Online detection method for crystallinity of PET (Polyethylene Terephthalate) film by utilizing dielectric property

Abstract

The invention discloses an online detection method for the crystallinity of a PET (Polyethylene Terephthalate) film by utilizing dielectric properties. The online detection method comprises the following steps: acquiring original dielectric data and real-time temperature data, acquiring space-time sampling data through parallel acquisition and processing, performing noise identification and removal processing on the data, and performing contact compensation processing to obtain high-quality dielectric characteristic data; performing temperature-frequency coupling effect separation and crystal form feature recognition on the high-quality dielectric feature data; establishing a multi-dimensional feature space through polarization mechanism quantization processing and microstructure feature extraction; carrying out crystallinity prediction by fusing a physical model and a deep learning method; and finally, obtaining a final crystallinity value through consistency check and dynamic correction. According to the method, high-precision online detection of the crystallinity of the PET film is realized through a multi-level data processing and model fusion strategy, and the method has relatively high anti-interference capability and prediction accuracy.

Description

Technical Field

[0001] The present invention belongs to the field of detecting the crystallinity of polymer materials, and particularly relates to an on-line detecting method for the crystallinity of a PET film by using dielectric properties. Background Art

[0002] Polyethylene terephthalate (PET) films are widely used in the fields of electronics, packaging, and optics due to their excellent mechanical properties, electrical insulation, and dimensional stability. The properties of PET films largely depend on their crystallinity, which not only affects the mechanical properties and thermal stability of the materials but also determines their gas barrier properties and optical transparency. Although traditional off-line detection methods such as differential scanning calorimetry (DSC) and X-ray diffraction (XRD) have high accuracy, they cannot meet the requirements of real-time quality control in modern production. Therefore, developing a reliable on-line detection method is of great significance for improving the production efficiency and product quality of PET films.

[0003] Currently, on-line detection methods based on dielectric properties have received extensive attention due to their non-contact and fast response characteristics. Existing technologies mainly adopt single-frequency point measurement or simple frequency scanning methods, and estimate by establishing an empirical relationship between the dielectric constant or the tangent value of the loss angle and the crystallinity. Some researchers use multi-frequency point measurement to improve the detection accuracy and introduce a temperature compensation algorithm to reduce the environmental impact. There are also studies that attempt to establish a correlation between dielectric response and crystal structure by establishing a simple physical model. However, these methods generally have problems such as low detection accuracy and limited applicable conditions.

[0004] The current technology has the following specific problems: First, in the data acquisition link, existing methods are difficult to effectively handle the measurement errors caused by the change of contact interface impedance. Especially on high-speed production lines, the dynamic change of the contact state between the film and the electrode seriously affects the measurement stability. Second, in terms of data processing, there is currently a lack of effective methods to separate the combined effects of temperature effect, orientation effect, and crystal form effect on dielectric response, resulting in the measurement results being interfered by multiple factors. Third, in terms of structural characterization, existing methods cannot accurately identify and quantitatively characterize the dynamic evolution process of α crystal form, β crystal form, and their mixed state. Especially when the crystallinity changes rapidly, it is difficult to capture the transient characteristics of the microstructure. Fourth, in terms of modeling prediction, a single empirical model or physical model is difficult to accurately describe the complex relationship between dielectric response and crystal structure. Especially when considering multi-scale structural characteristics, the prediction accuracy often cannot meet the industrial requirements. In addition, existing methods have insufficient robustness in dealing with abnormal working conditions and interference conditions, and lack a reliable prediction result evaluation and correction mechanism. Summary of the Invention

[0005] Objective of the invention: To provide an on-line detection method for the crystallinity of PET films using dielectric properties, in order to solve at least one technical problem existing in the prior art.

[0006] Technical solution: An on-line detection method for the crystallinity of PET films using dielectric properties, comprising the following steps:

[0007] S1. Obtain the original dielectric data and real-time temperature data, and obtain the spatio-temporal sampling data and sampling temperature data through parallel acquisition and processing; identify and remove noise from the spatio-temporal sampling data to obtain denoised data; perform contact compensation processing on the denoised data and sampling temperature data to obtain high-quality dielectric characteristic data;

[0008] S2. Separate the temperature-frequency coupling effect and identify the crystal form characteristics of the high-quality dielectric characteristic data to obtain a decoupled characteristic vector;

[0009] S3. Based on the decoupled characteristic vector, establish a multi-dimensional feature space through polarization mechanism quantization processing and microstructure feature extraction to obtain multi-dimensional feature space data;

[0010] S4. Based on the multi-dimensional feature space data, fuse a pre-configured physical model and a deep learning method for crystallinity prediction to obtain a preliminary crystallinity prediction value;

[0011] S5. Based on the preliminary crystallinity prediction value, obtain the final crystallinity value through consistency test and dynamic correction.

[0012] According to one aspect of the present application, step S1 is further as follows:

[0013] S11. Obtain the original dielectric data and real-time temperature data; construct sampling parameters according to the preset sampling interval and sampling position; perform parallel acquisition and processing on the original dielectric data and real-time temperature data to generate spatio-temporal sampling data and sampling temperature data;

[0014] S12. Based on the spatio-temporal sampling data, calculate the wavelet energy distribution characteristics, construct noise characteristic data; process the noise characteristic data using an adaptive threshold method to obtain the processed noise characteristic data; based on the processed noise characteristic data, perform denoising processing on the spatio-temporal sampling data to obtain denoised data;

[0015] S13. Obtain the contact state parameters in real time; based on the contact state parameters and sampling temperature data, construct a contact impedance model, calculate the compensation coefficient and combine with the denoised data to generate high-quality dielectric characteristic data.

[0016] According to one aspect of the present application, step S2 is further as follows:

[0017] S21. Based on high-quality dielectric characteristic data, construct a relationship matrix of temperature and frequency, calculate the temperature sensitivity coefficient and the frequency response function; based on the temperature sensitivity coefficient and the frequency response function, perform temperature-frequency coupling effect separation processing on the high-quality dielectric characteristic data to generate decoupled dielectric data;

[0018] S22. Based on the decoupled dielectric data, construct crystal form characteristic parameters, calculate the dielectric relaxation characteristics and the mixed state characteristics of the α crystal form and the β crystal form; according to the dielectric relaxation characteristics and the mixed state characteristics, calculate the crystal form distribution to generate crystal form characteristic data; where α and β represent different crystal forms;

[0019] S23. Based on the decoupled dielectric data and the crystal form characteristic data, calculate the molecular chain orientation function, construct a relationship model between orientation and dielectric properties; based on the relationship model, perform orientation effect separation processing on the decoupled dielectric data to generate decoupled eigenvectors.

[0020] According to one aspect of the present application, step S3 is further as follows:

[0021] S31. Based on the decoupled eigenvectors, construct a polarization component matrix including electronic polarization, ionic polarization and orientation polarization; calculate the weights and intensities of the polarization components in the polarization component matrix to generate polarization characteristic data;

[0022] S32. Based on the polarization characteristic data and the crystal form characteristic data, construct structural characteristic parameters, calculate the crystal size characteristics, the crystal integrity characteristics and the interface characteristics to generate microstructural characteristic data;

[0023] S33. Based on the microstructural characteristic data, construct a high-dimensional feature space and perform non-linear dimensionality reduction processing to generate multi-dimensional feature space data.

[0024] According to one aspect of the present application, step S4 is further as follows:

[0025] S41. Based on the multi-dimensional feature space data, construct and solve a physical constraint equation to generate a physical model prediction value;

[0026] S42. Based on the multi-dimensional feature space data, perform data processing through a deep neural network and an attention mechanism to generate a deep learning prediction value;

[0027] S43. Calculate the confidence weights of the physical model prediction value and the deep learning prediction value respectively; based on the confidence weights, perform weighted fusion processing on the physical model prediction value and the deep learning prediction value to generate a preliminary crystallinity prediction value.

[0028] According to one aspect of the present application, step S5 is further as follows:

[0029] S51. Calculate the historical consistency score and physical constraint score based on the preliminary crystallinity prediction value, and generate consistency scoring data through weighted processing;

[0030] S52. Based on the preliminary crystallinity prediction value and the consistency scoring data, construct a correction function; based on the correction function, perform dynamic correction processing on the preliminary crystallinity prediction value to generate the final crystallinity value.

