Paper plastic forming quality evaluation method fusing visual features and weighing data

Abstract

The application relates to a paper-plastic forming quality evaluation method fusing visual features and weighing data and belongs to the technical field of industrial automation detection. The method comprises the following steps: acquiring a continuous weight data sequence and a surface gray image of a paper-plastic product, and extracting a weight fluctuation curve and a surface defect feature vector respectively; when a specific abnormal condition is met, time domain and space domain filtering is respectively performed on the weight data and the visual features to obtain correction parameters; cross-correlation analysis is performed on the corrected weight peak-valley interval and the visual pixel width range, if the two present a synchronous change mode, a coupling abnormality index is calculated; finally, a quality deviation degree value is found in a segmented linear mapping table according to the coupling abnormality index and is compared with a qualified judgment limit, and a quality judgment result is output. The application can effectively solve the problems of quality judgment rigidity and high misjudgment rate under double uncertain conditions, and realizes continuous quality grade judgment.

Description

Technical Field

[0001] This application belongs to the field of industrial automation testing technology, specifically a method for evaluating the quality of paper-plastic molding that integrates visual features and weighing data. Background Technology

[0002] As the core carrier of food packaging and industrial cushioning materials, the quality stability of paper-plastic molded products directly affects product distribution safety and enterprise production efficiency. In large-scale automated production lines, the inspection time for a single product typically needs to be controlled within milliseconds. This requires the quality assessment system to complete defect identification and acceptance determination in an extremely short time. Currently, this field generally adopts a multi-sensor fusion technology approach. This involves simultaneously using weighing sensors to acquire the weight data of the product and using vision sensors to collect image information of the product's surface. The weight data and image data are processed independently, and fixed thresholds are set to determine whether the product's filling amount meets the standard and whether the surface quality is acceptable. Finally, the weight determination result and the vision determination result are logically combined to output a pass / fail conclusion.

[0003] However, existing evaluation schemes generally suffer from rigid decision boundaries. When test data falls within the critical range between acceptable and unacceptable, fixed-threshold binarization ignores the gradual and ambiguous nature of quality characteristics in actual production. In the paper-plastic molding process, the edges of surface defects often exhibit blurred transition zones due to changes in lighting, and traditional gradient operators cannot characterize the degree of uncertainty at the boundaries; weighing data also repeatedly jumps around the threshold due to equipment vibration or material density fluctuations. When the uncertainties of blurred visual edges and critical weighing data are superimposed, a slight deviation of sensor data from the threshold can trigger a completely opposite judgment. This abrupt judgment clearly contradicts the continuous and gradual characteristics of product quality, leading to an increased misjudgment rate for products in the critical range and failing to reflect the intermediate state between acceptable and unacceptable.

[0004] Therefore, how to establish an evaluation mechanism that can quantify the degree of uncertainty and achieve continuous quality grade determination under the dual uncertainty conditions of blurred visual feature boundaries and weighing data in the critical range has become a key issue in improving the accuracy and robustness of paper-plastic molding quality inspection. Summary of the Invention

[0005] To address the above issues, this application provides a paper-plastic molding quality assessment method that integrates visual features and weighing data. This method is used to solve the technical problem that it is difficult to establish an assessment mechanism that can quantify the degree of uncertainty and achieve continuous quality level determination under the dual uncertainty conditions of blurred visual feature boundaries and weighing data in the critical range.

[0006] To achieve the above objectives, the technical solution adopted in this application is as follows: This application provides a method for evaluating the quality of paper-plastic molding that integrates visual features and weighing data. The method includes: acquiring a continuous weight data sequence of the paper-plastic product within a static weighing stability window, and extracting a weight fluctuation curve; acquiring a grayscale image of the paper-plastic product surface, and extracting a surface defect feature vector; when the first peak-to-valley spacing of the weight fluctuation curve is less than a preset time interval threshold and the number of weight increase / decrease direction switching exceeds a preset number of frequent changes, performing time-domain filtering on the continuous weight data sequence to obtain corrected weight nonlinear drift trajectory parameters; when the number of edge contour breakpoints in the surface defect feature vector is greater than a preset breakage threshold and the pixel width of the blurred transition band exceeds a preset blurred width upper limit, further refining the surface defect feature vector... Linear spatial domain filtering is used to obtain the corrected local variance intensity parameter of grayscale value distribution. The second peak-valley spacing in the nonlinear drift trajectory parameter of weight is cross-correlation analysis with the pixel width range in the local variance intensity parameter of grayscale value distribution. If the shortening trend of the second peak-valley spacing and the expansion trend of the pixel width range show a synchronous change pattern, the ratio of the number of switching times of weight increase / decrease direction to the number of edge contour breakpoints is extracted. Combined with the product of instantaneous weight deviation amplitude and grayscale jump slope of defect edge, the coupling anomaly index is calculated. According to the coupling anomaly index, the corresponding quality deviation value is found in the piecewise linear mapping table. The quality deviation value is compared with the qualified judgment limit, and the product quality judgment result is output.

[0007] This application provides a method for evaluating the quality of paper-plastic molding that integrates visual features and weighing data. This method simultaneously acquires continuous weight data sequences of paper-plastic products within a static weighing stability window to extract weight fluctuation curves, and acquires grayscale images of the paper-plastic product surface to extract surface defect feature vectors, thus constructing a dual-modal data foundation. Based on this, when the first peak-to-valley spacing of the weight fluctuation curve is less than a preset time interval threshold and the number of weight increase / decrease direction switching exceeds a preset number of frequent changes, temporal filtering of the continuous weight data sequence eliminates interference from the sensor's dynamic response, thereby obtaining corrected weight nonlinear drift trajectory parameters. Simultaneously, when the number of edge contour breakpoints in the surface defect feature vector exceeds a preset breakpoint threshold and the pixel width of the blurred transition zone exceeds a preset blurred width upper limit, spatial domain filtering is applied to the surface defect feature vector to remove background texture interference, thereby obtaining the corrected local variance intensity parameter of the grayscale value distribution. Subsequently, the scheme performs cross-correlation analysis on the second peak-valley spacing in the corrected weight nonlinear drift trajectory parameter and the pixel width range in the corrected local variance intensity parameter of the grayscale value distribution. Once a synchronous change pattern is detected between the shortening trend of the second peak-valley spacing and the expansion trend of the pixel width range, the ratio of the number of switching times of weight increase / decrease direction to the number of edge contour breakpoints is extracted. Combined with the product of the instantaneous weight deviation amplitude and the grayscale jump slope of the defect edge, a coupling anomaly index that reflects the uneven molding pressure and abnormal material density distribution is calculated.

[0008] Finally, based on the coupling anomaly index, the corresponding quality deviation value is found in a pre-established piecewise linear mapping table, and this value is compared with the acceptance threshold to output the product quality judgment result. This process enables the originally isolated weight oscillation characteristics and visual ambiguity characteristics to achieve deep integration and mutual verification at the physical level, transforming the dual uncertainty into a quantifiable comprehensive index. This effectively solves the technical problem of establishing an evaluation mechanism that can quantify the degree of uncertainty and achieve continuous quality grade judgment under the dual uncertainty conditions of blurred visual feature boundaries and weighing data in the critical range. Therefore, it avoids the increased misjudgment rate and inability to reflect the continuous gradual change characteristics of product quality caused by traditional hard segmentation methods, significantly improving the accuracy, robustness, and ability to distinguish products in critical states in paper-plastic molding quality inspection. Attached Figure Description

[0009] Figure 1 A flowchart of a paper-plastic molding quality assessment method that integrates visual features and weighing data is provided in this application; Figure 2 A flowchart for extracting the weight fluctuation curve in the paper-plastic molding quality assessment method that integrates visual features and weighing data provided in this application; Figure 3 A flowchart illustrating the method for extracting surface defect feature vectors in the paper-plastic molding quality assessment method that integrates visual features and weighing data provided in this application; Figure 4 The flowchart shows the process of obtaining the corrected weight nonlinear drift trajectory parameters by performing time-domain filtering on a continuous weight data sequence in the paper-plastic molding quality assessment method that integrates visual features and weighing data provided in this application. Figure 5The flowchart illustrates the process of obtaining the corrected grayscale value distribution local variance intensity parameter by spatial domain filtering of the surface defect feature vector in the paper-plastic molding quality assessment method that integrates visual features and weighing data provided in this application; Figure 6 The flowchart shows the cross-correlation analysis of the second peak-valley spacing in the weight nonlinear drift trajectory parameter and the pixel width range in the local variance intensity parameter of gray value distribution in the paper-plastic molding quality assessment method that integrates visual features and weighing data provided in this application; Figure 7 A flowchart illustrating the method for constructing a piecewise linear mapping table in the paper-plastic molding quality assessment method that integrates visual features and weighing data provided in this application. Detailed Implementation

[0010] To enable those skilled in the art to better understand the technical solution, the present application will be described in detail below with reference to the embodiments. The description in this section is only exemplary and explanatory, and should not be used to limit the scope of protection of the present application in any way.

