each negative sample pair into the regularization term of the loss function to drive the optimization direction to push out the latent vectors of similar working conditions between different batches.

[0087] A joint loss function is constructed by combining the term that minimizes the Euclidean distance between positive sample pairs and the term that maximizes the Euclidean distance between negative sample pairs, and then linearly weighting them using weight coefficients.

[0088] The weights of the latent space mapping network are iteratively updated using the stochastic gradient descent method. In each iteration, the gradient of the joint loss function needs to be calculated and the parameters need to be adjusted in the optimizer to gradually compress the distribution range of positive sample pairs in the latent space and separate the distribution region of negative sample pairs.

[0089] The optimized latent space vectors are rewritten into a new set of latent vectors for process state fingerprints, and the corresponding compact cluster structure feature indexes are recorded to complete the latent space mapping transformation.

[0090] By using the constrained optimization method of the distance between positive and negative samples in the latent space, the original process state fingerprint vector from the previous step is transformed into an optimized process state fingerprint latent vector with compact cluster structure characteristics. This improves the aggregation degree of similar states and enhances the separability of dissimilar states, thereby achieving structural optimization of the process state fingerprint space and providing high-quality input for the next step of density peak clustering.

[0091] For example, in the monitoring of ACF (Anti-Coating Fluid) attachment conditions of LCD screens, positive sample pairs are latent vectors of state fingerprints from two consecutive sampling times within the same batch, while negative sample pairs are latent vectors from different batches but with similar pressure and temperature curve shapes. Let the average Euclidean distance of the positive sample pairs be... for The average Euclidean distance between negative sample pairs for The weighting coefficient α is set as Let β be... The combined loss is Calculated After performing 100 iterations of gradient descent, the Euclidean distance between positive sample pairs is reduced to... The Euclidean distance between negative samples increases to The latent vector cluster compactness is significantly improved and the inter-cluster separation is enhanced, which improves the accuracy of subsequent clustering. The final output optimized process state fingerprint latent vector shows a significantly improved process state discrimination ability on the clustering validation set.

[0092] S3.5: Based on the optimized process state fingerprint latent vector, the density peak clustering algorithm is applied to identify high-density clustered regions in the latent space, so as to automatically divide the process state clusters representing different typical stable operating conditions, and finally generate a process state fingerprint space with high discriminativeness and compactness.

[0093] like Figure 3 As shown, step S4 involves extracting the corresponding process state fingerprint from the time segments of the safe window in historical yield data that did not cause ACF warping, adhesive overflow, conductive particle breakage, or ITO circuit damage, and mapping them to various process state clusters in the process state fingerprint space to generate a safe boundary fingerprint library carrying a label of the maximum allowable compensation magnitude. Specifically, this includes: S4.1: Perform correlation and alignment processing on high-definition automatic optical inspection image data and equipment operation log data from historical production processes to identify safe process time periods in which no ACF warping, glue overflow, conductive particle breakage, or ITO circuit damage occurred, and convert the safe process time periods into a set of safe window time series.

[0094] When correlating and aligning high-definition automated optical inspection (AOI) image data and equipment operation log data from historical production processes, the input objects are high-resolution image sequences of the AOI system, marked with millisecond-level timestamps, collected from multiple batches of production tasks, and multi-source sensor operation log datasets recorded by the equipment's main control system. The conditions of unified time reference and parsable data format must be met. First, based on a unified time reference, the capture timestamps of the AOI image sequences are bidirectionally matched with the sensor sampling timestamps recorded in the operation logs to eliminate deviations caused by clock drift from different acquisition systems. Based on the matched data pairs, the AOI images are converted into grayscale matrix representations, and normalized encoding is performed on the sensor data in the operation logs, making the two types of data comparable within the same numerical range. Based on the pixel statistical features of the grayscale matrix and the normalized sensor vectors, a defect mode determination method is used to assess the process safety of each time slice. By setting a threshold matrix for four types of defects—edge warping, adhesive overflow, conductive particle breakage, and ITO circuit damage—time slices that do not trigger any threshold are marked as safe. For continuous safety state time slices, a time slice aggregation strategy is used to merge them into safety process time periods, and start and end time indices are recorded to generate a preliminary safety window set. Based on the preliminary safety window set, a time series smoothing screening method is used to remove short-term safety windows with durations below a preset lower limit, thereby improving the statistical stability of subsequent safety samples. Through the above processing, the results of the previous step are transformed into a structured safety window time series set, realizing the time range limitation for subsequent safety process state fingerprint extraction.

[0095] For example, the AOI image resolution is set to 4096×2160 pixels, the acquisition frame rate is 50fps, and the operation log includes four channels: pressure sensor, temperature sensor, tension meter, and accelerometer, with a sampling frequency of 1kHz. During time matching, the maximum allowed clock drift correction is ±2ms. The grayscale matrix normalization range is set to 0 to 1, and sensor value normalization is performed linearly scaled according to the historical maximum and minimum values ​​of each channel. In the defect judgment threshold matrix, edge warping detection is defined as an image edge length deviation of less than... The glue overflow detection indicates that the average brightness of the image edges is higher than that of the image itself. The area of ​​the region is smaller than the total area. The detection of conductive particle fracture showed that the particle density was lower than Particles / mm², ITO damage detection shows that the cumulative fracture length of the circuit is less than [value missing]. During safety status determination, each time slice that simultaneously meets all four conditions is marked. Time periods with consecutive safety states exceeding 200ms are merged into safety windows, and their start and end time indices are recorded. After removing windows shorter than 150ms using a smoothing filtering method, a set of safety window time series is output. This set is used for feature extraction and mapping of subsequent safety condition fingerprints. Validation results show that the generated window set significantly improves the stability and accuracy of boundary prediction in subsequent compensation boundary calculations.

[0096] S4.2: Based on the safety window time series set, extract multi-source heterogeneous signal segments with corresponding timestamps from the original data acquisition set, and use the lightweight state encoder that has been trained in the previous step to perform time-frequency joint embedding inference on the multi-source heterogeneous signal segments to generate a safety channel embedding vector group that represents the safety conditions.

[0097] Based on the set of time series within the safety window, the timestamp index ranges of five types of multi-source heterogeneous signals in the original data acquisition set corresponding to each time segment are determined. The pressure fluctuation sequence, temperature gradient time series, unwinding tension spectrum, vibration acceleration envelope, and subpixel displacement residual signal that match the index are then called in the data storage structure to form a set of data slices with temporal consistency.

[0098] The data slice set is input to the multi-channel input of the lightweight state encoder that has been trained and weighted in the previous step. While maintaining the independence of each physical quantity channel, time-frequency joint embedding processing is performed separately. The one-dimensional time series signal is transformed into a two-dimensional time-frequency matrix through short-time Fourier transform. Then, the significant frequency components are focused through convolutional feature extraction and temporal attention mechanism to output a high-dimensional feature tensor.