[0031] According to one aspect of the present application, step S11 is further as follows:

[0032] S111. Obtain the original dielectric data and real-time temperature data, calculate the difference sequence, construct the probability density distribution function, determine and optimize the sampling step size, and generate the sampling interval parameter;

[0033] S112. Divide the measurement space based on the sampling interval parameter and the original dielectric data, calculate the signal gradient and curvature, determine the basic grid and perform local encryption to generate the sampling position data;

[0034] S113. Based on the sampling position data, construct a task priority queue, group the acquisition tasks, calculate the execution time window; based on the execution time window, construct a clock synchronization benchmark, and perform multi-channel data acquisition to generate multi-channel acquisition data;

[0035] S114. Based on the multi-channel acquisition data, extract the timestamp information, construct a unified time benchmark, calculate the time deviation and use the cubic spline function to generate the aligned acquisition data;

[0036] S115. Calculate the local statistical features of the aligned acquisition data, identify the outliers, and construct the correlation matrix; integrate the correlation matrix with the real-time temperature data correspondingly to generate the spatio-temporal sampling data and the sampling temperature data.

[0037] According to one aspect of the present application, step S12 is further as follows:

[0038] S121. Group the spatio-temporal sampling data according to the frequency, perform multi-scale decomposition to obtain the scale coefficients; calculate the energy values under each scale coefficient and generate the energy distribution characteristics;

[0039] S122. Based on the energy distribution characteristics, calculate the local fluctuation index, identify the fluctuation abnormal interval, extract the time-frequency characteristic parameters, and generate the noise characteristic data;

[0040] S123. Perform clustering analysis on the noise characteristic data, calculate the energy concentration degree, construct the noise determination criterion, and generate the noise determination result;

[0041] S124. Based on the noise determination result, perform time-frequency decomposition on the spatio-temporal sampling data, calculate the local correlation coefficient, construct the signal reconstruction weight, and generate the preliminary noise reduction data;

[0042] S125. Calculate the local smoothness index of the signal based on the preliminary noise-reduced data, identify the residual noise points and perform local correction to obtain the correction result; integrate the preliminary noise-reduced data and the correction result to generate the noise-reduced data;

[0043] S126. Align the sampling temperature data with the noise-reduced data in terms of time, and output the aligned noise-reduced data and sampling temperature data.

[0044] According to one aspect of the present application, step S22 is further as follows:

[0045] S221. Based on the decoupled dielectric data, extract the characteristic frequency points of the dielectric spectrum, calculate the dielectric constant and loss value, and generate the crystal form parameter data; combine it with the decoupled dielectric data, calculate the shape parameter of the dielectric loss peak, construct the time-temperature equivalent curve of the loss peak, and generate the relaxation characteristic data;

[0046] S222. Based on the relaxation characteristic data, decompose the relaxation spectrum into basic relaxation units, calculate the orientation correlation function of the α crystal form, and generate the α crystal form characteristic data;

[0047] S223. Based on the relaxation characteristic data, identify the characteristic frequency interval of the β crystal form, construct the local order parameter, and generate the β crystal form characteristic data;

[0048] S224. Calculate the overlap degree of the α and β crystal form characteristic data, identify the mixed state region, construct the order parameter tensor and calculate the stability index, and generate the mixed state characteristic data;

[0049] S225. Based on the α crystal form, β crystal form and mixed state characteristic data, construct a multi-dimensional characteristic space, calculate the distribution density function, identify the characteristic clusters and calculate the transition probability, and generate the preliminary distribution data;

[0050] S226. Based on the preliminary distribution data, calculate the crystal form correlation degree, construct the free energy landscape, identify and optimize the local distribution abnormal regions, and obtain the optimized local distribution data; integrate the preliminary distribution data and the optimized local distribution data to generate the crystal form characteristic data.

[0051] According to one aspect of the present application, step S32 is further as follows:

[0052] S321. Based on the polarization characteristic data and the crystal form characteristic data, extract the polarization time constant sequence and calculate the polarization relaxation time distribution, and generate the structure parameter data;

[0053] S322. Based on the structure parameter data, calculate the characteristic length distribution at each scale, construct the scale correlation function and the hierarchical structure, and generate the size characteristic data;

[0054] S323. Calculate the order parameter of the local structure based on the size feature data and crystal form feature data, construct the spatial distribution function, extract the defect positions and types, and generate the integrity feature data;

[0055] S324. Calculate the interfacial energy based on the integrity feature data, construct the orientation distribution function and extract the interfacial curvature and roughness, and generate the interface feature data;

[0056] S325. Calculate the interfacial stress distribution based on the interface feature data, construct the stress-orientation coupling relationship, extract the stress concentration regions, and generate the interfacial stress data;

[0057] S326. Construct the correlation matrix of characteristic parameters based on the size feature data, integrity feature data, interface feature data and interfacial stress data, calculate the coupling strength, and generate the microstructure feature data.

[0058] Advantageous effects: The present invention ensures the quality of the original data, realizes the separation of multiple factors such as temperature, crystal form and orientation, and establishes a multi-dimensional characterization space including polarization characteristics and microstructure characteristics; integrates the advantages of physical models and deep learning methods, and ensures the reliability of the final results through consistency tests and dynamic corrections; not only solves the accuracy problem of online detection of the crystallinity of PET films, but also provides an interpretable structure-property relationship analysis method, which has important value for improving the level of product quality control. Description of the Drawings

[0059] Figure 1 is a flowchart of the present invention.

[0060] Figure 2 is a flowchart of step S1 of the present invention.

[0061] Figure 3 is a flowchart of step S2 of the present invention.

[0062] Figure 4 is a flowchart of step S3 of the present invention.

[0063] Figure 5 is a flowchart of step S4 of the present invention.

[0064] Figure 6 is a flowchart of step S5 of the present invention. Detailed Embodiments

[0065] The following describes the present application in more detail with reference to specific embodiments. As Figure 1 shown, the present application proposes an online detection method for the crystallinity of PET films using dielectric properties, including the following steps:

[0066] S1. Obtain the original dielectric data and real-time temperature data, and through parallel data acquisition and processing, obtain spatio-temporal sampling data and sampling temperature data; perform noise identification and removal on the spatio-temporal sampling data to obtain denoised data; adopt contact compensation to process the denoised data and sampling temperature data to obtain high-quality dielectric characteristic data; wherein the original dielectric data includes dielectric constant and dielectric loss data at a predetermined number of frequency points.

[0067] S2. Based on the high-quality dielectric characteristic data, construct the relationship between temperature and frequency; based on the relationship between temperature and frequency, separate the coupling effect of the high-quality dielectric characteristic data to obtain decoupled dielectric data; perform crystal form characteristic identification on the decoupled dielectric data to obtain crystal form characteristic data; based on the decoupled dielectric data and crystal form characteristic data, separate the orientation effect to obtain decoupled characteristic vectors.

[0068] S3. Based on the decoupled characteristic vectors, through polarization mechanism quantization processing, obtain polarization characteristic data; according to the polarization characteristic data and crystal form characteristic data, extract microstructure characteristics to obtain microstructure characteristic data; perform characteristic space reconstruction on the microstructure characteristic data to obtain multi-dimensional characteristic space data.

[0069] S4. Based on the multi-dimensional characteristic space data, calculate the physical model prediction value through a pre-configured physical model; based on the multi-dimensional characteristic space data, adopt a deep learning method to obtain the deep learning prediction value; fuse the physical model prediction value and the deep learning prediction value to obtain the preliminary crystallinity prediction value.

[0070] S5. Based on the preliminary crystallinity prediction value, perform consistency test to obtain consistency scoring data; according to the preliminary crystallinity prediction value and the consistency scoring data, perform dynamic correction to obtain the final crystallinity value.

[0071] As Figure 2 shown, according to one aspect of the present application, step S1 is further as follows:

[0072] S11. Obtain the original dielectric data and real-time temperature data; according to the preset sampling interval and sampling position, construct sampling parameters; based on the sampling parameters, perform parallel acquisition and processing on the original dielectric data and real-time temperature data to generate spatio-temporal sampling data and sampling temperature data; wherein the original dielectric data includes dielectric constant and dielectric loss data at a predetermined number of frequency points.

[0073] S12. Based on the spatio-temporal sampling data, calculate the wavelet energy distribution characteristics; according to the wavelet energy distribution characteristics, construct noise characteristic data; adopt an adaptive threshold method to process the noise characteristic data to obtain the processed noise characteristic data; based on the processed noise characteristic data, perform denoising processing on the spatio-temporal sampling data through a signal transformation method to obtain denoised data.

[0074] S13. Obtain the contact state parameters in real time; construct a contact impedance model based on the contact state parameters and the sampled temperature data, and calculate the compensation coefficient; generate high-quality dielectric characteristic data based on the compensation coefficient and the noise-reduced data.

