Paper finished product detection system based on frame paper production process

Abstract

The invention relates to the technical field of paper detection, in particular to a paper finished product detection system based on a framework paper production process. By extracting defect information, modeling based on a graph structure and carrying out graph convolution operation to smooth an image, a defect area is accurately identified. Large-scale trend and small-scale local anomaly are separated through a wavelet decomposition technology, errors caused by environmental factors are eliminated through a local adaptive regression method, and the accuracy of thickness deviation is ensured. The density uniformity is calculated through ultrasonic feature extraction, the hyperspectral reflectivity is corrected in combination with a physical compensation method of chemical residues, and the chemical residue index is accurately obtained. The comprehensive quality index of the paper is obtained through comprehensive index calculation, so that whether the finished paper is qualified or not is judged, and a basis is provided for paper production adjustment of a production line.

Description

Technical Field

[0001] The present invention relates to the technical field of paper detection, and in particular to a finished paper product detection system based on a frame paper production process. Background Art

[0002] In the frame paper production process, quality inspection is required after the paper is finished. Currently, this method relies on image or ultrasonic signal processing, but the following deficiencies exist in the finished product inspection:

[0003] In the surface defect detection process, most systems rely solely on edge detection and binary segmentation of optical imaging, which may ignore image noise and false cracks caused by environmental factors such as humidity, temperature, uneven lighting, and mechanical vibration on the production line.

[0004] In thickness measurement, global filtering or fixed models are usually used to denoise the depth map, without distinguishing the full-width bending caused by tension distribution from local bubbles or fiber agglomerations;

[0005] In density uniformity testing, the coupling effects of chemical residues and humidity changes on density measurement are not considered;

[0006] Finally, chemical residue detection relies more on offline laboratory testing or static dictionary sparse inversion, which cannot correct dynamic changes in process parameters such as retention agent dosage, slurry viscosity, and white water reuse in real time;

[0007] In summary, the above-mentioned links are difficult to conduct accurate and in-depth inspections on the quality of finished paper products because they do not take into account the actual application environment and specific affected conditions of the paper. Summary of the Invention

[0008] In response to the above-mentioned shortcomings of the existing technology, the present invention provides a paper product detection system based on the frame paper production process, which can effectively solve the problem that the existing technology does not take into account the actual application environment and specific affected conditions of the paper, making it difficult to accurately and deeply detect the quality of the finished paper.

[0009] To achieve the above objectives, the present invention is implemented through the following technical solutions:

[0010] The present invention provides a paper product detection system based on a frame paper production process, which at least includes:

[0011] The defect index analysis module creates an initial defect image of the paper, constructs a graph structure to calculate the neighborhood weights between pixels, uses graph convolution to smooth the image using the adjacency weights, and combines density information with mechanical vibration to determine surface defect indicators.

[0012] The thickness index analysis module obtains the paper depth map and separates it into different scale components, calculates the absolute deviation under the scale components, and fits the scale trend based on the environmental state, where:

[0013] Define the initial fitting residual and perform iterative update diffusion. Introduce production speed and paper basis weight as coupling factors to correct measurement deviation. Introduce humidity and temperature into the inverse compensation model to obtain thickness deviation residual. Combine the thickness deviation residual with the paper machine direction to determine the thickness deviation index.

[0014] The density index analysis module constructs a two-dimensional feature vector based on the propagation delay and amplitude attenuation of the ultrasonic wave, and determines the local anomaly score by combining the reconstruction error with the chemical index to define the density uniformity index;

[0015] Chemical index analysis module, which calculates chemical residue index by combining reflectivity, retention efficiency factor, and adsorption capacity ratio;

[0016] In response to the input surface defect index, thickness deviation index, density uniformity index and chemical residue index, the comprehensive quality index of the paper is output to complete the inspection of the finished paper product.

[0017] In summary, the present invention:

[0018] The quality of finished paper products is comprehensively analyzed through multiple indicators to ensure that the paper quality during the production process meets the standards and the production process can be dynamically adjusted.

[0019] Its core indicators include surface defect index, thickness deviation index, density uniformity index and chemical residue index. The calculation and analysis of these indicators take into account different factors (such as production speed, ambient humidity, light, etc.) to ensure accurate measurement results.

[0020] First, the paper surface image data is collected through a visual sensor, and image processing algorithms are applied to remove noise and extract defect information.

[0021] Then, based on graph structure modeling, graph convolution is performed to smooth the image and accurately identify defect areas. For thickness deviation, the solution uses wavelet decomposition technology to separate large-scale trends and small-scale local anomalies, and uses local adaptive regression methods to eliminate errors caused by environmental factors, ensuring the accuracy of thickness deviation.

[0022] In addition, the solution also calculates density uniformity through ultrasonic feature extraction and combines physical compensation methods for chemical residues to correct hyperspectral reflectance and accurately obtain chemical residue indicators;

[0023] Ultimately, the four indicators are calculated through weighted summation to obtain the comprehensive quality index of the paper, thereby judging whether the paper is qualified and providing a basis for adjusting the production line. This solution can monitor and optimize the quality of paper in real time during the production process, reduce the defect rate, and improve production efficiency.

