Visual identification and trend prediction method and system for wrinkles of continuous strip-shaped material

Abstract

The invention relates to a visual identification and trend prediction method and system for wrinkles of a continuous strip-shaped material, and the method comprises the steps: collecting a continuous image sequence of the surface of the continuous strip-shaped material, obtaining a standardized image frame, and synchronously obtaining a working condition state vector; inputting the standardized image frame into a wrinkle identification model, and outputting wrinkle region information, wrinkle category information and corresponding confidence; constructing a wrinkle index vector at the current moment according to the wrinkle region information, the wrinkle category information and the corresponding confidence coefficient; filtering and time sequence modeling are carried out on the wrinkle index vectors at the current and historical moments, and wrinkle index prediction values, trend labels and prediction uncertainty at multiple moments in the future are output; and constructing a future risk index and security constraint set according to the wrinkle index prediction value, the prediction uncertainty and the working condition state vector, solving an optimal parameter packet, and mapping the optimal parameter packet into an executable control instruction for output.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and in particular to a method and system for visual recognition and trend prediction of wrinkles in continuous strip materials. Background Technology

[0002] Continuous strip materials often exhibit surface defects such as edge wrinkles, transverse wrinkles, and periodic wrinkles during production, conveying, winding, coating, laminating, plating, printing, and slitting. Existing methods largely rely on manual parameter adjustment based on experience or post-production sampling inspections at the later stages of the production line. Regarding online inspection technology, some image recognition-based methods have been developed for detecting surface defects in strip materials. These methods typically employ edge detection, texture analysis, or machine learning models to identify defect areas. In terms of parameter adjustment, traditional solutions often use rule-based control strategies or pre-set parameter lookup tables to guide operators in making adjustments.

[0003] However, existing technologies have the following shortcomings: First, the formation of wrinkles is related to working conditions such as material properties, linear velocity, tension state, ambient light, and equipment status, exhibiting significant dynamic and time-varying characteristics, making it difficult to apply single static settings in the long term; second, online detection is easily affected by factors such as reflections, surface textures, seams, particles, scratches, and contamination points, leading to false alarms or missed alarms, which in turn causes the adjustment strategy to become inaccurate; third, even if wrinkles are detected, existing solutions generally lack closed-loop, interpretable, and continuously learning algorithmic mechanisms to provide effective and executable adjustment parameters without causing risks such as resonance, strip breakage, disturbance of quality-sensitive sections, and equipment overshoot; in addition, parameter combinations vary for different materials, different process stages, and different equipment statuses, and traditional rule bases have high maintenance costs, weak generalization capabilities, and are difficult to adapt to new working conditions.

[0004] Therefore, an online closed-loop solution is needed, which is based on visual recognition, trend prediction and safety optimization. It can reliably identify wrinkles and predict their development trend, automatically provide parameter and strategy suggestions while meeting safety constraints, and continuously learn and optimize itself based on feedback data during long-term operation. Summary of the Invention

[0005] In view of the shortcomings of the prior art described above, the purpose of this invention is to provide a visual recognition and trend prediction method and system for wrinkles in continuous strip materials. By combining online visual inspection, short-term trend prediction, risk assessment and safety self-optimization parameter recommendation, the invention achieves real-time reliable identification of wrinkle defects, accurate prediction of development trends, and closed-loop adaptive parameter adjustment under safety constraints. This effectively reduces false alarm and false negative rates, improves production line response speed, reduces manual intervention costs, and has continuous self-learning capabilities to adapt to dynamic changes in different materials, processes and equipment conditions.

[0006] To achieve the above objectives, the present invention adopts the following technical solution.

[0007] In a first aspect, the present invention provides a method for visual recognition and trend prediction of wrinkles in continuous strip materials, which adopts the following technical solution: A method for visual recognition and trend prediction of wrinkles in continuous strip materials, comprising: A continuous image sequence of the surface of a continuous strip material is acquired, and a working condition state vector is acquired simultaneously. The working condition state vector includes one or more of the following: velocity, position, temperature, material batch, equipment status, and tension status. The image frames in the continuous image sequence are preprocessed to obtain standardized image frames. The preprocessing includes at least one of illumination normalization, reflection suppression, and noise reduction. The standardized image frame is input into the wrinkle recognition model, which outputs wrinkle region information, wrinkle category information and corresponding confidence level, wherein the wrinkle category information includes one or more of edge wrinkles, transverse wrinkles, periodic wrinkles and pseudo-wrinkles. Based on the wrinkle region information, wrinkle category information and corresponding confidence level, a wrinkle index vector is constructed for the current moment. The wrinkle index vector includes at least one of the following: severity index, direction index, density index, area ratio index, and persistence index. The wrinkle index vectors at current and historical moments are filtered and time-series modeled to output predicted wrinkle index values, trend labels, and prediction uncertainties for multiple future moments. The trend labels include one of rising, stable, and falling trends. Based on the predicted value of the wrinkle index, the prediction uncertainty, and the working condition vector, a set of future risk indicators and safety constraints is constructed, and the optimal parameter package is solved within the set of safety constraints. The optimal parameter package is then mapped into an executable control command output.

[0008] Furthermore, the method described above, in constructing the wrinkle index vector at the current moment, also includes: Multiple wrinkle features are fused into a unified wrinkle index, which is calculated using the following formula: ; in Indicators representing the percentage of area Indicates contrast or grayscale difference characteristics. Represents the texture energy index. Indicates a persistent indicator. , , , The weighting coefficient is adaptively adjusted according to the operating condition state vector.