[0075] In an embodiment of the present application, an original dielectric data sequence D(t) is obtained, which includes {f[i], ε'[i], tanΔ[i]}, i = 1...n, and a sampling matrix H(t) is constructed, which includes the following elements: a frequency sampling interval vector h1, a time sampling interval vector h2, and a spatial sampling position vector h3, to obtain a spatio-temporal sampling data matrix M(t). Based on the spatio-temporal sampling data matrix M(t), the wavelet energy entropy E = -∑(p[i]log(p[i])) is calculated, where p[i] is the energy probability of the i-th scale; a noise characteristic map N(f, t) is constructed, the noise is identified by applying an adaptive threshold method, and the Hilbert-Huang transform is used to remove the noise, to obtain a noise-reduced data matrix Mc(t). Based on the noise-reduced data matrix Mc(t) and the contact state parameter λ(t), a contact impedance model Z(t) = R(t) + jX(t) is established, the compensation coefficient K(t) = f(Z(t), λ(t)) is calculated, and compensation is applied: Mf(t) = Mc(t) × K(t), and high-quality dielectric characteristic data Mf(t) is output. Where f[i] is the frequency point, ε'[i] is the real part of the dielectric constant corresponding to the frequency point, tanΔ[i] is the tangent of the dielectric loss angle corresponding to the frequency point, R(t) is the resistance component in the contact impedance model, j is the imaginary unit, X(t) is the reactance component in the contact impedance model, and f( ) is the function of the compensation coefficient.

[0076] In another embodiment of the present application, the noise-reduced data is Y(t, s) = X(t, s) - N(t, s)·M(t, s), where Y(t, s) is the noise-reduced data, N(t, s) is the noise characteristic function, and M(t, s) is the noise mask function.

[0077] The contact impedance compensation coefficient is C(x, y, t) = α(T)·R(p)·Z(f); where C(x, y, t) is the compensation coefficient at the position (x, y) and time t; α(T) = exp(-E_a / kT) is the temperature correction term; R(p) = (1 - exp(-βp)) / (1 + γp) is the pressure response function; Z(f) = Z_0 / (1 + jωτ) is the frequency characteristic function; E_a is the contact activation energy; k is the Boltzmann constant; T is the absolute temperature; p is the contact pressure; β, γ are the pressure coefficients; Z_0 is the standard contact impedance; τ is the characteristic relaxation time; ω is the angular frequency.

[0078] Contact compensation processing: Z(t, s) = Y(t, s)·C(p, T, f); where Z(t, s) is the compensated high-quality dielectric characteristic data; C(p, T, f) is the compensation coefficient function; p is the contact pressure; T is the temperature; f is the measurement frequency.

[0079] In this embodiment, high-quality dielectric characteristic data is obtained through parallel data acquisition and multi-level data processing strategies. Specifically, first, the original dielectric data and temperature data are processed through parallel acquisition to ensure the spatio-temporal synchronization and integrity of the data; then, the wavelet energy distribution characteristics are used for adaptive noise identification and removal to improve the signal-to-noise ratio of the signal; finally, compensation processing is performed through the contact impedance model to effectively eliminate the measurement errors caused by changes in the contact state. This embodiment not only improves the reliability of data acquisition but also enhances the measurement accuracy through compensation processing. Especially in the face of various interference factors during the dynamic production process of PET films, it can adaptively adjust the processing strategy to ensure the quality of the output data and provide a reliable data basis for subsequent crystallinity analysis.

[0080] According to one aspect of the present application, step S11 is further:

[0081] S111. Obtain the original dielectric data and real-time temperature data, calculate the difference sequence of the data at adjacent time points; construct the probability density distribution function of the difference sequence; determine the initial sampling step according to the variance of the probability density distribution function; adjust the initial sampling step through an iterative optimization method until the correlation coefficient between adjacent sampling points is less than a preset threshold to generate the sampling interval parameter;

[0082] S112. Based on the sampling interval parameter and the original dielectric data, divide the measurement space into a predetermined number of sub-regions; calculate the gradient value and curvature value of the signal in each sub-region; determine the basic grid size according to the gradient value to obtain the preliminary grid; perform local encryption and sparse processing on the preliminary grid using the curvature value to obtain the processed grid; based on the processed grid, perform secondary division on the encrypted region to obtain the refined grid; integrate the information of the preliminary grid, the processed grid, and the refined grid to generate the sampling position data;

[0083] S113. Based on the sampling position data and the original dielectric data, construct a task priority queue; group the acquisition tasks in the task priority queue according to the spatial proximity of the sampling position data to obtain the grouped acquisition tasks; based on the grouped acquisition tasks, calculate the execution time window for each group of tasks to generate the task scheduling sequence; construct a clock synchronization benchmark between multiple channels, and based on the clock synchronization benchmark and the task scheduling sequence, perform data acquisition to generate multi-channel acquisition data;

[0084] S114. Extract the timestamp information of each data point based on the multi-channel acquisition data and real-time temperature data to construct a unified time reference; calculate the time deviation between the data of each channel based on the unified time reference, and construct a piecewise continuous cubic spline function; calculate the data values at the unified time point using the cubic spline function to generate aligned acquisition data.

[0085] S115. Calculate the local statistical features of the aligned acquisition data, including the mean and standard deviation; identify outliers based on the local statistical features, and construct an association matrix in the time and space dimensions; integrate the association matrix with the real-time temperature data correspondingly to generate spatio-temporal sampling data and sampling temperature data.

[0086] In an embodiment of the present application, the spatio-temporal sampling data is X(t, s) = ∑[D(t_i, s_j)·W(t - t_i, s - s_j)]; where D(t_i, s_j) is the original dielectric data; W(t, s) is the sampling weight function; t_i is the time sampling point; s_j is the space sampling point.

[0087] This embodiment realizes efficient and reliable data acquisition. Specifically, first, determine the optimal sampling interval by analyzing the probability density distribution of the data difference sequence; then, perform spatial grid division based on the gradient value and curvature value to realize the optimal configuration of the sampling position; then, ensure the timing accuracy of data acquisition through the task priority queue and multi-channel clock synchronization; finally, generate high-quality spatio-temporal sampling data through spline function interpolation and outlier identification. This embodiment not only improves the efficiency of data acquisition, but also ensures the representativeness and reliability of the sampling data through multiple optimizations.

[0088] According to an aspect of the present application, step S12 is further as follows:

[0089] S121. Group the spatio-temporal sampling data according to the frequency to obtain grouped data; perform multi-scale decomposition on each group of grouped data to obtain different scale coefficients; calculate the energy values under each scale coefficient, and construct an energy distribution matrix; calculate the energy proportion of each scale coefficient based on the energy distribution matrix to generate energy distribution features.

[0090] S122. Calculate the local fluctuation index of each scale coefficient based on the energy distribution features; construct the cumulative distribution function of the local fluctuation index to identify the fluctuation abnormal interval; extract the time-frequency characteristic parameters of the fluctuation abnormal interval, and construct the mapping relationship between the time-frequency characteristic parameters and the preset background noise to generate noise characteristic data.

[0091] S123. Perform clustering analysis on the noise feature data to obtain noise categories; based on the noise categories, calculate the energy concentration degree of each type of noise; construct a noise determination criterion based on the energy concentration degree, calculate the noise probability of each frequency point according to the noise determination criterion, and generate a noise determination result;

[0092] S124. Based on the noise determination result, perform time-frequency decomposition on the spatio-temporal sampling data to obtain a time-frequency decomposition result; based on the time-frequency decomposition result, calculate the local correlation coefficient of each time-frequency point; based on the local correlation coefficient, construct a signal reconstruction weight; apply the signal reconstruction weight to the time-frequency decomposition result, perform inverse transformation to reconstruct the signal, and generate preliminary noise reduction data;

[0093] S125. Based on the preliminary noise reduction data, calculate the local smoothness index of the signal, identify residual noise points; construct a correction function containing neighborhood information, perform local correction on the residual noise points to obtain a correction result; integrate the preliminary noise reduction data and the correction result to generate noise reduction data;

[0094] S126. Align the sampled temperature data with the noise reduction data in time, keep the sampled temperature data unchanged, and output the aligned noise reduction data and the sampled temperature data.

[0095] In this embodiment, through multi-scale energy analysis and adaptive noise reduction processing, efficient noise identification and removal are achieved. First, calculate the energy distribution characteristics through multi-scale decomposition; then identify the noise characteristics based on the fluctuation index; then determine the noise categories and determination criteria through clustering analysis; then perform signal reconstruction through time-frequency decomposition and local correlation coefficient calculation; finally, eliminate the residual noise through local correction. This embodiment can not only effectively identify and remove different types of noise, but also maintain the integrity of the effective signal through a multi-level processing strategy, improving the accuracy of subsequent analysis.

[0096] According to one aspect of the present application, step S13 is further as follows:

[0097] S131. Receive the contact state parameter and the sampled temperature data, calculate the spatial uniformity index of the pressure distribution, establish the correspondence between the pressure distribution and the contact area, discretize the contact area into multiple micro-region units, calculate the effective contact rate of each micro-region, and generate contact feature data.

