Pure adhesive film quality evaluation method based on big data

Abstract

The invention relates to the technical field of pure adhesive film quality evaluation, and discloses a pure adhesive film quality evaluation method based on big data. The method comprises the following steps: acquiring a spectrum sequence, a surface texture image and environment temperature and humidity data acquired by a multi-source sensor in a production process, aligning different sampling frequency data sources through a multi-dimensional data fusion field, and performing abnormal marking on original feature data; the abnormal marking comprises adhesive film reference quality curve analysis, single-source data abnormal identification and composite verification, and the composite verification comprises time sequence fluctuation verification and spatial distribution verification; inputting the feature data into a feature quantization module to extract depth features, binding timestamps, inputting the depth features into a multi-stage evaluation transition module, performing iterative evaluation through a hierarchical analysis strategy, and outputting a result; and the dynamic adjusting and optimizing unit receives the result, adjusts and analyzes the step length parameter in real time, and judges whether the adhesive film slitting interval and the main production line need physical isolation or not.

Description

Technical Field

[0001] The present invention relates to the technical field of pure adhesive film quality evaluation, and specifically to a pure adhesive film quality evaluation method based on big data. Background Art

[0002] As a key material widely used in electronics, packaging, healthcare, and other fields, the quality stability of pure adhesive film directly impacts the performance and reliability of downstream products. With the expansion of industrial production scale and the increase in technological complexity, traditional pure adhesive film quality evaluation methods have gradually exposed significant limitations. Currently, the industry relies heavily on manual sampling and testing or single-sensor data analysis, which makes it difficult to fully reflect quality fluctuations during the film production process.

[0003] Traditional methods typically test only a single physical property of the film (such as thickness or light transmittance), ignoring the interrelationships between these properties and the combined impact of environmental factors (such as temperature and humidity) on quality. For example, during the film extrusion stage, a slight change in ambient temperature can simultaneously lead to a decrease in thickness uniformity and an increase in surface defects. Single-metric testing cannot capture this multi-dimensional quality correlation. Furthermore, differences in sampling frequencies between different sensors make it difficult to effectively integrate multi-source data, resulting in delayed and incomplete quality assessments.

[0004] Existing technologies often rely on simple threshold judgments, making it difficult to distinguish normal process fluctuations from substantive quality anomalies. For example, film thickness can fluctuate slightly during production due to equipment operating conditions. Using only fixed thresholds for judgment can easily lead to false positives or omissions. Furthermore, quality evaluation results are often static, making it impossible to dynamically adjust evaluation strategies based on real-time production data. This makes it difficult to meet the real-time quality monitoring requirements of continuous production. These limitations lead to delayed detection of quality issues during production, resulting in waste of raw materials and reduced production efficiency. Summary of the Invention

[0005] The purpose of the present invention is to provide a pure film quality evaluation method based on big data to solve the problems raised in the above background technology.

[0006] To achieve the above objectives, the present invention provides a method for evaluating the quality of pure adhesive films based on big data, the method comprising: Acquire spectral sequence data, surface texture image data, and ambient temperature and humidity data collected by multi-source sensors during the film production process. Align data sources with different sampling frequencies by establishing a multi-dimensional data fusion field, and mark abnormalities in the original film feature data. The anomaly marking includes film benchmark quality curve analysis, single-source data anomaly identification and composite verification process; The composite verification process includes time series fluctuation verification and spatial distribution verification. The time series fluctuation verification is used to detect the mutation characteristics, trend imbalance and signal decoupling phenomenon of the film thickness data. The spatial distribution verification is used to extract the correlation imbalance map between the film transmittance data and the surface defect data. The film feature data is input into the feature quantification module for deep feature extraction. The extracted feature vector is bound with a timestamp and then input into the multi-level evaluation and transition module. The multi-level evaluation and transition module configures a hierarchical analysis strategy to iteratively evaluate the film quality features and outputs the quality evaluation results. The dynamic tuning unit receives the quality evaluation result, adjusts the analysis step parameters of the multi-level evaluation shift module in real time, and determines whether the film slitting area needs to be physically isolated from the main production line.

[0007] Preferably, the multi-dimensional data fusion field is established to include a spectral distribution field and a texture energy field, and the time-varying correlation between the spectral signal and the texture feature is determined by spatiotemporal coding alignment. The spectral distribution field is constructed based on a non-uniform sampling grid discretization method to map the film transmittance data into a three-dimensional radiation field. When a sudden change in spectral intensity occurs in the spectral distribution field or an abnormality in high-frequency texture energy occurs in the texture energy field, the grid resolution level is dynamically optimized by dividing the high-resolution data grid into blocks and extracting the local neighborhood feature matrix.

[0008] Preferably, the film benchmark quality curve analysis is used to compare the thickness data, light transmittance data and surface defect data of the film to be tested to see if they are within a preset quality threshold range, and trigger an abnormal flag if any data exceeds the quality threshold range; The single-source data anomaly recognition is used to detect whether there is a characteristic mutation, distribution imbalance or data decoupling in thickness data, transmittance data or surface defect data. When a single-source data anomaly is identified, an independent anomaly report is generated and the working status of the corresponding sensor acquisition unit is verified.

[0009] Preferably, the multi-level evaluation shift module includes a thickness shift unit, a transmittance shift unit and a defect evolution unit; The thickness shifting unit receives a thickness feature vector with a timestamp, monitors thickness feature fluctuations within continuous timestamps, calculates the offset between historical thickness features and current thickness features, and the offset between predicted thickness features and current thickness features, and activates a continuous production batch abnormality marking mechanism.

[0010] Preferably, the transmittance shifting unit is configured with a transmittance prediction model, and transmittance values ​​of subsequent production nodes are predicted by the transmittance prediction model, and a gradient mutation detection algorithm is used to mark the material degradation risk; The gradient mutation detection includes predicting whether the gradient change rate of the transmittance and the actual measured value exceeds the mutation threshold, applying wavelet packet transform to decompose the transmittance signal and identify the area where the high-frequency component increases abnormally; The defect evolution unit integrates a defect classification model and a defect feature enhancement unit.

