Deep learning-based impregnated paper production quality detection method and system

Abstract

The invention relates to an impregnated paper production quality detection method and system based on deep learning, and the method comprises the following steps: S1, obtaining an industrial camera image flow, sensor and process parameters, and event and work order information, and constructing a synchronous data packet; s2, preprocessing the synchronous data packet to obtain preprocessed image data, process data and an ROI mask; s3, obtaining a detection frame set and a segmentation mask set based on multi-model AI detection and segmentation according to the preprocessed image data, the process data and the ROI mask; s4, segmenting the mask set according to the detection frame set, and performing track clustering along the paper feeding direction to form a defect event flow; s5, on the basis of the defect event stream, performing online quality prediction and root cause analysis, and obtaining future interval quality prediction and key factor sorting; and S6, according to future interval quality prediction and key factor sorting, adaptive process optimization is carried out, and an optimal control instruction is obtained. According to the invention, the detection accuracy and fineness are effectively improved, and intelligent identification and cross-frame tracking of defects are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent manufacturing, and in particular to a deep learning-based impregnated paper production quality detection method and system. BACKGROUND

[0002] With the accelerated promotion of intelligent manufacturing and Industry 4.0, impregnated paper, as a key intermediate product of plywood, engineered wood, and home flooring, etc., has higher requirements for its quality stability and defect control. Traditional quality detection relies on manual visual inspection and fixed-point sampling methods, which have low defect recognition rate, large response delay, and subjective results, etc., and are difficult to meet the strict requirements of high-end applications for process continuity and high-quality rate. In addition, the impregnated paper production process involves coating, impregnation, drying, and paper running, etc. Process parameter fluctuations, equipment state changes, and environmental disturbances can easily cause multiple types of defects (such as creases, damage, speckles, uneven coating, etc.), and these defects often have strong randomness, variable morphology, and fuzzy boundaries, etc., which bring great challenges to real-time detection and root cause tracing. SUMMARY

[0003] To solve the above problems, the purpose of the present application is to provide a deep learning-based impregnated paper production quality detection method and system, which effectively improves the detection accuracy and precision, and realizes intelligent identification and cross-frame tracking of defects.

[0004] To achieve the above purpose, the present application adopts the following technical solutions:

[0005] A deep learning-based impregnated paper production quality detection method, comprising the following steps

[0006] S1: Obtain an industrial camera image stream, sensor and process parameters, event and work order information, and construct a synchronous data packet;

[0007] S2: Preprocess the synchronous data packet to obtain preprocessed image data, process data, and ROI mask;

[0008] S3: Based on the preprocessed image data, process data, and ROI mask, obtain a detection frame set and a segmentation mask set based on multi-model AI detection and segmentation;

[0009] S4: According to the detection frame set and the segmentation mask set, perform trajectory clustering along the paper running direction to form a defect event stream;

[0010] S5: Based on the defect event stream, perform online quality prediction and root cause analysis to obtain future interval quality prediction and key factor sorting;

[0011] S6: According to the future interval quality prediction and key factor sorting, perform adaptive process optimization to obtain optimal control instructions.

[0012] Further, the synchronization data packet is constructed, specifically as follows: first, aligning all source data based on a unified clock, correcting the camera timestamp by NTP, and using the encoder pulse as the main time base to map each frame of image to the paper length coordinate system; then, loading and applying the camera internal and external parameters for geometric correction and pixel ratio calibration to establish a linear or segmented affine mapping from pixels to physical coordinates, if it is a multi-camera splicing scene, performing feature matching and color gain balancing in the overlapping area, and outputting a consistent full-frame picture across cameras and a unified coordinate system synchronization data packet.

[0013] Further, the image data is preprocessed, specifically as follows:

[0014] Noise suppression and illumination field equalization are performed on the industrial camera image, bilatera denoising is used, for the reflection and highlight areas, polarization imaging is combined with reflection component separation to achieve shadow and highlight suppression; for color-related defects, first do white balance and gain calibration with a gray card, then do adaptive histogram equalization, and limit the upper limit of local contrast enhancement to prevent amplifying texture noise and paper texture artifacts;

[0015] The lens distortion is corrected based on the calibration matrix, and the geometric scale and direction of the image are unified, the block resampling and seam stitching correction are used for wide-width paper to remove the brightness jump of the splicing seam, for decorative paper with periodic patterns, template alignment is performed: phase registration is performed with a gold sample reference image to obtain a normalized layout;

[0016] A multi-scale pyramid is constructed, and Gabor filtering enhancement is applied to each layer to highlight linear and point-like disturbances; for wrinkles and bubbles, morphological top-hat / black-hat is used to enhance local protrusions / recess features; for color difference detection paths, after color constancy, output an image copy in the linear color space as a dedicated input view for the subsequent color difference regression branch;

[0017] Invalid area rejection and ROI mask construction, according to the mechanical structure and field of view, construct a static mask to shield non-product areas, and superimpose a dynamic mask to handle temporary occlusion or contamination; for the paper edge, use adaptive edge tracking to obtain the actual effective width, and generate a frame-by-frame effective area contour; finally, synthesize a binary ROI mask M roi , whose value is 1 indicating the valid paper surface pixels allowed to be detected, and 0 for the ignored area; and output the minimum circumscribed polygon matching it, which is convenient for subsequent sliding window reasoning and cropping;

[0018] Process data Missing value processing and outlier removal are performed, followed by resampling at a uniform sampling rate and alignment with image frame timestamps. Short-window statistical features and hysteresis terms are calculated for key process quantities to capture the hysteretic effects on surface defects. These features, along with the original quantities, constitute... Multi-component representation;

[0019] Finally, the preprocessed image data is output. Process data and ROI mask To adapt to subsequent multi-task networks, image data The image set contains three types of views: a high-resolution RGB view for object detection, an equalization intensity enhancement view for segmentation, and a Lab view for color difference estimation. The transformation chain and calibration parameters of the three are recorded through metadata to ensure that the coordinates are consistent across the views, and all outputs are encapsulated in a unified data structure.