[0011] like Figure 1 The diagram shown is a flowchart of a paper-plastic molding quality assessment method that integrates visual features and weighing data, provided in an embodiment of this application. The method specifically includes: Step S110: Obtain the continuous weight data sequence of the paper-plastic product within the static weighing stability window, and extract the weight fluctuation curve; and obtain the grayscale image of the surface of the paper-plastic product, and extract the surface defect feature vector.

[0012] This step involves simultaneously acquiring weight distribution and surface morphology information of the paper-plastic product using multimodal sensors to construct a joint input data space. First, a continuous weight data sequence of the paper-plastic product within a static weighing stability window is obtained from weighing sensors deployed at the end of the production line. This static weighing stability window is defined as the time interval between the paper-plastic product entering the weighing platform and the mechanical vibration decaying to a preset range, reflecting the product's true weight distribution characteristics. Based on the continuous weight data sequence composed of high-frequency sampling points, a weight fluctuation curve is extracted to characterize the dynamic trajectory of the weight value on the time axis, capturing nonlinear drift caused by uneven material density or molding pressure fluctuations.

[0013] Simultaneously, raw grayscale images of the paper-plastic product surface are acquired using an industrial camera. Visual inspection technology is then used to process these grayscale images, focusing on changes in surface texture and light reflection intensity. Surface defect feature vectors are extracted from the grayscale images. These vectors contain multidimensional parameters describing surface geometric anomalies and irregular grayscale distribution, such as edge contour continuity, grayscale jump rate in defect areas, and width of blurred transition zones.

[0014] The aforementioned weight fluctuation curves and surface defect feature vectors together constitute an initial dataset that integrates the internal material filling state and external surface structure features, providing data support for solving the misjudgment problem caused by boundary blurring, illumination changes, or data critical jumps in single-modal visual inspection.

[0015] Specifically, in one alternative implementation, such as Figure 2 The diagram shown is a flowchart of a method for extracting weight fluctuation curves according to an embodiment of this application. The method specifically includes: Step S210: Perform frequency domain analysis on the continuous weight data sequence to obtain the oscillation frequency sequence and deviation amplitude sequence within the sampling period.

[0016] The acquired continuous weight data sequence is transformed from the time domain to the frequency domain for analysis. The aim is to decompose the original weight signal into a superposition of different frequency components, thereby identifying specific oscillation modes caused by mechanical vibration, hydrodynamic fluctuations, or external disturbances. For example, the Fast Fourier Transform (FFT) algorithm is used, with the duration of the static weighing stability window as the calculation window for the FFT. The continuous weight data sequence within the window is then transformed in the frequency domain, converting the time series signal into a frequency-amplitude spectrum.

[0017] The input data for the FFT algorithm is a continuous weight data sequence within a static weighing stability window. The intrinsic connection between this input data and the quality inspection of paper-plastic molding lies in the fact that the weight data sequence of paper-plastic products is essentially a signal formed by the change in the mass of pulp fibers deposited in the mold cavity over time. Under ideal molding conditions, the mass of fibers deposited on the molding surface per unit time should remain constant, and the weight data sequence exhibits stable characteristics. When there are periodic fluctuations in molding pressure, these fluctuations change the instantaneous accumulation rate of fibers in the cavity, resulting in the superposition of periodic components corresponding to the pressure fluctuation frequency in the weight data sequence. Therefore, the input data implicitly contains frequency information reflecting the stability of the molding pressure.

[0018] The FFT algorithm outputs an oscillation frequency sequence and a deviation amplitude sequence. The oscillation frequency sequence corresponds to the frequency values ​​of different periodic components in the weight data. Its intrinsic connection to paper-plastic molding quality inspection lies in the fact that these frequency values ​​can be traced back to the operating characteristic frequencies of specific mechanical components. For example, an abnormally increased frequency in a hydraulic pump plunger may indicate wear in the hydraulic system, and a side band in the meshing frequency of transmission gears may indicate gear eccentricity or bearing damage. These mechanical anomalies can affect the filling uniformity of paper-plastic products through the pressure transmission path. The deviation amplitude sequence corresponds to the energy contribution of each frequency component in the weight signal. Its intrinsic connection to quality inspection lies in the fact that when the deviation amplitude corresponding to a certain frequency is significantly higher than the normal baseline, it indicates that the interference factor from that frequency source has had a substantial impact on the fiber stacking quality. The larger the deviation amplitude corresponding to this interference factor, the more severe the damage to the product weight consistency. By comparing the frequency values ​​in the output data with a database of equipment fault characteristic frequencies, reverse tracing from the weight signal to the equipment status can be achieved.

[0019] In actual paper-plastic production lines, a sampling frequency of 1000Hz can be selected, and 2000 continuous weight data points can be collected within a 2-second static weighing stabilization window as the analysis sample for one FFT. By analyzing this spectrum, the main frequency components with concentrated energy within the sampling period are extracted to form an oscillation frequency sequence. This sequence characterizes the intensity of the weight signal oscillation at different frequency points. At the same time, the signal strength information at the corresponding frequency points is extracted to form a deviation amplitude sequence, which quantifies the degree of weight deviation from the reference value at each frequency component.

[0020] For example, in a real paper-plastic production line, if there is mechanical vibration at 45Hz, frequency domain analysis will accurately capture this frequency and output the corresponding deviation amplitude. By mapping weight data to the frequency domain, it is possible to effectively distinguish the sensor's true load response signal from environmental noise, providing a precise frequency basis for subsequent filtering processing.

[0021] Step S220: Calculate the standard deviation of the oscillation frequency sequence under multiple weighings. If the standard deviation is greater than the preset standard deviation threshold, then perform smoothing filtering on the deviation amplitude sequence.

[0022] The statistical stability of the oscillation frequency sequence during multiple weighing processes was analyzed. The standard deviation of the oscillation frequency sequence is a key indicator of whether the system's oscillation pattern fluctuates drastically over time. A larger standard deviation indicates greater instability in the oscillation frequency during weighing, potentially indicating random interference or unsteady-state conditions. The oscillation frequency sequences corresponding to multiple weighing operations were statistically analyzed, and their standard deviations were calculated. The determination of the preset standard deviation threshold specifically included: during the offline calibration phase, under standard steady-state conditions on the production line, the oscillation frequency sequences of multiple batches of qualified paper and plastic products were continuously collected, and the standard deviations were calculated. The mean of the standard deviations plus three times the standard deviation (i.e., the 3σ principle) was used as the initial threshold; during the online adaptive adjustment phase, a sliding window algorithm was used to continuously record the standard deviation of the oscillation frequency under qualified conditions, dynamically updating the preset standard deviation threshold to adapt to systematic baseline drift.

[0023] If the calculated standard deviation is greater than the preset standard deviation threshold, it indicates that the current weight data contains significant random high-frequency interference, and direct fitting will lead to curve distortion. In this case, a smoothing filter can be applied to the deviation amplitude sequence, for example, using a moving average algorithm, a low-pass filter algorithm, or a Savitzky-Golay filter. This smoothing operation aims to suppress spikes and glitches in the deviation amplitude sequence, while retaining low-frequency components that reflect the overall trend of the system, thus providing a cleaner and more reliable data foundation for subsequent trajectory extraction.

[0024] Step S230: Based on the smoothed and filtered deviation amplitude sequence, extract the nonlinear drift trajectory of the weight value changing with time, and use a polynomial regression algorithm to fit the nonlinear drift trajectory to generate a fluctuation curve that reflects the sensor load response delay, which serves as the weight fluctuation curve.

[0025] Based on the smoothed and filtered deviation amplitude sequence, the macroscopic trend of weight change over time is reconstructed. During the paper-plastic molding process, due to the sensor's inherent physical characteristics (such as creep of elastic elements) and the delay of the signal conditioning circuit, weight readings often do not stabilize instantaneously, but rather exhibit a nonlinear, gradual change—a nonlinear drift trajectory. From the smoothed and filtered deviation amplitude sequence, the overall trend of weight value evolution over time is extracted, removing transient high-frequency oscillations and focusing on extracting the low-frequency drift components.

[0026] To mathematically model the drift process, a polynomial regression algorithm was used to fit the extracted nonlinear drift trajectory. Polynomial regression, through optimization strategies such as least squares, constructs a polynomial function model that best approximates the distribution trend of data points. The coefficients of this model directly reflect the sensor's response characteristics at different stages. By mapping the fitted polynomial curve back to the time axis, a fluctuation curve accurately reflecting the sensor's load response delay characteristics was generated and defined as the weight fluctuation curve. This curve not only eliminates high-frequency noise interference but also explicitly expresses the load response delay mechanism mathematically. This allows subsequent calculations of features such as peak-to-valley spacing and switching frequency to be based on a trajectory with clear physical meaning, thus significantly improving the robustness and accuracy of the quality assessment results.

[0027] Furthermore, the aforementioned technical process effectively eliminated the influence of random noise on the weight signal by combining frequency domain analysis and statistical testing, and accurately modeled the nonlinear response characteristics of the sensor using polynomial regression. This processing method, which integrates data-driven approaches with physical models, ensures that the extracted weight fluctuation curve can truly reflect the weight change pattern of paper-plastic products, laying a solid data foundation for subsequent use of this curve to determine peak-valley spacing and identify anomalies.