[0011] Preferably, the defect classification model processes the defect feature map output by the defect feature enhancement unit, extracts defect morphological features through spectrum texture analysis, and classifies the defects into bubble defects, impurity inclusions, or mechanical damage categories based on feature matching results; The defect feature enhancement unit is used to fuse defect image data in a multi-light source environment and perform feature optimization. The multi-light source environment includes a natural light interference pattern and an artificial lighting interference pattern. The feature optimization refers to multi-scale enhancement processing of the original defect features.

[0012] Preferably, the method further comprises collecting thickness change rate, transmittance gradient change and defect spectrum feature vector in continuous production batches; The thickness change rate is used to drive the thickness pushing unit to adjust the analysis step. When the thickness change rate per unit production cycle exceeds the maximum weight threshold, the thickness pushing unit step parameter is compressed. If it does not exceed the maximum weight threshold, the system enters the standby tuning state. The step span is adjusted according to the acceleration of the transmittance gradient change. When it is detected that the acceleration increment of the transmittance change of adjacent timestamps is greater than the historical acceleration increment and exceeds the minimum sensitivity threshold, the step ratio compression mechanism is activated; The enhanced defect features output by the defect feature enhancement unit are compared with the defect spectrum library and the step size adjustment is activated. When it is identified that the defect spectrum in a specific frequency band meets the characteristics of bubble burst or impurity eddy current, it is determined to be material degradation and the analysis step size of the defect evolution unit is adjusted.

[0013] Preferably, the method further comprises setting a step boundary controller to receive a defect evolution unit step parameter, a transmittance compression step ratio, and a thickness shift unit step parameter, and configuring a sensitivity grading strategy; By adjusting the step size boundary parameters, the buffer fault tolerance layer is activated to temporarily accommodate step size changes. If the step size exceeds the rated boundary instantaneously but the total cumulative step size does not exceed the upper limit of the buffer fault tolerance layer, the current state is maintained. When the duration of the excess limit reaches the critical threshold, it immediately switches to the minimum safe step size mode.

[0014] Preferably, the method further includes continuing to optimize the step size parameters of the multi-level evaluation shift module based on real-time production data if the adjusted thickness change rate, transmittance gradient change amount, and defect spectrum characteristics are within the stable range of the buffer fault tolerance layer; If the total amount of step size adjustment of the multi-level evaluation push module does not exceed the upper limit of the buffer fault tolerance layer and the duration exceeds the fault tolerance threshold, the step size freezing mechanism is triggered.

[0015] Preferably, the determination of whether the film slitting area needs to be physically isolated from the main production line is performed, the sealing device of the slitting area isolation door is activated to enter a closed state, a high-precision position sensor is deployed at the connection between the slitting area and the main production line, the sensor converts the mechanical control signal into a digital instruction and transmits it to the central controller, which executes the separation operation of the mechanical connector and the electrical connector of the slitting area; After the film slitting section receives the isolation instruction output by the multi-level evaluation push module, it triggers the physical isolation of the slitting section from the main production line, simultaneously closes the section ventilation duct, activates the external inflatable sealing ring expansion device, starts the gravity balance system to adjust the longitudinal and transverse counterweight blocks to maintain a horizontal state, and finally completes the slitting section isolation operation.

[0016] Compared with the prior art, the present invention has the following beneficial effects: This method achieves a comprehensive assessment of film quality by fusing data from multiple sources of sensors. By incorporating spectral sequence data, surface texture image data, and ambient temperature and humidity data into the analysis, it transcends the limitations of traditional single-metric evaluation and captures the inherent correlations between different quality characteristics. For example, changes in ambient temperature and humidity can simultaneously affect a film's light transmittance and surface smoothness. Fusion analysis of multi-source data can simultaneously reflect these correlated changes, providing a more comprehensive picture of film quality.

[0017] The anomaly marking process combines baseline quality curve analysis, single-source data anomaly identification, and composite verification to improve the accuracy of anomaly judgment. The baseline quality curve provides a dynamic reference standard for quality evaluation, avoiding misjudgments caused by fixed thresholds; single-source data anomaly identification can initially screen out potential problems; the time series fluctuation verification and spatial distribution verification in the composite verification further distinguish between normal process fluctuations and substantive quality anomalies. The time series fluctuation verification analyzes the sudden change characteristics, trend imbalance, and signal decoupling of film thickness data, and can identify quality fluctuations caused by abnormal equipment operation. The spatial distribution verification extracts the imbalance map of the correlation between transmittance and surface defects, which can identify quality problems caused by changes in local production conditions.

[0018] The combination of the Feature Quantification Module and the Multi-Level Evaluation Progression Module achieves in-depth and dynamic quality evaluation. The Feature Quantification Module deeply extracts multi-source data, imbuing quality features with richer connotations. The timestamp-bound feature vectors preserve the temporal information of quality changes. The Multi-Level Evaluation Progression Module iteratively evaluates quality features using a hierarchical analysis strategy, adjusting evaluation priorities based on the characteristics of different production stages to ensure that the evaluation results are more closely aligned with the actual production process. This hierarchical iterative evaluation method enables progressively refined quality analysis, comprehensively analyzing the film quality status from overall trends to local details.