[0024] The method for determining surface defect indicators is:

[0025] Calculate the adjacency weight between each pixel and its neighboring pixels;

[0026] The defect confidence of each pixel is calculated by weighted averaging the information of neighboring pixels through the graph convolution method:

[0027] Calculate surface defect indicators.

[0028] Furthermore, the large-scale deviation obtained after scale decomposition includes the trend drift caused by environmental factors. Based on local adaptive regression, the environmental state is integrated into the weight, the large-scale trend of thickness is fitted, and the global drift caused by tension / humidity is eliminated, then:

[0029] A weighted quadratic polynomial is fitted on the window for each center to obtain the trend obtained by local regression fitting.

[0030] Furthermore, the thickness deviation residual is combined with the paper machine direction to determine the thickness deviation index as follows:

[0031] Add a double inverse compensation model for humidity and temperature;

[0032] Considering the different sensitivities to thickness fluctuations in the machine direction and cross direction of the paper:

[0033] Based on the cosine weighting of the main fiber direction and the barrel direction, the same amplitude deviation in different directions can be treated differently and the thickness deviation index can be calculated.

[0034] Furthermore, the method for determining the density uniformity index is:

[0035] Calculate the mean and covariance matrix of all eigenvectors;

[0036] Perform eigendecomposition on the covariance matrix and take the first l corresponding eigenvector matrices;

[0037] Project and reconstruct to obtain the reconstructed density vector;

[0038] Calculate the reconstruction error;

[0039] Considering the influence of chemical residues, the reconstruction error is amplified by the index:

[0040] Get the score representing the local abnormality;

[0041] Global aggregation calculates density uniformity indicators.

[0042] Furthermore, the method for determining the chemical residue index is:

[0043] Considering that thickness changes the optical path length, cracks cause local light scattering, and uneven density affects the absorption depth:

[0044] Perform physical coupling compensation on the original reflectivity vector to obtain the corrected reflectivity;

[0045] Through coupled modeling, slurry rheology and process parameters are mapped together into retention efficiency factors;

[0046] Chemical reaction kinetics were embedded in the model to determine the adsorption rate;

[0047] Calculate the chemical residual concentration index based on reflectivity, retention efficiency factor, and adsorption capacity ratio;

[0048] Determine chemical residue indices based on multi-point chemical residue concentration indices.

[0049] Compared with the prior art, the technical solution provided by the present invention has the following beneficial effects:

[0050] By constructing a multi-state coupled graph structure for surface defect detection, hyperspectral reflectance differences, thickness variations, local illumination differences, and mechanical vibration intensity are incorporated into the adjacency weight calculation. This allows noise nodes and real crack nodes to be smoothed and isolated to varying degrees in the graph convolution, enabling accurate calculation of surface defect indicators.

[0051] By separating the depth map into low-frequency, large-scale components and high-frequency, local components, and analyzing each inspection point based on tension, humidity, and thickness deviation, thickness drift caused by macro trends and environmental conditions is dynamically eliminated. True local mutation boundaries are retained based on anisotropic diffusion, accurately capturing true thickness anomalies in areas with tiny bubbles or fiber aggregations, and avoiding misjudging full-width bends as defects.

[0052] Pre-eliminate the interference of physical state on the spectrum, and then accurately estimate the chemical retention through the retention efficiency and adsorption equilibrium model. This ensures that the residual concentration prediction can be immediately corrected after each feeding or process adjustment. Quality feedback is quickly transmitted to the retention agent dosing, white water circulation or temperature control system, improving the accuracy of residue detection and achieving the determination of chemical residue indicators.

[0053] Based on the obtained surface defect index, thickness deviation index, density uniformity index, and chemical residue index, comprehensive quality index of paper is established to achieve multi-angle safety monitoring and control optimization of the quality of paper and paper production line. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. Those skilled in the art can also derive other drawings based on these drawings without inventive effort.

[0055] Figure 1 It is the overall module block diagram of the present invention. DETAILED DESCRIPTION

[0056] To make the purpose, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. 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 any creative efforts shall fall within the scope of protection of the present invention.

[0057] In the existing frame paper production and online testing practices, there are the following deficiencies:

[0058] Traditional visual inspection systems rely solely on a single optical image or structured light depth, ignoring systemic interference with pixel confidence from factors such as barrel vibration, uneven lighting, and humidity / temperature fluctuations. This often leads to false positives due to noise as defects. Furthermore, global filtering weakens the true crack signal, making it difficult to detect small cracks or fiber agglomerations.

[0059] Ultrasonic or hyperspectral testing also lacks compensation for environmental jitter and process conditions, making it difficult to distinguish anomalies between high-density and low-density areas. False positives (noise is amplified) and false negatives (weak true defects are suppressed) often coexist, seriously restricting the reliability and accuracy of detection.