[0020] Furthermore, the multi-model AI detection and segmentation model adopts a shared backbone and multi-branch architecture, which includes a detection branch, a segmentation branch, a color difference regression auxiliary head, and an unsupervised anomaly detection branch, as detailed below:

[0021] P and Q are extracted through a shared backbone and pyramid feature extraction network. The shared backbone network adopts CSP-Darknet and outputs multi-scale features {P3, P4, P5}. These features are then fused using FPN to construct a multi-scale feature set for use by the detection head. This facilitates multi-scale prediction and is represented as:

[0022] ;

[0023] Where Backbone represents the backbone network; I det For a preprocessed high-resolution RGB view; I seg Enhance the view to equalize intensity; For FPN networks, the term is used to perform top-down and horizontal fusion of multi-layer features of the shared backbone;

[0024] Constructing a high-resolution path Q in parallel for splitting branches:

[0025]

[0026] SegNeck represents the neck structure of the split branches;

[0027] The detection branch outputs candidate boxes and scores, and anchor boxes are used to output positions. With category probability :

[0028] ;

[0029] Among them, P l is the feature map from the l-th scale of the feature pyramid; k is the position index on the feature map at that scale. For bounding box regression heads, predict box parameters; As the classification header, output the logits for each category; θ det The trainable parameters for the detection branch; softmax is the activation function;

[0030] Training loss:

[0031] ;

[0032] Where, L CIoU For bounding box regression loss, o represents the prediction of the existence of the target; o represents the binary truth label for the existence of the target. For binary cross-entropy loss; For classification probability prediction;

[0033] y represents the one-hot true label of the category; For multi-class cross-entropy loss; λ box ,λ obj ,λ cls The weights of the three losses are hyperparameters used to balance the relative importance of bounding box regression, target existence, and classification losses, respectively.

[0034] The splitting branches use a soft mask output by U-Net++:

[0035] ;

[0036] in, Split the network for UNet++; θ seg These are the trainable parameters for the splitting branches; It is a function; For the predicted soft mask;

[0037] Training loss:

[0038] ;

[0039] Where M is the pixel-level real mask; It is an element-wise product; λ is the numerically stable term. bce λ represents the weights of the BCE segmentation loss. dice The weights for the Dice loss;

[0040] A color difference regression auxiliary head predicts pixel-level ΔE in Lab view. 2000Heat maps enhance the confidence and grading robustness of color difference defects:

[0041] ;

[0042] Among them, I lab Input the preprocessed Lab color space image; For color difference regression network; θ color This is the set of trainable parameters for the color difference regression head; For the predicted color difference heatmap;

[0043] The unsupervised anomaly detection branch uses PatchCore and outputs anomaly score graph A:

[0044] ;

[0045] in, This is an unsupervised anomaly detection model; θ ano Parameters for abnormal branches;

[0046] Overall training objectives:

[0047] ;

[0048] α, β, γ, δ are set through uncertainty weighting or grid search.

[0049] Furthermore, the multi-model AI detection and segmentation model applies a ROI mask before the pixel-level output of any branch to avoid false detections of non-product areas:

[0050] ;

[0051] in, For ROI binary mask; A roi The effective region results after applying the ROI;

[0052] For the detection frame, IoU filtering with ROI is applied:

[0053] ;

[0054] Where η∈[0,1] is the retention threshold; b i Let be the i-th candidate bounding box; ROI is the geometric region of ROI; The area function is 'keep'; 'keep' represents the filtering function.

[0055] Post-detection processing involves applying NMS to each class to attenuate the confidence level of overlapping targets.

[0056] ;

[0057] Among them, s i Let IoU(b) be the confidence score of the i-th candidate box; i ,b j ) represents the intersection-union ratio of the two boxes; τ represents the IoU suppression threshold of NMS;

[0058] Throttling and morphological cleansing of soft masks:

[0059] ;

[0060] in, This is an indicator function; the condition is 1 if true, and 0 otherwise; θ m The binarization threshold for segmenting the soft mask; For morphological closing operations; Opening operations for morphology;

[0061] Extract geometric properties from connected components to generate instance-level segmentation objects;

[0062] For low-confidence detection or segmentation candidates, if their region mean outlier... Exceeding the quantile threshold q 0.9 Increase its confidence level:

[0063] ;

[0064] Where s is the confidence level of the original candidate; s′ is the updated confidence level after fusing the anomaly scores; q 0.9 ,q 0.99 These are the 90th and 99th percentiles of anomaly heatmap A in the current frame;

[0065] The instance mask generated from the segmentation is used to generate the minimum bounding box b. mask Perform IoU matching with the set of detection boxes, where IoU > τ. m When merging into a single object, the category and confidence are weighted and combined:

[0066] ;

[0067] Among them, b det The detection branch provides the bounding box; bmask is the bounding box generated by the mask of the segmentation instance;

[0068] The geometric merging strategy for the two frames; s det To detect the confidence scores of branch candidates; s mask The confidence score for the split branch instance; α s The weights for confidence fusion; p det(c) represents the posterior probability of the detection branch for class c; p mask (c) represents the posterior probability of the splitting branch for class c; α c Weights for category fusion; The final category label after merging;

[0069] The final confidence score after fusion; The final bounding box after merging;

[0070] Finally, we obtain the set of detection boxes. and segmentation mask set .

[0071] Furthermore, based on the detection box set and the segmentation mask set, trajectory clustering is performed along the paper feed direction to form a defect event flow, as detailed below:

[0072] Based on the preset paper feed direction vector and pixel-physical mapping relationship, the centroid of each fusion candidate is projected onto the longitudinal arc length s - lateral offset r parameterized coordinate system, where s is accumulated along the paper movement direction and r is the lateral relative position. Combining timestamp and linear velocity information, a state vector is constructed for each candidate. A multi-target tracking association algorithm is used to maintain a Kalman filter for each existing trajectory to predict the state at the next moment, defining a gated Mahalanobis distance and a comprehensive cost function. The optimal match is solved using the Hungarian algorithm, and new observations are assigned to existing trajectories or initialized into new trajectories. Trajectories that are not matched for consecutive Kmiss frames enter a suspended state, allowing discontinuity caused by occlusion or missed detection, and the trajectory is terminated after timeout. The trajectory update uses an exponential moving average to fuse class probabilities and confidence levels, making the trajectory-level attributes more robust and able to reflect temporal consistency.