[0019] The dynamic tuning unit makes the evaluation system adaptive. By receiving quality evaluation results and adjusting the analysis step parameters of the multi-level evaluation transition module in real time, the evaluation frequency and depth can be optimized based on production rhythm and quality fluctuations. Furthermore, based on the evaluation results, the need for physical isolation between the film slitting area and the main production line can be determined. This allows for the timely isolation of products with quality issues and reduces the possibility of defective products being passed on to subsequent processes. This dynamic adjustment mechanism closes the loop between quality evaluation and the production process, enhancing adaptability to continuous production. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is a working principle diagram of the pure film quality evaluation method based on big data according to the present invention; Figure 2 Flowchart constructed for multi-dimensional data fusion field; Figure 3 Flowchart for thickness push unit work; Figure 4 Flowchart for defect classification and enhancement. DETAILED DESCRIPTION

[0021] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0022] See also Figure 1 The present invention provides a pure film quality evaluation method based on big data, and the method includes:

[0023] Multi-source sensors in the film production process collect spectral sequence data, surface texture image data, and ambient temperature and humidity data. A multi-dimensional data fusion field is constructed to align data sources with different sampling frequencies: the spectral sequence is mapped into a spatiotemporally continuous three-dimensional radiation field, and the surface texture data is converted into an energy distribution field, with environmental data serving as an auxiliary correction factor. The fusion process uses timestamp interpolation and spatial grid resampling techniques to eliminate data lag, and a spatial encoder is used to correlate the time-varying characteristics of the spectrum and texture. When flagging the raw film feature data for anomalies, a film benchmark quality curve analysis is performed: preset thickness thresholds (±0.5μm), transmittance thresholds (≥90%), and surface defect density limits are defined; any data exceeding these limits triggers an anomaly flag. Single-source data anomaly detection uses feature mutation detection algorithms (such as the differential threshold method) to identify data transition points, Kullback-Leibler divergence analysis to detect distribution imbalances, and cointegration tests to determine data decoupling. The complex verification process must be performed simultaneously: Temporal fluctuation verification uses wavelet transforms to detect frequency mutations in thickness data, and Hodrick-Prescott filtering to decompose trend and periodic terms to identify imbalances. Spatial distribution verification constructs a transmittance-defect correlation map and quantifies the degree of regional imbalance using the Pearson correlation coefficient. Correlation failure is determined when |r| < 0.7. The feature quantification module uses a deep convolutional network to extract fused feature vectors from multi-source data, which are then bound to millisecond-level timestamps and fed into the multi-level evaluation transition module. This module's hierarchical analysis strategy implements iterative evaluation through a series of thickness transition units, transmittance transition units, and defect evolution units. The output of the previous unit serves as the input for the subsequent units, ultimately generating a four-dimensional quality score vector (thickness stability, transmittance decay rate, defect level, and comprehensive risk index). The dynamic tuning unit receives the score vector in real time and adjusts the analysis step parameters of each unit based on the score fluctuation (for example, shortening the step size from the default 500ms to 200ms). If the comprehensive risk index continuously exceeds the limit, a cut and isolation instruction is issued.

[0024] Example 1: See Figure 2The construction of the multi-dimensional data fusion field is based on the collaborative mapping mechanism of the spectral distribution field and the texture energy field. The spectral distribution field is spatially modeled through a non-uniform sampling grid discretization method: the transmission direction of the film production line is used as the time reference axis, the visible spectrum range of 400 nanometers to 700 nanometers is used as the wavelength axis, and the transmittance measurement value is used as the intensity axis to form a three-dimensional radiation field coordinate system. The non-uniformly spaced raw data points collected by the spectral sensor are reconstructed into a continuous field model using a cubic spline interpolation algorithm to eliminate breakpoints caused by differences in sensor sampling rates. The texture energy field conversion process uses a multi-scale Gabor filter bank to decompose the surface texture image in 8 directions and 16 scales, extracting the high-frequency energy components at each scale, and using their modulus values ​​as the energy field intensity values. The spatiotemporal coding alignment adopts a dual-stream architecture: the spectral timestamp sequence and the texture image acquisition frame rate are synchronized at the millisecond level through the Lagrange interpolation method. The physical space position relies on the preset laser positioning markers to construct a unified coordinate system. The markers are evenly distributed at the edge of the production line at an interval of 20 mm. The coordinate transformation model calculates the Euclidean distance between the spectral acquisition point and the adjacent marker points and maps it to the corresponding pixel area of ​​the texture image.

[0025] When the spectral distribution field detects a sudden change in intensity within a local wavelength interval, the mutation is determined based on a threshold where the intensity difference between adjacent sampling points exceeds 15% of the baseline value. The trigger condition for high-frequency energy anomalies in the texture energy field is set to exceed 2 standard deviations of the historical mean. For such abnormal areas, high-resolution data grid blocks are implemented with an initial block accuracy of 0.1 mm × 0.1 mm. The spectral intensity gradient vectors and texture energy Hessian matrix eigenvalues ​​within the blocks are extracted. The grid resolution optimization rule is based on the matrix condition number: when the condition number exceeds 10^4, the grid accuracy is increased to 0.05 mm, and the area is recursively segmented through 12 layers until the condition number drops to a reasonable range.

[0026] The film benchmark quality curve analysis is modeled based on steady-state production line data. The thickness fluctuation benchmark curve utilizes dynamic boundary control: the root mean square (RMS) value of 100 consecutive batches of thickness data is calculated, and the floating threshold range is set within the mean ±3 standard deviation. An aging compensation factor is introduced within this range based on equipment operation time (a 0.1% threshold width is added every 8 hours of production). The transmittance benchmark model uses a piecewise linear decay function, dividing the test into segments every 10 meters of film length, with the lower transmittance threshold for each segment set at 90%. A defect distribution benchmark heatmap is generated using a Gaussian kernel density estimation algorithm, with a maximum defect density of 3 per 1 square centimeter. During real-time quality assessment, an anomaly is flagged when the dynamic time warping distance between the thickness data and the benchmark curve exceeds a tolerance of 0.3. An alarm is triggered when the measured transmittance value falls below the lower threshold for the segment. A distribution imbalance is identified when the KL divergence between the defect density heatmap and the benchmark exceeds 0.5.

[0027] Single-source data anomaly detection incorporates a three-layer parallel detection mechanism. Thickness data anomaly detection utilizes a chained Z-score validation method: 50 consecutive sampling points are standardized, and a feature mutation is determined when the absolute Z-score value of five consecutive points exceeds 3. Transmittance data distribution is verified through goodness-of-fit analysis, establishing an ideal transmittance distribution model (gamma distribution α=9, β=0.5). Distribution imbalance is determined when the coefficient of determination between the actual measured data and the model is less than 0.85. Surface defect data decoupling verification utilizes a modified Granger causality test, calculating the conditional probability of defect data using thickness and transmittance as independent variables. Data decoupling is confirmed when the significance level p-value is greater than 0.05.