[0009] Furthermore, the filtering and time-series modeling of the wrinkle index vectors at the current and historical moments in the above method includes: Kalman filtering is applied to the unified wrinkle index to obtain a smoothed state estimate; and The wrinkle index at the next K time points is predicted using ARIMA, LSTM, TCN, or Transformer models to obtain the future wrinkle index prediction sequence.

[0010] Furthermore, in the above method, the future risk indicator incorporates at least one of the following factors: Risk of continued deterioration of wrinkles; Equipment disturbance risk; Quality-sensitive window risk; and Model uncertainty risk.

[0011] Furthermore, in the above method, the set of security constraints includes at least one of the following constraints: Quality-sensitive window constraints prohibit the execution of certain parameter combinations or only allow the execution of degradation strategies within a preset sensitive window; Safety boundary constraints limit frequency, amplitude, rate of change, duration, or energy budget; Uncertainty constraints allow only conservative strategies to be chosen when the prediction uncertainty exceeds a threshold; and Equipment status constraints restrict high-risk parameter packages when the equipment is in an abnormal, derated, or maintenance state.

[0012] Furthermore, in the above method, solving for the optimal parameter package within the set of safety constraints includes: Construct the objective function ,in Indicates the benefits of wrinkle improvement. Indicates execution risk, Indicates execution cost, and For weighting coefficients; and The optimal parameter set that maximizes the objective function is solved within the set of security constraints using a Bayesian optimizer, a secure reinforcement learning policy network, or a hybrid optimizer that combines rules and learning.

[0013] Furthermore, in the above method, the wrinkle improvement benefit Calculated using the following formula: ; in The uniform wrinkle index at the current moment, To execute parameter package The expected wrinkle index after the event. Indicates the first The expected improvement in the amount of wrinkle characteristics. For the corresponding weights.

[0014] Furthermore, in the above method, the execution risk... Calculated using the following formula: ; in Indicates the execution parameter package Time The probability of occurrence of risk events. This indicates the quantified value of the loss from the risk event, which includes resonance risk, band breakage risk, quality-sensitive section disturbance risk, and / or equipment overshoot risk.

[0015] Furthermore, in the above method, the execution cost Calculated using the following formula: ; in Indicates the tension adjustment range. Indicates the speed adjustment range. Indicates the number of adjustments. Indicates the execution time. , , , This represents the corresponding cost coefficient.

[0016] Secondly, the present invention provides a visual recognition and trend prediction system for wrinkles in continuous strip materials, which adopts the following technical solution: The image acquisition module is used to acquire a continuous sequence of images of the surface of a continuous strip material. The working condition acquisition module is used to synchronously acquire the working condition status vector, which includes one or more of the following: speed, position, temperature, material batch, equipment status, and tension status. The preprocessing module is used to preprocess the image frames in the continuous image sequence to obtain standardized image frames. The preprocessing includes at least one of illumination normalization, reflection suppression, and noise reduction. The wrinkle recognition module is used to input the standardized image frame into the wrinkle recognition model and output wrinkle region information, wrinkle category information and corresponding confidence level, wherein the wrinkle category information includes one or more of edge wrinkles, transverse wrinkles, periodic wrinkles and pseudo wrinkles. The indicator construction module is used to construct a wrinkle indicator vector at the current moment based on the wrinkle region information, wrinkle category information and corresponding confidence level. The wrinkle indicator vector includes at least one of severity indicator, direction indicator, density indicator, area ratio indicator and persistence indicator. The trend prediction module filters and models the wrinkle index vectors at current and historical moments, outputting predicted wrinkle index values, trend labels, and prediction uncertainties for multiple future moments. The trend labels include one of rising, stable, and falling trends. The parameter optimization module is used to construct a set of future risk indicators and safety constraints based on the predicted value of the wrinkle index, the prediction uncertainty and the working condition vector, solve for the optimal parameter package within the set of safety constraints, and map the optimal parameter package into an executable control command output.

[0017] In summary, compared with the prior art, the present invention includes at least one of the following beneficial technical effects: The present invention provides a method for visual recognition and trend prediction of wrinkles in continuous strip materials. By acquiring continuous image sequences and simultaneously obtaining working condition state vectors, and combining preprocessing and wrinkle recognition models to output wrinkle region and category information, a wrinkle index vector containing multi-dimensional indicators is constructed. The historical and current wrinkle index vectors are filtered and time-series modeled to output predicted values, trend labels, and uncertainties for multiple future moments. Based on the prediction results and working condition information, a set of risk indicators and safety constraints is constructed, and the optimal parameter package is solved and mapped into control commands. This method achieves real-time and accurate identification of wrinkle defects, reliable prediction of development trends, and closed-loop adaptive parameter adjustment under safety constraints. It effectively solves the problems of static setting failure caused by dynamic changes in working conditions, false alarms and missed alarms caused by interference factors, and lack of closed-loop interpretable optimization mechanisms in existing technologies, significantly improving the intelligence level and production efficiency of production lines. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart of an embodiment of the visual recognition and trend prediction method for continuous strip material wrinkles of the present invention is shown.

[0020] Figure 2 A flowchart of an embodiment of the wrinkle index construction and trend prediction method of the present invention is shown.

[0021] Figure 3 The diagram shows a hierarchical composition node diagram of an embodiment of the risk assessment and security constraints of the present invention.