[0099] A fully connected projection layer is used to perform uniform length dimensionality reduction mapping on each high-dimensional feature tensor, compressing the high-dimensional tensors of different channels into low-dimensional vectors of fixed length. This ensures that the embedded vectors between channels can be directly spliced ​​and fused in dimensionality, while retaining the main dynamic features of the corresponding safety conditions.

[0100] During the embedding process, the energy of the time-frequency matrix is ​​normalized to eliminate signal amplitude differences.

[0101] The feature stability test is performed on the low-dimensional vector after dimensionality reduction within the channel. The mean and variance of the embedding vector of each channel are calculated by sliding window. If the variance is lower than the set threshold, it is marked as a high-stability channel feature, which is used to enhance the clustering consistency of subsequent security state fingerprints.

[0102] Through the above chained processing, the multi-source heterogeneous signal segments corresponding to the safety window time series are transformed into a set of safety channel embedding vectors that characterize the safety conditions. This achieves a standardized mapping of the safety state in the embedding space, providing a highly consistent and discriminative input for subsequent safety fingerprint fusion and spatial projection.

[0103] For example, in the ACF bonding process of an LCD screen, the safety window time series set contains 120 timestamp segments. The sampling frequency of the pressure fluctuation sequence in the original data acquisition set is 5kHz, the sampling frequency of the temperature gradient time series is 2kHz, the sampling resolution of the unwinding tension spectrum is 0.5Hz, the sampling frequency of the vibration acceleration envelope is 10kHz, and the sampling resolution of the displacement residual is 0.1μm. When each signal segment is input into the lightweight state encoder, the short-time Fourier transform window length is set to 256 points, the frame shift is 128 points, the number of filters in the convolutional feature extraction layer is 32, the kernel size of each filter is 3×3, the number of attention heads in the temporal attention mechanism is 4, and the output dimension of the fully connected projection layer is fixed at 64. The normalization process calculates the normalized energy value to ensure that the energy characteristics of different physical quantity channels are on the same scale. The sliding window variance test threshold is set to 0.01. The results show that 89% of the channel embedding vector variances are below the threshold and are marked as high-stability features. The secure channel embedding vector set output in this step significantly improves the cluster compactness of the secure state fingerprint in the subsequent fusion process, making the distribution of secure sample points projected into the process state fingerprint space significantly concentrated, and improving the robustness and accuracy of the dynamic compensation parameter boundary generation.

[0104] S4.3: Perform a splicing and fusion operation on the embedded vector group of the security channel according to the timestamp to generate a security process state fingerprint sequence that corresponds one-to-one with the security window, and project the security process state fingerprint sequence into the process state fingerprint space constructed by the unsupervised contrastive learning strategy.

[0105] S4.4: Based on the compact cluster structure in the process state fingerprint space, calculate the Euclidean distance between each safe process state fingerprint and the center of each process state cluster, and classify the safe process state fingerprint into the nearest target process state cluster to form an intra-cluster safe sample subset containing multiple safe sample points.

[0106] Based on the previously generated safety process state fingerprint sequence and the constructed process state fingerprint space, the cluster center coordinate matrix is ​​loaded as a matching reference vector set to ensure that each cluster center point is a high-density peak coordinate extracted during the latent space clustering process. The Euclidean distance calculation module is then invoked to perform distance metric calculations between each safety process state fingerprint and each cluster center point. The individual safety fingerprint is calculated using the following formula. With cluster center Euclidean distance: in, This represents the i-th dimension component of the secure fingerprint. Let represent the i-th dimension of the corresponding cluster center, and p be the dimension of the feature vector. Sort the distance values ​​output by the above formula according to the cluster index, and select the cluster center index with the smallest distance as the cluster identifier to which the safety fingerprint belongs. Form a mapping pair between the cluster identifier and the safety fingerprint, and write it into the intra-cluster safety sample index table to achieve the binding of the safety process state fingerprint to the cluster structure. Perform an index merging operation on multiple safety sample points within the same cluster to generate an intra-cluster safety sample subset containing all safety sample points. Through the above distance calculation and classification processing, the result of the previous step is transformed into a data set with cluster affiliation characteristics, realizing the automatic correspondence between safety state samples and stable operating conditions in the latent space.

[0107] For example, on an LCD screen ACF bonding production line, the safety process status fingerprint sequence contains 120 low-dimensional feature vectors, each with a length of 64 dimensions. The process status fingerprint space is clustered into 15 clusters, with the cluster center coordinate matrix having a dimension of 15×64. Using an Euclidean distance calculation module, for a specific safety fingerprint... Calculate the distance to each cluster center, including the distance to the cluster center. The distance is the smallest, and the value is... If the sample is found to be positive, it is classified into cluster 8. This process is repeated to generate a safe sample subset containing 47 sample points for cluster 8. During verification, fluctuations in compensation parameters during historical production did not trigger defects such as edge warping, glue overflow, or conductive particle breakage. This indicates that the operating conditions covered by the safe samples within this cluster are stable and can be used to calculate the maximum permissible compensation range. The final output safe sample subset within the cluster provides a complete data source for subsequent S4.5 extreme value statistics.

[0108] S4.5: Perform extreme value statistical analysis on the compensation parameter values ​​actually used by all samples in the safe sample subset within the cluster at historical moments, extract the maximum parameter amplitude that does not trigger process defects, define the maximum parameter amplitude as the maximum allowable compensation amplitude label and bind it to the corresponding target process state cluster to generate a safe boundary fingerprint database carrying dynamic safe boundary attributes.

[0109] Step S5: During the real-time attachment process, a new current operating condition fingerprint is generated every preset time interval, and a nearest neighbor search is performed in the security boundary fingerprint database to locate the target process state cluster to which the current operating condition fingerprint belongs, so as to obtain the dynamic compensation parameter boundary value bound to the target process state cluster. Specifically, this includes: S5.1: Based on the preset time interval trigger signal, the five types of multi-source heterogeneous raw data, namely the synchronously acquired attachment head pressure fluctuation sequence, the pressure head temperature gradient time sequence, the ACF tape unwinding tension spectrum, the platform vibration acceleration envelope, and the sub-pixel displacement residual, are timestamped to generate a multi-channel fused raw data frame with a unified time base, providing standardized input conditions for the real-time generation of subsequent state fingerprints.

[0110] The input conditions include five types of multi-source heterogeneous raw data: a trigger signal with a preset time interval, and synchronously acquired data during real-time attachment, including the attachment head pressure fluctuation sequence, pressure head temperature gradient time series, ACF tape unwinding tension spectrum, platform vibration acceleration envelope, and sub-pixel displacement residual. Based on the arrival of the trigger signal, a unified acquisition controller is invoked to perform buffer refresh operations on the data from each sensor, ensuring that all signal segments correspond to the same time base. A unified time base mapping method is applied to each type of signal stream, converting the original acquisition time stamps to a global reference coordinate system. Nanosecond-level interpolation compensation is used during the mapping process to fill time defects caused by acquisition delays. Using cross-channel synchronization constraints, frame-level alignment is performed between different signal sources to ensure that all physical quantities within each fused data frame correspond to the same process time. This constraint is achieved by minimizing the time difference between the boundary moments of each channel. After time mapping and frame alignment, the data undergoes format standardization processing, converting all pressure, temperature, tension, vibration, and displacement residual signals into a unified matrix format. Matrix rows correspond to time steps, and columns correspond to physical quantity channels, thus generating raw data frames for multi-channel fusion. By using timestamp alignment and format standardization methods, the result of the previous step is transformed into a multi-channel fused raw data frame with a unified time base and structured format, thereby standardizing the input conditions for the subsequent state fingerprint generation step.