[0028] After the independent exception report is generated, the sensor diagnostic protocol is initiated. The thickness sensor zero drift calibration is performed in three steps: first, 30 seconds of zero-point data is collected in a no-load state. If the standard deviation exceeds 0.03 microns, the compensation circuit is automatically activated. The transmittance sensor reference plate is calibrated using a standard reflective plate overlay method. When the reference plate reflectance value deviates from the baseline by ±1.5%, the sensor calibration parameters are rewritten. The defect camera white balance is corrected using a three-color calibration plate. When the color difference ΔE value calculated in the CIELab color space exceeds 4, the optical system refocuses and the gamma curve is reset. All diagnostic processes are completed between equipment operations, and a single diagnostic cycle does not exceed 500 milliseconds.

[0029] Physical isolation of abnormal areas is initiated simultaneously. When the spectral distribution field identifies a localized abnormality, control instructions drive the optical stabilization platform to the coordinate location, activating the high-resolution re-inspection module (spectral resolution increased to 1 nanometer). Abnormal areas in the texture energy field trigger the darkroom shading system, deploying a three-panel shading panel within 300 milliseconds to surround the detection area and eliminate ambient light interference. During data grid optimization, sub-pixel positioning compensation is performed on the boundary points of the reconstructed grid, keeping the position error within ±0.15 mm.

[0030] A mapping relationship is established between sensor diagnostic results and anomaly types. When a sudden change in the thickness Z-score is detected, accompanied by sensor zero drift, a sensor hardware fault indicator is added to the anomaly report. When an imbalance in transmittance distribution is associated with a reference plate calibration failure, the optical window is deemed contaminated and the ultrasonic cleaning device is activated. When defect data decoupling and white balance anomalies occur simultaneously, the ambient light interference type is noted in the report. All anomaly report data packets are encapsulated in ISO-GP encoding format and transmitted to a central database for batch traceability.

[0031] Example 2: See Figure 3The thickness shift unit receives a thickness feature vector with millisecond-level timestamps. This vector consists of 12-dimensional feature data, including time-domain statistical indicators (mean, variance, and skewness), frequency-domain transform components (amplitude of the first six harmonics of the Fourier transform), and morphological features (density of local curvature extreme points). A triple offset analysis is performed within the observation window formed by consecutive timestamps: the historical offset calculates the Euclidean distance between the current thickness vector and the average of the feature vectors of the previous 30 production batches, using a sliding window approach to update the reference value every 5 seconds. The instantaneous offset measures the directional change of the feature vectors between adjacent timestamps, quantifying the intensity of the fluctuation using cosine similarity differences. The predicted offset, based on an ARIMA(2,1,1) model, predicts the feature value of the next sampling point based on the previous 60 seconds of thickness data, using the Manhattan distance from the actual measured value as the deviation criterion. The offset thresholds are set to 1.2 for the historical offset, 0.4 for the instantaneous offset, and 0.8 for the predicted offset. Exceeding any of these thresholds activates the continuous anomaly flagging mechanism, generating a binary event code containing the batch number, offset type, and magnitude of the offset.

[0032] The transmittance shift unit incorporates a long-short-term memory prediction network, with the input layer processing normalized spectral features. Three channels (450 nm, 550 nm, and 650 nm) are selected as dominant wavelengths, and intensity values ​​are scaled to the [0, 1] range using a min-max method. The network structure comprises 128 hidden units, with a time step of 10 seconds. It outputs a transmittance prediction sequence for the next five production nodes (at 2-second intervals). Gradient abrupt change detection utilizes a two-stage verification mechanism. The primary module calculates the derivative difference between the predicted and measured curves per unit time. A primary alarm is generated when the absolute difference exceeds a threshold of 0.05 per millisecond for three consecutive seconds. Secondary verification utilizes wavelet packet transform analysis, employing a three-layer decomposition using the Daubechies4 wavelet basis, to extract the energy integral of detail components with frequencies above 100 Hz. When the high-frequency energy increases by 200% of the baseline level within a 5-second observation window, a spectral high-frequency spike event is flagged and the coordinates of the abnormal region are located.

[0033] The defect evolution unit simultaneously runs a classification model and feature enhancement module. The multi-light source environment processing mechanism distinguishes between two interference modes: Under natural light interference (illuminance meter readings exceeding 50,000 lux), an adaptive histogram equalization algorithm is used to reconstruct the image brightness distribution, dynamically adjusting the grayscale compression curve with a 0.5-second cycle. Under artificial sodium light interference, a digital bandpass filter with a central wavelength of 590 nanometers and a bandwidth of 20 nanometers is deployed to suppress optical noise interference in specific frequency bands. Multi-scale feature enhancement is achieved through three-channel parallel convolution: the original defect image is simultaneously fed with Sobel edge detection kernels at three scales: 5×5 pixels, 10×10 pixels, and 20×20 pixels. The output feature maps are weighted and fused to generate a high dynamic range feature image. The weight coefficients are dynamically assigned based on the image signal-to-noise ratio (SNR < 30dB): large scale weight 0.7, SNR ≥ 30dB, small scale weight 0.6.

[0034] The transmittance prediction model is trained using a time series cross-validation strategy. Data is partitioned into 8-hour production periods, with 70% of the period data used as the training set, 20% as the validation set, and 10% as the test set. Network training parameters include a learning rate of 0.001, a batch size of 64, and a maximum number of iterations of 500. Early stopping is activated after the validation set loss shows no decrease for 10 consecutive iterations. The prediction output format is a 5×3 matrix, with row vectors corresponding to time nodes and column vectors containing the predicted value and the lower and upper bounds of the confidence interval.

[0035] When the defect classification model processes high-dynamic-range feature images, a hybrid analysis method is used to extract spectral texture features. Spatial spectrum analysis performs a fast Fourier transform on the defect area, calculates the radial energy distribution function, and records the mainlobe peak position. The gray-level co-occurrence matrix constructs displacement vectors at 0, 45, and 90 degrees, extracting the second-order angular moment and contrast eigenvalues. Morphological contour recognition uses an 8-neighborhood boundary tracking algorithm to generate closed curves and calculate the curvature change rate and area-to-perimeter ratio. The feature matching process uses a preset template library: the bubble defect template requires the gradient variance of the annular contour to be greater than 300 and the fundamental spectrum component to account for more than 65%; the impurity doping template matches irregular shapes with an area dispersion greater than 0.35; and the mechanical damage template requires the phase angle standard deviation of the Fourier descriptor corresponding to linear edges to be less than 0.1.