[0022] Figure 4 A flowchart of an embodiment of the safety constraint parameter optimization method of the present invention is shown.

[0023] Figure 5 A structural block diagram of an embodiment of the visual recognition and trend prediction system of the present invention is shown. Detailed Implementation

[0024] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application. Furthermore, it should be understood that the specific embodiments described herein are only for illustration and explanation of this application and are not intended to limit this application.

[0025] It should be noted that the order of description of the following embodiments is not intended to limit the preferred order of the embodiments of this application. Furthermore, the descriptions of each embodiment in the following embodiments have their own emphasis; for parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0026] The method steps described in this embodiment of the invention can be executed in the order described in the specific implementation, or the execution order of each step can be adjusted according to actual needs, provided that the technical problem can be solved. These are not listed one by one here.

[0027] The present invention will be further described in detail below with reference to the accompanying drawings.

[0028] Reference Figure 1 , Figure 1 This diagram illustrates the overall flowchart of a visual recognition and trend prediction method 100 for wrinkles in continuous strip materials according to an embodiment of the present invention. Method 100 includes steps 102, 104, 106, 108, 110, and 112. Continuous strip materials may include types such as rolled-up coated materials, paper, metal foil, fabric, and composite strips. By sequentially executing the above steps, method 100 achieves online visual recognition, short-term trend prediction, and safety self-optimization parameter recommendation for wrinkle defects on the surface of continuous strip materials.

[0029] In step 102, a continuous image sequence of the continuous strip material surface is acquired, and a working condition state vector is simultaneously obtained. The working condition state vector includes one or more of the following: velocity, position, temperature, material batch, equipment status, and tension status. In some embodiments, the working condition state vector may also include historical execution status and historical adjustment records. By simultaneously acquiring the working condition state vector, method 100 can correlate image information with the current production condition, providing a working condition context for subsequent trend prediction and parameter optimization.

[0030] Continue to refer to Figure 1 In step 104, image frames in the continuous image sequence are preprocessed to obtain standardized image frames. Preprocessing includes at least one of illumination normalization, reflection suppression, and noise reduction. Through preprocessing, the impact of illumination fluctuations, surface reflections, and dynamic blur on subsequent wrinkle recognition can be reduced, resulting in standardized image frames with stable image quality.

[0031] In step 106, a standardized image frame is input into the wrinkle recognition model, which outputs wrinkle region information, wrinkle category information, and corresponding confidence scores. Wrinkle category information includes one or more of edge wrinkles, transverse wrinkles, periodic wrinkles, and pseudo-wrinkles. Wrinkle region information can be a segmentation mask, bounding box, or contour set. The confidence score characterizes the reliability of the wrinkle recognition result.

[0032] like Figure 1 As shown, in step 108, a wrinkle index vector for the current moment is constructed based on the wrinkle region information, wrinkle category information, and corresponding confidence levels. The wrinkle index vector includes at least one of the following: severity index, direction index, density index, area proportion index, and persistence index. The wrinkle index vector converts the identification results into a quantifiable numerical form to facilitate subsequent time-series modeling and trend prediction.

[0033] In step 110, the wrinkle index vectors at the current and historical moments are filtered and time-series modeled to output predicted wrinkle index values, trend labels, and prediction uncertainties for multiple future moments. The trend label includes one of three states: rising, stable, or falling. The prediction uncertainty characterizes the confidence level of the prediction result. By performing time-series prediction of the wrinkle index, method 100 can perceive the development trend of wrinkles in advance, achieving early intervention rather than post-event correction.

[0034] Continue to refer to Figure 1 In step 112, based on the predicted value of the wrinkle index, the prediction uncertainty, and the operating condition vector, a set of future risk indicators and safety constraints is constructed. The optimal parameter package is then solved within this set of safety constraints and mapped to an executable control command output. By optimizing parameters within the set of safety constraints, method 100 can provide parameter recommendations while meeting equipment, process, and quality constraints, avoiding the risks associated with unconstrained trial and error. The executable control command can be output to the upper-level control system, actuator, or operator parameter adjustment interface to achieve early intervention in wrinkle defects and controlled risk adjustment.

[0035] Furthermore, in step 104 above, the preprocessing operation has a direct impact on obtaining high-quality standardized image frames. The specific implementation of the preprocessing operation will be described in detail below.

[0036] Illumination normalization is used to eliminate brightness differences caused by uneven light sources or variations in illumination during image acquisition. In some implementations, illumination normalization employs a flat-field correction method, which pre-acquires a uniform reference image and calculates correction coefficients to perform pixel-by-pixel correction on the acquired image frames, thereby eliminating the effects of uneven spatial distribution of the light source. In other implementations, illumination normalization employs a gamma correction method, which adjusts the grayscale response curve of the image through nonlinear transformation, ensuring that the image exhibits consistent brightness distribution characteristics under different illumination conditions.

[0037] Reflection suppression is used to reduce the interference of reflections from continuous strip-shaped material surfaces on wrinkle recognition. In some embodiments, reflection suppression employs a polarization difference method, which involves acquiring multiple images under different polarization directions and performing difference operations to separate the reflective and diffuse reflection components, thereby suppressing the influence of highlight areas. In some embodiments, reflection suppression employs a highlight area masking method, which detects areas in the image whose brightness exceeds a preset threshold and generates a mask, allowing for special processing or exclusion of highlight areas in subsequent processing. In some embodiments, reflection suppression employs an HDR fusion method, which acquires multiple images under different exposure conditions and fuses them to expand the dynamic range of the image, ensuring that effective texture information is preserved in both highlight and shadow areas.