[0111] For example, on an LCD screen mounting production line, the trigger interval is set to 20ms, the mounting head pressure sensor sampling rate is 100kHz, the pressure head temperature sensor sampling rate is 50kHz, the tension sensor sampling rate is 10kHz, the vibration sensor sampling rate is 5kHz, and the optical positioning system displacement residual output frequency is 1kHz. When the trigger signal arrives, the acquisition controller extracts continuous data segments from the most recent 20ms from the buffers of each sensor and maps the original timestamps of these segments to a unified nanosecond-level global time reference. Interpolation algorithms are used to fill in the time gaps that occur under different sampling rates. After frame alignment processing, the time difference between each channel within the same data frame is less than 5ns, meeting the process synchronization requirements. During format standardization, signals with different sampling rates are resampled into a matrix form with a uniform time step. The matrix size is 20×5, with rows corresponding to time steps and columns corresponding to five types of physical quantities. The processed fused data frame can be directly input into a lightweight state encoder, significantly improving synchronization and feature extraction stability during the state fingerprint generation stage.

[0112] S5.2: Using the trained lightweight state encoder, perform time-frequency joint embedding transformation processing independently on each physical quantity signal in the multi-channel fused original data frame, mapping the high-dimensional time-series physical quantity into a fixed-length low-dimensional feature vector to generate an independent channel embedding vector set that characterizes the characteristics of a single signal channel at the current instant.

[0113] Under the condition of multi-channel fusion of raw data frame input, for multi-source signals including the attachment head pressure fluctuation sequence, the pressure head temperature gradient time series, the ACF tape unwinding tension spectrum, the platform vibration acceleration envelope, and sub-pixel displacement residuals, a time-frequency joint embedding transformation process is performed with independent operation of each signal channel. For each physical quantity signal, the corresponding encoder input port is called, and the raw values ​​are converted into a matrix-based time series representation conforming to the network input format. The weight parameters of the convolutional feature extraction layer of the pre-trained lightweight state encoder are loaded to ensure that the signal preprocessing configuration of the current environment is consistent with that of the training environment. The short-time Fourier transform function module is called to generate a time-spectrum map of the input signal according to the set window length and frame shift. The window function type and overlap ratio are set to optimize energy leakage control. The spatial feature convolution and pooling operations are performed on the generated two-dimensional time-spectrum data through the convolutional feature extraction layer to extract local frequency changes and energy peak distribution, and retain important feature components that are sensitive to process conditions. The convolution output is processed through the time-series attention mechanism layer to strengthen the weights of key frequency bands in the time series, weaken redundant frequency band information, and form a high-dimensional feature tensor. Linear dimensionality reduction is performed using a fully connected projection layer, compressing the variable-length high-dimensional feature tensor into a fixed-length low-dimensional vector, which is then stored as an element of an independent channel embedding vector set. This process is repeated for each signal channel to obtain its corresponding low-dimensional embedding vector, ensuring that each element in the set accurately represents the physical characteristics of the single signal channel at the current instant. This processing method transforms the original multi-channel fusion data frame into a set of independent channel embedding vectors of uniform length that can be spliced ​​and fused, thus achieving stable input feature conditions for the real-time state fingerprint generation process.

[0114] S5.3: Based on the time synchronization mechanism, the independent channel embedded vector set is spliced ​​and fused in the channel order, and a comprehensive state vector containing complete current process state information is generated through the latent space projection operation, so as to form a current working condition state fingerprint that can uniquely identify the current attachment working condition.

[0115] S5.4: Call the pre-built security boundary fingerprint database, and use the approximate nearest neighbor search method to calculate the Euclidean distance or cosine similarity between the current operating condition fingerprint and the center of all historical process state clusters in the database, so as to quickly locate the target process state cluster index with the highest matching degree to the current operating condition fingerprint.

[0116] Using the current operating condition fingerprint generated in the previous steps as input, the system calls a pre-built security boundary fingerprint database and loads all historical process state cluster center vectors to establish a query index structure to support high-concurrency retrieval. An approximate nearest neighbor search method is employed, which reduces the number of database traversals and significantly improves retrieval speed by constructing multi-layered segmentation hyperplanes or hash bucket mappings. During the retrieval process, the Euclidean distance between the current operating condition fingerprint and the center vector of each process state cluster is calculated, and the distance metric is implemented using the following formula: in, Let i be the i-th dimension component of the current fingerprint. Let be the i-th component of the cluster center, and p be the dimension of the feature vector. The distance value is obtained by summing the squared differences and taking the square root. Cosine similarity is also supported as a supplementary metric, as shown in the following formula: The numerator is the sum of the products of corresponding components, and the denominator is the product of the magnitudes of the two vectors. This metric reflects directional consistency. Priority rules are set under both distance and similarity metrics, with the minimum Euclidean distance being the primary matching criterion. When the Euclidean distance difference is below a set threshold, the maximum cosine similarity is used for filtering, ensuring stable and reliable matching results. The cluster center index with the highest matching degree is output as the target process state cluster index, providing a unique location basis for downstream compensation boundary value extraction. Through a search based on approximate nearest neighbors and a dual-metric judgment processing method, the current operating condition fingerprint generated in the previous step is transformed into a target process state cluster index that can be directly mapped to the compensation boundary, achieving millisecond-level matching and positioning with guaranteed accuracy.

[0117] For example, in a high-precision LCD panel mounting production line, the security boundary fingerprint database contains 200 process state clusters, each with a 64-dimensional center vector. A hierarchical positioning tree structure is used for approximate nearest neighbor search, with a maximum of 20 center vectors for each cluster. The input current process state fingerprint is a 64-dimensional floating-point vector. During retrieval, Euclidean distance is calculated, and the square root of the sum of squared differences of the distance values ​​of each center vector is taken. When the minimum Euclidean distance value is... And the difference between the distance to the second smallest Euclidean distance is less than When the cosine similarity is determined, the highest calculated value is... The corresponding cluster index is cluster 145. This index is directly passed to the next sub-step to extract the maximum allowable compensation magnitude label value. Verification results show that the average response time in 100,000 real-time retrieval simulations is less than [missing information]. The millisecond accuracy is significantly improved, meeting the needs of high-cycle production for real-time compensation boundary dynamic anchoring.