[0060] Existing thickness measurements mostly use unified filtering or simple threshold segmentation, which can only capture large-scale full-width bending but cannot accurately locate local bubbles, fiber accumulation or indentations; ultrasonic density analysis is often independent of thickness measurement and relies solely on global covariance or simple low-rank decomposition, which cannot absorb thickness deviation information, resulting in confusion between thickness trends and local defects.

[0061] A single density model does not respond adequately to environmental coupling (temperature, humidity, tension, and chemical residues), has poor robustness on actual production lines, requires frequent recalibration, and cannot meet the production needs of high-speed and long-cycle operations.

[0062] Current hyperspectral or chemical analysis mostly relies on offline laboratory measurements, ignoring the combined effects of retention aid dosage, slurry viscosity, white water reuse rate, pH and temperature on residual adsorption equilibrium.

[0063] Furthermore, it is difficult to accurately analyze and test the quality of the finished paper.

[0064] To this end, the present invention is proposed.

[0065] The present invention will be further described below with reference to the embodiments.

[0066] Example 1 (see Figure 1 ): A finished paper product inspection system based on the frame paper production process, including at least:

[0067] The defect index analysis module is used to calculate the surface defect index of paper. The steps are as follows:

[0068] Vision sensors on the production line (such as high-resolution cameras or hyperspectral cameras) collect image data from the paper surface. This image data often contains noise and uneven lighting. Image processing algorithms are used to extract surface defect information. The surface defect index s1 is then calculated, including:

[0069] Perform image preprocessing on the acquired image data to reduce image noise caused by factors such as lighting, dust, and sensors;

[0070] Use edge detection method to extract the area where the paper surface structure changes significantly;

[0071] Threshold segmentation, cluster segmentation or image segmentation neural network method is used to separate the crack or defect area from the image background, thereby outputting an initial defect image x(p)∈{0,1}, indicating whether it is a defect area, 1 for yes and 0 for no;

[0072] On the basis of the initial defect map, in order to avoid segmentation roughness and noise interference, a graph structure G is constructed to model the entire image. For each pixel, the adjacency weight A between the pixel and its neighboring pixels is calculated. pq (s2,s5,L), the adjacency weight reflects the similarity of adjacent pixels in the image, then:

[0073]

[0074] Among them, R p 、R q Represents the spectral reflectance vector of pixel p, q respectively, Z p , Z q Represents the thickness values ​​of pixels p and q respectively (when the humidity of paper is too high, the reliability of thickness data decreases), L p 、Lq Represent the local illumination intensity of pixels p and q respectively, represents the normal reference value of the vibration state, s5 represents the vibration state value when the current sample is collected, s2 represents the surface humidity of the current sample (paper), L represents the global illumination level of the current environment, γ1, γ2(s2), γ3(L), ​​and γ4 are the spectral difference weight (controlling the influence of the reflectance spectrum difference on the adjacent weight), thickness difference weight (controlling the influence of thickness difference), illumination difference sensitivity coefficient (controlling the influence of local illumination difference on similarity), and vibration influence coefficient (controlling the penalty intensity of vibration abnormality on similarity), respectively.

[0075] According to the above obtained adjacency weight A pq (s2, s5, L), if it approaches 1, it means that the reflectivity spectra, thickness, illumination, and vibration states of p and q are very similar, and it is determined that they belong to the same material or area, and have the ability to transmit confidence.

[0076] Therefore, based on the above adjacency relationship, the initial defect map is smoothed to remove noise and enhance the recognition of cracks. The defect confidence of each pixel is calculated by weighted averaging the information of neighboring pixels through the graph convolution method.

[0077] Among them, N(p) represents the neighborhood set of pixel p, that is, the adjacent pixels around p, D pp 、D qq Denote the degrees of pixels p and q, respectively, and are the sum of the weights of the connections between pixels and their neighborhoods, ensuring that the smoothing effect between different pixels is not too strong or too weak. Wx(q) represents the original defect confidence value obtained from the image, which is used to calculate the weighted defect value of each pixel. σ(·) represents the activation function, which is used to introduce nonlinearity and enhance the expressiveness of the model. The image is smoothed using the adjacent weights through the graph convolution operation, thereby accurately extracting the defect information of each pixel.

[0078] Furthermore, after smoothing the initial defect map, the surface defect index s1 is calculated:

[0079]

[0080] Where Ω1 represents the set of all pixels in the paper area in the image, α represents the cross-modal fusion weight, which is used to balance the visual information and internal density information, and h(s3(p), s5) represents an amplification function used to adjust the density information s3(p) (such as fiber inhomogeneity, which is equal to the amplified correction). ) and the influence of mechanical vibration on defect indicators, δ1 and δ2 represent the corresponding weight coefficients.