[0098] S132. Receive the contact feature data and the sampled temperature data, calculate the equivalent circuit parameters of each micro-region, establish a parallel equivalent circuit including contact resistance and contact capacitance, calculate the influence coefficient of temperature on the circuit parameters, and integrate all parameters to generate impedance model parameters.

[0099] S133. Receive the impedance model parameters and noise reduction data, calculate the standard response curve under the ideal contact state, extract the deviation characteristics between the actual measurement data and the standard curve, construct the mapping relationship between the deviation characteristics and the compensation factor, and generate the preliminary compensation coefficient.

[0100] S134. Receive the preliminary compensation coefficient, calculate the spatial distribution characteristics of the compensation coefficient, identify the compensation abnormal area, construct a correction function based on physical constraints, correct the compensation coefficient of the abnormal area, and generate the optimized compensation coefficient.

[0101] S135. Receive the optimized compensation coefficient and noise reduction data, process the data by frequency segmentation, calculate the compensation weight for each frequency band, perform segmented compensation operations, and perform continuous smoothing processing on the compensation results to generate the compensated processed data.

[0102] S136. Receive the compensated processed data, calculate the quality evaluation indexes of the data, including the signal-to-noise ratio, consistency coefficient, and physical rationality parameter, screen the valid data according to the evaluation indexes, integrate all the processing results, and generate high-quality dielectric characteristic data.

[0103] In this embodiment, through accurate contact characteristic modeling and compensation processing, effective correction of measurement errors is achieved. First, the pressure distribution is analyzed to establish a micro-region contact model; then a parallel equivalent circuit considering the temperature influence is constructed; then the compensation coefficient is calculated through the standard response curve; then the compensation is optimized through physical constraints; finally, high-quality data is obtained through segmented compensation and smoothing processing. This embodiment not only considers the spatial distribution characteristics of the contact state, but also improves the measurement accuracy through temperature compensation, effectively solving the measurement error problem caused by the change of the contact state.

[0104] As Figure 3 shown, according to one aspect of the present application, step S2 is further:

[0105] S21. Based on the high-quality dielectric characteristic data, construct the relationship matrix between temperature and frequency, calculate the temperature sensitivity coefficient and the frequency response function; based on the temperature sensitivity coefficient and the frequency response function, perform temperature-frequency coupling effect separation processing on the high-quality dielectric characteristic data to generate decoupled dielectric data;

[0106] S22. Based on the decoupled dielectric data, construct the crystal form characteristic parameters, calculate the dielectric relaxation characteristics and the mixed state characteristics of the α crystal form and the β crystal form; according to the dielectric relaxation characteristics and the mixed state characteristics, calculate the crystal form distribution, and generate the crystal form characteristic data; where α and β represent different crystal forms;

[0107] S23. Calculate the molecular chain orientation function based on the decoupled dielectric data and crystal form characteristic data, and construct a relationship model between orientation and dielectric properties; based on the relationship model, perform orientation effect separation processing on the decoupled dielectric data to generate decoupled eigenvectors.

[0108] In an embodiment of the present application, high-quality dielectric characteristic data Mf(t) and temperature data T(t) are obtained, a temperature-frequency coupling matrix C(f, T) = {ε'(f, T), tanΔ(f, T)} is constructed, and the temperature sensitivity coefficient is calculated: ST = ΨC / ΨT, where Ψ is the partial derivative; the frequency response function is calculated: RF(f) = C(f, T) / C(f0, T); decoupling calculation is performed: Cd(f, T) = C(f, T) - ST·ΔT - RF(f), and the decoupled dielectric data Cd(f, T) is output. Based on the decoupled dielectric data Cd(f, T), a crystal form eigenvector is constructed: V(α, β) = {v1, v2, v3}, where v1 is the dielectric relaxation characteristic of the α crystal form; v2 is the dielectric relaxation characteristic of the β crystal form; v3 is the mixed state characteristic; the crystal form distribution function is calculated: P(α, β) = g(V(α, β)), and the crystal form characteristic data P(α, β) is output. Based on the decoupled dielectric data Cd(f, T) and the crystal form characteristic data P(α, β), the molecular chain orientation function is calculated: O(θ) = ∫P(θ)cos 2 θdθ; construct the orientation-dielectric relationship: D(O) = h(Cd, O(θ)), separate the orientation effect: Cf(f, T) = Cd(f, T) / D(O), and output the decoupled eigenvector F(t). Where ε'(f, T) is the real part of the dielectric constant at frequency f and temperature T; tanΔ(f, T) is the tangent value of the dielectric loss angle at frequency f and temperature T; C(f0, T) is the temperature-frequency coupling matrix at the reference frequency f0 and temperature T, S is the temperature sensitivity coefficient, ΔT is the temperature change, g( ) is a function for calculating the crystal form distribution function; P(θ) is the probability density function of the molecular chain orientation function, θ is the orientation angle of the molecular chain, dθ is the infinitesimal change of the orientation angle, h( ) is a function for constructing the orientation-dielectric relationship, and Cf(f, T) is the decoupled dielectric data at frequency f and temperature T.

[0109] In another embodiment of the present application, the temperature-frequency coupling effect separation processing is specifically: ε(T, f) = ε_0(f)·exp(-E(f) / kT); where ε(T, f) is the dielectric constant, ε_0(f) is the reference dielectric constant, E(f) is the frequency-dependent activation energy, k is the Boltzmann constant, and T is the absolute temperature.

[0110] The specific crystal form characteristic parameters are constructed as follows: χ(α, β) = K_α·F_α(f) + K_β·F_β(f); where χ(α, β) is the crystal form characteristic function; K_α and K_β are the crystal form weight coefficients; F_α(f) and F_β(f) are the α and β crystal form characteristic spectra.

[0111] In this embodiment, by establishing a temperature-frequency coupling effect separation model and a crystal form characteristic recognition method, the decoupling of various influencing factors in dielectric data is achieved. First, a relationship matrix of temperature and frequency is constructed, and by calculating the temperature sensitivity coefficient and the frequency response function, the temperature-frequency coupling effect is effectively separated; then, the dielectric relaxation characteristics and the characteristics of the mixed state of the α and β crystal forms are identified, and the crystal form distribution information is accurately obtained; finally, by establishing a relationship model between the molecular chain orientation and the dielectric properties, the separation of the orientation effect is achieved. This embodiment not only improves the accuracy of crystallinity measurement, but also can separately obtain the independent influences of factors such as temperature, crystal form, and orientation, providing important information for in-depth understanding of the crystallization behavior of PET films.

[0112] According to one aspect of the present application, step S22 is further as follows:

[0113] S221. Based on the decoupled dielectric data, extract the characteristic frequency points of the dielectric spectrum; calculate the dielectric constant and loss value at the characteristic frequency points, construct a parameter matrix including the crystal form discrimination threshold, and generate crystal form parameter data; based on the crystal form parameter data and the decoupled dielectric data, calculate the positions of the dielectric loss peaks at different frequencies; extract the shape parameters of the positions of the dielectric loss peaks, including the peak width and peak height, construct the time-temperature equivalent curve of the loss peak, and generate relaxation characteristic data;

[0114] S222. Based on the relaxation characteristic data, decompose the relaxation spectrum into a predetermined number of basic relaxation units; calculate the characteristic time and intensity of each basic relaxation unit, and extract the response data in the characteristic relaxation time interval of the α crystal form; based on the response data, calculate the orientation correlation function of the α crystal form, and generate α crystal form characteristic data;

[0115] S223. Based on the relaxation characteristic data, identify the characteristic frequency interval of the β crystal form, and calculate the relaxation intensity distribution within the characteristic frequency interval of the β crystal form; based on the relaxation intensity distribution, construct the local order parameter of the β crystal form, extract the conformational characteristic parameters of the β crystal form, and generate β crystal form characteristic data;

[0116] S224. Calculate the overlap degree of the α crystal form characteristic data and the β crystal form characteristic data, and based on the overlap degree, identify the mixed state region; construct the order parameter tensor of the mixed state region; based on the order parameter tensor, calculate the stability index of the mixed state, and generate mixed state characteristic data;

[0117] S225. Based on the α-crystal form characteristic data, β-crystal form characteristic data, and mixed-state characteristic data, construct a multi-dimensional characteristic space; calculate the distribution density function in the multi-dimensional characteristic space to identify the characteristic clusters of each crystal form; calculate the transition probability between the characteristic clusters to generate preliminary distribution data;

[0118] S226. Based on the preliminary distribution data, calculate the crystal form correlation degree of adjacent regions to construct the free energy landscape of crystal form transformation; based on the free energy landscape, identify and optimize the locally abnormal distribution regions to obtain the optimized local distribution data; integrate the preliminary distribution data and the optimized local distribution data to generate crystal form characteristic data.