[0073] The stable trajectory that meets the minimum duration is output as a defect event stream, and each event contains complete lifecycle attributes.

[0074] Furthermore, based on the defect event stream, online quality prediction and root cause analysis are performed to obtain future interval quality predictions and key factor rankings, as detailed below:

[0075] Using the defect event stream output by S4 as the main input, and simultaneously integrating process parameter time windows and batch information, interactive features are constructed, including temperature and humidity coupling effect, line speed-tension coupling, and hysteretic correlation between process parameters and event occurrence rate, forming a high-dimensional time series feature vector;

[0076] A hierarchical prediction system for different time scales is established. Short-term prediction uses lightweight XGBoost deployed on edge devices to achieve low-latency response. Medium-term prediction uses Temporal Fusion Transformer, which integrates known future inputs and historical sequences to output future interval quality KPI predictions. The prediction mean and confidence interval are output through deep ensemble, and probability calibration is performed by combining temperature scaling.

[0077] A real-time interpretability framework based on SHAP is established to decompose the feature importance of the prediction results in the current time window and generate a ranking of key factors. Combined with domain knowledge and causal discovery algorithms, the causal links between process parameters and defect events are identified, and the credibility of the causal effect is verified through historical intervention data or digital twin simulation.

[0078] Constructing a multi-dimensional risk scoring mechanism:

[0079] ;

[0080] in, The predicted defect rate; This is the target upper limit threshold for the defect rate; The standard deviation of the historical defect rate; is the cumulative distribution function of the standard normal distribution; w1 is the weighting coefficient for exceeding the defect rate limit;

[0081] w1 represents the predicted probability of exceeding the limits for quality indicators such as color difference; w2 represents the weighting coefficient for exceeding the limits for quality indicators. For indicator functions; w3 represents the predicted probability of a critical defect event occurring; w3 is the weighting coefficient for critical events. This represents the predicted rate of change in the defect rate. λ is the predicted acceleration for the defect rate. accel w4 is the acceleration penalty coefficient; w4 is the overall weight of the trend penalty; trend_penalty is the trend penalty term;

[0082] The system uses a three-tiered approach: green, yellow, and red, based on risk scores to establish early warning levels.

[0083] Furthermore, based on future interval quality predictions and key factor rankings, adaptive process optimization is performed to obtain the optimal control commands, as follows:

[0084] Using the future interval quality prediction and key factor ranking output from S5 as input, a multi-objective constrained optimization problem is constructed, and a controllable parameter space is defined:

[0085] u=[v(t),T1(t),T2(t),Ptension(t),Hhumidity(t)]

[0086] This includes key process parameters such as linear speed, temperature settings for each section, tension control, and humidity adjustment.

[0087] Sensitivity analysis results based on key factor ranking With lag time τ i Establish a parameter-quality response model:

[0088]

[0089] in, It can be a linearized local sensitivity model or a pre-trained surrogate model;

[0090] Design a multi-level constraint system: hard constraints include equipment physical limits, safe operating boundaries, and product specification requirements; soft constraints cover energy consumption optimization objectives, equipment life protection, and operational stability.

[0091] A hierarchical optimization architecture is adopted to address different risk levels and time scales: preventative fine-tuning is performed during green alerts, and model predictive control is used to minimize prediction quality bias within a rolling time window. The objective function is:

[0092] ;

[0093] Where Q is the quality weight matrix, R u The control weight matrix is ​​denoted by Δu, which represents the control increment.

[0094] When a yellow alert is issued, an active optimization mode is activated. Based on sensitivity analysis, high-impact factors are prioritized for adjustment. Constrained quadratic programming is used for solution, and robustness considerations are introduced to handle prediction uncertainties through worst-case scenario optimization.

[0095] When a red alert is issued, an emergency intervention protocol is triggered to execute a rule-based rapid response, while simultaneously initiating a parallel global re-optimization calculation to determine the optimal recovery path.

[0096] The optimization results are converted into a standardized sequence of control commands, with each control vector containing the target setpoint, rate of change constraint, execution priority, and fault tolerance range.

[0097] The deep learning-based impregnated paper production quality inspection system includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps in the deep learning-based impregnated paper production quality inspection method described above.

[0098] A computer storage medium storing a plurality of instructions adapted for loading by a processor and executing the method steps described above.

[0099] The present invention has the following beneficial effects:

[0100] 1. This invention acquires multi-source synchronous data packets, covering multi-dimensional information such as images, process parameters, and business events. Combined with ROI mask processing, it effectively distinguishes products from invalid areas. The biscuit can identify defects in all aspects through multi-model AI detection and segmentation strategies. At the same time, it integrates process data, scene conditions, and historical knowledge to achieve highly robust online defect detection and localization.

[0101] 2. The present invention uses defect trajectory clustering and event flow modeling to achieve cross-frame tracking and lifecycle merging of defects in the paper feeding direction, so that each defect can be tracked, classified and statistically analyzed, eliminating problems such as duplicate detection and missed masking. Based on the deep coupling of event flow and process data, it supports real-time prediction of indicators such as defect rate and risk level in future production intervals, and automatically identifies key factors affecting quality and their direction of action through SHAP causal analysis method.

[0102] 3. This invention adaptively defines optimal process control instructions, covering core parameters, taking into account process constraints and energy consumption costs, and employs a hierarchical response and self-learning mechanism. It can achieve preventive fine-tuning under normal conditions and also trigger emergency process intervention in a timely manner when major risk warnings are issued, minimizing scrap losses. Attached Figure Description

[0103] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0104] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0105] refer to Figure 1 In this embodiment, a deep learning-based method for detecting the production quality of impregnated paper is provided, including the following steps:

[0106] S1: Acquire image stream from industrial camera: X imgt (RGB / grayscale, multi-camera, line scan / area scan), sensor and process parameters: X proct (Linear speed, impregnation tank temperature, resin solids content, viscosity, anilox roller speed, drying section temperature and humidity, tension, thickness / basis weight), Event and work order information: X mest (Batch, shift, paper / resin batch number, target quality control indicators), construct a synchronous data package;

[0107] S2: Preprocess the synchronization data packets to obtain preprocessed image data, process data, and ROI mask;

[0108] S3: Based on the preprocessed image data, process data, and ROI mask, obtain the detection box set and segmentation mask set using multi-model AI detection and segmentation;

[0109] S4: Based on the detection box set and the segmentation mask set, perform trajectory clustering along the paper feed direction to form a defect event flow;

[0110] S5: Based on the defect event flow, perform online quality prediction and root cause analysis to obtain future interval quality prediction and key factor ranking;

[0111] S6: Based on future interval quality predictions and key factor ranking, perform adaptive process optimization to obtain the optimal control command.