[0028] Furthermore, such as Figure 3 The diagram shown is a flowchart of a method for extracting surface defect feature vectors according to an embodiment of this application. The method specifically includes: Step S310: Perform convolution operation on the grayscale image using an edge detection operator to obtain an edge response map.

[0029] In practical applications, edge detection operators are mathematical operators built based on the first or second derivative principle of local gray-level abrupt changes in an image. They are used to identify pixels in an image where the gray-level value changes drastically. In this embodiment, after acquiring the original gray-level image of the surface of a paper-plastic product through an industrial camera or vision sensor, the edge detection operator is invoked to perform a convolution operation on the gray-level image, calculating the gradient magnitude and gradient direction of each pixel in the horizontal and vertical directions, thereby generating an edge response map that reflects the edge intensity.

[0030] This process transforms originally smooth image data into a feature map that highlights geometric contours, allowing for the quantification and extraction of pixel position and intensity information of defect edges. Specifically, the edge detection operator can be the Sobel operator, Laplacian operator, or Cannibal operator, etc., which are merely illustrative examples here. In this embodiment, through convolution operations, high response value regions in the edge response map correspond to potential defect contours on the surface of paper-plastic products, such as cracks, scratches, or folded edges, thus laying a data foundation for subsequent extraction of accurate geometric features.

[0031] Step S320: Based on the edge response map, extract the gray-level jump rate of the defect edge, calculate the gray-level difference between adjacent pixels corresponding to the gray-level jump rate, and obtain the cumulative sum of gray-level differences.

[0032] Specifically, the grayscale jump rate can refer to the amount of grayscale change per unit distance along the edge gradient direction in the edge response map. Its physical essence reflects the steepness of the material contrast on both sides of the defect edge. In this embodiment, the pixel coordinates of the defect edge can be located based on the edge response map, and a local pixel sequence containing the edge can be extracted along the gradient direction. The grayscale jump rate of the defect edge can be extracted by calculating the ratio of the grayscale difference between the first and last pixels of the sequence to the distance.

[0033] Simultaneously, for this local pixel sequence, the absolute value of the grayscale difference between adjacent pixels is calculated point by point, and all differences are summed to obtain the cumulative grayscale difference. This indicator not only covers the total contrast of the edge but also performs integral statistics on pixel fluctuations within the edge width, effectively distinguishing between sharp mechanical cuts and gentle shadow boundaries. For example, for sharp crack edges, the grayscale jump rate is usually high and the cumulative grayscale difference is concentrated within a narrow pixel range; while for blurry water stain edges, the grayscale jump rate is low and the cumulative grayscale difference is distributed in a wider transition zone.

[0034] Through the above calculation process, vague visual perception can be transformed into measurable numerical features.

[0035] Step S330: Identify the pixel width of the blurred transition zone by accumulating the gray-level difference, and calculate the local variance of the gray-level values ​​within the pixel width.

[0036] In practical applications, the blurred transition zone refers to the pixel band in an image of a paper-plastic product surface where the grayscale value gradually changes and transitions non-abruptly between the defect area and the normal background area. It is a grayscale gradient region on the paper-plastic material surface caused by diffuse reflection of light or material wetting characteristics, and its pixel width is a key geometric parameter characterizing the degree of defect blurring. Based on the cumulative grayscale difference obtained in step S320, the search is expanded pixel-by-pixel along the edge gradient direction. During the expansion process, the rate of change of the cumulative grayscale difference relative to the number of steps is calculated at each location. When this rate of change continuously decreases to near zero or a clear inflection point appears, it indicates that the grayscale change has become gradual. This location is defined as the boundary of the transition zone, thus determining the start and end positions of the blurred transition zone and obtaining its pixel width.

[0037] Within a defined pixel width, the variance of all pixel grayscale values ​​within the region is further calculated to obtain the local variance value of the grayscale values. This statistic objectively reflects the dispersion of grayscale distribution within the transition zone; a smaller variance value indicates a smoother transition, while a larger variance value indicates a more complex texture. Through this step, the fuzzy boundaries that were originally difficult to accurately define using binarization segmentation are transformed into quantized features with clear width and intensity attributes, effectively solving the problem of traditional edge detection being insensitive to weak boundaries.

[0038] Step S340: Count the number of continuous break points of the edge contour. If the number of continuous break points is greater than the preset break threshold, extract the spatial distribution density value of the edge contour.

[0039] Specifically, the continuity breakpoints of the edge contour can refer to pixel nodes that occur during edge tracking, where the edge link is interrupted due to image noise or inherent defects. The edge response map is refined or binarized based on a threshold. A neighborhood search algorithm (such as eight-neighbor search) is used to track the connected paths of the edge contour. During the tracking process, the locations of nodes where the path is broken or cannot be connected are recorded, and the number of continuity breakpoints of the edge contour is counted.

[0040] When the number of break points obtained from the statistics exceeds a preset break threshold, it indicates that the defective area has severe fragmentation or texture interference. In this case, the spatial distribution density value of the edge contour is further extracted. This spatial distribution density value is obtained by calculating the proportion of edge pixels to total pixels within a sliding window of a preset size, and is used to characterize the degree of concentration of defects in a local area. For example, when there are many break points and the spatial distribution density value is low, it may mean that there is discrete noise interference; while when the density value is high, it points to a large area of ​​surface damage.

[0041] It should be noted that the preset fracture threshold can be determined in the following ways: acquire multiple grayscale images of defect-free products, count the number of fracture points in each image using the same method, and calculate the statistical upper quantile (e.g., the 99th percentile) of the number of fracture points as the preset fracture threshold to filter out false fracture points caused by environmental noise; or, select a product sample known to contain the smallest detectable defect, multiply its number of fracture points by a redundancy coefficient, and set it as the preset fracture threshold to ensure the system's sensitivity to detecting minute defects.

[0042] Step S350: Obtain the gradient direction consistency of the gradient region corresponding to the spatial distribution density value, and extract the grayscale fluctuation amplitude corresponding to the texture roughness.

[0043] In practical applications, gradient direction consistency in gradient regions is a metric describing the regularity of edge pixel gradient directions within their local neighborhoods, reflecting the orderliness of defect textures. Within a region of interest with a defined spatial distribution density, the gradient direction angle of each edge pixel is calculated, and a gradient direction histogram is constructed. The gradient direction consistency in gradient regions is quantified by statistically analyzing the proportion of the main peak in the gradient direction histogram or calculating the standard deviation of the direction angle. High direction consistency indicates that the edges are regularly arranged (e.g., regular cracks); low direction consistency indicates a chaotic texture (e.g., surface fiber fuzzing). Simultaneously, the grayscale fluctuation amplitude corresponding to the texture roughness is extracted. This is typically achieved by calculating the contrast parameter of the grayscale co-occurrence matrix or the difference in grayscale extreme values ​​in a specific direction, used to characterize the severity of surface micro-geometry undulations. By combining gradient direction consistency and grayscale fluctuation amplitude, the micro-morphological features of paper-plastic products can be comprehensively characterized from two dimensions: texture arrangement and surface undulation.

[0044] Step S360: Combine the gray-level jump rate, cumulative gray-level difference, local variance of gray-level value, number of continuity breakpoints, spatial distribution density, gradient direction consistency of the gradient region, and gray-level fluctuation amplitude into a surface defect feature vector.

[0045] In this embodiment, the surface defect feature vector is an ordered set of the aforementioned independent features. It is constructed by arranging scalar values ​​of multiple dimensions in a predetermined order to form a high-dimensional data structure. The seven feature components calculated in the above steps—grayscale jump rate, cumulative sum of grayscale differences, local variance of grayscale values, number of continuity breakpoints, spatial distribution density, gradient direction consistency of gradient regions, and grayscale fluctuation amplitude—are aligned and normalized to generate a seven-dimensional surface defect feature vector. This vector is not a simple list of data, but rather provides a comprehensive mathematical description of the surface quality of paper-plastic products from multiple perspectives, including edge sharpness, geometric shape, ambiguity, spatial distribution, texture order, and surface roughness. In the subsequent quality assessment process, this feature vector will serve as input data, supporting threshold judgments on the number of breakpoints and evaluations of the pixel width of the fuzzy transition zone, providing rich and accurate information for the final quality decision.

[0046] Step S120: When the first peak-to-valley spacing of the weight fluctuation curve is less than the preset time interval threshold and the number of times the weight increase / decrease direction switches exceeds the preset number of frequent changes, perform time-domain filtering on the continuous weight data sequence to obtain the corrected weight nonlinear drift trajectory parameters.

[0047] This step aims to eliminate noise interference and restore the true weight change trend when high-frequency oscillations and unstable fluctuations are detected in the weight data. The system performs feature analysis on the weight fluctuation curve, calculating the first peak-to-valley distance and the number of times the weight increase / decrease direction switches. The first peak-to-valley distance refers to the time span between adjacent maximum and minimum points in the waveform, reflecting the weight oscillation period; the number of times the weight increase / decrease direction switches statistically represents the frequency of weight data transitioning between upward and downward trends, characterizing data instability.