[0119] In an embodiment of the present application, the crystal form characteristic data χ(f, T) = w_α·S_α(f, T) + w_β·S_β(f, T); where χ(f, T) is the comprehensive crystal form characteristic index; S_α(f, T) = A_α·exp[-(f - f_α) 2 / 2σ_α 2 is the α-crystal form spectrum characteristic; S_β(f, T) = A_β·exp[-(f - f_β) 2 / 2σ_β 2 is the β-crystal form spectrum characteristic; w_α, w_β are weight coefficients; f_α, f_β are characteristic frequencies; σ_α, σ_β are spectral width parameters; A_α, A_β are amplitude coefficients.

[0120] This embodiment realizes the precise analysis of different crystal form components in the PET film. First, establish a crystal form discrimination parameter system through dielectric spectrum characteristic frequency analysis; then extract the characteristic information of the α-crystal form and β-crystal form respectively through relaxation spectrum decomposition; then analyze the overlapping regions of the two crystal forms to determine the mixed-state characteristics; finally, establish a crystal form distribution model through multi-dimensional characteristic space analysis. This embodiment realizes the quantitative characterization of the α-crystal form orientation correlation through the relaxation time distribution, depicts the conformational characteristics of the β-crystal form through the local order parameter, describes the stability of the mixed state through the order parameter tensor, and realizes the prediction of crystal form transformation through the free energy landscape analysis. This not only provides the accurate quantification of crystal form components, but also reveals the kinetic mechanism of crystal form transformation, providing theoretical guidance for process optimization.

[0121] As Figure 4 shown, according to one aspect of the present application, step S3 is further:

[0122] S31. Based on the decoupled eigenvectors, construct a polarization component matrix including electronic polarization, ionic polarization, and orientation polarization; calculate the weights and intensities of each polarization component in the polarization component matrix to generate polarization characteristic data;

[0123] S32. Based on the polarization feature data and crystal form feature data, construct structural feature parameters, calculate the crystal size feature, crystal integrity feature, and interface feature, and generate microstructure feature data;

[0124] S33. Based on the microstructure feature data, construct a high-dimensional feature space; based on the data in the high-dimensional feature space, perform non-linear dimensionality reduction processing to generate multi-dimensional feature space data.

[0125] In an embodiment of the present application, based on the decoupled feature vector F(t), construct a polarization component matrix: P(t) = {P1(t), P2(t), P3(t)}, where P1 is the electronic polarization component, P2 is the ionic polarization component, and P3 is the orientation polarization component; calculate the polarization intensity: I(P) = ∑wi·Pi(t), and output the polarization feature data I(P). Based on the polarization feature data I(P) and the crystal form feature data P(α, β), construct a structural feature tensor: S = {s1, s2, s3}, where s1 is the crystal size feature, s2 is the crystal integrity feature, and s3 is the interface feature; calculate the topological feature: T(S) = k(S, I(P)), and output the microstructure feature T(S). Based on the microstructure feature T(S), construct a high-dimensional feature space: H = {F(t), I(P), T(S)}, and apply non-linear dimensionality reduction: L = PCA(H) + t-SNE(H), and output the multi-dimensional feature space data L. Where wi is the weight coefficient, Pi(t) is the polarization component, k( ) is the topological feature function, PCA( ) is the principal component analysis, and t-SNE( ) is the t-distributed stochastic neighbor embedding.

[0126] In another embodiment of the present application, the polarization feature quantization is P(t) = P_e + P_i + P_o; where P(t) is the total polarization intensity, P_e is the electronic polarization component, P_i is the ionic polarization component, and P_o is the orientation polarization component; the microstructure feature S(r) = ∫[G(r, r')·ρ(r')]dr'; where S(r) is the structural feature function, G(r, r') is the structural correlation function, ρ(r') is the local density distribution at the position r', and dr' is the integration variable at the position r'.

[0127] In this embodiment, a multi-dimensional feature space representation method is established through polarization mechanism quantification and microstructure feature extraction. First, by constructing a polarization component matrix, quantitative characterization of electronic polarization, ionic polarization, and orientation polarization is achieved; then, combined with crystal form characteristic data, microstructure features such as crystal size, integrity, and interface are extracted; finally, through feature space reconstruction, a multi-dimensional feature space capable of comprehensively representing the crystallization state of the PET film is established. This embodiment not only provides complete information on the crystallization process but also retains key feature information through non-linear dimensionality reduction, providing reliable feature input for subsequent crystallinity prediction.

[0128] According to one aspect of the present application, step S32 is further as follows:

[0129] S321. Based on the polarization characteristic data and crystal form characteristic data, extract the polarization time constant sequence; calculate the polarization relaxation time distribution of the polarization time constant sequence; based on the polarization relaxation time distribution, construct the corresponding relationship between the relaxation time and the structural scale, and generate structural parameter data;

[0130] S322. Based on the structural parameter data, calculate the characteristic length distribution at each scale, and construct a scale correlation function; based on the scale correlation function, extract the hierarchical structure of the characteristic scale; establish the topological connection relationship of the hierarchical structure, and generate size characteristic data;

[0131] S323. Based on the size characteristic data and crystal form characteristic data, calculate the order parameter of the local structure, and construct the spatial distribution function of the order parameter; based on the spatial distribution function, extract the position and type information of the structural defects, construct the defect density distribution map, and generate integrity characteristic data;

[0132] S324. Based on the integrity characteristic data, calculate the interface energy between structural units, and construct the interface orientation distribution function; based on the interface orientation distribution function, extract the curvature and roughness parameters of the interface, construct the connectivity network of the interface, and generate interface characteristic data;

[0133] S325. Based on the interface characteristic data, calculate the stress distribution state of the interface, and construct the stress-orientation coupling relationship; based on the stress-orientation coupling relationship, extract the interface stress concentration region, and establish the stress relaxation parameter, and generate interface stress data;

[0134] S326. Based on the size characteristic data, integrity characteristic data, interface characteristic data, and interface stress data, construct the correlation matrix of the characteristic parameters, and calculate the coupling strength between the parameters; based on the coupling strength, construct the multi-scale structural feature space, and generate microstructure characteristic data.

[0135] In one embodiment of the present application, the polarization relaxation time distribution G(τ) = ∑[H_i·exp(-(ln(τ / τ_i)) 2 / 2B_i 2 )]; where G(τ) is the relaxation time distribution function, τ is the relaxation time, τ_i is the characteristic time of the i-th relaxation process, H_i is the intensity of the i-th relaxation process, and B_i is the distribution width parameter.

[0136] The integrity characteristic data I(r) = ∫[ρ(r)·Φ(r)]dr; where I(r) is the structural integrity index, r is the spatial position vector, ρ(r) is the local density distribution function, and Φ(r) is the order parameter field.

[0137] In this embodiment, a complete microscopic structure characterization system is established through multi-scale structural feature analysis and parameter coupling characterization. First, the structural scale correlation is established through the analysis of the polarization relaxation time distribution; then, the characteristic size distribution is determined through the hierarchical structure analysis; then, the structural integrity is characterized through the order parameter analysis; then, the interface characteristics are described through the interface energy and stress analysis; finally, a comprehensive characterization model is established through multi-parameter coupling. This embodiment establishes a quantitative correspondence between the polarization characteristics and the structural characteristics, reveals the structural evolution law through multi-scale analysis, and clarifies the structural stability mechanism through the interface characteristic analysis; it not only provides a comprehensive characterization of the structural characteristics, but also reveals the structure-property relationship through the parameter coupling analysis, providing a new idea for material design.

[0138] As Figure 5 shown, according to one aspect of the present application, step S4 is further as follows:

[0139] S41. Based on the multi-dimensional feature space data, construct a physical constraint equation; solve the non-linear equation system of the physical constraint equation to generate a physical model prediction value;

[0140] S42. Based on the multi-dimensional feature space data, perform data processing through a deep neural network and an attention mechanism to generate a deep learning prediction value;

[0141] S43. Calculate the confidence weights of the physical model prediction value and the deep learning prediction value respectively; based on the confidence weights, perform weighted fusion processing on the physical model prediction value and the deep learning prediction value to generate a preliminary crystallinity prediction value.

[0142] In one embodiment of the present application, based on the multi-dimensional feature space data L, a physical constraint equation is constructed: Φ(x) = {φ1(x), φ2(x), φ3(x)}, and a non-linear equation set is solved: x* = arg min||Φ(x)||, and the physical model prediction value x* is output. Based on the multi-dimensional feature space data L, a deep neural network is constructed: DNN(L) → y, an attention mechanism is applied: A(L) = softmax(W·L), and the prediction value is calculated: y = DNN(A(L)), and the deep learning prediction value y is output. Based on the physical model prediction value x* and the deep learning prediction value y, the confidence weights are calculated: w1 = f1(x*), w2 = f2(y); the prediction is fused: z = w1·x* + w2·y, and the preliminary crystallinity prediction value z is output. Where φ1(x), φ2(x), φ3(x) are different components in the physical constraint equation Φ(x); W is a weight matrix; DNN( ) is a deep neural network.