[0038] Denoising and detail enhancement are used to improve image quality and highlight wrinkle texture features. In some implementations, denoising employs bilateral filtering, which smooths noise while preserving image edge information and avoiding blurring of wrinkle boundaries. In other implementations, denoising uses nonlocal mean filtering, which removes noise while retaining image detail by using a weighted average based on information from similar blocks in the image. In still other implementations, detail enhancement employs desharpening masking, which enhances image edges and details by subtracting a blurred version from the original image and superimposing it back, making wrinkle textures clearer.

[0039] Motion compensation is used to reduce the impact of dynamic blurring generated during high-speed movement of continuous strip materials on wrinkle recognition. In some embodiments, motion compensation employs an optical flow-based inter-frame alignment method, which estimates pixel-level motion vectors by calculating the optical flow field between adjacent frames and uses this to perform alignment correction on image frames. In other embodiments, motion compensation employs an encoder synchronization-based inter-frame alignment method, which acquires encoder signals synchronized with image acquisition, calculates the material displacement based on the encoder signals, and uses this to perform position compensation on image frames.

[0040] Region cropping is used to extract regions of interest (ROIs) in the process, reducing the computational burden and interference of irrelevant areas on subsequent processing. In some implementations, ROI extraction is performed on the image based on preset ROI parameters, retaining only the portion of the image containing the wrinkle region to be detected. ROI extraction can be dynamically adjusted according to parameters such as material width, edge position, and process segment type to adapt to different production conditions.

[0041] After obtaining standardized image frames through preprocessing, method 100 proceeds to the wrinkle recognition stage. The goal of the wrinkle recognition stage is to accurately extract wrinkle regions, categories, and confidence information from the standardized image frames, providing reliable input data for subsequent indicator construction and trend prediction.

[0042] Before wrinkle recognition, directional filtering, frequency domain energy analysis, or texture enhancement can be performed on standardized image frames to obtain candidate wrinkle response maps. These candidate wrinkle response maps highlight wrinkle texture features and serve as auxiliary input to the wrinkle recognition model, improving robustness and computational efficiency. In some implementations, directional filtering employs Gabor filter banks or directional derivative filters, filtering the standardized image frame in multiple directions to extract wrinkle texture responses with specific directional features. In some implementations, frequency domain energy analysis calculates the energy distribution in a specific frequency band by performing Fourier transform or wavelet transform on the standardized image frame, thereby detecting wrinkle textures with periodic features. In some implementations, texture enhancement uses local contrast enhancement or morphological filtering methods to enhance the texture difference between wrinkle regions and background regions, making wrinkle features more prominent. In some implementations, the step of generating candidate wrinkle response maps can be omitted, and the standardized image frame can be directly input into the wrinkle recognition model.

[0043] The wrinkle recognition model can be a detection network, a segmentation network, a joint detection and classification network, or a joint segmentation and classification network. In some implementations, the wrinkle recognition model employs a detection network architecture, outputting bounding boxes of wrinkled regions along with their corresponding categories and confidence scores using object detection algorithms. In some implementations, the wrinkle recognition model employs a segmentation network architecture, outputting segmentation masks of wrinkled regions along with their corresponding categories and confidence scores using semantic segmentation or instance segmentation algorithms. In some implementations, the wrinkle recognition model employs a joint detection and classification network architecture, performing fine-grained classification of wrinkle categories while detecting wrinkled regions. In some implementations, the wrinkle recognition model employs a joint segmentation and classification network architecture, determining the category of each segmented region while outputting the segmentation mask. The wrinkle region information can be a segmentation mask, bounding boxes, or a set of contours, the specific form depending on the architecture of the recognition model used.

[0044] To improve the reliability of wrinkle recognition results, the confidence level output by the wrinkle recognition model can be calibrated. Confidence level calibration improves the consistency between the confidence level and the true accuracy, ensuring that the confidence level accurately reflects the reliability of the recognition results. In some implementations, confidence level calibration uses a temperature scaling method, scaling the model's output logits by learning temperature parameters on the validation set, making the calibrated confidence level distribution more closely match the true accuracy distribution. In other implementations, confidence level calibration uses an isometric regression method, dividing the confidence level interval into multiple bins and calibrating within each bin, ensuring that the prediction accuracy across different confidence level intervals remains consistent with the confidence level value.

[0045] To reduce the false alarm rate, pseudo-wrinkle exclusion classes can be established for easily confused targets such as seams, particles, scratches, printing textures, and reflective strips. These pseudo-wrinkle exclusion classes are used to distinguish these easily confused targets from genuine wrinkles during wrinkle recognition, avoiding misclassification of non-wrinkle defects as wrinkles. In some implementations, the pseudo-wrinkle exclusion classes are trained by labeling easily confused target samples in the training data and training them as independent categories, enabling the wrinkle recognition model to learn the feature differences that distinguish between genuine wrinkles and pseudo-wrinkles.

[0046] To further improve the stability of the recognition results, a temporal consistency determination can be introduced. This determination analyzes the recognition results across multiple consecutive frames, confirming the final wrinkle result only when these frames meet the criteria of consistent wrinkle morphology, continuous position, and a confidence level exceeding a threshold. In some implementations, the temporal consistency determination calculates the overlap, morphological similarity, and positional offset of wrinkle regions between adjacent frames to determine whether the wrinkles are continuous in the temporal dimension. This temporal consistency determination effectively suppresses false detections caused by transient noise, lighting flicker, or occasional interference, improving the temporal stability of the wrinkle recognition results.