[0048] The preset time interval threshold is used to define the shortest allowable oscillation period during normal weighing. It can be obtained by extracting the maximum allowable mechanical oscillation frequency based on the inherent response frequency of the weighing sensor and the highest interference frequency of the mechanical transmission of the production line, and multiplying the reciprocal of the maximum mechanical oscillation frequency by a safety factor, thereby accurately defining the shortest oscillation period during normal weighing. The preset number of frequent changes is used to define the upper limit of the allowable trend transformation per unit time. During the system initialization phase, a continuous weight data sequence of multiple batches of qualified paper and plastic products within the static weighing stability window is collected, the number of times the weight increase or decrease direction is switched for each batch is counted, and the mean and standard deviation of the number of switching are calculated. The result of adding three times the standard deviation (i.e., the 3σ principle) to the mean is set as the preset number of frequent changes to cover 99.7% of the normal trend transformation range.

[0049] When the first peak-to-valley spacing is less than a preset time interval threshold and the number of switching operations exceeds a preset number of frequent changes, it indicates that the current weight data is significantly affected by high-frequency noise interference or abnormal mechanical vibration. Under this condition, the system performs time-domain filtering on the continuous weight data sequence, such as using Kalman filtering or Finite Impulse Response (FIR) low-pass filtering algorithms, to filter out high-frequency oscillation components and retain low-frequency components that reflect the trend of material density changes, thereby obtaining the corrected weight nonlinear drift trajectory parameters. This parameter describes the slow, monotonous, or long-period nonlinear change trajectory of the weight value over time after noise interference is removed, accurately reflecting the uniformity of the material filling inside the paper-plastic product.

[0050] In one alternative implementation, such as Figure 4 As shown, this application also provides a flowchart of a method for performing time-domain filtering on a continuous weight data sequence to obtain corrected weight nonlinear drift trajectory parameters. This method specifically includes: Step S410: Obtain the load response delay time of the sensor, and use the load response delay time as a compensation parameter to introduce the state transition matrix of the Kalman filter algorithm. Perform time-domain filtering on the weight oscillation data within the continuous sampling period to obtain the weight data after compensation for the delay.

[0051] In practical applications, the Kalman filter algorithm is suitable for use in noisy and uncertain environments, achieving optimal estimation of dynamic data by predicting and updating the system state. The state transition matrix in this algorithm describes the evolution of the system state from the previous moment to the current moment, and is typically established based on the system's physical model or kinematic equations. To avoid phase lag and estimation bias caused by sensor load response delay, the load response delay time is introduced as a compensation parameter into the state transition matrix. Specifically, a time step offset term representing the delay can be added to the corresponding position in the state transition matrix, or the parameters of the state transition matrix can be adjusted to adapt to the delayed system response characteristics, thereby enabling the Kalman filter algorithm's prediction model to accurately reflect the sensor's true dynamic behavior under load changes.

[0052] Furthermore, the load response delay time of the sensor can be read from the sensor specifications or obtained through online calibration experiments. This delay time characterizes the time difference between the application of the weight of the paper / plastic product to the sensor and the output of a stable reading. This load response delay time is converted into the corresponding number of sampling periods and used as a correction factor to construct a new state transition matrix. Using this corrected state transition matrix, combined with the state estimate from the previous moment, a priori prediction of the weight state at the current moment is made. The actual weight oscillation data within the current sampling period is acquired, and the measurement residual, i.e., the difference between the actual observation and the predicted value, is calculated. Then, the Kalman gain is calculated, which determines the degree of confidence in the predicted and observed values ​​when updating the state estimate. Based on the Kalman gain and the measurement residual, the prior prediction is corrected to obtain the posterior state estimate, thus completing the time-domain filtering of the weight oscillation data within continuous sampling periods and outputting smooth, delayed-compensated weight data that eliminates the effects of delay.

[0053] Step S420: Extract the static weighing stability window, use a low-pass filter to filter the data outside the stability window, remove mechanical vibration noise interference outside the window duration, and obtain the denoised weight data.

[0054] In practical applications, the static weighing stability window refers to the period during which the paper / plastic product remains relatively stationary on the weighing platform and the sensor readings fluctuate minimally. Data within this period primarily reflects the product's true weight, while data outside the window typically includes high-frequency interference caused by the product loading / unloading process and robotic arm movements. A low-pass filter, as a signal processing tool, allows signals with frequencies below the cutoff frequency to pass through while attenuating signals with frequencies above the cutoff frequency, and is commonly used to filter out high-frequency noise. Filtering data outside the stability window using a low-pass filter aims to preserve the low-frequency trends reflecting the true weight changes while suppressing high-frequency jitter caused by mechanical vibrations.

[0055] The start time and duration of the static weighing stabilization window are set according to the production line cycle time and sensor response characteristics. For example, the window opens after a preset delay after the workpiece is placed in place and continues until the data acquisition ends. The preset delay can be determined based on the stabilization time measured in the weighing sensor step response experiment. For example, a weight equivalent to the weight of the paper-plastic product is applied to the sensor, and the stabilization time required from the application of the load to the sensor output value entering the ±0.5% error band is recorded. This stabilization time is used as the preset delay.

[0056] For weight data outside the aforementioned window period, a low-pass filter with a finite impulse response (IIR) or infinite impulse response (IIR) structure is used for processing. The cutoff frequency of the low-pass filter can be set, for example, to a value determined based on the fundamental frequency of mechanical vibration, such as between 5Hz and 10Hz, to ensure effective filtering of mechanical vibration noise interference introduced by conveyor belt vibration, motor operation, etc. The weight data outside the window is input into the low-pass filter, and the signal is smoothed through convolution or recursive operations to attenuate the amplitude of high-frequency components. After processing, the data within the window retains its original high-precision characteristics, while the noise components of the data outside the window are significantly suppressed. This removes mechanical vibration noise interference outside the window duration from the overall data sequence, laying the foundation for subsequent extraction of a clean drift trajectory.

[0057] Step S430: Perform polynomial fitting on the weight data after Kalman filtering and low-pass filtering to obtain the corrected weight nonlinear drift trajectory parameters.

[0058] After performing Kalman filtering and low-pass filtering on the weight data as described above, polynomial fitting is required. Polynomial fitting is a mathematical analysis method that approximates or fits the distribution trend of observed data points by constructing a polynomial function, often used to extract nonlinear variation patterns in the data. The corrected nonlinear drift trajectory parameters of the weight can refer to the characteristic quantities that can quantitatively describe the nonlinear drift change of weight over time after the above time-domain filtering and low-pass denoising processing, such as the coefficients of the polynomial.

[0059] Specifically, a polynomial fitting is performed on the purified weight data sequence after Kalman filtering and low-pass filtering. First, the order of the fitting polynomial is determined; for example, a quadratic or cubic polynomial can be chosen based on the complexity of the data distribution to balance fitting accuracy and computational complexity. Then, numerical optimization algorithms such as the least squares method are used to solve for the polynomial coefficients, minimizing the sum of squared residuals between the fitted curve and the filtered weight data sequence. This fitted curve characterizes the long-term trend of the weight data after noise removal and delay compensation. Finally, the calculated polynomial coefficients (such as quadratic coefficients, linear coefficients, and a constant term) are output as the corrected weight nonlinear drift trajectory parameters. These parameters accurately reflect the nonlinear drift characteristics of the sensor within the static weighing stability window, eliminating the masking effect of random noise and system delay, thus providing an accurate and reliable weight modal data foundation for subsequent peak-valley spacing analysis and cross-correlation analysis.

[0060] This application's embodiments effectively solve the estimation lag problem caused by the sensor's dynamic characteristics by introducing the sensor's load response delay time as a compensation parameter into the Kalman filtering algorithm; by truncating the static weighing stability window and applying a low-pass filter to eliminate mechanical vibration noise interference, the data environment is purified; finally, high-precision nonlinear drift trajectory parameters are extracted through polynomial fitting. This series of processing steps collectively ensures that the obtained nonlinear drift trajectory parameters accurately and objectively reflect the weight changes of paper-plastic products, improving the reliability and accuracy of weight dimension data in the quality assessment system.

[0061] Further, in step S130: when the number of edge contour break points in the surface defect feature vector is greater than the preset break threshold and the pixel width of the blurred transition zone exceeds the preset blurred width upper limit, spatial domain filtering is performed on the surface defect feature vector to obtain the corrected gray value distribution local variance intensity parameter.

[0062] This step is used to enhance the feature signal and remove background texture interference when surface defect features exhibit uncertainties such as blurred boundaries and texture breaks, thereby accurately quantifying the severity of the defect. The system analyzes the number of edge contour break points and the pixel width of the blurred transition zone in the surface defect feature vector. The number of edge contour break points refers to the total number of pixels in the binarized image after edge detection where the defect edge line, which should be continuous, is interrupted, reflecting the degree of damage to the integrity of the defect structure; the pixel width of the blurred transition zone refers to the width of the pixel area where the gray values ​​gradually transition between the defect area and the normal background area, characterizing the clarity of the defect boundary.

[0063] The preset fracture threshold is mainly used to distinguish between normal texture edges and real structural fractures. It is determined as follows: during the system initialization phase, multiple grayscale images of defect-free paper-plastic products are collected. The number of edge connectivity path fractures caused by background texture noise in each image is counted using the same method. The upper quantile (e.g., 99th percentile) of the number of fracture points is calculated as the preset fracture threshold to filter out false fracture points caused by environmental noise.