[0118] S5.5: Based on the target process state cluster index, extract the pre-bound maximum allowable compensation amplitude label value from the security boundary fingerprint database, parse the value into the dynamic compensation parameter boundary value at the current moment, and output it to the downstream execution module as a hard constraint condition to limit the output range of the compensation strategy.

[0119] Step S6: The boundary values ​​of the dynamic compensation parameters are input as hard upper limit constraints to the compensation strategy generation module to truncate or limit the preliminary compensation parameters calculated in real time, so as to generate the final execution compensation parameters that meet the safety requirements of the current process window. Specifically, this includes: S6.1: Construct a compensation strategy generation module to obtain the boundary values ​​of dynamic compensation parameters generated by the previous steps and the preliminary compensation parameters output in real time by the compensation strategy generation module. Parse the boundary values ​​of the dynamic compensation parameters into a safe threshold range containing the maximum allowable positive adjustment amount and the maximum allowable negative adjustment amount, and map the preliminary compensation parameters into the original adjustment vector to be verified, so as to form a parameter verification dataset containing constraints and the object to be optimized.

[0120] A compensation strategy generation module is constructed based on a deep neural network architecture to model the mapping from the process state fingerprint space to compensation parameters. This module uses the current operating condition fingerprint ft, output from step S5, as its core input, and optionally incorporates the historical operating condition fingerprint sequence {ft} from previous time steps. The vector k,…,ft} is used to capture the dynamic trend of process drift and output the initial compensation parameter vector Ct∈Rm, where m corresponds to the actuator adjustment dimensions such as pressure, displacement, and temperature.

[0121] The compensation strategy generation module employs an LSTM + attention encoder structure. First, if the input is a time-series sequence, evolutionary features in the time dimension are extracted through a two-layer stacked Long Short-Term Memory network to obtain the hidden state sequence {ht-k,...,ht}. If only the current fingerprint is input, dimensionality is directly increased through a fully connected layer. Subsequently, a multi-head self-attention mechanism is introduced to weighted aggregate the hidden states, enabling the network to adaptively focus on the feature dimensions most relevant to the compensation amount in the fingerprint space—for example, when an ACF adhesive layer thickness fluctuation is detected, the attention weights automatically shift towards the corresponding sensor channel, thereby improving the targeting of the compensation. The output of the attention layer, after batch normalization and Dropout, is fed into two parallel fully connected branches: the main regression branch uses a three-layer progressive dimensionality reduction structure, employing PReLU as the activation function to alleviate gradient vanishing, and finally outputs preliminary compensation parameters; the auxiliary branch outputs a confidence scalar to indicate the uncertainty of the current compensation prediction, for reference by the upper-level decision system.

[0122] The training phase dataset construction fully utilizes the historical yield data contained in the security boundary fingerprint database in S4. Specifically, paired samples (ft, ct) are extracted from the time segments corresponding to the security window. ), where ct The compensation parameter represents the actual successful execution, and ft represents the fingerprint of the current operating condition. The loss function is designed as a weighted sum of three terms: the mean squared error of the main regression branch ensures prediction accuracy; the negative log-likelihood loss of the auxiliary branch is used to calibrate the uncertainty estimate; in addition, a soft boundary constraint loss is introduced. Here, λ represents the soft boundary constraint loss value, and λ is the boundary loss weight coefficient. The compensation parameter vector predicted by the model. Let max(0,·) be the safety boundary vector, and let max(0,·) be the hinge function.

[0123] Penalizing predictions that exceed historical safety boundaries enables the network to possess boundary awareness during training, thereby reducing the truncation frequency of subsequent S6 clipping modules. The entire network employs the AdamW optimizer, uses a cosine annealing learning rate strategy, and introduces random noise during training as data augmentation to improve robustness.

[0124] Using the boundary values ​​of the dynamic compensation parameters generated in the preceding steps as the input data source, and combining them with the preliminary compensation parameters output in real time by the compensation strategy generation module, a parameter processing flow for boundary constraint verification is established. Based on the boundary values ​​of the dynamic compensation parameters, analytical operations are performed to decompose them into two independent threshold components, each containing a maximum allowable positive adjustment and a maximum allowable negative adjustment, and these components are stored in a structured data format as a safety threshold interval matrix. The mapping operation module is called to map the preliminary compensation parameters dimensionally to the original adjustment vector to be verified. This mapping process must maintain the consistency of the order of physical quantities and the integrity of numerical accuracy across all dimensions. A parameter verification dataset is constructed from the safety threshold interval matrix and the original adjustment vector. This dataset contains the constraint condition matrix and the matrix of the object to be optimized, and the dimension-wise correspondence in the parameter verification process is ensured through index binding. A vector merging operator is used to perform matrix alignment operations on the upper and lower bounds of the safety threshold interval matrix and the original adjustment vector, providing a precise benchmark framework for subsequent amplitude limiting determination. By using the above processing method, the results of the previous step are transformed into a parameter verification dataset with dual attributes of constraints and the object to be optimized, realizing a structured association between compensation parameters and dynamic boundary values, and providing a stable data foundation for the subsequent determination input of the bidirectional hard limiting method.

[0125] For example, in a high-precision operation on a certain LCD screen mounting production line, the boundary values ​​of the dynamic compensation parameters are a maximum allowable positive adjustment of 0.015 mm and a maximum allowable negative adjustment of -0.012 mm. After analysis, a safety threshold interval matrix is ​​formed. The initial compensation parameters output in real time by the compensation strategy generation module are a three-dimensional vector. After mapping, the dimension-by-dimensional binding relationship between the original adjustment vector and the safety threshold interval matrix is ​​obtained: the first dimension adjustment amount of 0.018 and the interval... Correspondingly, the second dimension adjustment of -0.010 corresponds to the same interval, and the third dimension adjustment of 0.007 corresponds to the same interval. In subsequent clipping processing, the first dimension will be marked as out of bounds due to exceeding the positive upper limit, while the second and third dimensions will be marked as compliant. By establishing this dataset, out-of-bounds dimensions can be accurately determined after nanosecond-level matrix alignment operations, significantly improving the truncation response speed and ensuring the safe adjustment range of physical quantities in each dimension, thereby enhancing adhesion consistency and material integrity.

[0126] S6.2: Based on the parameter verification dataset, the bidirectional hard limiting method is used to compare each dimension component in the original adjustment vector with the upper and lower bounds of the safety threshold interval one by one, identify the out-of-bounds components that exceed the safety threshold interval and mark their out-of-bounds direction, so as to generate a parameter state classification result containing a set of compliant components and a set of out-of-bounds component labels.

[0127] S6.3: Based on the parameter state classification results, keep the values ​​of the compliant component set unchanged, and forcibly assign the maximum allowable positive adjustment amount to the components pointing to positive boundary in the boundary component mark set, and forcibly assign the maximum allowable negative adjustment amount to the components pointing to negative boundary, so as to complete the boundary truncation processing of the original adjustment vector and generate a restricted compensation parameter vector.