[0047] Based on confidence calibration and anomaly removal, reliable identification results can be obtained. Reliable identification results refer to wrinkle identification results confirmed after confidence calibration, false wrinkle elimination, and temporal consistency judgment. In subsequent indicator construction steps, only indicators corresponding to reliable identification results are fed into subsequent steps to ensure the accuracy and reliability of the wrinkle indicator vector.

[0048] After obtaining reliable identification results, method 100 converts the wrinkle region information, wrinkle category information, and corresponding confidence levels into quantifiable wrinkle index vectors. The construction of these wrinkle index vectors provides a numerical input basis for subsequent time-series modeling and trend prediction. (Refer to...) Figure 2 , Figure 2 A detailed flowchart of the wrinkle index construction and trend prediction method 200 is shown. Method 200 includes steps 202, 204, 206, 208, and 210.

[0049] As mentioned earlier, the wrinkle index vector includes at least one of the following: severity index, orientation index, density index, area ratio index, and persistence index. The wrinkle index vector may also include a texture energy index, used to characterize the texture intensity in a specific orientation or frequency band. The texture energy index is obtained by calculating the cumulative energy value of the wrinkled region in a specific orientation filter response or in a specific frequency band, and can reflect the salience of the wrinkled texture.

[0050] Continue to refer to Figure 2 In step 202, multiple wrinkle features are fused into a unified wrinkle index. The unified wrinkle index is calculated using the following formula: ; in Indicators representing the percentage of area Indicates contrast or grayscale difference characteristics. Represents the texture energy index. Indicates a persistent indicator. , , , This is a weighting coefficient. Area percentage indicator. Used to characterize the proportion of wrinkled areas within the target area. Contrast or grayscale difference features. Used to characterize the degree of grayscale difference between wrinkled regions and the surrounding background regions. Texture energy index Used to characterize the texture intensity of wrinkled regions in a specific direction or frequency band. Persistence index Used to characterize the persistence of wrinkles in consecutive frames.

[0051] like Figure 2 As shown, in step 204, the weighting coefficients , , , The system adaptively adjusts based on the operating condition state vector. This vector includes one or more of the following: velocity, position, temperature, material batch, equipment status, and tension status. Under different operating conditions, the contribution of each wrinkling feature to the wrinkling severity varies. In some implementations, when the material velocity is high, the persistence index... weight This can be increased to better reflect the temporal stability of wrinkles under high-speed motion conditions. In some implementations, when the material surface texture is complex, the texture energy index... weight This can be reduced to minimize the interference of background texture on the uniform wrinkle index. In some implementations, the weighting coefficients are obtained by looking up a table based on the discretized index of the operating condition vector. In other implementations, the weighting coefficients are calculated using a regression model based on the continuous values ​​of the operating condition vector.

[0052] By fusing multiple wrinkle features into a unified wrinkle index The identification results can be incorporated into subsequent prediction and optimization processes in a continuous numerical form, avoiding insufficient control information caused by relying solely on discrete category labels. (Unified Wrinkle Index) It comprehensively reflects the overall state of the wrinkles at the current moment, providing a unified quantization input for subsequent filtering and time series modeling.

[0053] In obtaining a uniform wrinkle index Subsequently, Method 200 filters the unified wrinkle index to eliminate the effects of measurement noise and instantaneous fluctuations, and predicts the trend of wrinkle index changes at multiple future times through time series modeling.

[0054] Continue to refer to Figure 2 In step 206, a Kalman filter is applied to the unified wrinkle index to obtain a smoothed state estimate. Kalman filtering is a recursive state estimation method that estimates the true state of the system by combining the system dynamic model and observation data under the minimum mean square error criterion. In the application scenario of wrinkle index filtering, Kalman filtering calculates the unified wrinkle index at the current moment. As an observation, combined with the temporal dynamics of the wrinkle index, a smoothed state estimate is output. Smoothed state estimation It can effectively suppress the wrinkle index jump caused by image acquisition noise, fluctuations in recognition model output, or transient interference, making the wrinkle index sequence more stable and providing more reliable input data for subsequent time series prediction.

[0055] In some implementations, the Kalman filter employs a linear state-space model to model the dynamic changes of the wrinkle index as a first- or second-order autoregressive process. In other implementations, the Kalman filter uses extended Kalman filtering or unscented Kalman filtering to handle the nonlinear characteristics of the dynamic changes in the wrinkle index. In still other implementations, the process noise covariance and observation noise covariance of the Kalman filter are set based on the statistical characteristics of historical data, or adaptively adjusted based on the operating condition state vector.

[0056] In step 208, a time series model is used to predict the wrinkling index for the next K time points. The time series model can be ARIMA, LSTM, TCN, or Transformer. The ARIMA model is a classic time series forecasting method that models time series data using three components: autoregression, differencing, and moving average. The LSTM model is a Long Short-Term Memory network that captures long-term dependencies in time series data through a gating mechanism. The TCN model is a temporal convolutional network that models time series data through causal convolution and dilated convolution structures, and is characterized by high parallel computation efficiency. The Transformer model models global dependencies in time series data through a self-attention mechanism, enabling it to capture long-distance temporal correlations.