[0036] A dynamic parameter feedback mechanism continuously optimizes the analysis process. The root mean square (RMS) change in thickness is calculated per minute. When the instantaneous rate exceeds the 0.3 microns per second threshold, the thickness shift unit analysis step size is compressed from the baseline value of 1 second to 0.3 seconds. The transmittance gradient acceleration is input using the difference in the maximum slope between the previous and subsequent batches. If the acceleration increment exceeds 0.02 per millisecond squared and reaches 1.5 times the historical average, the transmittance module step size is proportionally reduced by 50%. The defect spectrum feature vector performs 12th-order linear predictive coding on the frequency band below 200 Hz. When the energy proportion of the 150-200 Hz subband suddenly increases by 25 percentage points, the defect analysis step size is locked to 0.5 seconds, and the material degradation diagnosis subroutine is initiated.

[0037] System maintenance protocols are automatically executed during equipment idle periods. The transmittance prediction network undergoes weekly model updates, loading the latest seven days of production data to fine-tune network parameters, with fine-tuning time limited to 30 minutes. The defect template library undergoes monthly version iterations, with newly added abnormal samples annotated by experts and imported into the template library. The version number is updated synchronously with the production batch number. Sobel convolution kernel parameters are calibrated quarterly, with edge detection accuracy verified using standard test patterns. If deviation exceeds 5%, the convolution weight matrix is ​​rewritten. All maintenance records are stored in a secure storage area as encrypted logs for at least three years.

[0038] Example 3: See Figure 4 The defect classification model processes high dynamic range image data generated by the feature enhancement unit, with an input image spatial resolution of 2048×2048 pixels. The model performs a three-order feature extraction process: spatial spectrum analysis selects a 128×128 pixel window from the defect area for fast Fourier transform, calculates the integral value of the radial energy distribution function, and records the main lobe peak frequency coordinates; the grayscale co-occurrence matrix features construct displacement vectors in the three directions of 0 degrees, 45 degrees, and 90 degrees, with a fixed moving step of 2 pixels, and calculates the angular second moment parameter, contrast parameter, and inverse disparity parameter; morphological contour extraction uses an 8-neighborhood boundary tracking algorithm to generate a closed curve chain code, and calculates the curvature change rate parameter, area-to-perimeter ratio parameter, and convex hull defect depth of the closed curve. Feature fusion is achieved through the weighted matching formula: ; In the formula Indicates the comprehensive matching degree, is the frequency domain similarity (correlation coefficient between the actual spectrum and the template spectrum), is the shape matching degree (the inverse of the difference in contour curvature), is the texture correlation (cosine distance of grayscale feature vectors). Represents spectral weight, morphological weight and texture weight respectively, with the default values ​​set to 0.4, 0.3 and 0.3. When the value exceeds 0.85, it is judged as this type of defect.

[0039] The template library construction rules are based on historical defect sample statistics: the bubble defect template requires a ring structure with a diameter of 5-50 pixels, a contour gradient variance greater than 300, and a fundamental spectrum component (0-50Hz) energy contribution exceeding 65%. The impurity doping template corresponds to an irregular shape of 2-100 pixels, an area dispersion index greater than 0.35, and a spectrum exhibiting broadband noise characteristics (within the 100-500Hz band, a standard deviation of fluctuation greater than 15dB). The mechanical damage template constraints require a linear structure length exceeding 20 pixels, an aspect ratio greater than 8:1, and a Fourier descriptor phase angle standard deviation less than 0.1 radian. The template matching process uses a sliding window search strategy, moving the detection window in 64-pixel steps across the entire image, and performing NMS non-maximum suppression on overlapping areas.

[0040] The defect feature enhancement unit simultaneously receives 3D dynamic parameters from successive production batches: The thickness change rate parameter is calculated by dividing the root mean square change in thickness of the current batch by the sampling interval, with a minimum calculation period of 60 seconds. The transmittance gradient change parameter is derived from the maximum slope difference between the transmittance curves of adjacent batches. The defect spectrum feature vector consists of 20 coefficients, and the frequency domain components are extracted through a fast Fourier transform within a 25 millisecond time window. These parameters form a feedback control loop: The thickness shift unit receives the thickness change rate parameter and, when the instantaneous rate exceeds the critical value of 0.3 microns per second, compresses the analysis step length parameter from the baseline value of 1 second to 0.3 seconds. When the rate parameter is in the range of 0.1-0.3 microns per second, the dynamic monitoring mode is maintained, and the step length is recalibrated every 5 seconds.

[0041] The transmittance shift unit responds to the transmittance gradient change acceleration parameter (i.e. the second-order derivative of the gradient change). The parameter acquisition cycle is set to 15 seconds. If the acceleration increment in adjacent cycles is satisfy If the acceleration exceeds 1.5 times the historical average acceleration, the step-size ratio compression mechanism is activated, reducing the current step size by a factor of 0.5. Synchronously activate high-frequency component verification: In the wavelet packet decomposition layer of the transmittance signal, energy envelope detection is performed on the 100-200Hz sub-band. If the energy proportion in this frequency band suddenly increases by more than 25 percentage points, it is marked as a potential signal of material degradation.

[0042] The defect evolution unit adjusts its operating mode based on the defect spectrum feature vector. Spectral feature analysis focuses on the 150-200Hz frequency band. When the integrated energy value in this band exceeds 30% of the total energy in the entire frequency band, feature pattern matching is performed: the bubble burst feature requires the spectrum peak full width at half maximum to be greater than 50Hz and no harmonic attenuation; the impurity eddy current feature requires the second harmonic amplitude attenuation slope to exceed 12dB / oct. After a successful match, the analysis step size is locked to 0.5 seconds and the degradation diagnosis subroutine is initiated. This program rescans the defect area three times within 0.2 seconds and verifies the signal authenticity by comparing the rescan results.