[0143] In another embodiment of the present application, the physical model prediction X_p = f(S, P, T)·[1 + Δ(t)]; where X_p is the physical model prediction value, f(S, P, T) is the physical model function, and Δ(t) is a correction term.

[0144] Deep learning prediction: X_d = ∑[w_i·h_i(F)]; where X_d is the deep learning prediction value, w_i is the neural network weight, h_i is the hidden layer activation function, and F is the feature input vector.

[0145] According to another aspect of the present application, the construction of the physical constraint equation is specifically: the basic balance equation of crystallinity: X(t) = X_a(t) + X_b(t) + X_m(t); where X(t) is the total crystallinity, X_a(t) is the α-crystalline form crystallinity, X_b(t) is the β-crystalline form crystallinity, X_m(t) is the amorphous region content, and the constraint condition is satisfied: X_a(t) + X_b(t) + X_m(t) = 1. The crystallization kinetics equation: dX_a / dt = k_a(T)·[1 - X_a(t)]·f_a(σ) dX_b / dt = k_b(T)·[1 -X_b(t)]·f_b(σ); where k_a(T), k_b(T) are temperature-dependent crystallization rate constants; f_a(σ), f_b(σ) are stress-dependent functions; σ is the local stress field. The dielectric response equation: ε(ω) = ε_∞ + (ε_s - ε_∞)·[X_a·φ_a(ω) + X_b·φ_b(ω)]; where ε(ω) is the complex dielectric constant; ε_∞ is the high-frequency dielectric constant; ε_s is the static dielectric constant; φ_a(ω), φ_b(ω) are crystalline form characteristic response functions.

[0146] The specific non - linear equations for solving the physical constraint equations are as follows: Construct the residual functions: \(R_1(X)=\vert X_a + X_b+ X_m - 1\vert\), \(R_2(X)=\vert\varepsilon_{measured}-\varepsilon_{calculated}\vert\), \(R_3(X)=\vert\frac{dX}{dt}-F(X,T,\sigma)\vert\); where \(R_1\) is the mass - conservation residual, \(R_2\) is the dielectric - response residual, \(R_3\) is the kinetic residual, \(\varepsilon_{measured}\) is the measured dielectric constant, \(\varepsilon_{calculated}\) is the calculated dielectric constant, and \(F(X,T,\sigma)\) is the kinetic function representing the kinetic behavior of the system at state \(X\), temperature \(T\), and stress \(\sigma\). Optimize the objective function: \(J(X)=w_1\cdot R_1(X)+w_2\cdot R_2(X)+w_3\cdot R_3(X)\); where \(w_1\), \(w_2\), and \(w_3\) are weight coefficients. The solution strategy is to use the improved Levenberg - Marquardt algorithm: \(X_{(k + 1)}=X_k-[H+\lambda I]\) -1 \(\cdot\nabla J(X_k)\); where \(H\) is the Hessian matrix, \(\lambda\) is the damping factor, \(\nabla J\) is the gradient vector, and \(I\) is the identity matrix.

[0147] Generate the predicted values of the physical model specifically as follows: Crystallinity prediction equation: \(X_{predict}=g(X_a,X_b,S)\); where \(g\) is the prediction function and \(S\) is the structural - feature parameter. Correction of the predicted value: \(X_{final}=X_{predict}\cdot[1 + \gamma(T)\cdot\Delta(\sigma)]\); where \(\gamma(T)\) is the temperature - correction function and \(\Delta(\sigma)\) is the stress - correction function.

[0148] In this embodiment, through the integration of the physical model and the deep - learning method, high - precision crystallinity prediction is achieved. First, the predicted values of the physical model are obtained by solving the physical constraint equations, ensuring the physical rationality of the prediction results; then, the deep - learning predicted values are obtained by processing through the deep neural network and the attention mechanism, making full use of the implicit laws in the data; finally, the two methods are fused through confidence weights, integrating the advantages of both methods. This embodiment not only improves the accuracy of the prediction but also ensures the interpretability and reliability of the prediction results through the constraints of the physical model.

[0149] According to one aspect of the present application, step S42 is further as follows:

[0150] S421. Receive multi - dimensional feature - space data, calculate the correlation matrix between feature vectors, extract the temporal dependence of features, construct a multi - scale feature pyramid, calculate the feature weights for each scale level, and generate hierarchical feature data.

[0151] S422. Receive hierarchical feature data, construct long-range and short-range dependency graphs, calculate the information flow intensity between nodes, extract key information transfer paths, establish a feature propagation network, and generate feature dependency data.

[0152] S423. Receive feature dependency data, calculate the importance scores of each feature node, construct an adaptive attention weight matrix, extract the association patterns of cross-scale features, establish a dynamic attention mapping, and generate attention feature data.

[0153] S424. Receive attention feature data, construct a multi-layer perceptron scene graph, calculate the activation states of scene nodes, extract the conversion rules between scenes, establish a scene evolution model, and generate scene representation data.

[0154] S425. Receive scene representation data, establish a multi-channel predictor group, calculate the prediction results of each channel, extract the uncertainty distribution of the prediction results, construct a confidence evaluation model, and generate multi-channel prediction data.

[0155] S426. Receive multi-channel prediction data, calculate the consistency index of each channel's prediction, construct a prediction fusion weight, extract the optimal prediction path, establish a prediction calibration model, and generate deep learning prediction values.

[0156] This embodiment realizes high-precision crystallinity prediction. First, hierarchical feature representations are established through a multi-scale feature pyramid; then, a feature propagation network is established through dependency graph analysis; next, feature weight optimization is achieved through an adaptive attention mechanism; then, a dynamic prediction model is established through scene graph analysis; finally, the final prediction result is obtained through multi-channel fusion. This embodiment effectively integrates multi-scale information through the feature pyramid, captures the long- and short-range associations between features through the dependency graph, improves the weights of key features through the dynamic attention mechanism, and realizes the dynamic update of the prediction through the scene evolution model; not only improves the accuracy of the prediction, but also enhances the reliability of the prediction through the multi-channel design, providing a practical solution for industrial applications.

[0157] As Figure 6 shown, according to one aspect of the present application, step S5 is further as follows:

[0158] S51. Based on the preliminary crystallinity prediction value, calculate the historical consistency score and the physical constraint score; based on the historical consistency score and the physical constraint score, generate consistency scoring data through weighted processing;

[0159] S52. Based on the preliminary crystallinity prediction value and the consistency scoring data, construct a correction function; based on the correction function, perform dynamic correction processing on the preliminary crystallinity prediction value to generate the final crystallinity value.

[0160] In an embodiment of the present application, based on the preliminary crystallinity prediction value z, the historical consistency score is calculated: H(z)= d(z, z_hist); the physical constraint score is calculated: P(z) = c(z, Φ); the comprehensive score is calculated: S(z) = w1·H(z) + w2·P(z), and the consistency score S(z) is output. Based on the preliminary crystallinity prediction value z and the consistency score S(z), a correction function is constructed: R(z) = g(z, S(z)), and the correction is applied: zf = R(z), and the final crystallinity value zf is output. Where z_hist is the historical crystallinity data, d( ) is the distance function, c( ) is the physical constraint function, Φ is the physical constraint condition, w1 and w2 are different weight coefficients, and g( ) is the correction function.

[0161] In another embodiment of the present application, the consistency score Q(t) = α·H(t) + β·V(t); where Q(t) is the consistency score, H(t) is the historical consistency component, and V(t) is the physical constraint component; α and β are weight coefficients. The final crystallinity is calculated as: X_f = γ·X_p + (1 - γ)·X_d + K·Q(t); where X_f is the final crystallinity value, γ is the model fusion coefficient, K is the correction coefficient, X_p is the physical model prediction, and X_d is the deep learning prediction.

[0162] According to another aspect of the present application, the calculation of the historical consistency score is specifically: the time series consistency function: H_t(X) = exp[-α·(X(t) - X_pred(t)) 2 / σ_t 2 ; where H_t(X) is the time consistency score, X(t) is the current prediction value, X_pred(t) is the prediction value based on historical data, α is the attenuation coefficient, and σ_t is the standard deviation of time fluctuation. The spatial consistency function: H_s(X) = exp[-β·∑(X(r) - X(r + Δr)) 2 / σ_s 2 ; where H_s(X) is the spatial consistency score, X(r) is the prediction value at position r, Δr is the spatial displacement vector, β is the spatial weight coefficient, and σ_s is the standard deviation of spatial fluctuation. The comprehensive historical consistency score: H(X) = w_t·H_t(X) + w_s·H_s(X); where w_t and w_s are the spatio-temporal weight coefficients.