[0057] In some implementations, the time series model uses a smoothed state estimation sequence. As input, L represents the historical window length. In some implementations, the time series model also uses the operating condition vector as an auxiliary input to improve prediction accuracy. In some implementations, the time series model employs a multi-task learning architecture, simultaneously outputting the wrinkle index prediction and prediction uncertainty.

[0058] In step 210, the future wrinkle index prediction sequence is output. The future wrinkle index prediction sequence is represented as follows: Where K represents the prediction time domain length. The future wrinkle index prediction sequence reflects the expected change trend of the wrinkle index over the next K time points. In some implementations, the prediction time domain length K is set based on production line speed, adjustment response time, and process requirements. In some implementations, the prediction time domain length K can be dynamically adjusted based on the current operating condition state vector.

[0059] In some implementations, the time series model also outputs a trend label and prediction uncertainty. The trend label, including rising, stable, and falling values, characterizes the overall direction of future wrinkle index changes. The trend label can be obtained by linear fitting or difference analysis of the predicted future wrinkle index sequence. Prediction uncertainty characterizes the confidence level of the prediction result and can be estimated using Bayesian neural networks, Monte Carlo dropout, or ensemble learning methods. Higher prediction uncertainty indicates lower reliability of the prediction result, allowing for more conservative strategies in subsequent risk assessment and parameter optimization.

[0060] After obtaining the future fold index prediction sequence and prediction uncertainty, the prediction results need to be transformed into risk assessments and constraints that can be used for parameter optimization. (Refer to...) Figure 3 , Figure 3The diagram shows the hierarchical structure of Risk Assessment and Security Constraints 300. Risk Assessment and Security Constraints 300 comprises two main categories: Future Risk Indicators 302 and a Set of Security Constraints 312.

[0061] Future risk indicator 302 integrates at least one of the following factors: risk of continued deterioration of wrinkles 304, equipment disturbance risk 306, quality sensitivity window risk 308, and model uncertainty risk 310. Future risk indicator 302 is used to quantify various risks that may be faced after executing a specific parameter package, providing a risk assessment basis for subsequent parameter optimization.

[0062] Continue to refer to Figure 3 The risk of continued deterioration of wrinkles 304 is used to characterize the risk that the wrinkle index will remain at a high level or continue to rise after the parameter package is executed. The risk of continued deterioration of wrinkles 304 is calculated based on a future wrinkle index prediction sequence. When the prediction sequence shows that the wrinkle index is trending upwards or remains at a high level, the value of the risk of continued deterioration of wrinkles 304 increases accordingly. In some embodiments, the risk of continued deterioration of wrinkles 304 is obtained by calculating the deviation between the mean, peak, or cumulative value of the future wrinkle index prediction sequence and a preset threshold. In some embodiments, the risk of continued deterioration of wrinkles 304 also considers the rate of change of the wrinkle index; when the wrinkle index rises rapidly, the value of the risk of continued deterioration of wrinkles 304 increases further.

[0063] Equipment disturbance risk 306 characterizes the risk that parameter adjustments may cause resonance, overshoot, band breakage, or instability. Equipment disturbance risk 306 is related to adjustment parameters such as frequency, amplitude, and phase changes in the parameter package. In some embodiments, equipment disturbance risk 306 assesses whether the adjustment parameters in the parameter package are close to the sensitive area of ​​the equipment by consulting the equipment characteristic curve or resonance frequency distribution diagram. In some embodiments, equipment disturbance risk 306 also considers the parameter change rate; when the parameter changes too rapidly, the value of equipment disturbance risk 306 increases accordingly.

[0064] like Figure 3 As shown, the quality-sensitive window risk 308 is used to characterize the risk that adjustments made in a specific process segment, splicing segment, inspection-sensitive segment, or critical processing area may cause additional quality fluctuations. The quality-sensitive window risk 308 is related to the position information and process segment type in the current operating condition state vector. In some embodiments, the quality-sensitive window risk 308 is evaluated by determining whether the current position is within a preset quality-sensitive window; when within the quality-sensitive window, the value of the quality-sensitive window risk 308 increases significantly. In some embodiments, the quality-sensitive window risk 308 also considers the magnitude of the adjustment parameter; when a large adjustment is performed within the quality-sensitive window, the value of the quality-sensitive window risk 308 increases further.

[0065] Model uncertainty risk 310 characterizes the risk of reducing the priority of aggressive adjustments when prediction uncertainty is too high. Model uncertainty risk 310 is directly related to the prediction uncertainty of the trend forecast output. High prediction uncertainty indicates low reliability of the prediction results, and aggressive parameter adjustments may lead to unforeseen consequences. In some implementations, model uncertainty risk 310 is obtained by comparing the prediction uncertainty with a preset threshold; when the prediction uncertainty exceeds the threshold, the value of model uncertainty risk 310 increases accordingly. In some implementations, model uncertainty risk 310 also considers the trend of prediction uncertainty; when prediction uncertainty continues to rise, the value of model uncertainty risk 310 further increases.

[0066] Continue to refer to Figure 3 The set of safety constraints 312 is used to define the feasible region boundary for parameter optimization, ensuring that parameter optimization is performed within a safe range. The set of safety constraints 312 includes at least one of the following constraints: quality sensitive window constraint 314, safety boundary constraint 316, uncertainty constraint 318, and equipment state constraint 320.