[0112] In this embodiment, a synchronization data packet is constructed as follows: First, all source data are aligned based on a unified clock, and the camera timestamps are corrected using NTP timing. The encoder pulse is used as the primary time base to map each frame of image to the paper feed length coordinate system. Then, camera intrinsic and extrinsic parameters are loaded and applied for geometric correction and pixel ratio calibration to establish a linear or piecewise affine mapping from pixels to physical coordinates (millimeters). If it is a multi-camera stitching scene, overlapping area feature matching and color gain equalization are performed to output a synchronization data packet with a consistent full-frame image and unified coordinate system across cameras.

[0113] In this embodiment, the image data preprocessing is as follows:

[0114] Noise suppression and illumination field equalization are performed on industrial camera images. Bilatera denoising is used to target reflective and bright areas. Combined with the separation of reflection components in polarization imaging, shadow and highlight suppression is achieved. For color-related defects, white balance and gain calibration are first performed using a gray card. Then adaptive histogram equalization is performed, and the upper limit of local contrast enhancement is limited to prevent amplification of texture noise and paper texture artifacts.

[0115] Lens distortion is corrected based on calibration matrix, and the geometric scale and orientation of the image are unified (e.g., to ensure that the paper feed direction is consistent in the vertical direction). For wide paper, block resampling and seam correction are used to remove brightness jumps at the seams. For decorative paper with periodic patterns, template alignment is performed: phase registration is performed using the gold sample reference image to obtain the normalized image of the page.

[0116] To improve the visibility of small and low-contrast defects, a multi-scale pyramid is constructed, and Gabor filtering is applied to each layer to enhance linear and point-like perturbations. For wrinkle and bubble-like defects, morphological top caps / black caps are used to enhance local bulge / depression features. For the color difference detection path, color constraining (e.g., Gray-World or Shades-of-Gray) is performed, and then an image copy in a linear color space (e.g., Lab) is output as a dedicated input view for the subsequent color difference regression branch.

[0117] Invalid area removal and ROI mask construction: Based on the mechanical structure and field of view, a static mask is constructed to shield non-product areas (such as clamping edges, guide rollers, borders, and rulers), and a dynamic mask is superimposed to handle temporary obstructions or contamination (e.g., detecting the time when the guide roller enters the image based on changes in light intensity and edge density); adaptive edge tracking is used on the paper edges to obtain the actual effective width, generating frame-by-frame effective area contours; finally, a binary ROI mask M is synthesized. roi A value of 1 indicates that the valid paper pixels allowed for detection are allowed, and 0 indicates that the region is ignored; and the minimum bounding polyline that matches it is output, which is convenient for subsequent sliding window inference and clipping;

[0118] Process data Missing value processing (forward filling / spline interpolation selected according to variable type) and outlier removal (based on IQR / Isolation Forest) are performed. Subsequently, resampling is performed at a uniform sampling rate and aligned with image frame timestamps. Short-window statistical features (mean, variance, slope) and hysteresis terms are calculated for key process quantities (temperature, viscosity, tension, linear velocity, etc.) to capture the hysteretic effects on surface defects. These features, along with the original quantities, constitute... Multi-component representation;

[0119] Finally, the preprocessed image data is output. Process data and ROI mask To adapt to subsequent multi-task networks, image data The image set contains three types of views: a high-resolution RGB view for object detection, an equalization intensity enhancement view for segmentation, and a Lab view for color difference estimation. The transformation chain and calibration parameters of the three are recorded through metadata to ensure that the coordinates are consistent across the views, and all outputs are encapsulated in a unified data structure.

[0120] Each batch of preprocessing generates a process log and quality summary, including noise level estimates, illumination uniformity metrics, effective swath ratio, shielded area ratio, and potential anomalies (such as light source attenuation or lens contamination). These metrics, along with versioned preprocessing configurations, are written to the data lake to ensure that subsequent model performance fluctuations can be replayed and located. If illumination or focus exceeds thresholds, an alarm is triggered, and recalibration or cleaning maintenance is requested.

[0121] In this embodiment, the multi-model AI detection and segmentation model adopts an architecture with a shared backbone and multiple branches. The branches include a detection branch, a segmentation branch, a color difference regression auxiliary head, and an unsupervised anomaly detection branch, as detailed below:

[0122] P and Q are extracted through a shared backbone and pyramid feature extraction network. The shared backbone network adopts CSP-Darknet and outputs multi-scale features {P3, P4, P5}. These features are then fused using FPN to construct a multi-scale feature set for use by the detection head. This facilitates multi-scale prediction and is represented as:

[0123] ;

[0124] Where Backbone represents the backbone network; I det For a preprocessed high-resolution RGB view; I seg Enhance the view to equalize intensity; For FPN networks, the term is used to perform top-down and horizontal fusion of multi-layer features of the shared backbone;

[0125] Constructing a high-resolution path Q in parallel for splitting branches:

[0126]

[0127] SegNeck represents the neck structure of a segmentation branch (such as jump connection fusion in U-Net, branch aggregation in HRNet, ASPP, etc.).

[0128] The detection branch outputs candidate boxes and scores, and uses anchor boxes (such as YOLOv8 / RT-DETR) to output the position. With category probability :

[0129] ;

[0130] Among them, P l is the feature map from the l-th scale of the feature pyramid; k is the position index on the feature map at that scale. For the bounding box regression head, predict the bounding box parameters (such as center point, width and height, or with rotation angle added); As the classification header, output the logits for each category; θ det The trainable parameters for the detection branch; softmax is the activation function;

[0131] Training loss:

[0132] ;

[0133] Where, L CIoU For bounding box regression loss, o represents the prediction of the existence of the target; o represents the binary truth label for the existence of the target. For binary cross-entropy loss; For classification probability prediction (after softmax);

[0134] y represents the one-hot true label of the category; For multi-class cross-entropy loss; λ box ,λ obj ,λ cls The weights of the three losses are hyperparameters used to balance the relative importance of bounding box regression, target existence, and classification losses, respectively.