[0081] On actual production lines, environmental factors such as uneven lighting, barrel vibration, high temperature and humidity can introduce a large amount of noise and false signals into visual and spectral data, leading to high false positives or false negatives in traditional detection methods that rely solely on images or spectra. Through the above implementation, a multi-state coupled graph structure is constructed for surface defect detection. Hyperspectral reflectivity differences, thickness variations, local illumination differences, and mechanical vibration intensity are all incorporated into the adjacency weight calculation, resulting in varying degrees of smoothing and isolation between noise nodes and true crack nodes in the graph convolution.

[0082] For example, when a certain area suddenly produces a jitter pseudo-crack due to machine vibration, the addition of the vibration state will quickly reduce the connection weight between the area and the surrounding points, preventing the spread of noise in the graph convolution; and the real crack area, due to the synchronous amplification of the signal of sudden thickness change or uneven internal density, can maintain its high confidence during the smoothing process, thereby reducing the impact of environmental interference on surface defect judgment.

[0083] Thickness index analysis module is used to calculate thickness deviation index. The specific method is as follows:

[0084] On a high-speed paper production line, paper web thickness variations include both macroscopic overall fluctuations (e.g., full-width bending due to tension differences, barrel shrinkage, or conveyor lines) and microscopic local anomalies (e.g., bubbles, fiber agglomeration, or small holes). Considering that thickness deviations include both large-scale trends (whole-roll bending, tension distribution) and small-scale mutations (bubbles, fiber agglomeration), we obtain a depth map Z(i, j) of the paper surface and separate it into large-scale and small-scale components, corresponding to the large-roll bending trend and local micro-defects, respectively. Thus, we have:

[0085] Perform two-dimensional discrete wavelet decomposition (Haar, Daubechies, etc.) on the depth map to obtain:

[0086] Z(i,j)=Z coarse (i,j)+Z fine (i,j), Z coarse (i,j) represents the two-dimensional wavelet processing of the depth map

[0087] The low-frequency components after decomposition reflect the overall deformation and trend fluctuation of the paper. fine (i, j) represents the high-frequency components after two-dimensional wavelet decomposition of the depth map, including microstructure differences and local abnormal changes;

[0088] Compute the absolute deviation of the two parts:

[0089] d C (i,j)=|Z coarse (i,j)-μ C |,d F(i,j)=|Z fine (i,j)|,d C (i,j) represents the low-frequency component

[0090] With its global average μ C The degree of deviation indicates the thickness fluctuation of each point in the overall trend, d F (i, j) represents the absolute value of the high-frequency component, describing the intensity of the local abnormal fluctuation at each point;

[0091] Furthermore, the large scale is used to detect the overall bending trend, and the small scale accurately captures local anomalies without mixing in the global drift.

[0092] Furthermore, the large-scale deviation d obtained after multi-scale decomposition C (i, j) still contains trend drift caused by environmental factors (such as tension and humidity). Traditional global filtering or fixed models are difficult to adapt to non-stationary and nonlinear fluctuations in the production line. Therefore:

[0093] By integrating the environmental state into the weight based on local adaptive regression, fitting the large-scale trend of thickness, and eliminating the global drift caused by tension / humidity, we have:

[0094] Perform a weighted quadratic polynomial fit on each center on the window to obtain the trend obtained by local regression fitting

[0095]

[0096] Among them, N R (i, j) represents the pixel set with radius R at the center (i, j), represents the row and column index in the neighborhood, represents the coarse-scale deviation, a, b, c represent the local fitting coefficients of the quadratic polynomial, ki, Represents neighborhood points Row and column offsets relative to the center, Represents the weight coefficient of local regression;

[0097] For the weight coefficient:

[0098]

[0099] Among them, K represents the kernel function, represents the standardized spatial distance, λτ, λ H They represent the attenuation sensitivity coefficients of tension difference and humidity difference to weight, τij, Indicates the tension sensor at pixel (i, j) and The tension value measured at ij 、 Represents the humidity sensor at pixel (i, j) and The humidity value measured at the pixel is 1 / 2. Therefore, by adaptively fitting a quadratic polynomial surface to each pixel within its radius neighborhood, it is possible to flexibly capture local thickness trends of arbitrary shapes. Furthermore, by introducing tension difference and humidity difference into the weights, the environmental state and spatial distance jointly determine the fitting contribution: when the tension / humidity of a neighborhood location differs significantly from that of the center point, its weight is exponentially suppressed, avoiding misleading trend estimation using inconsistent environmental data.

[0100] In the above example, the residual error after fitting contains both random noise and real defect signals (such as local bubbles or indentations). It is difficult to distinguish between noise peaks and real mutations by directly using the threshold. Therefore:

[0101] Let the initial fitting residual The residual contains both structural thickness anomaly signals (such as bubbles, fiber agglomeration, local collapse, etc.) and a large amount of random noise (sensor noise, quantization error, small environmental interference, etc.). If the residual field is directly thresholded or simply filtered, the following will occur:

[0102] Mis-smoothing structural mutations will weaken or lose the real defect signal and leave residual noise pseudo peaks, resulting in an increase in false positives. Therefore, iterative updates are performed:

[0103]