[0064] The preset blur width upper limit is used to define the acceptable degree of edge blur. It is determined by obtaining the point spread function (PSF) parameter of the optical imaging system, calculating the system's limit blur pixel span under preset lighting conditions, and multiplying the limit blur pixel span by a safety factor (such as 1.5) as the preset blur width upper limit, thereby accurately defining the boundary between normal optical diffuse reflection and abnormal defect blur.

[0065] When the number of edge contour breakpoints exceeds a preset break threshold, and the pixel width of the blurred transition band exceeds a preset blur width limit, it indicates that the current surface defect features are affected by uneven illumination, dust interference, or material surface roughness, resulting in significant blurring and breakage. Under these conditions, the system performs spatial domain filtering on the surface defect feature vector. For example, it performs Gaussian smoothing on the grayscale image to suppress high-frequency noise, or uses morphological closing operations to connect the broken edge contours. After processing, the system calculates the local variance intensity parameter of the corrected grayscale value distribution. This parameter quantifies the dispersion and local fluctuation intensity of grayscale values ​​in a specific region of the filtered image, effectively distinguishing between real material surface defects and background texture noise, providing reliable feature indicators for subsequent correlation analysis.

[0066] In one alternative implementation, such as Figure 5 As shown, this application also provides a flowchart of a method for obtaining the corrected local variance intensity parameter of grayscale value distribution by spatial domain filtering of surface defect feature vectors. This method specifically includes: Step S510: Determine the target defect area.

[0067] In this embodiment, if the number of contour breakpoints exceeds a preset break threshold and the pixel width of the blurred transition band exceeds a preset blurred width upper limit, that is, when both conditions are met simultaneously, it indicates that the edge not only has many physical breaks but also exhibits significant blurred transition characteristics visually. Therefore, it is determined that the current detection area contains a target defect region requiring special processing. The process of determining the target defect region includes marking connected regions that meet the above conditions in the original grayscale image and generating corresponding binary masks or coordinate sets. Through this dual-condition judgment, defect regions with complex deformation and blurred feature coupling can be accurately screened, avoiding unnecessary over-processing of clear and complete edges, thereby improving the targeting of subsequent filtering operations.

[0068] Step S520: Based on the target defect area, extract the consistency between the gray-level jump slope of the defect edge and the gradient direction of the gradient area, and perform weighted fusion of the gray-level jump slope and the gradient direction consistency to obtain weighted calculation features.

[0069] Furthermore, within the defined target defect area, the consistency between the gray-level transition slope of the defect edge and the gradient direction of the gradient region is calculated. The gray-level transition slope refers to the rate at which the gray-level value changes with distance along the normal direction of the defect edge, reflecting the sharpness of the edge; a larger value indicates higher contrast on both sides of the edge. Gradient direction consistency refers to the dispersion of the gradient directions of each pixel within the blurred transition zone, reflecting the regularity of edge changes; a value closer to 1 indicates more consistent gradient directions. The extraction process includes extracting a gray-level profile along the normal direction of the edge points, calculating the maximum value of the first derivative of the profile curve as the gray-level transition slope, and statistically analyzing the gradient directions of all pixels within the transition zone, calculating the peak percentage of the direction histogram as the gradient direction consistency.

[0070] Subsequently, the two features with different physical dimensions are mapped to a unified evaluation space for weighted fusion, and the formula can be expressed as: (1) in, For weighted calculation of features, The slope of the grayscale transition. To ensure gradient direction consistency, These are the preset weighting coefficients.

[0071] In practical applications, preset weighting coefficients The determination method specifically includes: obtaining a historical sample dataset, and using principal component analysis (PCA) to calculate the variance contribution rate of the consistency between the gray-level jump slope and the gradient direction to the defect classification result; and normalizing and setting weight coefficients based on the variance contribution rate of the gray-level jump slope. The value of is chosen so that the weighted calculation features can adaptively represent the depth and geometric information of the edges.

[0072] Step S530: Calculate the edge detection responsivity based on the weighted calculation features, and extract the spatial distribution density corresponding to the edge detection responsivity.

[0073] Using the weighted features obtained in the previous step as the core input, the edge detection responsivity of each pixel or local neighborhood is calculated. Edge detection responsivity is a quantitative indicator characterizing the probability or intensity of the presence of edges in a local region of an image. It is typically calculated further based on weighted features and is used to measure the confidence level of the region as a defect edge. The calculation process includes mapping the weighted features to a response function, or directly using them as response intensity values. Subsequently, the distribution of these response intensities within the target defect region is statistically analyzed, and the spatial distribution density is extracted.

[0074] Spatial distribution density refers to the proportion of pixels with edge detection responsivity exceeding a specific threshold within a unit area or a specific window, reflecting the density of defect distribution. The extraction process involves dividing the target defect region into a grid, calculating the average response intensity within each grid, and then statistically analyzing the proportion of high-response grids. By calculating the edge detection responsivity and its spatial distribution density, the region with the highest concentration of defect energy can be located, providing a spatial basis for the adaptive adjustment of subsequent filtering parameters and ensuring that the filtering effect is mainly concentrated in key defect regions.

[0075] Step S540: Based on the spatial distribution density, perform Gaussian filtering on the cumulative sum of gray-level differences between adjacent pixels to obtain the filtering result.

[0076] The cumulative sum of gray-level differences between adjacent pixels is a statistic describing the total amount of local detail changes in an image. The core of this step lies in constructing an adaptive filtering strategy, which uses the previously extracted spatial distribution density to dynamically control the smoothing intensity of the Gaussian filter. Specifically, the system uses the spatial distribution density as the mapping basis for the standard deviation parameter (or filter kernel size parameter) of the Gaussian filter.

[0077] When the spatial distribution density is high, it indicates that the pixel distribution at the defect edge is compact and the structural features are obvious. The system automatically maps a smaller standard deviation parameter, making the weight of the Gaussian filter more concentrated on the central pixel, thereby smoothing noise while preserving the sharp details of the defect to the greatest extent. Conversely, when the spatial distribution density is low, it indicates that the defect distribution is more discrete or the background texture interference is heavier. The system automatically maps a larger standard deviation parameter, making the weight distribution of the filter more gradual, thereby enhancing the suppression and smoothing effect on background noise.

[0078] Through the above mechanism, the Gaussian filter uses a spatial weight matrix generated by a two-dimensional Gaussian function to perform a local weighted average operation on the cumulative sum of gray-level differences between adjacent pixels. This strategy of dynamically adjusting the filtering parameters based on spatial distribution density effectively overcomes the shortcomings of traditional fixed-parameter filtering, which easily leads to blurred edges or insufficient smoothness. While accurately removing high-frequency noise, it completely preserves the effective structural information of the defects, thus obtaining high-quality filtering results.

[0079] Step S550: Remove background texture interference from the filtering results and calculate the local variance intensity parameter of the corrected grayscale value distribution.

[0080] The filtering results are analyzed to identify and separate background texture interference items. Background texture interference items can refer to grayscale variations caused by non-defect factors such as normal machining textures on the paper-plastic product surface and the inherent graininess of the material. The removal process includes using frequency domain analysis to transform the filtering results to the frequency domain and filter out specific frequency components representing background texture; or using morphological operations to erode fine texture details using structuring elements. After removing interference items, the corrected local variance intensity parameter of the grayscale value distribution is calculated based on the processed clean data. This parameter can be a statistical measure that reflects only the severity of grayscale fluctuations in the actual defect area after removing background interference; it is the core parameter for cross-correlation analysis in the above embodiments. The calculation process includes setting a sliding window within the target defect area, calculating the variance of pixel grayscale values ​​within the window, and using the mean or maximum variance of all windows as the final parameter. Through this series of operations, the masking effect of background texture on feature extraction is eliminated, ensuring that the final obtained local variance intensity parameter of the grayscale value distribution can truly and accurately reflect the fuzzy characteristics and severity of surface defects in paper-plastic products.

[0081] Step S140: Perform cross-correlation analysis between the second peak-valley spacing in the weight nonlinear drift trajectory parameters and the pixel width range in the grayscale value distribution local variance intensity parameters.

[0082] This step establishes a spatiotemporal mapping relationship between weight modal features and visual modal features, and identifies whether there is an inherent physical coupling between the two through cross-correlation analysis. The system extracts the second peak-valley spacing from the corrected weight nonlinear drift trajectory parameters, i.e., the time span between the main peaks and troughs, reflecting the macroscopic cycle of material settling or pressure release during the forming process. Its numerical change characterizes the frequency and amplitude evolution trend of weight oscillations—a shorter peak-valley spacing means increased weight oscillations and a shorter cycle, which can serve as a quantitative indicator of weight stability. Simultaneously, the pixel width range is extracted from the local variance intensity parameter of the grayscale value distribution, characterizing the spatial distribution breadth of surface defect areas or the geometric size of the main defect band; its numerical expansion indicates an increased defect influence range.