[0043] The light source interference suppression system implements a hierarchical process. In natural light overexposure mode (illuminance meter reading ≥ 50,000 lux), non-uniform brightness compensation is activated: the image is divided into 16×16 grid blocks, and adaptive histogram equalization is performed independently on each block. The gamma correction coefficient is dynamically calculated based on the regional median brightness. Sodium lamp interference suppression uses a two-stage filtering structure: the pre-optical filter has a physical cutoff wavelength of less than 580nm, and the post-digital bandpass filter has a passband range of 590±10nm and a stopband attenuation of no less than 40dB. A noise assessment module is added to the multi-scale feature enhancement process: when the image signal-to-noise ratio is less than 30dB, the large-scale convolution weight of the Sobel edge detection kernel is increased to 0.7 to suppress high-frequency noise interference.

[0044] The system's parameter linkage update mechanism operates on a fixed cycle. The thickness change rate parameter is recalculated every 30 seconds, driving the thickness shift unit step size calibration. The transmittance acceleration parameter and the defect spectrum parameter are synchronously updated every 15 seconds, linking the transmittance shift unit and the defect evolution unit. During each linkage update, the central controller verifies parameter validity: if the thickness sensor signal is interrupted, the rate parameter update is frozen; if the transmittance signal is abnormal, the historical mean prediction mode is switched; and if the defect image is missing, the previous state analysis step size is maintained. The calibration log records the timestamps, parameter values, and control decisions of all linkage events. The data blocks are stored using AES-256 encryption, and the retention period is synchronized with the production batch lifecycle.

[0045] Example 4: The step boundary controller receives adjustment parameters input by three units: the step parameter of the defect evolution unit is identified as δ1, the compression ratio parameter of the transmittance unit is δ2, and the step parameter of the thickness shift unit is δ3. The parameter sampling period is fixed at 1 second, and the central processing unit records the time series values ​​of each parameter. The sensitivity grading strategy is based on the instantaneous change amplitude of the parameter relative to the baseline value: when the δ change is less than 20%, it is marked as a first-level sensitive state, and only the parameter log is stored; 20% to 50% changes are classified as a second-level sensitive state, triggering the parameter review mechanism; when the change exceeds 50%, it enters a third-level sensitive state, and the boundary control protocol is immediately activated. The following table shows the evolution of some parameter states recorded during continuous production: .

[0046] Boundary control operates within preset parameter ranges: thickness step δ3 is limited to the range of [0.2s, 1.5s], transmittance compression ratio δ2 is maintained in the range of [0.3, 1], and defect analysis step δ1 is controlled within [0.4s, 2s]. A two-level protection mechanism is implemented in the buffer fault tolerance layer. In the transient offset tolerance mechanism, if a parameter briefly exceeds a boundary but the accumulated absolute offset within five consecutive sampling periods does not exceed the upper limit (0.5s for thickness, 0.3s for transmittance, and 0.6s for defects), the system maintains its current operating state. Sustained offset protection uses a critical duration threshold (15s for thickness, 10s for transmittance, and 8s for defects). Upon timeout, the system is forced to switch to the preset minimum safe step mode (δ3 = 0.2s, δ2 = 0.3s, and δ1 = 0.4s).

[0047] While the buffer tolerance layer is stable (within a parameter change rate of ±5% per minute), the step size optimization module continuously fine-tunes the thickness step size. The thickness step size adjustment is positively correlated with the thickness change acceleration; every 0.1 micron per square second of acceleration corresponds to a 0.005-second step size change. The transmittance compression ratio is dynamically adjusted based on real-time prediction accuracy. When the coefficient of determination between the predicted and measured values ​​exceeds 0.8, the compression ratio is increased by 0.02. The defect step size is adjusted inversely based on the highest classification confidence level, decreasing by 0.03 seconds for every 10 percentage point increase in confidence. All adjustment parameters are updated every 2 seconds.

[0048] The step size adjustment total monitor maintains a running buffer, continuously calculating the cumulative parameter change over a time period of Δt. This buffer is subject to two conditions: the cumulative total parameter change (|Δδ1| + |Δδ2| + |Δδ3|) does not exceed 1.5, and the stable state lasts for 300 seconds. When both conditions are met, the step size freeze mechanism is immediately triggered: all parameter adjustments are suspended for 120 seconds, during which time the sensor secondary calibration process is activated. This calibration process involves three parallel operations: the thickness sensor performs a three-point calibration (measuring 0μm / 50μm / 100μm standard blocks), the transmittance unit verifies the reflectance of a reference plate, and the defect camera captures a standard test chart for MTF resolution testing.

[0049] The parameter review mechanism is activated in the second-level sensitivity state. The review process first searches the historical database, comparing the current parameter combination with historical anomaly patterns. If a similar parameter combination is found in the past 24 hours of data (δ fluctuation trend similarity >85%), the corresponding historical treatment plan is loaded. If no matching record is found, a triple verification is initiated: the median offset of the previous 60 seconds is recalculated for the thickness unit, the prediction model is rerun twice for the transmittance unit, and the feature vector is re-extracted for the defect unit. The review takes less than 500 milliseconds, the output result is marked as "Verified Parameters", and operation is resumed.

[0050] Three levels of sensitivity trigger a coordinated response from the equipment. When the transmittance unit δ2 exceeds the lower limit of 0.3, the lighting system immediately increases the illumination by 2000 lux to improve the signal-to-noise ratio. When the thickness unit δ3 exceeds 1.3 seconds, the production line transmission speed is automatically reduced by 20%. When the defect unit δ1 reaches 1.8 seconds, the optical image stabilization system is triggered to lock the camera platform. The coordinated signal is transmitted via industrial Ethernet, with response latency less than 100 milliseconds. All state transition records are written to a ring storage area in chronological order, with the most recent 2000 records maintained in non-volatile storage.