[0135] The splitting branches use a soft mask output by U-Net++:

[0136] ;

[0137] in, Split the network for UNet++; θ seg These are the trainable parameters for the splitting branches; It is a function; For the predicted soft mask;

[0138] Training loss:

[0139] ;

[0140] Where M is the pixel-level real mask; It is an element-wise product; λ is the numerically stable term. bce λ represents the weights of the BCE segmentation loss. dice The weights for the Dice loss;

[0141] A color difference regression auxiliary head predicts pixel-level ΔE in Lab view. 2000 Heat maps enhance the confidence and grading robustness of color difference defects:

[0142] ;

[0143] Among them, I labInput the preprocessed Lab color space image; For color difference regression network; θ color This is the set of trainable parameters for the color difference regression head; For the predicted color difference heatmap;

[0144] The unsupervised anomaly detection branch uses PatchCore and outputs anomaly score graph A:

[0145] ;

[0146] in, This is an unsupervised anomaly detection model; θ ano Parameters for abnormal branches;

[0147] Overall training objectives:

[0148] ;

[0149] α, β, γ, δ are set through uncertainty weighting or grid search.

[0150] In this embodiment, the multi-model AI detection and segmentation model applies a ROI mask before the pixel-level output of any branch to avoid false detections of non-product areas:

[0151] ;

[0152] in, For ROI binary mask; A roi The effective region results after applying ROI are used for subsequent thresholding / fusion, which can suppress false detections in non-product regions;

[0153] For the detection frame, IoU filtering with ROI is applied:

[0154] ;

[0155] Where η∈[0,1] is the retention threshold; b i Let be the i-th candidate bounding box; ROI is the geometric region of ROI; The area function is 'keep'; 'keep' represents the filtering function.

[0156] Post-detection processing involves applying NMS to each class to attenuate the confidence level of overlapping targets.

[0157] ;

[0158] Among them, s i Let IoU(b) be the confidence score of the i-th candidate box;i ,b j ) represents the intersection-union ratio (IoU) of the two boxes; τ is the IoU suppression threshold of NMS. If the IoU between the two boxes exceeds τ, the one with the higher score is retained.

[0159] Throttling and morphological cleansing of soft masks:

[0160] ;

[0161] in, This is an indicator function; the condition is 1 if true, and 0 otherwise; θ m The binarization threshold for segmenting the soft mask; For morphological closing operations (erosion after dilation); For morphological opening operations (dilation after erosion);

[0162] Extract geometric properties (area, perimeter, slenderness ratio, principal axis direction) from connected components to generate instance-level segmentation objects;

[0163] For low-confidence detection or segmentation candidates, if their region mean outlier... Exceeding the quantile threshold q 0.9 Increase its confidence level:

[0164] ;

[0165] Where s is the confidence level of the original candidate; s′ is the updated confidence level after fusing the anomaly scores; q 0.9 ,q 0.99 These are the 90th and 99th percentiles of anomaly heatmap A in the current frame;

[0166] The instance mask generated from the segmentation is used to generate the minimum bounding box b. mask Perform IoU matching with the set of detection boxes, where IoU > τ. m When merging into a single object, the category and confidence are weighted and combined:

[0167] ;

[0168] Among them, b det The detection branch provides the bounding box; bmask is the bounding box generated by the mask of the segmentation instance;

[0169] For the geometric merging strategy of the two frames, a weighted average (weighted by area or confidence level, center and size); s det To detect the confidence scores of branch candidates; s mask The confidence score for the split branch instance; α s The weights for confidence fusion; p det(c) represents the posterior probability of the detection branch for class c; p mask (c) represents the posterior probability of the splitting branch for class c; α c Weights for category fusion; The final category label after merging;

[0170] The final confidence score after fusion; The final bounding box after merging;

[0171] Finally, we obtain the set of detection boxes. and segmentation mask set .

[0172] In this embodiment, based on the detection box set and the segmentation mask set, trajectory clustering is performed along the paper feed direction to form a defect event flow, as detailed below:

[0173] Based on the preset paper feed direction vector and pixel-physical mapping relationship, the centroid of each fusion candidate is projected onto the vertical arc length s - horizontal offset r parameterized coordinate system, where s is accumulated along the paper movement direction and r is the horizontal relative position; combining timestamp and linear velocity information, a state vector z = is constructed for each candidate. sconf (geometric properties), where velocity components (Relatively stable laterally); A multi-target tracking association algorithm is adopted to maintain a Kalman filter for each existing trajectory to predict the state at the next moment. Gated Mahalanobis distance and a comprehensive cost function (spatial distance, class consistency, shape IoU, confidence difference) are defined. The optimal match is solved by the Hungarian algorithm, and new observations are assigned to existing trajectories or new trajectories are initialized. Trajectories that are not matched for consecutive Kmiss frames (typically 3-5) are put into a suspended state, allowing discontinuity caused by occlusion or missed detection. The trajectory is terminated after timeout. The trajectory update adopts an exponential moving average to fuse class probability and confidence, making the trajectory-level attributes more robust and able to reflect temporal consistency.

[0174] Stable trajectories that meet the minimum duration (≥ 10-20 mm along the paper feed direction or ≥ 3 frames) are output as defect event streams. Each event contains complete lifecycle attributes: basic identifiers (category, start and end times [tstart, tent], location interval [sstart, send], lateral range [rmin, rmax]), geometric statistics (cumulative area, maximum length and width, mask coverage heatmap summary), dynamic behavior (area expansion rate, lateral drift amplitude, number of trajectory discontinuities), and confidence distribution (mean, 95th percentile, lowest value interval).