[0083] Cross-modal correlation analysis calculates the temporal correlation between the second peak-valley spacing and the pixel width range to determine whether they follow a common evolutionary pattern. The second peak-valley spacing serves as a temporal reference, providing a periodic scale for weight fluctuations, while the pixel width range serves as a spatial reference, providing a scale for defect expansion. The correlation analysis is used to detect whether there is a synchronicity between the periodic changes in weight distribution and the spatial expansion of surface defects. For example, the system can calculate the Pearson correlation coefficient or cross-correlation function between two parameter sequences to detect whether there is a synchronicity between the periodic changes in weight distribution and the spatial expansion of surface defects, i.e., whether there is a physical causal relationship between internal material density fluctuations and surface structural anomalies. When both exhibit a synchronous change pattern, it indicates that weight oscillations and defect expansion share a common physical driving source (such as molding pressure fluctuations), and their coupling constitutes a typical characteristic of quality degradation in paper-plastic products. This cross-modal correlation analysis can effectively eliminate single-modal anomalies caused by random disturbances, improving the robustness of quality assessment.

[0084] Specifically, such as Figure 6 The diagram shown is a flowchart of a method for cross-correlation analysis between the second peak-valley spacing in the weight nonlinear drift trajectory parameters and the pixel width range in the local variance intensity parameters of the grayscale value distribution, according to an embodiment of this application. The method includes: Step S610: Obtain the changing trend of the second peak-valley spacing and the changing trend of the pixel width range within the continuous sampling period.

[0085] Within a continuous sampling period, the second peak-valley spacing sequence in the corrected weight nonlinear drift trajectory parameters and the pixel width range sequence in the corrected grayscale value distribution local variance intensity parameters are monitored and recorded in real time. The second peak-valley spacing characterizes the oscillation density of the weight data in the time domain; a smaller value indicates more frequent weight fluctuations per unit time. The pixel width range characterizes the spatial coverage of the blurred transition band in the visual image; a larger value indicates more severe edge blurring of surface defects.

[0086] Specifically, the sliding window regression algorithm or the first-order difference algorithm can be used to calculate the slope change rate of the two sequences in the continuous sampling period respectively; when the slope change rate of the second peak-valley spacing sequence is negative, it is determined that it shows a shortening trend; when the slope change rate of the pixel width range sequence is positive, it is determined that it shows an expanding trend.

[0087] Step S620: If the second peak-valley spacing shows a shortening trend and the pixel width range shows an expanding trend, then calculate the linear correlation coefficient between the changing trend of the second peak-valley spacing and the changing trend of the pixel width range.

[0088] When the aforementioned steps simultaneously detect a shortening trend in the second peak-to-valley spacing and an expanding trend in the pixel width range, it indicates that the weight stability of the product is decreasing and the surface blurriness is increasing, suggesting a potential negative correlation between the two in their respective directions of change. To quantify the statistical strength of this cross-modal correlation, it is necessary to further calculate the linear correlation coefficient between these two trends.

[0089] Specifically, the evolution sequence of the second peak-valley spacing extracted in the previous steps is aligned with the evolution sequence of the pixel width range. The Pearson correlation coefficient algorithm is used to calculate the ratio of the product of their covariance and their respective standard deviations, yielding a linear correlation coefficient. Since the second peak-valley spacing shows a shortening trend while the pixel width range shows an expanding trend, this linear correlation coefficient is numerically negative. The system directly extracts the absolute value of this linear correlation coefficient as a correlation metric parameter characterizing the intensity of synchronous changes across modal data. This processing transforms isolated weight and visual data into a continuous, dynamic correlation vector, effectively filtering out the interference of random errors from a single sensor on the analysis results and providing a reliable statistical basis for subsequent calculation of the coupling anomaly index.

[0090] Step S630: If the linear correlation coefficient is negative and its absolute value is greater than the preset correlation threshold, it is determined that the shortening trend of the second peak-valley spacing and the expansion trend of the pixel width range show a synchronous change pattern.

[0091] If the linear correlation coefficient calculated based on the aforementioned steps is negative, it indicates that the change in the second peak-valley spacing is negatively correlated with the change in the pixel width range. That is, when the second peak-valley spacing gradually shortens, the pixel width range tends to gradually expand, which is consistent with the inherent mechanism of physical weight oscillation intensification accompanied by surface texture blurring.

[0092] The absolute value of the linear correlation coefficient is further compared with a preset correlation threshold. The preset correlation threshold is determined as follows: During system initialization, historical data from multiple batches under normal steady-state production conditions are collected. The linear correlation coefficient between the second peak-to-valley distance and the pixel width range is calculated for each batch, and the mean and standard deviation of these correlation coefficients are calculated. The mean minus three times the standard deviation (i.e., based on the 3σ principle) is set as the preset correlation threshold. This threshold is used to statistically define whether there is a significant abnormal linkage between weight oscillation and surface ambiguity, thereby eliminating spurious correlations caused by normal production fluctuations.

[0093] If the linear correlation coefficient is negative and its absolute value is greater than the preset correlation threshold, it indicates that this negative correlation has significant statistical meaning and is not an accidental phenomenon caused by random fluctuations. Under this condition, it is determined that the shortening trend of the second peak-valley spacing and the expanding trend of the pixel width range show a synchronous change pattern, that is, it is confirmed that the weight features and visual features have undergone strong coupling and synergistic deterioration in the current time period, and the calculation instruction of the coupling anomaly index is triggered, thereby providing a strict mathematical basis for subsequent steps.

[0094] Step S150: If the shortening trend of the second peak-valley spacing and the expansion trend of the pixel width range show a synchronous change pattern, then extract the ratio of the number of switching times of weight increase / decrease direction to the number of edge contour break points, and combine it with the product of the instantaneous weight deviation amplitude and the gray-scale jump slope of the defect edge to calculate the coupling anomaly index.

[0095] This step, upon confirming the synchronous evolution pattern of weight and visual features, constructs a quantitative coupling anomaly index to comprehensively assess the degree of quality deviation of the paper-plastic product. If cross-correlation analysis shows that the second peak-to-valley spacing is shortening (weight oscillation cycle shortens and fluctuations intensify), while the pixel width range is expanding (the area affected by surface defects becomes larger), and both exhibit a synchronous change pattern, then the paper-plastic product is determined to be in a critical state of coupling between uneven molding pressure and abnormal fiber distribution.

[0096] In this state, the system extracts the number of times the weight increase / decrease direction switches and the number of edge contour break points, and calculates the ratio between the two. This ratio represents the weight oscillation frequency under unit structural breakage. By using the ratio form, the influence of dimensions caused by differences in data sampling rate or image resolution is eliminated, thereby accurately representing the weight instability corresponding to the unit visual degree of breakage.

[0097] Simultaneously, by combining the instantaneous weight deviation amplitude (the absolute value of the difference between the current weight value and the standard weight value) and the gray-scale jump slope at the defect edge (the rate of change of the gray-scale value at the defect edge with spatial position), the product of the two is calculated. This product characterizes the superimposed effect of the visual salience of the defect and its contribution to the overall weight. This processing method ensures that a single weight deviation or a single visual blur will not cause a significant increase in this product term, thus effectively capturing the intrinsic relationship between uneven molding pressure and abnormal material density distribution.

[0098] Finally, the system multiplies the above ratios and products to obtain the coupling anomaly index. This dimensionless parameter comprehensively reflects the degree of abnormal coupling in both weight distribution and surface morphology of the paper-plastic product. Through the nonlinear combination of multidimensional features, the coupling anomaly index can sensitively capture the fuzzy state near the boundary between acceptable and unacceptable, transforming dual uncertainties into a measurable numerical index. This index is a dimensionless comprehensive evaluation index used to quantify the coupling strength between "molding pressure fluctuations" and "material microstructure defects." Its value directly reflects the tightness of fiber interweaving and the stability of macroscopic geometric dimensions within the paper-plastic product. A higher index means that the product has been subjected to the dual effects of unstable pressure impact (leading to weight fluctuations) and uneven fiber distribution (leading to abnormal visual texture) during the molding process.

[0099] The theoretical range of the coupling anomaly index: After normalization, the theoretical range of this index is 0 to 100. The corresponding intervals are as follows: 0-30: corresponds to "steady-state molding". At this point, the weight peak-to-valley spacing is stable, and the visual edges are clear, indicating good mold closure and uniform slurry adsorption. 30-60: corresponds to "metastable molding". At this point, slight fluctuations occur in the weight or visual single modality, indicating possible slight insufficient vacuum adsorption or localized fluctuations in slurry concentration. 60-100: corresponds to "unstable molding". At this point, severe weight fluctuations and obvious visual defects indicate possible blockages, wear in the molding mold, or malfunctions in the mechanical transmission, requiring immediate shutdown and repair.

[0100] Step S160: Based on the coupling anomaly index, find the corresponding quality deviation value in the piecewise linear mapping table, compare the quality deviation value with the pass / fail judgment limit, and output the product quality judgment result.

[0101] This step converts the coupling anomaly index into an intuitive numerical value representing the degree of quality deviation, and outputs the final judgment result through boundary comparison. A piecewise linear mapping table predefines the correspondence between the coupling anomaly index and the numerical value of the quality deviation, dividing the range of the coupling anomaly index into three intervals: a safe zone, a vague zone, and a defective zone, using different linear mapping slopes in different intervals. Particularly in the vague zone (critical interval), the piecewise linear mapping table uses a mapping slope greater than that in the safe and defective zones to amplify subtle differences in the coupling anomaly index within this region, improving the ability to distinguish critical states.