[0163] The calculation of the physical constraint score is specifically as follows: Thermodynamic constraint function: P_t(X) = exp[-γ·|ΔG(X)| / RT]; where P_t(X) is the thermodynamic score, ΔG(X) is the change in crystallization free energy, R is the gas constant, T is the absolute temperature, and γ is the scaling coefficient. Kinetic constraint function: P_k(X) = exp[-λ·|dX / dt - f(X, T)| 2 ; where P_k(X) is the kinetic score, f(X, T) is the crystallization kinetic function, and λ is the kinetic weight coefficient. Comprehensive physical constraint score: P(X) = w_th·P_t(X) + w_k·P_k(X); where w_th and w_k are the physical constraint weights. The consistency scoring data is specifically as follows: Q(X) = μ_h·H(X) + μ_p·P(X); where Q(X) is the final consistency score; μ_h and μ_p are the overall weight coefficients.

[0164] The process of generating the final crystallinity value is specifically as follows: Construct a correction function: The basic correction function is C(X, Q) = X·[1 + η(Q)·Δ(T)]; where C(X, Q) is the correction function, η(Q) is the score-dependent function, and Δ(T) is the temperature correction term; The adaptive weight function is W(Q) = (1 + exp[-k(Q - Q_0)]) -1 ; where W(Q) is the weight function, k is the steepness coefficient, and Q_0 is the threshold parameter. Perform dynamic correction processing and construct a time smoothing function: S(t) = ∫[X(τ)·K(t - τ)]dτ; where S(t) is the smoothing result, K(t) is the kernel function, and τ is the integration variable; Calculate the final crystallinity: X_final = W(Q)·C(X, Q) + [1 - W(Q)]·S(t); where X_final is the final crystallinity value, W(Q) is the adaptive weight, C(X, Q) is the correction result, and S(t) is the smoothing result.

[0165] In this embodiment, through consistency testing and dynamic correction, the optimization and correction of the crystallinity prediction results are achieved. First, the historical consistency score and physical constraint score are calculated to evaluate the reliability of the prediction results; then, based on these scores, dynamic correction is performed to improve the stability and accuracy of the prediction results. This embodiment can not only identify and correct abnormal prediction results, but also improve the robustness of the prediction through the analysis of historical data, ensuring the reliability of the final crystallinity value.

[0166] According to one aspect of the present application, an on-line detection device for the crystallinity of a PET film using dielectric properties includes a dielectric property measurement unit, a temperature measurement unit, a film surface state monitoring unit, a synchronous control unit, a data acquisition and preprocessing unit, and an auxiliary measurement unit. The dielectric property measurement unit includes: a multi-frequency array electrode system composed of multiple groups of parallel plate electrodes arranged along the movement direction of the PET film for realizing multi-point parallel measurement; a high-precision impedance analyzer capable of simultaneously measuring the dielectric constant and dielectric loss at multiple frequency points (such as 100 Hz - 1 MHz); an electrode contact pressure control device including a precision pressure sensor and a pressure regulating mechanism for monitoring and controlling the contact state between the electrode and the PET film; and a signal conditioning circuit for amplifying, filtering, and preprocessing the original dielectric signal. The temperature measurement unit includes: a non-contact infrared temperature sensor array arranged along the width direction of the film surface for realizing the measurement of the temperature field distribution; a contact thermocouple arranged at key positions for calibrating the infrared temperature measurement; and a temperature signal acquisition module for synchronously acquiring multi-channel temperature signals. The film surface state monitoring unit includes: a high-speed industrial camera for monitoring the surface state and movement characteristics of the PET film; a laser displacement sensor for measuring the flatness and thickness fluctuation of the film surface; and a tension sensor for monitoring the tension distribution state of the film surface. The synchronous control unit includes: a high-precision encoder for accurately measuring the movement speed and position of the film material; a clock synchronization module for ensuring the timing consistency of multi-channel data acquisition; and a trigger control circuit for realizing the synchronous sampling of multiple sensors. The data acquisition and preprocessing unit includes: a high-speed multi-channel data acquisition card supporting the parallel acquisition of various signals such as dielectric and temperature; a real-time data cache module for temporarily storing the original data acquired at high speed; and a field bus communication module for realizing the data exchange between each measurement unit and the control system. The auxiliary measurement unit includes: an environmental parameter monitoring device for measuring parameters such as environmental temperature and humidity; a production parameter acquisition interface for obtaining process parameters such as linear velocity and draw ratio; and a marking system for automatically marking abnormal points or characteristic points.

[0167] The performance index requirements for dielectric measurement accuracy are as follows: the relative error of the dielectric constant ≤ 0.1%, and the accuracy of the tangent value of the loss angle ≤ 0.0001; the performance index requirements for temperature measurement accuracy are: ±0.1°C; the performance index requirements for the sampling frequency are: for dielectric measurement ≥ 1 kHz, and for temperature measurement ≥ 100 Hz; the performance index requirements for spatial resolution are: the measuring point spacing along the film width direction ≤ 50 mm. The system integration requirements for modular design are: each functional unit can be independently installed and debugged; the system integration requirements for the protection level are: the key measurement unit reaches the IP65 protection level; the system integration requirements for anti-interference are: electromagnetic shielding and anti-interference design are available; the system integration requirements for the calibration function are: online calibration and compensation functions are supported. The synchronization requirements for time synchronization accuracy are: better than 1 ms; the synchronization requirements for spatial positioning accuracy are: better than ±1 mm; the synchronization requirements for unified data time stamps are: GPS or industrial Ethernet clock synchronization is adopted.

[0168] In view of the problem of impedance change at the contact interface, the present invention proposes a contact compensation processing method. Specifically, first, an index of spatial uniformity of pressure distribution is established to achieve accurate characterization of the contact characteristics in micro-regions; then, a parallel equivalent circuit including contact resistance and contact capacitance is constructed, and a temperature influence coefficient is introduced to achieve accurate modeling of contact impedance; finally, dynamic compensation based on physical constraints is realized, and through segmented compensation operations and continuous smoothing processing, the measurement errors caused by changes in the contact state are effectively eliminated. This multi-level compensation strategy improves the stability of measurements on high-speed production lines. Regarding the problem of decoupling multiple effect couplings, the systematic separation of temperature effects, orientation effects, and crystal form effects is achieved. Specifically, by constructing a relationship matrix between temperature and frequency, calculating the temperature sensitivity coefficient and frequency response function, the accurate separation of the temperature-frequency coupling effect is realized; the accurate identification of α-crystal form and β-crystal form is achieved by analyzing the dielectric relaxation characteristics; the separation of the orientation effect is realized by establishing a relationship model between orientation and dielectric properties. This systematic decoupling method effectively solves the problem of interference from multiple factors. Regarding the problem of characterizing structural dynamic evolution, a complete structural characterization system is constructed. Specifically, by analyzing the characteristic frequency of the dielectric spectrum, the relaxation time distribution, and the order parameter tensor, the dynamic characterization of α-crystal form, β-crystal form, and their mixed states is realized; by establishing a multi-scale structural feature space, combining polarization characteristics and crystal form characteristics, the accurate characterization of the evolution of the microstructure is realized; especially, by constructing the free energy landscape of crystal form transformation, the dynamic prediction of the structural evolution process is realized. Regarding the problem of the prediction accuracy of the model, a fusion strategy of physical models and deep learning is proposed. Specifically, the physical rationality of the prediction results is ensured through physical constraint equations; the prediction accuracy of the deep learning model is improved through multi-scale feature pyramids, dependency graphs, and dynamic attention mechanisms; the optimal fusion of the two methods is realized through the dynamic optimization of confidence weights. This hybrid prediction strategy improves the accuracy and reliability of predictions. Regarding the problem of the robustness of the prediction results, a complete result evaluation and correction mechanism is established. Specifically, the reliability evaluation of the prediction results is realized by calculating the historical consistency score and the physical constraint score; the real-time correction of abnormal results is realized by constructing a dynamic correction function; this post-processing strategy improves the robustness of the system under complex working conditions.

[0169] The present invention discloses an on-line detection method for the crystallinity of PET films using dielectric properties, including: obtaining original dielectric data and real-time temperature data, obtaining spatio-temporal sampling data through parallel acquisition and processing, performing noise identification and removal processing on the data, and obtaining high-quality dielectric characteristic data through contact compensation processing; separating the temperature-frequency coupling effect and identifying the crystal form characteristics of the high-quality dielectric characteristic data; establishing a multi-dimensional feature space through polarization mechanism quantization processing and microstructure feature extraction; fusing physical models and deep learning methods for crystallinity prediction; and finally obtaining the final crystallinity value through consistency testing and dynamic correction. The present invention has established a complete on-line detection method for the crystallinity of PET films through systematic data processing, feature extraction, and model fusion strategies. Adopting a five-layer architecture of "data acquisition - feature decoupling - feature reconstruction - model prediction - result optimization", each layer provides a solution for specific technical difficulties. In the data acquisition layer, the quality of the original data is ensured through parallel acquisition and multiple processing; in the feature decoupling layer, the separation of multiple factors such as temperature, crystal form, and orientation is achieved; in the feature reconstruction layer, a multi-dimensional representation space including polarization characteristics and microstructure characteristics is established; in the model prediction layer, the advantages of physical models and deep learning methods are fused; in the result optimization layer, the reliability of the final result is ensured through consistency testing and dynamic correction. The present invention not only solves the accuracy problem of on-line detection of the crystallinity of PET films, but also provides an interpretable structure-property relationship analysis method, which has important value for improving the product quality control level.