[0128] Based on the parameter state classification results generated in step S6.2, the compliant component set and the out-of-bounds component marker set are used as input objects for subsequent processing. A constant value preservation operation is performed on the compliant component set to ensure that its value does not change in amplitude under the current process state, thus maintaining the normal adjustment capability of the original compensation strategy. For components in the out-of-bounds component marker set pointing to positive out-of-bounds, the maximum permissible positive adjustment amount parameter is called, and a substitution assignment operation is used to overwrite their original values ​​with this safe upper limit value, thereby eliminating the risk of exceeding the positive amplitude of the process window. For components in the out-of-bounds component marker set pointing to negative out-of-bounds, the maximum permissible negative adjustment amount parameter is called, and the same substitution assignment mechanism is used to overwrite their values ​​with this safe lower limit value, thereby eliminating the risk of exceeding the negative amplitude of the process window. The compliant component set and the corrected component set after the above processing are assembled into a new constrained compensation parameter vector in dimensional order, establishing a one-to-one correspondence with the original adjustment vector and completing boundary truncation.

[0129] By using the above processing method, the result of the previous step is transformed into a constrained compensation parameter vector with safety boundary constraints, thereby realizing dynamic amplitude limiting control of each dimension component during the real-time execution of the compensation strategy.

[0130] S6.4: Based on the constrained compensation parameter vector, a smooth transition filter is used to perform differential suppression processing on parameter jumps near the cutoff point, eliminating the step change signal caused by hard cutoff, and generating an intermediate execution compensation parameter sequence with continuity and smoothness to prevent mechanical shock from the actuator due to instruction change.

[0131] S6.5: Perform final encapsulation processing on the intermediate execution compensation parameter sequence, convert it into a standard control command format that can be recognized by the attachment actuator, output the final execution compensation parameters that meet the safety requirements of the current process window, and drive the attachment head to complete high-precision position or pressure fine-tuning actions.

[0132] The intermediate execution compensation parameter sequence is loaded with precise encapsulation task input conditions, including a multidimensional constrained compensation parameter vector processed by a smooth transition filter and a communication protocol description file for the attached actuator interface. The multidimensional constrained compensation parameter vector is reconstructed in bit width and mapped to data types according to protocol requirements, ensuring that each parameter component conforms to the numerical domain constraints and encoding format specifications of the control command. Byte order adjustment and frame structure partitioning operations are performed on the mapped data, assigning different control components to designated field positions in the protocol frame, and inserting a synchronization identifier at the beginning of the frame to ensure correct command recognition by the actuator under the high-speed bus. A CRC cyclic redundancy check code generation operation is performed based on the frame structure data. The check value is inserted at the end of the frame and the entire data is encapsulated into a standard control command data packet in a complete format. The interface driver program inserts the data packet into the actuator's command buffer, triggering an execution signal for high-precision attachment head position or pressure fine-tuning. Through the above encapsulation and transmission processing, the intermediate execution compensation parameter sequence from the previous step is transformed into standard control commands that meet process window safety requirements and can directly drive the actuator, achieving a closed-loop adaptation effect between parameter constraints and execution actions.

[0133] For example, in the implementation of an LCD screen ACF bonding production line, the intermediate execution compensation parameter sequence is a three-dimensional vector [0.012mm, [0.015mm, 0.035N], protocol is CANopen, bit width is 16-bit fixed-point format, during mapping, the position component is set to an integer value of millimeters multiplied by 100, and the pressure component is set to an integer value of Newtons multiplied by 1000. The mapping result is [1.2, [1.5, 35], byte order is arranged in high-order order, frame field is position X, position Y, pressure. Synchronization code is 0xAA55, CRC uses 16-bit polynomial. The initial register value is set to 0xFFFF, and the calculated CRC result is 0xB32F, which is placed at the end of the frame. The completed instruction packet is sent to the attachment actuator via the CANopen bus. The actuator responds in milliseconds to complete the fine-tuning actions of 0.012mm positive X displacement, 0.015mm negative Y displacement, and 0.035N pressure increase. Production line verification results show that the attachment position deviation is significantly reduced and the pressure fluctuation remains within the safe process window. The system response delay remains stable and does not exceed the preset safety threshold.

[0134] Step S7: Based on the final execution compensation parameters, the attachment actuator is driven to adjust the position or pressure of the attachment head. Simultaneously, it monitors whether the continuously generated current operating condition fingerprint falls into an unknown cluster without a safety label or whether the frequency of triggering the compensation parameter exceeding the threshold, in order to generate a boundary drift warning signal. Specifically, this includes: S7.1: Obtain the mapping relationship between the final execution compensation parameters and the control interface of the attachment actuator, and use the pulse width modulation signal generation method to perform digital-to-analog conversion processing on the final execution compensation parameters to generate an analog control voltage signal that drives the attachment head to adjust its position or apply pressure.

[0135] Based on the final execution compensation parameters encapsulated in the S6.5 sub-step, the attached actuator control protocol parsing module is invoked to extract the mapping relationship between each physical quantity component in the parameter vector and the actuator control interface port, clarifying the position control channel or pressure control channel port index corresponding to each component to be adjusted. The mapping relationship is then associated with the parameter vector to form a drive instruction element table containing the control port index, target adjustment amplitude, and control type. For each control port target adjustment amplitude in the drive instruction element table, a pulse width modulation signal generation method is used at a preset PWM base frequency. Under Hz conditions, the corresponding duty cycle value is calculated based on the product of the target adjustment amplitude and the actuator proportional coefficient. Based on the calculated duty cycle value, the digital PWM signal is converted into an analog voltage waveform via a digital-to-analog converter (DAC). A bipolar mapping strategy is then applied to map the analog voltage signal corresponding to the positive adjustment component to the positive drive polarity, and the analog voltage signal corresponding to the negative adjustment component to the reverse drive polarity. The analog control voltage signal, processed by DAC and polarity mapping, is distributed to the corresponding drive channel of the attachment actuator according to the control port index, realizing the injection of drive signals required for position fine-tuning or pressure application. Through the above DAC and PWM signal generation processing method, the final execution compensation parameters from the previous step are converted into high-precision analog control voltages that can be directly applied to the actuator, achieving precise drive control of the attachment head position or pressure.

[0136] For example, in a high-precision LCD assembly line application, the final compensation parameters include a position adjustment component of 0.15 mm and a pressure adjustment component. MPa, the control interface proportional coefficient is set to the position channel respectively. mm / V, pressure channel MPa / V, the maximum allowable control voltage value is V. At the PWM base frequency At Hz, the duty cycle calculation result is for the position channel: = Pressure channel: = After digital-to-analog conversion and polarity mapping, the position channel generates a PWM positive drive waveform with a duty cycle of 0.06 and converts it into a 0.3 V analog voltage, while the pressure channel generates a PWM positive drive waveform with a duty cycle of 0.000025 and converts it into a 0.000125 V analog voltage. Feedback tests show that the position control response delay is lower than... ms, pressure control response latency is less than The drive amplitude is stable and there is no obvious overshoot, which enables the attachment head to make precise pose and pressure adjustment under the constraints of dynamic compensation parameters.