[0175] In this embodiment, based on the defect event stream, online quality prediction and root cause analysis are performed to obtain future interval quality predictions and key factor rankings, as detailed below:

[0176] Using the defect event stream output by S4 as the main input, and simultaneously integrating process parameter time windows and batch information, interactive features are constructed, including temperature and humidity coupling effect, line speed-tension coupling, and hysteretic correlation between process parameters and event occurrence rate, forming a high-dimensional time series feature vector;

[0177] A hierarchical forecasting system for different time scales is established. Short-term forecasts (1-5 minutes) are deployed on edge devices using lightweight XGBoost to achieve low-latency response. Medium-term forecasts (10-30 minutes) use Temporal FusionTransformer, which integrates known future inputs (planned line velocity, set temperature) with historical sequences to output future interval quality KPI forecasts. The forecast mean and confidence interval are output through deep ensemble, and probability calibration is performed by combining temperature scaling.

[0178] A real-time interpretability framework based on SHAP is established to decompose the feature importance of the prediction results in the current time window, generate a ranking of key factors, and clarify the contribution, positive or negative direction, and typical lag time of each influencing factor to the quality prediction. In addition, by combining domain knowledge and causal discovery algorithms (such as PCMCI), the causal links between process parameters and defect events are identified, and the credibility of the causal effect is verified by historical intervention data or digital twin simulation.

[0179] Constructing a multi-dimensional risk scoring mechanism:

[0180] ;

[0181] in, The predicted defect rate; This is the target upper limit threshold for the defect rate; The standard deviation of the historical defect rate; is the cumulative distribution function of the standard normal distribution; w1 is the weighting coefficient for exceeding the defect rate limit;

[0182] w1 represents the predicted probability of exceeding the limits for quality indicators such as color difference; w2 represents the weighting coefficient for exceeding the limits for quality indicators. For indicator functions; w3 represents the predicted probability of a critical defect event occurring; w3 is the weighting coefficient for critical events. This represents the predicted rate of change in the defect rate. λ is the predicted acceleration for the defect rate. accel w4 is the acceleration penalty coefficient; w4 is the overall weight of the trend penalty; trend_penalty is the trend penalty term;

[0183] The system uses a three-tiered warning system based on risk scores: green (Risk < 0.3, normal monitoring), yellow (0.3 ≤ Risk < 0.6, optimization and adjustment recommended), and red (Risk ≥ 0.6, immediate intervention and possible speed limit / enhanced removal).

[0184] In this embodiment, adaptive process optimization is performed based on future interval quality prediction and key factor ranking to obtain the optimal control command, as follows:

[0185] Using the future interval quality prediction and key factor ranking output from S5 as input, a multi-objective constrained optimization problem is constructed, and a controllable parameter space is defined:

[0186] u=[v(t),T1(t),T2(t),Ptension(t),Hhumidity(t)]

[0187] This includes key process parameters such as linear speed, temperature settings for each section, tension control, and humidity adjustment.

[0188] Sensitivity analysis results based on key factor ranking With lag time τ i Establish a parameter-quality response model:

[0189]

[0190] in, It can be a linearized locality sensitivity model or a pre-trained surrogate model (such as Gaussian Process regression).

[0191] Design a multi-level constraint system: hard constraints include equipment physical limits ( ), safe operating boundaries (including temperature change rate limits) Soft constraints include tension fluctuation constraints and product specification requirements (minimum line speed guarantee, critical temperature window); soft constraints cover energy consumption optimization goals, equipment life protection (reducing frequent adjustments) and operational smoothness (minimizing control action amplitude).

[0192] A hierarchical optimization architecture is adopted to address different risk levels and time scales: preventative fine-tuning is performed during green alerts, and model predictive control is used to minimize prediction quality bias within a rolling time window. The objective function is:

[0193] ;

[0194] Where Q is the quality weight matrix, R u The control weight matrix is ​​denoted by Δu, which represents the control increment.

[0195] When a yellow alert is issued, an active optimization mode is activated. Based on sensitivity analysis, high-impact factors are prioritized for adjustment. Constrained quadratic programming is used for solution, and robustness considerations are introduced to handle prediction uncertainties through worst-case scenario optimization.

[0196] When a red alert is issued, an emergency intervention protocol is triggered to execute a rule-based rapid response (such as immediately limiting the rate to a safe threshold, strengthening the elimination strategy, and emergency adjustment of critical temperatures), while simultaneously initiating a parallel global re-optimization calculation to determine the optimal recovery path.

[0197] The optimization results are converted into a standardized sequence of control commands, with each control vector containing the target setpoint, rate of change constraint, execution priority, and fault tolerance range.

[0198] The deep learning-based impregnated paper production quality inspection system includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps in the deep learning-based impregnated paper production quality inspection method described above.

[0199] A computer storage medium storing a plurality of instructions adapted for loading by a processor and executing the method steps described above.

[0200] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0201] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0202] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0203] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0204] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A deep learning-based method for quality inspection in impregnated paper production, characterized in that, Includes the following steps S1: Acquire industrial camera image streams, sensor and process parameters, event and work order information, and construct synchronization data packets; S2: Preprocess the synchronization data packets to obtain preprocessed image data, process data, and ROI mask; S3: Based on the preprocessed image data, process data, and ROI mask, obtain the detection box set and segmentation mask set using multi-model AI detection and segmentation; S4: Based on the detection box set and the segmentation mask set, perform trajectory clustering along the paper feed direction to form a defect event flow; S5: Based on the defect event flow, perform online quality prediction and root cause analysis to obtain future interval quality prediction and key factor ranking; S6: Based on future interval quality predictions and key factor ranking, perform adaptive process optimization to obtain the optimal control command.

2. The method for quality inspection of impregnated paper production based on deep learning according to claim 1, characterized in that, The construction of the synchronization data packet is as follows: First, all source data are aligned based on a unified clock, and the camera timestamps are corrected using NTP time synchronization. The encoder pulse is used as the main time base to map each frame image to the paper feed length coordinate system. Then, camera intrinsic and extrinsic parameters are loaded and applied to perform geometric correction and pixel ratio calibration, and a linear or piecewise affine mapping from pixels to physical coordinates is established. If it is a multi-camera stitching scene, overlapping area feature matching and color gain equalization are performed to output a synchronization data packet with a consistent full-frame image and unified coordinate system across cameras.