[0051] Example 5: The cutting interval isolation system responds to the isolation instruction output by the multi-level evaluation push module. The isolation instruction contains a 12-bit binary operation code and a risk level identifier. After the instruction arrives at the isolation control unit, it triggers the position sensor array to start. Six groups of high-precision laser ranging sensors deployed at the connection between the cutting interval and the main production line perform synchronous scanning. The sensor spacing is configured to be 250 mm, and the measurement accuracy reaches ±0.1 mm. The scanning data generates three types of control instructions: the displacement encoding converts the three-dimensional space coordinates into 12-bit binary data words; the velocity vector instruction calculates the XYZ three-axis velocity components according to the mechanical separation direction; the emergency braking instruction generates an 8-bit braking code when an excessive displacement is detected, and the code bitmap corresponds to different braking levels.

[0052] After receiving the command, the central controller executes the mechanical connector separation operation. The hydraulic drive unit retracts the connecting pin at a constant speed of 0.5 m / s, and the pin position is monitored in real time by a magnetostrictive displacement sensor. During the separation process, each pin is equipped with a dual feedback mechanism: the compensation oil circuit is activated when the actual displacement curve deviates from the theoretical path by more than 0.3 mm; after the pin is fully separated, the mechanical locking mechanism is triggered to fix the separation state. The simultaneous disconnection of the electrical connector uses zero-current switching technology: the thyristor device disconnects the 380V power circuit at the zero crossing of the sine wave, and the arc generated during the switching process lasts no longer than 0.5 milliseconds. The uninterruptible power supply system completes the power supply switch within 20 milliseconds after disconnection, and the output voltage ripple is controlled within ±2%.

[0053] The ventilation duct shutoff system utilizes a coordinated dual-valve design. A stepper motor drives the main butterfly valve to rotate 90 degrees to the closed position within 0.8 seconds, with a valve core displacement encoder verifying the rotation angle in real time. An auxiliary solenoid valve maintains pipeline pressure balance during the main valve's operation and closes synchronously with the main valve's position. After the valve seal is closed, the valve seal is pressurized to 0.5 MPa, and a pressure sensor continuously monitors the pressure distribution on the sealing surface.

[0054] The external inflatable sealing system operates at the joint. The polyurethane sealing ring expands to a diameter of 150 mm in 0.2 seconds, maintaining a stable internal pressure within a range of 1.2 ± 0.1 MPa. The inflation process utilizes graded pressure control: initially, inflation is performed at a rate of 50 kPa / ms to 80% of the target pressure, followed by PID control to gradually approach the set pressure. Eight pressure monitoring points are evenly distributed around the circumference of the sealing ring; a localized air replenishment mechanism is triggered when the pressure difference between any two points exceeds 10%.

[0055] The gravity balance system operates continuously during the isolation process. The XY-axis servo motor drives the longitudinal counterweight to move within a travel range of ±500 mm, and the movement speed is dynamically adjusted according to the rate of change of the inclination angle: when the horizontal inclination sensor reading deviation reaches 0.005 degrees, the counterweight moves at a speed of 5 mm / s; when the deviation exceeds 0.008 degrees, the speed increases to 10 mm / s. The Z-axis lateral counterweight system is driven by a hydraulic push rod with a displacement resolution of 2 microns. The system collects 32 inclination data of each corner point of the platform every 50 milliseconds and generates a counterweight leveling solution through a weighted algorithm. The final horizontal control target of the platform is a full-plane error of less than 0.01 degrees.

[0056] The pressure maintenance system within the slitting zone automatically activates upon isolation. A centrifugal fan operates at constant power, injecting clean air into the zone through a high-efficiency air filter. A differential pressure sensor continuously monitors the pressure difference between the zone and the external environment. The controller dynamically adjusts the fan speed via a frequency converter to maintain a positive pressure differential within the 20-25 Pa range. If the particle counter detects an excess concentration of 0.5-micron particles, the fan speed is increased by 20% for forced ventilation.

[0057] The status of all actuators is transmitted in real time via the industrial bus. The control unit collects 128-byte status data packets from each actuator every 20 milliseconds, encapsulates them into a unified format after CRC32 verification, and transmits them to the central control console. The data display interface is divided into three monitoring areas: the mechanical connection status area displays the latch displacement curve and the locking mechanism status icon; the fluid pressure area dynamically updates the pressure values ​​of the sealing ring points; the environmental parameter area updates key parameters such as horizontality and air pressure difference in real time. The system has a three-level alarm mechanism: when a single parameter deviates from the normal value by 10%, a yellow warning is triggered; when it deviates by 20%, it turns to an orange alarm; and when all three parameters are abnormal, a red alarm is triggered and the emergency backup system is activated.

[0058] The system activates emergency protocols to handle abnormal situations. If mechanical separation is blocked for more than three seconds, the backup hydraulic unit automatically takes over. If the inflation delay of the inflatable seal reaches 0.5 seconds, the auxiliary air source is switched. If the horizontal balance adjustment timeout occurs, the platform is locked and the vibration compensation mechanism is activated. A complete event log is generated for each isolation operation, detailing the execution time points and equipment response parameters of each stage. The log data is retained for at least the entire production line lifecycle.

[0059] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0060] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for evaluating the quality of pure adhesive films based on big data, characterized in that: The following steps are included: Acquire spectral sequence data, surface texture image data, and ambient temperature and humidity data collected by multi-source sensors during the film production process. Align data sources with different sampling frequencies by establishing a multi-dimensional data fusion field, and mark abnormalities in the original film feature data. The anomaly marking includes film benchmark quality curve analysis, single-source data anomaly identification and composite verification process; The composite verification process includes time series fluctuation verification and spatial distribution verification. The time series fluctuation verification is used to detect the mutation characteristics, trend imbalance and signal decoupling phenomenon of the film thickness data. The spatial distribution verification is used to extract the correlation imbalance map between the film transmittance data and the surface defect data. The film feature data is input into the feature quantification module for deep feature extraction. The extracted feature vector is bound with a timestamp and then input into the multi-level evaluation and transition module. The multi-level evaluation and transition module configures a hierarchical analysis strategy to iteratively evaluate the film quality features and outputs the quality evaluation results. The dynamic tuning unit receives the quality evaluation result, adjusts the analysis step parameters of the multi-level evaluation shift module in real time, and determines whether the film slitting area needs to be physically isolated from the main production line.