[0102] A mechanism employing a larger mapping slope for the fuzzy region aims to address the issue of weak critical state differentiation. During the configuration of the piecewise linear mapping table, the slope value of the linear mapping function for the fuzzy region is set to be greater than the slope values ​​for the safe and defective regions. For example, the safe region can use a linear function... ,in For example, it can be set to 0.5, indicating that minor anomalies have a small impact on overall quality, and the system adopts a tolerant strategy; the defect area can be represented by a linear function. ,in For example, it can be set to 1.5 to directly map severe defects. The intercept of the linear mapping function for the defect area is used to adjust the overall offset of the quality deviation benchmark value within that interval. Its value is determined based on the connection relationship between the starting point of the defect area (i.e., when the coupling anomaly index is 60) and the ending point of the fuzzy area; for example, it can be set to -15. Within the fuzzy area, a linear function is used. and set (e.g., 2.0) greater than and , This represents the intercept of the linear mapping function in the fuzzy region, used to adjust the overall offset of the quality deviation benchmark value within this region. Its value is determined based on the connection between the mapping value at the start of the fuzzy region (i.e., when the coupling anomaly index is 30) and the end of the safe region; for example, it can be set to -45. When the actually calculated coupling anomaly index falls into this fuzzy region (e.g., a value of 45), due to the slope... A larger slope means that small changes in the input value (e.g., from 45 to 46) will be mapped to significant changes in the output quality deviation (e.g., from 80 to 82). This differentiated slope configuration amplifies subtle differences in the coupling anomaly index within the ambiguity region, improving the system's sensitivity in the critical range and preventing the missed detection of defective products on the verge of passing. This significantly improves the resolution of quality assessment while maintaining overall robustness.

[0103] The system retrieves the corresponding quality deviation value from a table based on the calculated coupling anomaly index. This continuously changing value precisely characterizes the degree to which the current product deviates from the standard quality. Subsequently, this value is compared with a preset acceptance threshold (a threshold determined based on product quality standards and customer tolerance). If the quality deviation value is greater than the acceptance threshold, the material filling uniformity of the paper-plastic product is deemed unqualified; if it is less than or equal to the acceptance threshold, it is deemed qualified. Finally, the system outputs the product quality judgment result and stores the result, along with the corresponding coupling anomaly index, in the quality traceability database, providing data support for production process optimization and quality backtracking.

[0104] Furthermore, such as Figure 7 As shown, this application also provides a specific implementation process for constructing a piecewise linear mapping table. This method aims to establish a piecewise linear mapping relationship that accurately reflects the quality degradation law of paper and plastic products through a data-driven automated calibration process, thereby replacing the traditional method that relies on manual experience and improving the generalization ability and deployment efficiency of the quality assessment system. The construction method includes the following steps: Step S710: Obtain multiple sets of paper-plastic product samples under different combinations of molding pressure and slurry concentration, and score the quality of each set of samples to obtain the true value of the degree of quality deviation.

[0105] Constructing a high-quality training dataset covering all production conditions of paper-plastic products is fundamental to implementing this method. Molding pressure and slurry concentration are key process parameters that determine the final molding quality and physical properties of the products. Different combinations of these two parameters directly lead to differences in the internal fiber interweaving structure of the material and fluctuations in the macroscopic geometric dimensions of the products. To ensure that the subsequently constructed mapping model has broad adaptability and robustness, orthogonal experimental design or full factorial experimental design is adopted to conduct combined trial production at different levels of molding pressure (e.g., low, medium, and high pressure levels) and slurry concentration (e.g., thin, medium, and thick concentration levels), systematically collecting multiple sets of paper-plastic product samples.

[0106] For each collected sample, a refined quality score was performed by combining destructive physical testing and manual re-inspection. Structural integrity was assessed through visual inspection or penetrant staining, thickness uniformity was measured using a multi-point thickness gauge, and compressive strength was obtained through compression testing using a universal testing machine. This scoring process is based on a pre-defined quality standard system, quantifying multi-dimensional physical indicators such as structural integrity, surface defect severity, thickness uniformity, and compressive strength into specific numerical scores. These scores are then normalized to represent the true value of quality deviation within a pre-defined range. The normalization process uses a min-max normalization method, mapping the comprehensive score of all samples to a continuous range of 0 to 100. Higher values ​​indicate poorer quality (i.e., a greater deviation from the acceptable state), while lower values ​​indicate better quality. This true value serves as a supervisory signal for subsequent model training, accurately representing the actual quality state of the samples under the current combination of process parameters, providing a solid physical truth basis for the construction of the mapping table.

[0107] Step S720: Perform K-means cluster analysis on the coupling anomaly index calculated for each group of samples, and automatically divide the value range of the coupling anomaly index into a safe zone, an ambiguous zone, and a defective zone according to the cluster center.

[0108] After obtaining paired data containing the coupling anomaly index and the true values ​​of quality deviation, the K-means clustering algorithm is used to deeply analyze the data distribution characteristics of the coupling anomaly index. As an unsupervised machine learning algorithm, K-means clustering automatically divides an unlabeled dataset into K distinct clusters based on the inherent similarity of data samples in the feature space. In this embodiment, the number of clusters K is set to 3, corresponding to the three ideal quality state categories: safe zone, ambiguous zone, and defective zone. Through iterative calculation, the algorithm automatically finds three cluster centers that minimize the sum of squared distances from each sample point to its respective cluster center. This process abandons the practice of subjectively setting interval boundary thresholds based on human experience, determining the split points entirely based on the actual density and sparsity of the coupling anomaly index in the sample space. For example, the clustering results naturally classify samples with low and concentrated index values ​​as the safe zone, samples with medium and relatively dispersed index values ​​as the ambiguous zone, and samples with high and concentrated index values ​​as the defective zone. This data-driven adaptive partitioning method not only eliminates the subjective blindness of manually setting thresholds, but also effectively captures the dynamic changes in data distribution under different production batches and process conditions, ensuring the objectivity and scientific nature of interval partitioning.

[0109] Step S730: For the sample data in each interval, perform linear regression fitting with the coupling anomaly index as the independent variable and the true value of the quality deviation as the dependent variable to determine the linear mapping function corresponding to each interval, thereby generating a piecewise linear mapping table.

[0110] After dividing the data into intervals, for each interval (safe zone, ambiguous zone, defective zone), independent linear regression fitting is performed using sample data falling within that interval. The goal of linear regression fitting is to establish a mathematical model using the coupling anomaly index within the interval as the independent variable and the corresponding true value of the quality deviation as the dependent variable. The parameters (slope and intercept) of the best-fit line are solved by minimizing the sum of squared errors. Here, the independent variable is the input variable in the regression model, corresponding to the coupling anomaly index in the sample data in this step; this value is used as a known quantity in the model calculation. The dependent variable is the output variable in the regression model, corresponding to the true value of the quality deviation in the sample data in this step; this value is the target value that the model attempts to fit. The purpose of linear regression is to establish a mapping relationship between the independent and dependent variables, so that given any coupling anomaly index, the corresponding quality deviation can be predicted through this mapping relationship.

[0111] Given that the impact mechanism of changes in the coupling anomaly index on the degree of quality deviation differs significantly across different quality intervals—for example, small fluctuations in the index may have a minor impact on overall quality within the safe zone, while the same fluctuations may indicate a sharp deterioration in quality within the ambiguous zone—a piecewise fitting approach can accurately capture specific nonlinear variation characteristics within each interval. After fitting, each interval is assigned a unique linear mapping function, for example… ,in The coupling anomaly index (independent variable) is used. The degree of quality deviation (dependent variable). Parameter The slope of the fitted straight line reflects the sensitivity of the coupling anomaly index change to the degree of quality deviation. The larger the value, the greater the change in quality deviation caused by the change in the unit coupling anomaly index, and the higher the sensitivity of the judgment in this interval; parameter The intercept of the fitted line represents the baseline offset of the quality deviation when the coupling anomaly index is zero. In practical applications, it is mainly used to ensure the continuous transition of the mapping value at the boundary of adjacent intervals and to avoid abrupt changes in the quality judgment result due to the breakpoint jump.

[0112] Finally, the system integrates and stores the linear mapping functions of these three intervals and their corresponding exponential range boundaries, forming a complete piecewise linear mapping table. In practical applications, by simply inputting the coupling anomaly index of the product to be tested, the system can quickly locate its corresponding interval and call the appropriate mapping function to calculate the degree of quality deviation, thereby achieving efficient and accurate quality assessment.

[0113] This application's embodiments simultaneously analyze weight fluctuation characteristics and surface defect characteristics, utilizing temporal and spatial domain filtering to eliminate random interference and effectively extract key parameters reflecting the uniformity of material filling. Through cross-correlation analysis, the synchronous evolution pattern of weight oscillation cycles and defect spatial distribution is identified, constructing a coupled anomaly index that integrates temporal and spatial domain characteristics, achieving accurate quantification of quality status under dual uncertainties. The application of a piecewise linear mapping table solves the problem of insufficient differentiation ability in critical intervals by traditional hard threshold judgments. By designing differentiated slopes to amplify the differences in features within the fuzzy region, the accuracy and robustness of paper-plastic molding quality assessment are significantly improved, reducing the false judgment rate.