[0104] in, represents the residual after the t+1th iteration, η represents the diffusion rate, N(i,j) represents the neighborhood set, g represents the diffusion function, and the smoothing intensity is adjusted according to the residual difference. Indicates the pixel at the tth iteration The residual at , it should be noted that the diffusion function Δ represents the residual difference between adjacent pixels, η1 represents the edge sensitivity parameter, and η2 represents the decay rate exponent, which controls the high-order sensitivity of the diffusion function to the difference;

[0105] Through the above, in areas with fewer defects, iterative diffusion can quickly suppress random noise and make the residual distribution smoother. When the residual difference exceeds the threshold (controlled by η1 and η2), the diffusion function is almost close to 0, avoiding over-smoothing of the real defect edge. Secondly, the diffusion intensity is dynamically calculated for the residual amplitude of each neighborhood to avoid manually setting a fixed filter that is not suitable for the production noise characteristics.

[0106] Building on the above, different paper grades (basis weight ω) and production speeds v will change the tension-thickness coupling mechanism and measurement inertia error. At high speeds, the mismatch between the sensor sampling time window and the actual thickness change rate will cause measurement delay or ambiguity. Different basis weights will affect the dynamic response of the paper web. Therefore, speed and basis weight are added as coupling factors to the residual amplification to correct the measurement deviation, resulting in:

[0107]

[0108] in, represents the corrected residual, represents the residual error at pixel (i, j) after anisotropic diffusion denoising after the Tth iteration, v0 represents the target velocity, ω0 represents the target basis weight, μ ω Represents the basis weight coupling coefficient, which determines the sensitivity of basis weight deviation to residual amplification, μ v It represents the velocity coupling coefficient and describes the sensitivity of linear velocity deviation to the residual error. Therefore, when v deviates from v0, the residual is amplified to compensate for the error caused by measurement delay, allowing the control system to respond faster. When the basis weight deviates significantly from the calibrated ω0, the detection sensitivity is adjusted accordingly to avoid misjudging inherent thickness fluctuations as defects.

[0109] Paper fibers will expand or shrink when humidity and temperature change, causing false thickness deviation. This effect is particularly obvious at the microscopic scale. If it is not compensated, it will lead to a large number of false positives. By adding a double inverse compensation model for humidity and temperature, the false increase caused by fiber expansion and contraction is automatically deducted at high humidity or high temperature, and the following is obtained:

[0110] f inv (H ij ,T ij )=1-θ H tanh(H ij -H0)-θ T tanh(T ij -T0);

[0111]

[0112] Among them, f inv (H ij ,T ij ) represents the reverse fiber expansion and contraction compensation coefficient, which is used to deduct the false deviation caused by fiber expansion or contraction caused by humidity / temperature. ij represents the ambient humidity at pixel (i, j), T ij represents the ambient temperature at pixel (i, j), θ HRepresents the humidity compensation coefficient, which controls the compensation amplitude when the humidity deviates from the reference. Tanh represents the hyperbolic tangent function. After inputting a real number, the output is between (-1, 1) and has a smooth saturation characteristic. H0 represents the humidity reference value. θ T It represents the temperature compensation coefficient, which controls the compensation amplitude when the temperature deviates from the reference. T0 represents the temperature reference value. Represents the thickness deviation residual after final correction. In contrast to conventional compensation, the expansion and contraction components are deducted through the tanh term, allowing the residual to focus more on irreversible structural thickness anomalies. The tanh function is linear for small differences and saturates for large differences, which can avoid overcompensation.

[0113] Finally, the machine direction of the paper machine (MD, the direction the paper runs on the paper machine, such as longitudinal direction, and CD, the cross-width direction of the paper) has different sensitivities to thickness fluctuations. Fiber stretching is common in the axial direction of the ink roller, while warping waves are more likely to occur in the cross direction. Therefore:

[0114] Based on the cosine weighting of the main fiber direction (cross direction CD) and the barrel direction (ink roller axial direction MD), the same amplitude deviation in different directions can be treated differently, and the thickness deviation index s2 can be determined:

[0115] g ij =1+δcos(2(θ ij -θ0))

[0116] Among them, Ω2 represents the set of all pixel pairs representing the paper area in the image, g ij Represents the directional weighted coefficient, which determines the weight of the pixel according to the angle between the main direction of the fiber and the direction of the barrel. N (i, j) represents the thickness standard deviation of the local neighborhood N(i, j) of pixel (i, j), which is used for local normalization. ε represents a constant to avoid division by 0. δ represents the directional sensitivity coefficient, which controls the influence of the directional weight on the overall index. θ ij It represents the main direction angle of the fiber at pixel (i, j), and θ0 represents the axis direction angle of the barrel or production line. Therefore, the deviation on MD is amplified by the positive weight, and the deviation on CD is appropriately suppressed. Both contribute fairly to the final s2, reflecting the true quality. Dividing by the thickness standard deviation prevents excessive amplification of high-texture areas, making the indicator equally stable in high / low-texture areas.