[0137] S7.2: Based on the analog control voltage signal, drive the attachment actuator to complete the physical action. At the same time, collect the sub-pixel level displacement residual feedback from the optical positioning system and the attachment head pressure fluctuation sequence output by the pressure sensor. Use the sliding time window interception method to perform time-series alignment processing on the multi-source heterogeneous signals to generate real-time data segments containing the latest process status information.

[0138] Based on the analog control voltage signal received by the attachment actuator, the drive mechanism generates corresponding physical output force or displacement commands, causing the attachment head to perform fine-tuning actions in position or pressure according to preset compensation parameters. Sub-pixel-level displacement residual data output in real time by the optical positioning system during this process is acquired, and the timing signal of the attachment head pressure fluctuation measured by the pressure sensor is acquired simultaneously. Both types of data are recorded in continuous sampling mode. The two heterogeneous signals are input into the sliding time window interception module. The length of the time window and the sliding step size are designed based on the response characteristics of the attachment mechanism and the process cycle, ensuring that the window covers a complete attachment adjustment cycle. The timestamps of the optical displacement residual signal and the pressure fluctuation signal are aligned. A multi-source signal fusion alignment operator is used to map data with different sampling frequencies to a unified time axis, eliminating timing deviations caused by sampling rate or transmission delay. Synchronous interception and index mapping of signals within the execution window are performed, and the displacement residual sequence and pressure fluctuation sequence are combined in a matrix according to time synchronization to form a multi-dimensional data segment containing the latest process state physical quantities. By using the above-mentioned window capture and timing alignment processing method, the physical feedback signal generated by the output action of the actuator is transformed into a real-time process status data fragment with complete time consistency and physical correlation, thereby realizing the high-precision input conditions required for subsequent status fingerprint generation.

[0139] For example, on a high-precision LCD screen ACF bonding production line, the bonding actuator is composed of a piezoelectric drive component with a response time of 2ms, a process cycle of 500ms, a pressure sensor sampling frequency of 2kHz, and an optical positioning system sampling frequency of 1kHz. The analog control voltage signal amplitude range is set to 0~5V to adjust the bonding head position by ±25μm. The sliding time window length is configured to 50ms, with a step size of 10ms. Each window contains 100 points of pressure sensor data and 50 points of optical displacement data. During window alignment, the pressure signal is downsampled to 1kHz to match the optical signal using linear interpolation, and then timestamped. The aligned data is constructed in matrix form, where the first column is the displacement residual value (unit: μm), and the second column is the pressure fluctuation value (unit: N). Within a certain execution cycle, the average displacement residual matrix of the windowed data is 0.08 μm, and the average pressure fluctuation matrix is ​​0.12 N. Through the above alignment and truncation processing, this data segment can accurately reflect the synchronous changes of the attachment head fine-tuning action and process status within the cycle, providing an input signal matrix with spatiotemporal consistency and process relevance for the next step of state fingerprint encoding, significantly improving the stability and accuracy of fingerprint generation.

[0140] S7.3: Based on the real-time data fragment, the constructed lightweight state encoder is invoked to perform time-frequency joint embedding operation, which maps multidimensional physical quantities into low-dimensional vectors of a unified dimension, and generates a real-time operating condition fingerprint that represents the operating condition characteristics at the current moment through vector splicing and fusion operation.

[0141] S7.4: Based on the real-time operating condition fingerprint, perform Euclidean distance nearest neighbor search operation in the security boundary fingerprint database to calculate the distance metric between the real-time operating condition fingerprint and the center of each process state cluster, so as to determine whether the real-time operating condition fingerprint falls into the unknown cluster region without a security label and generate an unknown cluster attribution determination result.

[0142] S7.5: Based on the unknown cluster attribution determination result and the historical compensation parameter boundary record, the boundary crossing frequency within a consecutive preset number of time steps is accumulated using a sliding counter, and the boundary crossing frequency is compared with a preset threshold to generate a boundary drift warning signal indicating that the process window has drifted.

[0143] Based on the unknown cluster affiliation determination result and the historical compensation parameter out-of-bounds record, the out-of-bounds state label set of continuous time steps is imported as the input data source of the sliding counter.

[0144] The sliding time window mechanism is used to limit the number of time steps for cumulative calculation, and the window width is set to a preset continuous step size parameter to ensure that the statistical frequency covers the short-term dynamic range related to real-time process drift.

[0145] Perform step-by-step counting on the out-of-bounds status labels within the window, using the out-of-bounds event corresponding to each label as a counting unit, and sum them up to form the total out-of-bounds frequency of the current window.

[0146] A threshold comparison operation is used to compare the total value of the out-of-bounds frequency with a preset safety threshold. The safety threshold is set based on the statistical results of the maximum out-of-bounds tolerance frequency during the historical process stability period.

[0147] The result of the trigger condition being met is encapsulated into a standard format warning signal data packet and output to the boundary drift monitoring module to initiate the subsequent boundary label freezing and incremental calibration process.

[0148] By using sliding count statistics and threshold comparison processing, the unknown cluster determination results from the previous step are transformed into a quantitative frequency index that can be used for drift risk identification, thereby achieving the expected technical effect of real-time detection of process window drift trends.

[0149] For example, in a high-speed ACF (Anti-Fluorescent Film) attachment production line for LCD screens, the preset number of continuous time steps is 50, and the safety threshold is set to a tolerance frequency of 15 out-of-bounds occurrences. It is known that the out-of-bounds frequency during historically stable production phases is less than or equal to 14 times. During real-time operation, the sliding window covers the current moment and the previous 49 steps, and the cumulative frequency of the out-of-bounds status label is 17 times. ,get =2, greater than zero, meeting the trigger condition. The system encapsulates this judgment result into a binary warning frame and transmits it to the process monitoring host via the Ethernet interface, initiating the freezing of the target process status cluster label. The quality inspection module then retrieves high-definition AOI images and multi-source sensor logs from the last 200 seconds for small-sample relabeling. In this scenario, the warning signal proactively identifies the risk of process window drift caused by abnormal platform vibration, significantly improving process stability and material protection capabilities.

[0150] Step S8: If the boundary drift warning signal is detected, the safety boundary fingerprint database label of the current target process state cluster is automatically frozen, and the high-resolution AOI image of the most recent sample and the equipment log are retrieved for small-sample re-labeling to update the maximum allowable compensation amplitude label and complete the incremental calibration of the dynamic compensation parameter boundary. Specifically, this includes: S8.1: Based on the boundary drift warning signal, a locking command is triggered to perform instantaneous freezing processing on the security boundary fingerprint library tag associated with the current target process state cluster, so as to generate a process state cluster to be calibrated in a read-only protection state, to prevent the new compensation strategy from calling the wrong boundary value before the relabeling is completed.