2. The method for evaluating the quality of pure adhesive films based on big data according to claim 1, characterized in that ,The establishment of the multi-dimensional data fusion field needs to include the spectral distribution field and the texture energy field. The time-varying correlation between the spectral signal and the texture feature is determined through the ,spatial coding alignment. The spectral distribution field is constructed based on the ,nonuniform sampling grid discretization method to map the film transmittance data into a ,three-dimensional radiation field. When a sudden change in spectral intensity occurs in the spectral distribution field or an abnormality in high-frequency texture energy occurs in the texture energy field, the grid resolution level is dynamically optimized by dividing the high-resolution data grid into blocks and extracting the local neighborhood feature matrix.

3. The method for evaluating the quality of pure adhesive films based on big data according to claim 2, characterized in that ,The film benchmark quality curve analysis is used to compare the thickness data, light transmittance data and surface defect data of the ,film to be tested to see if they are within a preset quality threshold ,interval. If any data exceeds the quality threshold interval, an abnormal mark ,is triggered; The single-source data anomaly recognition is used to detect whether there is a characteristic mutation, distribution imbalance or data decoupling in thickness data, transmittance data or surface defect data. When a single-source data anomaly is identified, an independent anomaly report is generated and the working status of the corresponding sensor acquisition unit is verified.

4. The method for evaluating the quality of pure adhesive films based on big data according to claim 3, characterized in that ,The multi-level evaluation shift module includes a thickness shift unit, a transmittance shift unit and a defect evolution unit; The thickness shifting unit receives a thickness feature vector with a timestamp, monitors thickness feature fluctuations within continuous timestamps, calculates the offset between historical thickness features and current thickness features, and the offset between predicted thickness features and current thickness features, and activates a continuous production batch abnormality marking mechanism.

5. The method for evaluating the quality of pure adhesive films based on big data according to claim 4, characterized in that The transmittance shift unit is configured with a transmittance prediction model, which predicts the transmittance values ​​of subsequent production nodes through the transmittance prediction model, and uses a gradient mutation detection algorithm to mark the material degradation risk; The gradient mutation detection includes predicting whether the gradient change rate of the transmittance and the actual measured value exceeds the mutation threshold, applying wavelet packet transform to decompose the transmittance signal and identify the area where the high-frequency component increases abnormally; The defect evolution unit integrates a defect classification model and a defect feature enhancement unit.

6. The method for evaluating the quality of pure adhesive films based on big data according to claim 5, characterized in that ,The defect classification model processes the defect feature map output by the defect feature ,enhancement unit, extracts the defect morphological features through ,spectral texture analysis, and classifies it into bubble defects, ,impurity inclusion or mechanical damage categories based on the ,feature matching results; The defect feature enhancement unit is used to fuse defect image data in a multi-light source environment and perform feature optimization. The multi-light source environment includes a natural light interference pattern and an artificial lighting interference pattern. The feature optimization refers to multi-scale enhancement processing of the original defect features.

7. The method for evaluating the quality of pure adhesive films based on big data according to claim 6, characterized in that ,It also includes collecting the thickness change rate, transmittance gradient change and defect spectrum feature vector in continuous production batches; The thickness change rate is used to drive the thickness pushing unit to adjust the analysis step. When the thickness change rate per unit production cycle exceeds the maximum weight threshold, the thickness pushing unit step parameter is compressed. If it does not exceed the maximum weight threshold, the system enters the standby tuning state. The step span is adjusted according to the acceleration of the transmittance gradient change. When it is detected that the acceleration increment of the transmittance change of adjacent timestamps is greater than the historical acceleration increment and exceeds the minimum sensitivity threshold, the step ratio compression mechanism is activated; The enhanced defect features output by the defect feature enhancement unit are compared with the defect spectrum library and the step size adjustment is activated. When it is identified that the defect spectrum in a specific frequency band meets the characteristics of bubble burst or impurity eddy current, it is determined to be material degradation and the analysis step size of the defect evolution unit is adjusted.

8. The method for evaluating the quality of pure adhesive films based on big data according to claim 1, characterized in that ,It also includes setting the step boundary controller to receive the defect evolution unit step parameters, transmittance compression step ratio and thickness shift unit step parameters, and configuring the sensitivity grading strategy; By adjusting the step size boundary parameters, the buffer fault tolerance layer is activated to temporarily accommodate step size changes. If the step size exceeds the rated boundary instantaneously but the total cumulative step size does not exceed the upper limit of the buffer fault tolerance layer, the current state is maintained. When the duration of the excess limit reaches the critical threshold, it immediately switches to the minimum safe step size mode.

9. The method for evaluating the quality of pure adhesive films based on big data according to claim 8, characterized in that , also includes if the adjusted thickness change rate, transmittance gradient change and defect spectrum characteristics are in the stable range of the buffer fault tolerance layer, then continue to optimize the step parameters of the multi-level evaluation push module based on real-time production data; If the total amount of step size adjustment of the multi-level evaluation push module does not exceed the upper limit of the buffer fault tolerance layer and the duration exceeds the fault tolerance threshold, the step size freezing mechanism is triggered.

10. The method for evaluating the quality of pure adhesive films based on big data according to claim 9, characterized in that , the judgment is made as to whether the film slitting area and the main production line need to be physically isolated, the slitting area isolation door starts the sealing device to enter the closed state, and a high-precision position sensor is deployed at the connection between the slitting area and the main production line. The sensor converts the mechanical control signal into a digital instruction and transmits it to the central controller, which executes the separation operation of the mechanical connector and the electrical connector of the slitting area; After the film slitting section receives the isolation instruction output by the multi-level evaluation push module, it triggers the physical isolation of the slitting section from the main production line, simultaneously closes the section ventilation duct, activates the external inflatable sealing ring expansion device, starts the gravity balance system to adjust the longitudinal and transverse counterweight blocks to maintain a horizontal state, and finally completes the slitting section isolation operation.

Images