3. The method for quality inspection of impregnated paper production based on deep learning according to claim 2, characterized in that, The image data preprocessing is as follows: Noise suppression and illumination field equalization are performed on industrial camera images. Bilatera is used for noise reduction. For reflective and bright areas, the reflection component of polarization imaging is combined to separate shadows and highlights. For color-related defects, white balance and gain calibration are first performed using a gray card, and then adaptive histogram equalization is performed. Lens distortion is corrected based on calibration matrix, and the geometric scale and orientation of the image are unified. For wide-format paper, block resampling and seam correction are used to remove brightness jumps at the seams. For decorative paper with periodic patterns, template alignment is performed: phase registration is performed using the gold sample reference image to obtain the normalized image of the layout. A multi-scale pyramid is constructed, and Gabor filtering is applied to each layer to enhance linear and point-like perturbations. For wrinkles and bubble-like defects, morphological caps are used to enhance local bulge features. For the color difference detection path, color constantization is performed, and then a copy of the image in linear color space is output as a dedicated input view for the subsequent color difference regression branch. Invalid region removal and ROI mask construction: Based on the mechanical structure and field of view, a static mask is constructed to shield non-product areas, and a dynamic mask is superimposed to handle temporary occlusion or contamination; adaptive edge tracking is used on the paper edges to obtain the actual effective width, generating frame-by-frame effective area contours; finally, a binary ROI mask M is synthesized. roi A value of 1 indicates that the valid paper pixels allowed for detection are allowed, and 0 indicates that the region is ignored; and the minimum bounding polyline that matches it is output, which is convenient for subsequent sliding window inference and clipping; Process data Perform missing value processing and outlier removal, then resample at a uniform sampling rate and align with the image frame timestamps; Short-window statistical characteristics and lag terms are calculated for key process quantities to capture the hysteretic effects on surface defects. These characteristics, along with the original quantities, constitute... Multi-component representation; Finally, the preprocessed image data is output. Process data and ROI mask To adapt to subsequent multi-task networks, image data An image set containing three types of views: a high-resolution RGB view for object detection, an equalization intensity enhancement view for segmentation, and a Lab view for color difference estimation; All three elements record the transformation chain and calibration parameters through metadata to ensure consistent coordinates across views, and all outputs are encapsulated in a unified data structure.

4. The method for quality inspection of impregnated paper production based on deep learning according to claim 1, characterized in that, The multi-model AI detection and segmentation model adopts a shared backbone and multi-branch architecture. The branches include a detection branch, a segmentation branch, a color difference regression auxiliary head, and an unsupervised anomaly detection branch, as detailed below: P and Q are extracted using a shared backbone and pyramid feature extraction method. The shared backbone network employs CSP-Darknet, outputting multi-scale features {P3, P4, P5}. These features are then fused using FPN to construct a multi-scale feature set for use by the detection head. This facilitates multi-scale prediction and is represented as: ； Where Backbone represents the backbone network; I det For a preprocessed high-resolution RGB view; I seg Enhance the view to equalize intensity; For FPN networks, the term is used to perform top-down and horizontal fusion of multi-layer features of the shared backbone; Constructing a high-resolution path Q in parallel for splitting branches: ； SegNeck represents the neck structure of the split branches; The detection branch outputs candidate boxes and scores, and anchor boxes are used to output positions. With category probability : ； Among them, P l is the feature map from the l-th scale of the feature pyramid; k is the position index on the feature map at that scale. For bounding box regression heads, predict box parameters; As the classification header, output the logits for each category; θ det The trainable parameters for the detection branch; softmax is the activation function; Training loss: ; Where, L CIoU For bounding box regression loss, o represents the prediction of the existence of the target; o represents the binary truth label for the existence of the target. For binary cross-entropy loss; For classification probability prediction; y represents the one-hot true label of the category; For multi-class cross-entropy loss; λ box ,λ obj ,λ cls The weights of the three losses are hyperparameters used to balance the relative importance of bounding box regression, target existence, and classification losses, respectively. The splitting branches use a soft mask output by U-Net++: ； in, Split the network for UNet++; θ seg These are the trainable parameters for the splitting branches; It is a function; For the predicted soft mask; Training loss: ; Where M is the pixel-level real mask; It is an element-wise product; λ is the numerically stable term. bce λ represents the weights of the BCE segmentation loss. dice The weights for the Dice loss; A color difference regression auxiliary head predicts pixel-level ΔE in Lab view. 2000 Heat maps enhance the confidence and grading robustness of color difference defects: ； Among them, I lab Input the preprocessed Lab color space image; For color difference regression network; θ color This is the set of trainable parameters for the color difference regression head; For the predicted color difference heatmap; The unsupervised anomaly detection branch uses PatchCore and outputs anomaly score graph A: ； in, This is an unsupervised anomaly detection model; θ ano Parameters for abnormal branches; Overall training objectives: ； α, β, γ, δ are set through uncertainty weighting or grid search.

5. The method for quality inspection of impregnated paper production based on deep learning according to claim 4, characterized in that, The multi-model AI detection and segmentation model applies a ROI mask before the pixel-level output of any branch to avoid false detections of non-product areas. ； in, For ROI binary mask; A roi The effective region results after applying the ROI; For the detection frame, IoU filtering with ROI is applied: ； Where η∈[0,1] is the retention threshold; b i Let be the i-th candidate bounding box; ROI is the geometric region of ROI; The area function is 'keep'; 'keep' represents the filtering function. Post-detection processing involves applying NMS to each class to attenuate the confidence level of overlapping targets. ; Among them, s i Let IoU(b) be the confidence score of the i-th candidate box; i ,b j ) represents the intersection-union ratio of the two boxes; τ is the IoU suppression threshold of NMS; Throttling and morphological cleansing of soft masks: ； in, This is an indicator function; the condition is 1 if true, and 0 otherwise; θ m The binarization threshold for segmenting the soft mask; For morphological closing operations; Opening operations for morphology; Extract geometric properties from connected components to generate instance-level segmentation objects; For low-confidence detection or segmentation candidates, if their region mean outlier... Exceeding the quantile threshold q 0.9 Increase its confidence level: ; Where s is the confidence level of the original candidate; s′ is the updated confidence level after fusing the anomaly scores; q 0.9 ,q 0.99 These are the 90th and 99th percentiles of anomaly heatmap A in the current frame; The instance mask generated from the segmentation is used to generate the minimum bounding box b. mask Perform an IoU match with the set of detection boxes, where IoU > τ. m When merging into a single object, the category and confidence are weighted and combined: ; Among them, b det The detection branch provides the bounding box; bmask is the bounding box generated by the mask of the segmentation instance; The geometric merging strategy for the two frames; s det To detect the confidence scores of branch candidates; s mask The confidence score for the split branch instance; α s The weights for confidence fusion; p det (c) represents the posterior probability of the detection branch for class c; p mask (c) represents the posterior probability of the splitting branch for class c; α c Weights for category fusion; The final category label after merging; The final confidence score after fusion; The final bounding box after merging; Finally, we obtain the set of detection boxes. and segmentation mask set .