[0151] S8.2: Based on the time index range of the process state cluster to be calibrated, the most recent preset number of sub-pixel level defect image sequences and multi-source sensor time-series records are synchronously extracted from the storage area of ​​the high-definition automatic optical inspection system and the equipment operation log database to construct a re-annotated original dataset containing visual representations and physical quantity trajectories.

[0152] S8.3: Based on the re-annotated original dataset, the sub-pixel level defect image sequence and the time-series records of multi-source sensors are spatiotemporally aligned and displayed using the human-machine collaborative quality inspection interface, and the critical safety state judgment result input by the quality inspection expert is received to generate a small sample re-annotated sample set carrying manual verification marks.

[0153] Based on the re-annotated original dataset, a quality inspection visualization interface module with human-computer interaction is invoked to synchronously align sub-pixel-level defect image sequences and multi-source sensor time-series records according to a unified time benchmark, forming a space-time mapping matrix that can be interpreted by experts. The aligned defect image sequences are loaded into a high-resolution image rendering engine, using a pixel-level difference enhancement filter to highlight process anomaly areas and simultaneously marking inflection points of sensor signal changes on the time axis, enabling experts to comprehensively determine the correlation between image features and physical quantity fluctuations. The rendered image matrix and sensor curve data are displayed in multiple layers on the same interface, employing transparency adjustment and color mapping enhancement functions to make the feature information of different data domains visually distinguishable and interconnected, thus facilitating the determination of critical safety states. Combining expert mouse clicks, region selection, and semantic label input operations, manual verification marks are established for each frame of defect image and its corresponding sensor data segment, and these marks are bound to the original data to form small-sample re-annotated data units with safety state determination fields. All data units are aggregated according to time order and operating condition category to form a small-sample re-annotated sample set, providing accurate safety state boundary criteria for subsequent extreme value statistical analysis of compensation amplitude. Through the aforementioned human-machine collaborative display and labeling processing method, the re-labeled raw data from the previous step is transformed into a small-sample re-labeled sample set with a unified three-dimensional description of time, space, and state, realizing a high-precision safety status determination data source required for compensation boundary incremental calibration. For example, in the LCD screen ACF bonding production line, the quality inspection visualization interface is called to align sub-pixel-level defect images from the most recent 50 time steps with five-channel sensor signals (pressure, temperature, tension, vibration, and displacement residuals) at a 0.5-millisecond sampling interval. During rendering, a 5×5 high-pass convolution kernel is used to enhance abnormal regions, and color gradient mapping is used on the sensor curves to distinguish signal domains. Experts select 12 image frames showing signs of edge lifting in the interface and mark the intervals where the rate of change exceeds a threshold on the corresponding pressure curves. After the system records these markers, they are bound to the original timestamps to form data units, ultimately resulting in a small-sample re-labeled sample set containing 12 critical samples and 38 safe samples. This sample set was used in subsequent steps to statistically analyze the extreme values ​​of the compensation parameters. The results significantly improved the accuracy of boundary updates and avoided inconsistencies in attachment caused by historical boundary lag.

[0154] S8.4: Based on the subset of samples marked as critical safety states in the small sample relabeled sample set, statistically analyze the extreme value distribution characteristics of their corresponding historical compensation parameters to recalculate and update the maximum allowable compensation amplitude label bound to the process state cluster to be calibrated.

[0155] S8.5: Based on the updated maximum allowable compensation amplitude label, remove the read-only protection status of the process state cluster to be calibrated, and write the new label into the security boundary fingerprint database to replace the old value, so as to complete the incremental calibration of the dynamic compensation parameter boundary and restore the boundary constraint function in the real-time attachment process.

[0156] Based on the updated maximum allowable compensation range label, the security boundary fingerprint database write interface is invoked to remove permissions from the read-only protected process state cluster to be calibrated, switching its label storage area status from read-only mode to writable mode. For the label index of this process state cluster, the updated maximum allowable compensation range label value is loaded, and the data consistency verification module confirms that the label value format and precision conform to the security boundary fingerprint database's entry specifications. A transactional write mechanism is used to overwrite the previously stored maximum allowable compensation range value with the updated label value, ensuring atomicity and uninterruptibility of the write operation in a multi-threaded environment. A label cache refresh operation is performed, replacing the old labels in the memory cache with the new labels, and the fingerprint mapping table is refreshed synchronously, allowing subsequent real-time compensation strategy calls to directly reference the latest boundary values. Label change records generated during the update process are appended to the process state cluster version control log, forming a traceable label evolution history for rollback or review when necessary. Through the above processing, the relabeling result of the previous step is transformed into a dynamic security boundary label that can be used in the real-time attachment process, realizing incremental calibration of the process state cluster and restoring boundary constraints.

[0157] For example, in an LCD screen ACF bonding production line, the original maximum allowable compensation range label value for a certain process status cluster was 0.35N, corresponding to a position adjustment of 14μm. After small-sample relabeling, the updated value is 0.38N, corresponding to a position adjustment of 15.2μm. The permission removal process switches the protection bit of the status cluster label storage area from "01" to "00", indicating writable mode. The data consistency verification module detects that the input value 0.38N conforms to the library's floating-point precision specification (two decimal places) and matches the corresponding process status cluster index. The transactional write mechanism, in a multi-threaded compensation strategy request environment, overwrites the original value with the new value and locks the write process, ensuring atomicity. After the label cache is refreshed, the real-time compensation call interface directly references the new label value of 0.38N in the next cycle, synchronously updating the position adjustment boundary. The label change record is written to the process status cluster version log. When subsequent compensation strategies call the new label value and the bonding head performs fine-tuning actions, the boundary control function significantly improves adaptability and stability.

[0158] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0159] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0160] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A real-time compensation method for ACF (Anchor Fluorescent Foliar Fabric) attachment accuracy used in LCD panel assembly, specifically including: S1: Acquire multi-source heterogeneous signals during the ACF bonding process of the LCD screen to form a raw data acquisition set containing complete process status information; S2: Based on the original data acquisition set, time-frequency joint embedding is performed independently for each type of signal, mapping physical quantities of different dimensions into low-dimensional vectors of uniform length, and generating independent channel embedding vectors that characterize the characteristics of a single signal channel. S3: After aligning the embedded vectors of independent channels by timestamp, they are spliced ​​and fused to bring the vector distance between adjacent times within the same batch closer and push the vector distance between similar working conditions between different batches closer in the latent space, thus constructing a process state fingerprint space with a compact cluster structure. S4: Based on the time segments corresponding to the safe windows in the historical yield data that do not cause ACF warping, glue overflow, conductive particle breakage or ITO circuit damage, extract the corresponding process state fingerprints and map them to each process state cluster in the process state fingerprint space to generate a safe boundary fingerprint library carrying the label of the maximum allowable compensation amplitude. S5: During the real-time attachment process, a new current operating condition fingerprint is generated every preset time interval, and the nearest neighbor search is performed in the security boundary fingerprint library to locate the target process state cluster to which the current operating condition fingerprint belongs, and the boundary value of the dynamic compensation parameter bound to the target process state cluster is obtained. S6: Input the boundary values ​​of the dynamic compensation parameters as hard upper limit constraints into the compensation strategy generation module, truncate or limit the preliminary compensation parameters calculated in real time, and generate the final execution compensation parameters that meet the safety requirements of the current process window.