[0114] In addition, through comparative testing on a paper-plastic packaging production line (sample size N=10,000 pieces), the following quantitative results were obtained: Improved detection accuracy: Compared with pure visual inspection methods, the false detection rate of this application is 0.4%, which is significantly lower than the 3.5% of the pure visual approach. In particular, when detecting "hidden indentation" defects, the false detection rate of the pure visual approach is 12%, while the false detection rate of this application is reduced to 0.8% by using weight data to assist in verification.

[0115] Criticality judgment capability: In the judgment of products in the "critical range" with a weight deviation of ±1.5% and blurred visual features, the judgment accuracy of this application reaches 96.5%, which effectively solves the problem of "different grades for the same quality" caused by traditional binary judgment.

[0116] Real-time performance indicators: This application adopts piecewise linear mapping and fast search algorithm, and the comprehensive detection time of a single product (including weighing stabilization waiting and image processing) is controlled within 150 milliseconds, which fully meets the real-time detection cycle requirements of industrial production lines with a capacity of 200-300 pieces per minute.

[0117] It should be noted that, in this document, the terms "comprising," "including," and any other variations are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Specific examples have been used in this document to illustrate the principles and implementation methods of the technical solutions of this application. The above examples are only for the purpose of helping to understand the methods and core ideas of this application. The above descriptions are merely preferred embodiments of this application. It should be pointed out that, due to the limitations of written expression and the objective existence of infinite specific structures, those skilled in the art can make several improvements, modifications, or changes without departing from the principles of this application, and can also combine the above technical features in an appropriate manner; these improvements, modifications, changes, or combinations, or the direct application of the concept and technical solutions of this application to other situations without modification, should all be considered within the scope of protection of this application.

Claims

1. A method for evaluating the quality of paper-plastic molding that integrates visual features and weighing data, characterized in that, The method includes: Obtain the continuous weight data sequence of paper-plastic products within the static weighing stability window and extract the weight fluctuation curve; and obtain the grayscale image of the surface of the paper-plastic products and extract the surface defect feature vector. When the first peak-to-valley spacing of the weight fluctuation curve is less than a preset time interval threshold and the number of times the weight increase / decrease direction changes exceeds a preset number of frequent changes, the continuous weight data sequence is subjected to time-domain filtering to obtain the corrected weight nonlinear drift trajectory parameters. When the number of edge contour break points in the surface defect feature vector is greater than the preset break threshold and the pixel width of the blurred transition band exceeds the preset blurred width upper limit, the surface defect feature vector is subjected to spatial domain filtering to obtain the corrected gray value distribution local variance intensity parameter. The second peak-valley spacing in the weight nonlinear drift trajectory parameters is cross-correlation analyzed with the pixel width range in the gray value distribution local variance intensity parameters. If the shortening trend of the second peak-valley spacing and the expanding trend of the pixel width range show a synchronous change pattern, then extract the ratio of the number of switching times of the weight increase / decrease direction to the number of edge contour break points, and combine it with the product of the instantaneous weight deviation amplitude and the gray-scale jump slope of the defect edge to calculate the coupling anomaly index. The corresponding quality deviation value is found in the piecewise linear mapping table based on the coupling anomaly index. The quality deviation value is compared with the pass / fail threshold, and the product quality judgment result is output.

2. The method according to claim 1, characterized in that, The extracted weight fluctuation curve includes: Frequency domain analysis was performed on the continuous weight data sequence to obtain the oscillation frequency sequence and deviation amplitude sequence within the sampling period; Calculate the standard deviation of the oscillation frequency sequence under multiple weighings. If the standard deviation is greater than a preset standard deviation threshold, then perform smoothing filtering on the deviation amplitude sequence. Based on the smoothed and filtered deviation amplitude sequence, the nonlinear drift trajectory of the weight value over time is extracted. A polynomial regression algorithm is used to fit the nonlinear drift trajectory to generate a fluctuation curve that reflects the sensor load response delay, which is used as the weight fluctuation curve.

3. The method according to claim 1, characterized in that, The extraction of surface defect feature vectors includes: An edge response map is obtained by performing a convolution operation on the grayscale image using an edge detection operator. Based on the edge response map, the gray-level jump rate of the defect edge is extracted, the gray-level difference between adjacent pixels corresponding to the gray-level jump rate is calculated, and the cumulative sum of gray-level differences is obtained. By accumulating the grayscale differences, the pixel width of the blurred transition zone is identified, and the local variance of the grayscale values ​​within the pixel width is calculated. The number of continuity break points of the edge contour is counted. If the number of continuity break points is greater than a preset break threshold, the spatial distribution density value of the edge contour is extracted. Obtain the gradient direction consistency of the gradient region corresponding to the spatial distribution density value, and extract the grayscale fluctuation amplitude corresponding to the texture roughness. The grayscale jump rate, the cumulative sum of grayscale differences, the local variance of grayscale values, the number of continuity breakpoints, the spatial distribution density, the consistency of gradient direction in the gradient region, and the grayscale fluctuation amplitude are combined to form the surface defect feature vector.

4. The method according to claim 1, characterized in that, The step of performing time-domain filtering on the continuous weight data sequence to obtain the corrected weight nonlinear drift trajectory parameters includes: The load response delay time of the sensor is obtained and used as a compensation parameter in the state transition matrix of the Kalman filter algorithm. The weight oscillation data in the continuous sampling period is filtered in the time domain to obtain the weight data after compensation for the delay. A static weighing stability window is extracted, and a low-pass filter is used to filter the weight data outside the stability window to remove mechanical vibration noise interference outside the window duration, thereby obtaining the denoised weight data. Polynomial fitting is performed on the weight data after Kalman filtering and low-pass filtering to obtain the corrected weight nonlinear drift trajectory parameters.

5. The method according to claim 1, characterized in that, The step of performing spatial domain filtering on the surface defect feature vector to obtain the corrected local variance intensity parameter of the gray value distribution includes: Identify the target defect area; Based on the target defect region, the gray-level jump slope of the defect edge is extracted and the gradient direction of the gradient region is consistent. The gray-level jump slope and the gradient direction consistency are weighted and fused to obtain weighted calculation features. Based on the weighted calculation features, the edge detection responsivity is calculated, and the spatial distribution density corresponding to the edge detection responsivity is extracted. Based on the spatial distribution density, Gaussian filtering is applied to the cumulative sum of gray-level differences between adjacent pixels to obtain the filtering result; Remove background texture interference from the filtering results and calculate the local variance intensity parameter of the corrected grayscale value distribution.

6. The method according to claim 1, characterized in that, The step of performing a cross-correlation analysis between the second peak-valley spacing in the weight nonlinear drift trajectory parameters and the pixel width range in the grayscale value distribution local variance intensity parameters includes: Obtain the changing trend of the second peak-valley spacing and the changing trend of the pixel width range within a continuous sampling period; If the second peak-valley spacing shows a shortening trend and the pixel width range shows an expanding trend, then calculate the linear correlation coefficient between the changing trend of the second peak-valley spacing and the changing trend of the pixel width range; If the linear correlation coefficient is negative and its absolute value is greater than a preset correlation threshold, it is determined that the shortening trend of the second peak-valley spacing and the expanding trend of the pixel width range exhibit a synchronous change pattern.

7. The method according to claim 1, characterized in that, The ratio of the number of times the weight increase / decrease direction changes to the number of edge contour break points is extracted, and combined with the product of the instantaneous weight deviation amplitude and the gray-scale jump slope of the defect edge, the coupling anomaly index is calculated, including: The coupling anomaly index is obtained by multiplying the ratio by the product.

8. The method according to claim 1, characterized in that, When constructing the piecewise linear mapping table, the value of the coupling anomaly index is divided into three intervals: a safe zone, a fuzzy zone, and a defective zone. Within the fuzzy region, the piecewise linear mapping table employs a mapping slope greater than that of the safe region and the defective region to amplify subtle differences in the coupling anomaly index within the fuzzy region.

9. The method according to claim 8, characterized in that, The method for constructing the piecewise linear mapping table includes: Multiple sets of paper-plastic product samples under different combinations of molding pressure and slurry concentration were obtained, and the quality of each set of samples was scored to obtain the true value of the degree of quality deviation. K-means clustering analysis is performed on the coupling anomaly index calculated for each group of samples. Based on the cluster centers, the range of values ​​of the coupling anomaly index is automatically divided into the safe zone, the ambiguous zone, and the defective zone. For the sample data within each interval, a linear regression is performed using the coupling anomaly index as the independent variable and the true value of the quality deviation as the dependent variable to determine the linear mapping function corresponding to each interval, thereby generating the piecewise linear mapping table.

10. The method according to claim 1, characterized in that, The step of comparing the quality deviation value with the acceptance threshold and outputting the product quality judgment result includes: If the quality deviation value is greater than the acceptance threshold, the material filling uniformity of the paper-plastic product is marked as unqualified. If the quality deviation value is less than or equal to the acceptance threshold, the material filling uniformity of the paper-plastic product is marked as qualified. The product quality judgment result is associated with the corresponding coupling anomaly index record and stored in the quality traceability database.

Images