[0067] The quality-sensitive window constraint 314 is used to prohibit the execution of certain parameter combinations or only allow the execution of degradation strategies within a preset sensitive window. The quality-sensitive window constraint 314 determines whether it is within the quality-sensitive window based on the position information in the current operating condition state vector, and accordingly limits the range of selectable parameter combinations. In some embodiments, the quality-sensitive window constraint 314 prohibits high-amplitude parameter adjustments within splicing sections, detection-sensitive sections, or critical processing areas. In some embodiments, the quality-sensitive window constraint 314 only allows the execution of degradation strategies within the quality-sensitive window; degradation strategies refer to parameter combinations with smaller adjustment amplitudes and lower risks.

[0068] like Figure 3As shown, safety boundary constraint 316 is used to limit frequency, amplitude, rate of change, duration, or energy budget. Safety boundary constraint 316 defines the allowable range of each adjustment parameter, ensuring that the parameter optimization results do not exceed the safe operating boundaries of the equipment. In some embodiments, safety boundary constraint 316 limits the upper and lower limits of the adjustment frequency to prevent the adjustment frequency from entering the equipment's resonant frequency band. In some embodiments, safety boundary constraint 316 limits the maximum value of the adjustment amplitude to prevent excessive adjustment amplitude from causing equipment overshoot or material damage. In some embodiments, safety boundary constraint 316 limits the maximum value of the parameter change rate to prevent excessively rapid parameter changes from causing unstable equipment response. In some embodiments, safety boundary constraint 316 limits the maximum value of the adjustment duration to prevent prolonged adjustment operations from causing continuous interference to the production process. In some embodiments, safety boundary constraint 316 limits the maximum value of the energy budget to prevent adjustment operations from consuming excessive energy resources. Safety boundary constraint 316 can also limit the prohibited frequency band and the maximum disturbance duration. The prohibited frequency band refers to the frequency range that should be avoided during equipment operation, and the maximum disturbance duration refers to the longest allowed duration of a single adjustment operation.

[0069] Uncertainty constraint 318 is used to allow only conservative strategies when the prediction uncertainty exceeds a threshold. Uncertainty constraint 318 determines the reliability of the current prediction result based on the prediction uncertainty of the trend prediction output and accordingly limits the range of selectable parameter combinations. In some embodiments, uncertainty constraint 318 sets a prediction uncertainty threshold; when the prediction uncertainty exceeds the threshold, only conservative strategies with smaller adjustment ranges and lower risks are allowed. In some embodiments, uncertainty constraint 318 dynamically adjusts the range of selectable parameter combinations based on the value of the prediction uncertainty; the higher the prediction uncertainty, the narrower the range of selectable parameter combinations.

[0070] Equipment status constraint 320 is used to restrict high-risk parameter packages when the equipment is in an abnormal, derating, or maintenance state. Equipment status constraint 320 determines the current operating state of the equipment based on equipment status information in the operating condition state vector and accordingly restricts the range of selectable parameter combinations. In some embodiments, equipment status constraint 320 prohibits any parameter adjustment operations when the equipment is in an abnormal state. In some embodiments, equipment status constraint 320 allows only low-risk parameter combinations when the equipment is in a derating state. In some embodiments, equipment status constraint 320 restricts the use of some adjustment functions when the equipment is in a maintenance state. In some embodiments, equipment status constraint 320 also considers the equipment's stability index; when the equipment stability is low, it restricts the selection of high-risk parameter packages.

[0071] After completing the risk assessment and constructing the set of security constraints, it is necessary to solve for the optimal parameter set within the set of security constraints that balances the benefits of wrinkle improvement, execution risk, and execution cost. (Refer to...) Figure 4 , Figure 4 A flowchart of a safety constraint parameter optimization method 400 is shown. Method 400 includes steps 402, 404, 406, 408, 410, and 412.

[0072] In step 402, the wrinkle improvement benefit is calculated. Wrinkle improvement benefits Used to characterize the execution parameter package The extent to which the wrinkle index decreases in the future. Benefits of wrinkle improvement. Calculated using the following formula: ; in The uniform wrinkle index at the current moment, To execute parameter package The expected wrinkle index after the event. Indicates the first The expected improvement in the amount of wrinkle characteristics. For corresponding weights. Unified wrinkle index. As mentioned earlier, this is obtained by weighted fusion of area ratio, contrast or grayscale difference features, texture energy, and persistence metrics. Expected improvement amount. Based on historical execution data and predictive model estimation, it reflects the expected change in each wrinkle feature after executing a specific parameter package.

[0073] Continue to refer to Figure 4 In step 404, the execution risk is calculated. Execution risk Used for quantization execution parameter package The combined losses from various potential risk events. Execution risk. Calculated using the following formula: ; in Indicates the execution parameter package Time The probability of occurrence of risk events. This represents the quantified loss value of the risk event. Risk events include resonance risk, band breakage risk, quality-sensitive section disturbance risk, and equipment overshoot risk. Resonance risk refers to the mechanical resonance that may occur when the frequency of the adjusted parameter approaches the equipment's natural frequency. Band breakage risk refers to the material fracture that may occur when the adjustment amplitude is too large or the change is too rapid. Quality-sensitive section disturbance risk refers to the product quality fluctuation that may be caused when adjustments are made within the quality-sensitive window. Equipment overshoot risk refers to the risk that the equipment response will exceed the expected range when the parameter changes too rapidly. In some implementations, the probability of occurrence of the risk event is... This is obtained based on statistical analysis of historical execution data. In some implementation methods, the probability of occurrence of risk events is... Based on the parameter package using the risk assessment model It is obtained by calculating the current operating condition state vector.