2. The real-time compensation method for ACF bonding accuracy in LCD screen assembly according to claim 1, characterized in that, The raw data acquisition set includes five types of multi-source heterogeneous signals: the pressure fluctuation sequence of the attachment head, the time sequence of the pressure head temperature gradient, the ACF tape unwinding tension spectrum, the platform vibration acceleration envelope, and the sub-pixel displacement residual feedback from the optical positioning system.

3. The real-time compensation method for ACF bonding accuracy in LCD screen assembly according to claim 1, characterized in that, The compensation strategy generation module adopts an LSTM+attention encoder structure.

4. The method for real-time compensation of ACF bonding accuracy for LCD screen assembly according to claim 1, characterized in that, Step S6 is followed by: S7: Based on the final execution compensation parameters, drive the attachment execution mechanism to adjust the position or pressure of the attachment head, and at the same time monitor whether the continuously generated current working condition fingerprint falls into an unknown cluster without a safety label or triggers the compensation parameter to exceed the threshold frequency, so as to generate a boundary drift warning signal. S8: If the boundary drift warning signal is detected, the security boundary fingerprint library label of the current target process state cluster is automatically frozen and the high-definition AOI image and equipment log of the most recent sample are retrieved for small sample re-labeling to update the maximum allowable compensation amplitude label and complete the incremental calibration of the dynamic compensation parameter boundary.

5. The method for real-time compensation of ACF bonding accuracy for LCD screen assembly according to claim 1, characterized in that, Step S3 specifically includes: Based on the independent channel embedding vectors generated in the previous steps, timestamp alignment is performed on the independent channel embedding vectors to eliminate timing deviations in the multi-source heterogeneous signal acquisition process and generate a timing-aligned embedding vector set. Based on the time-aligned embedded vector set, the low-dimensional vectors of each channel are dimensionally concatenated and fused to integrate the multimodal feature information of pressure, temperature, tension, vibration and displacement residuals, and generate the original process state fingerprint vector that represents the complete instantaneous process state. Based on the original process state fingerprint vector, positive and negative sample pairs are constructed. The original process state fingerprint vectors of adjacent times within the same batch are defined as positive sample pairs, and the original process state fingerprint vectors of similar working conditions between different batches are defined as negative sample pairs. A set of sample constraint relationships for latent space optimization is established. Generate an optimized process state fingerprint latent vector based on the sample constraint relationship set; Based on the optimized process state fingerprint latent vector, high-density clustered regions in the latent space are identified, and process state clusters representing different typical stable operating conditions are automatically divided to generate a process state fingerprint space with high discriminativeness and compactness.

6. The method for real-time compensation of ACF bonding accuracy for LCD screen assembly according to claim 5, characterized in that, The process of generating an optimized process state fingerprint latent vector based on a set of sample constraint relationships specifically involves performing a latent space mapping transformation based on the set of sample constraint relationships. This is achieved by minimizing the Euclidean distance between positive sample pairs and maximizing the Euclidean distance between negative sample pairs in the latent space, thereby compressing the distribution range of similar process states and separating the distribution regions of dissimilar process states, thus generating an optimized process state fingerprint latent vector with compact cluster structure characteristics.

7. The method for real-time compensation of ACF bonding accuracy for LCD screen assembly according to claim 1, characterized in that, Step S4 specifically includes: By correlating and aligning high-definition automatic optical inspection image data from historical production processes with equipment operation log data, safe process time periods without ACF warping, glue overflow, conductive particle breakage, or ITO circuit damage are identified and converted into a set of safe window time series. Based on the safety window time series set, multi-source heterogeneous signal segments with corresponding timestamps are extracted from the original data collection set, and time-frequency joint embedding inference is performed on them to generate a safety channel embedding vector group that represents the characteristics of safety conditions. The embedded vector groups of the safety channel are spliced ​​and fused according to the timestamp to generate a safety process state fingerprint sequence that corresponds one-to-one with the safety window, and then projected into the process state fingerprint space. A subset of secure samples within a cluster is formed based on the fingerprint space of the process state. Extreme value statistical analysis is performed on the compensation parameter values ​​actually used in the historical time of all samples in the safe sample subset within the cluster. The maximum parameter amplitude that does not trigger process defects is extracted, defined as the maximum allowable compensation amplitude label, and bound to the corresponding target process state cluster to generate a safe boundary fingerprint library carrying dynamic safe boundary attributes.

8. The real-time compensation method for ACF bonding accuracy in LCD screen assembly according to claim 7, characterized in that, The process of generating a subset of secure samples within a cluster based on the process state fingerprint space specifically involves calculating the Euclidean distance between each secure process state fingerprint and the center of each process state cluster based on the compact cluster structure in the process state fingerprint space, classifying the secure process state fingerprints into the nearest target process state cluster, and forming a subset of secure samples within a cluster containing multiple secure sample points.

9. The method for real-time compensation of ACF bonding accuracy for LCD screen assembly according to claim 1, characterized in that, Step S5 specifically includes: Based on the preset time interval trigger signal, the five types of multi-source heterogeneous raw data, namely the pressure fluctuation sequence of the attachment head, the time sequence of the pressure head temperature gradient, the spectrum of ACF tape unwinding tension, the platform vibration acceleration envelope, and the sub-pixel displacement residual, are time-stamped and aligned to generate a multi-channel fused raw data frame. Perform time-frequency joint embedding transformation on each physical quantity signal in the multi-channel fused original data frame to map the high-dimensional time-series physical quantity into a low-dimensional feature vector of fixed length, and generate an independent channel embedding vector set that characterizes the characteristics of a single signal channel at the current instant. The current operating condition fingerprint is formed based on the set of independent channel embedding vectors. Call the security boundary fingerprint database, calculate the Euclidean distance or cosine similarity between the current operating condition fingerprint and the center of all historical process state clusters in the database, and locate the index of the target process state cluster with the highest matching degree. Based on the target process state cluster index, the maximum allowable compensation amplitude label value is extracted from the safety boundary fingerprint database. This value is then parsed into the dynamic compensation parameter boundary value at the current moment and output as a hard constraint condition to limit the output range of the compensation strategy to the downstream execution module.

10. The real-time compensation method for ACF bonding accuracy in LCD screen assembly according to claim 9, characterized in that, The process of forming a current operating condition fingerprint based on an independent channel embedded vector set involves using a time synchronization mechanism to perform vector splicing and fusion of the independent channel embedded vector set in channel order, and generating a comprehensive state vector containing complete current process state information through latent space projection operation, thus forming a current operating condition fingerprint that can uniquely identify the current attachment condition.

Images