[0117] Traditional thickness detection often applies unified filtering or global fitting across the entire image, making it difficult to simultaneously account for both the large-scale bending trend caused by barrel compression and the local thickness fluctuations caused by fiber agglomeration or bubbles. This solution uses wavelet decomposition to separate the depth map into low-frequency, large-scale components and high-frequency, local components. Analysis is then performed at each detection point based on tension, humidity, and thickness deviation, dynamically eliminating thickness drift caused by macroscopic trends and environmental conditions.

[0118] Subsequently, the true local mutation boundary is retained based on anisotropic diffusion, which can not only accurately capture the true thickness anomalies in the area of ​​tiny bubbles or fiber aggregation, but also avoid misjudging the full-width bending as a defect.

[0119] Density index analysis module, used to calculate the density uniformity index of paper, including:

[0120] Collect ultrasonic features. For each position (which can be a pixel or an ultrasonic probe sampling point), collect the ultrasonic propagation delay and ultrasonic amplitude attenuation and combine them into a two-dimensional feature vector V f ;

[0121] Assume that the total number of samples is P and calculate the mean of all eigenvectors That is V f The mean of , and calculate the covariance matrix based on the mean Covariance coupling considers the relationship between delay and attenuation;

[0122] Perform eigendecomposition on the covariance matrix and take the eigenvector matrix U corresponding to the first l largest eigenvalues l ∈R 2×2 , for V f Project and reconstruct:

[0123] represents the reconstruction operator that returns to the original space after dimensionality reduction, f represents the ultrasonic sampling point index, represents the reconstructed density vector;

[0124] According to V f and Calculate the reconstruction error, which can highlight atypical abnormal blocks and improve the detection sensitivity of small density unevenness;

[0125] Considering the contribution of chemical residue to density inhomogeneity, the reconstruction error is amplified by the index:

[0126] represents the local abnormality score, e f represents the reconstruction error, η3 represents the chemical amplification factor, S 4,f Represents the chemical residue concentration index at the fth sampling point;

[0127] Calculate density uniformity index through global aggregation

[0128] Ultrasonic delay and attenuation appear as a strongly correlated low-dimensional subspace in uniform regions; however, local density anomalies (such as fiber accumulation and bubbles) deviate from this subspace in the two-dimensional feature space. Low-rank reconstruction is used to extract background patterns into density vectors, maximizing the reconstruction error corresponding to the anomaly.

[0129] Chemical residues can cause uneven expansion of fibers after moisture absorption, and the changes in their ultrasonic delay and attenuation characteristics are slight but important. f Identification is easily masked by physical noise. The introduction of chemical indicators for multiplicative amplification allows chemically dominated density anomalies to obtain higher weights in the final score, thereby increasing the sensitivity to chemically induced defects and achieving accurate analysis of density uniformity.

[0130] Furthermore, it also includes:

[0131] The chemical index analysis module is used to determine the chemical residue index. The steps are as follows:

[0132] Considering that paper thickness deviation, surface defect, and density uniformity will directly distort hyperspectral or ultrasonic measurements: thickness changes the optical path length, cracks cause local light scattering, and density unevenness affects the absorption depth, if these are not corrected first, the subsequent chemical residue inversion model will confuse physical interference with the true chemical signal, resulting in false positives or missed negatives. Therefore:

[0133] Perform physical coupling compensation on the original reflectivity (or transmittance) vector:

[0134] R phys (λ)=R raw (λ)exp[-α1s1-α2s2-α3s3]

[0135] Among them, R phys (λ) represents the corrected reflectivity, R raw (λ) represents the original reflectivity (the original reflectivity at wavelength λ is collected by the hyperspectral sensor installed on the production line), α1, α2, and α3 are the corresponding interference attenuation coefficients, which are calibrated according to empirical methods or experiments and will not be repeated here. λ represents the spectral band index, which strips off the signal offset and distortion caused by thickness, cracks, and density fluctuations, and normalizes the hyperspectral data to a physically consistent domain, ensuring that its subsequent inversion focuses more on chemical absorption characteristics, greatly reducing the false positive / false negative rate.

[0136] The dispersion and adhesion efficiency of chemical retention aids in slurry are strongly affected by slurry viscosity, retention agent dosage and white water reuse concentration. Through coupling modeling, slurry rheology (slurry viscosity) and process parameters (retention agent dosage, white water reuse concentration) are mapped together into retention efficiency factors (coupling effects are likely to exist between variables, such as:

[0137] High viscosity + high dosage can easily lead to excessive aggregation, resulting in chemical deposition and uneven distribution;

[0138] Low viscosity + high white water concentration easily produces a dilution effect, resulting in insufficient effective retention aid;

[0139] Viscosity fluctuation + white water reuse → prone to regional retention instability, affecting residual concentration detection), then:

[0140] R eff =R0+β V V+β ret D ret -β BW C BW

[0141] Among them, R eff represents the retention efficiency factor, R0 represents the baseline retention efficiency, β V , β ret , β BW They represent the corresponding coupling coefficients respectively, V represents the slurry viscosity, which describes the viscous characteristics of the slurry in the flowing state. The higher the viscosity, the slower the retention agent disperses and the more likely it is to aggregate. If it is too low, the dispersion is fast but unstable. ret Indicates the dosage of retention agent, C BW Indicates the white water reuse concentration (considering the impact of reflux);

[0142] This allows quantification of chemical loss / retention along the production line from slurry to finished product, ensuring that subsequent concentration predictions are based on true retention rather than just the original dosage.