[0170] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all belong to the protection scope of the present invention.

Claims

1. A method for online detection of PET film crystallinity using dielectric properties, characterized in that: The steps include: S1. Obtaining raw dielectric data and real-time temperature data, and obtaining high-quality dielectric characteristic data through parallel acquisition processing, noise identification and removal, and contact compensation processing; S2, separating the temperature-frequency coupling effect and identifying the crystal characteristics of the high-quality dielectric characteristic data to obtain a decoupled characteristic vector; S3, based on the decoupled feature vector, a multi-dimensional feature space is established through the quantification of the polarization mechanism and the extraction of microstructure features to obtain multi-dimensional feature space data; S4. Based on the multi-dimensional feature space data, the pre-configured physical model and deep learning method are integrated to predict the crystallinity and obtain a preliminary crystallinity prediction value; S5. Based on the preliminary crystallinity prediction value, consistency check and dynamic correction are performed to obtain the final crystallinity value.

2. The method for online detection of PET film crystallinity using dielectric properties according to claim 1, characterized in that: Step S1 is further as follows: S11, acquiring original dielectric data and real-time temperature data; and performing parallel acquisition processing on them according to pre-built sampling parameters to generate spatiotemporal sampling data and sampling temperature data; S12, based on the spatiotemporal sampling data, calculating the wavelet energy distribution characteristics and constructing the noise characteristic data; Adopting an adaptive threshold method to process noise characteristic data, and obtaining processed noise characteristic data; Based on the processed noise characteristic data, the spatiotemporal sampling data is subjected to noise reduction processing to obtain noise-reduced data; S13, obtaining contact state parameters in real time; Based on the contact state parameters and sampled temperature data, a contact impedance model is constructed, the compensation coefficient is calculated and combined with the noise reduction data to generate high-quality dielectric characteristic data.

3. The method for online detection of PET film crystallinity using dielectric properties according to claim 2, characterized in that: Step S2 is further as follows: S21, constructing a temperature-frequency relationship matrix for the high-quality dielectric characteristic data, calculating the temperature sensitivity coefficient and the frequency response function, and performing temperature-frequency coupling effect separation processing on the high-quality dielectric characteristic data accordingly, to generate decoupled dielectric data; S22, based on the decoupled dielectric data, the dielectric relaxation characteristics and mixed state characteristics of the α crystal form and the β crystal form are calculated and the crystal form distribution is calculated accordingly to generate crystal form characteristic data; S23. Based on the decoupled dielectric data and crystal characteristic data, the molecular chain orientation function is calculated to construct a relationship model between orientation and dielectric properties; based on the relationship model, the orientation effect of the decoupled dielectric data is separated to generate a decoupled characteristic vector.

4. The method for online detection of PET film crystallinity using dielectric properties according to claim 3, characterized in that: Step S3 is further as follows: S31. Based on the decoupled eigenvector, construct a polarization component matrix including electronic polarization, ionic polarization and orientation polarization; calculate the weight and strength of each polarization component in the polarization component matrix to generate polarization characteristic data; S32, based on the polarization characteristic data and the crystal form characteristic data, constructing structural characteristic parameters, calculating the crystal size characteristics, crystal integrity characteristics and interface characteristics, and generating microstructure characteristic data; S33. Based on the microstructure feature data, a high-dimensional feature space is constructed and nonlinear dimensionality reduction processing is performed to generate multi-dimensional feature space data.

5. The method for online detection of PET film crystallinity using dielectric properties according to claim 4, characterized in that: Step S4 is further as follows: S41. Based on the multi-dimensional feature space data, construct and solve the physical constraint equation to generate the physical model prediction value; S42, based on multi-dimensional feature space data, data processing is performed through deep neural network and attention mechanism to generate deep learning prediction value; S43. Calculate the confidence weights of the physical model prediction value and the deep learning prediction value respectively; based on the confidence weights, perform weighted fusion processing on the physical model prediction value and the deep learning prediction value to generate a preliminary crystallinity prediction value.

6. The method for online detection of PET film crystallinity using dielectric properties according to claim 5, characterized in that: Step S5 is further as follows: S51. Based on the preliminary crystallinity prediction value, the historical consistency score and the physical constraint score are calculated and weighted to generate consistency score data; S52. Construct a correction function by combining the preliminary predicted crystallinity value and the consistency score data; based on the correction function, dynamically correct the preliminary predicted crystallinity value to generate a final crystallinity value.

7. The method for online detection of PET film crystallinity using dielectric properties according to claim 6, characterized in that: Step S11 is further as follows: S111, acquiring original dielectric data and real-time temperature data, and generating sampling interval parameters; S112, determining a basic grid based on a sampling interval parameter, and generating sampling position data; S113, performing multi-channel data acquisition based on the sampling position data to generate multi-channel acquisition data; S114, based on the multi-channel acquisition data, calculating the time deviation and generating the aligned acquisition data using a cubic spline function; S115, calculating local statistical features of the aligned collected data, identifying outliers, and constructing a correlation matrix; The correlation matrix is ​​correspondingly integrated with the real-time temperature data to generate spatiotemporal sampling data and sampling temperature data.

8. The method for online detection of crystallinity of PET film using dielectric properties according to claim 6, characterized in that: Step S12 is further as follows: S121, grouping the spatiotemporal sampling data according to frequency, performing multi-scale decomposition, calculating the energy value at each scale, and generating energy distribution characteristics; S122, based on the energy distribution characteristics, identifying the fluctuation abnormal interval, extracting the time-frequency characteristic parameters, and generating the noise characteristic data; S123, performing cluster analysis on the noise feature data, calculating energy concentration, and generating a noise determination result; S124, performing time-frequency decomposition on the spatiotemporal sampling data based on the noise determination result, calculating the local correlation coefficient, constructing the signal reconstruction weight, and generating preliminary noise reduction data; S125, based on the preliminary noise reduction data, calculating the local smoothness index of the signal, identifying the residual noise points and performing local correction to obtain a correction result; Integrate preliminary noise reduction data and correction results to generate noise reduction data; S126: Time-align the sampled temperature data with the noise reduction data, and output the aligned noise reduction data and sampled temperature data.

9. The method for online detection of crystallinity of PET film using dielectric properties according to claim 6, characterized in that: Step S22 is further as follows: S221. Based on the decoupled dielectric data, calculate the dielectric loss peak shape parameters, construct a time-temperature equivalent curve of the loss peak, and generate relaxation characteristic data; S222, calculating the orientation correlation function of the α crystal form based on the relaxation characteristic data, and generating α crystal form characteristic data; S223, constructing local order parameters based on the relaxation characteristic data, and generating β crystal characteristic data; S224, calculating the overlap of α and β crystal form characteristic data, identifying the mixed state region, constructing the order parameter tensor and calculating the stability index, and generating mixed state characteristic data; S225. Construct a multidimensional feature space for the characteristic data of the α-crystal form, the β-crystal form and the mixed state, calculate the distribution density function, identify the characteristic clusters and calculate the transition probability, and generate preliminary distribution data; S226. Based on the preliminary distribution data, calculate the crystal form correlation, identify and optimize the local distribution abnormal area, and generate crystal form characteristic data.

10. The method for online detection of PET film crystallinity using dielectric properties according to claim 6, characterized in that: Step S32 is further as follows: S321, based on the polarization characteristic data and the crystal characteristic data, extract the polarization time constant sequence and calculate the polarization relaxation time distribution to generate the structural parameter data; S322, calculating characteristic length distribution at each scale to obtain dimension characteristic data; S323, based on the size characteristic data and the crystal form characteristic data, calculating the order parameter of the local structure, extracting the defect position and type, and generating integrity characteristic data; S324, calculating the interface energy, extracting the interface curvature and roughness, and obtaining interface characteristic data; S325, calculating interface stress distribution, extracting stress concentration areas, and generating interface stress data; S326. Based on the size characteristic data, integrity characteristic data, interface characteristic data and interface stress data, a characteristic parameter correlation matrix is ​​constructed, the coupling strength is calculated, and microstructure characteristic data is generated.

Images