6. The method for quality inspection of impregnated paper production based on deep learning according to claim 5, characterized in that, The process of clustering the trajectory along the paper feed direction based on the detection box set and the segmentation mask set to form a defect event flow is as follows: Based on the preset paper feed direction vector and pixel-physical mapping relationship, the centroid of each fusion candidate is projected onto the longitudinal arc length s - lateral offset r parameterized coordinate system, where s is accumulated along the paper movement direction and r is the lateral relative position. Combining timestamp and linear velocity information, a state vector is constructed for each candidate. A multi-target tracking association algorithm is used to maintain a Kalman filter for each existing trajectory to predict the state at the next moment, defining a gated Mahalanobis distance and a comprehensive cost function. The optimal match is solved using the Hungarian algorithm, and new observations are assigned to existing trajectories or initialized into new trajectories. Trajectories that are not matched for consecutive Kmiss frames enter a suspended state, allowing discontinuity caused by occlusion or missed detection, and the trajectory is terminated after timeout. The trajectory update uses an exponential moving average to fuse class probabilities and confidence levels, making the trajectory-level attributes more robust and able to reflect temporal consistency. The stable trajectory that meets the minimum duration is output as a defect event stream, and each event contains complete lifecycle attributes.

7. The deep learning-based method for quality inspection of impregnated paper production as described in claim 6, characterized in that, The process involves online quality prediction and root cause analysis based on defect event streams to obtain future interval quality predictions and key factor rankings, as detailed below: Using the defect event stream output by S4 as the main input, and simultaneously integrating process parameter time windows and batch information, interactive features are constructed, including temperature and humidity coupling effect, line speed-tension coupling, and hysteretic correlation between process parameters and event occurrence rate, forming a high-dimensional time series feature vector; A hierarchical prediction system for different time scales is established. Short-term prediction uses lightweight XGBoost deployed on edge devices to achieve low-latency response. Medium-term prediction uses Temporal Fusion Transformer, which integrates known future inputs and historical sequences to output future interval quality KPI predictions. The prediction mean and confidence interval are output through deep ensemble, and probability calibration is performed by combining temperature scaling. A real-time interpretability framework based on SHAP is established to decompose the feature importance of the prediction results for the current time window and generate a ranking of key factors. By combining domain knowledge with causal discovery algorithms, the causal links between process parameters and defect events are identified, and the credibility of the causal effect is verified through historical intervention data or digital twin simulation. Constructing a multi-dimensional risk scoring mechanism: ； in, The predicted defect rate; This is the target upper limit threshold for the defect rate; The standard deviation of the historical defect rate; is the cumulative distribution function of the standard normal distribution; w1 is the weighting coefficient for exceeding the defect rate limit; w1 represents the predicted probability of exceeding the limits for quality indicators such as color difference; w2 represents the weighting coefficient for exceeding the limits for quality indicators. For indicator functions; w3 represents the predicted probability of a critical defect event occurring; w3 is the weighting coefficient for critical events. This represents the predicted rate of change in the defect rate. λ is the predicted acceleration for the defect rate. accel w4 is the acceleration penalty coefficient; w4 is the overall weight of the trend penalty; trend_penalty is the trend penalty term; The system uses a three-tiered approach: green, yellow, and red, based on risk scores to establish early warning levels.

8. The method for quality inspection of impregnated paper production based on deep learning according to claim 7, characterized in that, The adaptive process optimization based on future interval quality prediction and key factor ranking to obtain the optimal control command is as follows: Using the future interval quality prediction and key factor ranking output from S5 as input, a multi-objective constrained optimization problem is constructed, and a controllable parameter space is defined: u=[v(t),T1(t),T2(t),Ptension(t),Hhumidity(t)] This includes key process parameters such as linear speed, temperature settings for each section, tension control, and humidity adjustment. Sensitivity analysis results based on key factor ranking With lag time τ i Establish a parameter-quality response model: ； in, It can be a linearized local sensitivity model or a pre-trained surrogate model; Design a multi-level constraint system: hard constraints include equipment physical limits, safe operating boundaries, and product specification requirements; soft constraints cover energy consumption optimization objectives, equipment life protection, and operational stability. A hierarchical optimization architecture is adopted to address different risk levels and time scales: preventative fine-tuning is performed during green alerts, and model predictive control is used to minimize prediction quality bias within a rolling time window. The objective function is: ； Where Q is the quality weight matrix, R u The control weight matrix is ​​denoted by Δu, which represents the control increment. When a yellow alert is issued, an active optimization mode is activated. Based on sensitivity analysis, high-impact factors are prioritized for adjustment. Constrained quadratic programming is used for solution, and robustness considerations are introduced to handle prediction uncertainties through worst-case scenario optimization. When a red alert is issued, an emergency intervention protocol is triggered to execute a rule-based rapid response, while simultaneously initiating a parallel global re-optimization calculation to determine the optimal recovery path. The optimization results are converted into a standardized sequence of control commands, with each control vector containing the target setpoint, rate of change constraint, execution priority, and fault tolerance range.

9. A deep learning-based quality inspection system for impregnated paper production, characterized in that, It includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps in the deep learning-based impregnated paper production quality inspection method as described in any one of claims 1-8.

10. A computer storage medium, characterized in that, The computer storage medium stores a plurality of instructions, which are adapted to be loaded by a processor and executed as method steps as claimed in any one of claims 1 to 8.

Images