[0143] Furthermore, the adsorption equilibrium of chemical residues is not only related to the solid phase retention efficiency, but also varies with pH and temperature:

[0144] Different pH values ​​will change the molecular charge and surface potential, while temperature will change the activation energy E a Changing the dynamic equilibrium constant K(Θ), therefore, embedding the chemical reaction kinetics into the model accurately captures the dynamic adjustment effect of pH and temperature on the residue, avoiding overestimation or underestimation of concentration under high temperature or extreme pH conditions, then:

[0145] K(Θ)=K0exp[-E a / (R′Θ)]

[0146] Among them, Q ads Indicates the adsorption capacity ratio per unit fiber, C solrepresents the equivalent solution concentration (based on the calibrated spectrum), K0 represents the initial equilibrium constant, R′ represents the gas constant, Θ represents the ambient temperature of the production site, K(Θ) represents the adsorption equilibrium constant, which is the affinity between the adsorbent and the adsorbate at a given temperature. A larger affinity indicates a stronger adsorption capacity. exp represents an exponential function.

[0147] Therefore, based on the corrected reflectivity R obtained above phys (λ), retention efficiency factor R eff , adsorption ratio Q ads Calculate the chemical residual concentration index S 4,f =ω1||R phys ||1+ω2R eff Q ads , ω1 and ω2 represent the fusion weights respectively;

[0148] From this, the final chemical residue index is calculated

[0149] Chemical residue detection not only relies on spectral signals but is also affected by the combined effects of retention agent dosage, slurry rheology, white water reuse concentration, pH, and temperature. Therefore, by pre-eliminating the interference of physical state on the spectrum, a retention efficiency and adsorption equilibrium model is established to accurately estimate the chemical retention amount, ensuring that the residual concentration prediction can be immediately corrected after each feeding or process adjustment. This full-link design can reflect the impact of production parameter changes on residues in real time, and quickly transmit quality feedback to the retention agent dosing, white water circulation, or temperature control system, thereby improving the accuracy of residue detection and enabling the production line to automatically and intelligently adjust the chemical addition strategy, significantly reducing chemical waste and environmental emission risks.

[0150] In summary, the surface defect index s1, thickness deviation index s2, density uniformity index s3, and chemical residue index s4 are mapped to [0,1] through normalization processing, and then the corresponding weight coefficients are assigned to them. The comprehensive quality index Q of the mi paper is obtained by weighted summation. mi , obtain the quality index threshold corresponding to the paper. If the comprehensive quality index is less than the quality index threshold, the current finished paper quality is defective, that is, the production line production process and other operations need to be adjusted. Otherwise, it is a good product and no corresponding adjustment is required. In this way, the quality safety monitoring and control optimization of paper and paper production lines are achieved.

[0151] Finally, the present invention also provides:

[0152] A method for detecting finished paper products is implemented according to the finished paper product detection system based on the frame paper production process.

[0153] A computer device includes a memory and a processor, wherein the memory stores a computer program and the processor implements the system when executing the computer program.

[0154] A computer-readable storage medium stores a computer program, which implements the system when executed by a processor.

[0155] For the above, please refer to the system of the present invention and will not be described in detail here.

[0156] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements will not cause the essence of the corresponding technical solutions to deviate from the protection scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A finished paper product inspection system based on the frame paper production process, characterized in that: include: The defect index analysis module creates an initial defect image of the paper, constructs a graph structure to calculate the neighborhood weights between pixels, uses graph convolution to smooth the image using the adjacency weights, and combines density information with mechanical vibration to determine surface defect indicators. The thickness index analysis module obtains the paper depth map and separates it into different scale components, calculates the absolute deviation under the scale components, and fits the scale trend based on the environmental state, where: Define the initial fitting residual and perform iterative update diffusion. Introduce production speed and paper basis weight as coupling factors to correct measurement deviation. Introduce humidity and temperature into the inverse compensation model to obtain thickness deviation residual. Combine the thickness deviation residual with the paper machine direction to determine the thickness deviation index. The density index analysis module constructs a two-dimensional feature vector based on the propagation delay and amplitude attenuation of the ultrasonic wave, and determines the local anomaly score by combining the reconstruction error with the chemical index to define the density uniformity index; Chemical index analysis module, which calculates chemical residue index by combining reflectivity, retention efficiency factor, and adsorption capacity ratio; In response to the input surface defect index, thickness deviation index, density uniformity index and chemical residue index, the comprehensive quality index of the paper is output to complete the inspection of the finished paper product.

2. The paper product detection system based on the frame paper production process according to claim 1 is characterized in that: The method for determining the surface defect index is: Calculate the adjacency weight between each pixel and its neighboring pixels; The defect confidence of each pixel is calculated by weighted averaging the information of neighboring pixels through the graph convolution method Among them, N(p) represents the neighborhood set of pixel p, D pp 、D qq denote the degree of pixel p and pixel q respectively, Wx(q) denotes the original defect confidence value obtained from the image, σ(·) denotes the activation function, s2 denotes the humidity or temperature value of the paper surface, and L denotes the global illumination level of the current environment; Calculate the surface defect index s1: Among them, Ω1 represents the set of all pixels in the paper area in the image, α represents the cross-modal fusion weight, h(s3(p),s5) represents the magnification function, s3(p) represents the density information, and s5 represents the vibration state value of the paper.

3. The paper product detection system based on the frame paper production process according to claim 1 is characterized in that: The method of fitting scale trend in combination with environmental state is: The large-scale deviation obtained after scale decomposition includes the trend drift caused by environmental factors. Based on local adaptive regression, the environmental state is incorporated into the weight, the large-scale trend of thickness is fitted, and the global drift caused by tension / humidity is eliminated. Then, we have: Perform a weighted quadratic polynomial fit on each center on the window to obtain the trend obtained by local regression fitting Among them, N R (i, j) represents the pixel set with radius R at the center (i, j), (k, l) represents the row and column index in the neighborhood, d C (k, l) represents the coarse scale deviation, a, b, c represent the quadratic polynomial local fitting coefficients, ki, lj represent the row and column offsets of the neighborhood point (k, l) relative to the center, Represents the weight coefficient of local regression.

4. The paper product detection system based on the frame paper production process according to claim 1 is characterized in that: The method for determining the thickness deviation index by combining the thickness deviation residual with the paper machine direction is as follows: Adding a double inverse compensation model to humidity and temperature, we have: f inv (H ij ,T ij )=1-θ H tanh(H ij -H0)-θ T tanh(T ij -T0); Among them, f inv (H ij ,T ij ) represents the reverse fiber expansion and contraction compensation coefficient, T ij represents the ambient temperature at pixel (i, j), θ H represents the humidity compensation coefficient, tanh represents the hyperbolic tangent function, H0 represents the humidity reference value, θ T represents the temperature compensation coefficient, T0 represents the temperature reference value, represents the thickness deviation residual after correction, represents the corrected residual; Considering the different sensitivities to thickness fluctuations in the machine direction and cross direction of the paper: Based on the cosine weighting of the main fiber direction and the barrel direction, the same amplitude deviation in different directions can be treated differently, and the thickness deviation index s2 is calculated: Among them, Ω2 represents the set of all pixel pairs representing the paper area in the image, g ij represents the directional weighted coefficient, which determines the weight of the pixel according to the angle between the main direction of the fiber and the direction of the barrel. σN(i,j) represents the thickness standard deviation of the local neighborhood N(i,j) where the pixel (i,j) is located. ε represents a constant. δ represents the directional sensitivity coefficient. θ ij It represents the main direction angle of the fiber at pixel (i, j), and θ0 represents the axis direction angle of the barrel or production line.

5. The paper product detection system based on the frame paper production process according to claim 1 is characterized in that: The method for determining the density uniformity index is: Calculate the mean and covariance matrix of all eigenvectors; Perform eigendecomposition on the covariance matrix and take the first l corresponding eigenvector matrices; Project and reconstruct to obtain the reconstructed density vector; Calculate the reconstruction error; Considering the influence of chemical residues, the reconstruction error is amplified by the index: Get the score representing the local abnormality; Global aggregation calculates density uniformity indicators.

6. The paper product detection system based on the frame paper production process according to claim 1 is characterized in that: The method for determining the chemical residue index is: Considering that thickness changes the optical path length, cracks cause local light scattering, and uneven density affects the absorption depth: Perform physical coupling compensation on the original reflectivity vector to obtain the corrected reflectivity; Through coupled modeling, slurry rheology and process parameters are mapped together into retention efficiency factors; Chemical reaction kinetics were embedded in the model to determine the adsorption rate; Calculate the chemical residual concentration index based on reflectivity, retention efficiency factor, and adsorption capacity ratio; Determine chemical residue indices based on multi-point chemical residue concentration indices.

7. The paper product detection system based on the frame paper production process according to claim 6 is characterized in that: The adsorption ratio is obtained according to the following relationship: Among them, Q ads Indicates the adsorption capacity ratio per unit fiber, C sol represents the equivalent solution concentration, K0 represents the initial equilibrium constant, R′ represents the gas constant, Θ represents the ambient temperature, and K(Θ) represents the adsorption equilibrium constant.

8. A method for detecting finished paper products, characterized in that: The method is implemented according to the finished paper product detection system based on the frame paper production process according to any one of claims 1 to 7.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the system according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the system according to any one of claims 1 to 7 is implemented.

Images