[0074] like Figure 4 As shown, in step 406, the execution cost is calculated. Execution costs Used for quantization execution parameter package The resource costs required. Execution costs. Calculated using the following formula: ; in Indicates the tension adjustment range. Indicates the speed adjustment range. Indicates the number of adjustments. Indicates the execution time. , , , This corresponds to the cost coefficient. Tension adjustment range. Reflects the parameter package The change in tension setting relative to the current value. Speed ​​adjustment range. Reflects the parameter package The change in centerline velocity correction relative to the current value. Number of adjustments. Reflects the execution parameter package Number of parameter adjustment operations required. Execution time. Reflects the completion of the parameter package Execution time required. Cost factor. , , , The settings are determined based on production process requirements and equipment characteristics.

[0075] Continue to refer to Figure 4 In step 408, the objective function is constructed. Objective function Taking into account the benefits of wrinkle improvement, implementation risks, and implementation costs, this method is used to evaluate the overall merits of candidate parameter packages. Objective function Constructed using the following formula: ; in Indicates the benefits of wrinkle improvement. Indicates execution risk, Indicates execution cost, and These are the weighting coefficients. Weighting coefficients are used to adjust the relative weight of execution risk in the objective function. Used to adjust the relative weight of execution cost in the objective function. In some implementations, the weighting coefficients... and The weighting coefficients are set based on production process requirements and risk appetite. In some implementations, the weighting coefficients... and Adaptive adjustment is made based on the current operating condition state vector.

[0076] In step 410, the optimizer solves the objective function within the set of safety constraints. The optimal parameter package is maximized. As mentioned earlier, the set of safety constraints includes at least one of the following: quality-sensitive window constraints, safety boundary constraints, uncertainty constraints, and device state constraints. The optimizer can be a Bayesian optimizer, a safety reinforcement learning policy network, or a hybrid rule-and-learning optimizer. A Bayesian optimizer finds the optimal parameter package within a finite number of evaluations by constructing a surrogate model of the objective function and using a collection function to guide parameter search. A safety reinforcement learning policy network learns parameter selection strategies under safety constraints, maximizing cumulative gains while ensuring safety. A hybrid rule-and-learning optimizer combines preset rules and a learning algorithm, using the learning algorithm to optimize parameters under rule constraints.

[0077] like Figure 4 As shown, when employing an online learning strategy, the search, exploration, and updating of all candidate parameters are confined within the set of safety constraints, and undergo secondary verification via a safety filter before execution. The safety filter performs a compliance check on the parameter package before output, ensuring that the parameter package satisfies all constraints in the set of safety constraints. This secondary verification by the safety filter avoids parameter out-of-bounds issues caused by optimizer output errors or changes in boundary conditions, ensuring that online exploration does not exceed the safety boundaries.

[0078] In step 412, the optimal parameter package is mapped to executable control command output. The optimal parameter package is converted into actual executable control commands through the device parameter mapping relationship. The device parameter mapping relationship establishes a correspondence between algorithm parameters and execution objects, which may include one or more of the following: tension setpoint, roller system adjustment amount, oscillation mechanism parameters, linear speed correction amount, execution window control amount, and threshold switching signal. Through the device parameter mapping relationship, the parameter package output by method 400 can directly act on the actual actuator, rather than being an abstract value detached from the device.

[0079] Before obtaining the optimal parameter package and mapping it to executable control instructions, the specific configuration content of the parameter package needs to be defined and organized. The parameter package contains configuration information in multiple dimensions to guide subsequent parameter execution and policy output.

[0080] The parameter package may include a start-stop window strategy. The start-stop window strategy defines the start conditions, stop conditions, and timing constraints for parameter adjustment. In some implementations, the start-stop window strategy employs a dual-threshold hysteresis mechanism, setting two different thresholds—a start threshold and a stop threshold—to avoid frequent start-stops caused by fluctuations in the wrinkle index around the threshold. Parameter adjustment is triggered when the wrinkle index exceeds the start threshold and ends when the wrinkle index falls below the stop threshold. The difference between the start threshold and the stop threshold forms a hysteresis interval, giving parameter adjustment stable start-stop characteristics. In some implementations, the start-stop window strategy includes a minimum hold time constraint, where the minimum hold time refers to the minimum duration after parameter adjustment begins, used to avoid insufficient effects due to excessively short adjustment times. In some implementations, the start-stop window strategy includes a cooling time constraint, where the cooling time is the minimum interval between two parameter adjustments, used to avoid the cumulative effects on equipment and materials caused by continuous and frequent adjustments.

[0081] The parameter package may include a set of adjustment parameters. The set of adjustment parameters defines the specific values ​​and configurations for parameter adjustments. In some embodiments, the set of adjustment parameters includes a frequency parameter, which specifies the operating frequency of the oscillation mechanism or periodic adjustment action. In some embodiments, the set of adjustment parameters includes an amplitude parameter, which specifies the magnitude of change in tension adjustment, speed correction, or other adjustment actions. In some embodiments, the set of adjustment parameters includes a duration parameter, which specifies the execution duration of a single adjustment action. In some embodiments, the set of adjustment parameters includes a phase parameter, which specifies the initial phase or phase offset of the periodic adjustment action. In some embodiments, the set of adjustment parameters includes a scan step size parameter, which specifies the step increment during parameter search or progressive adjustment. In some embodiments, the set of adjustment parameters includes a threshold strategy parameter, which specifies the threshold setting method for trigger conditions, decision conditions, or switching conditions.

[0082] The parameter package may include a prohibited set. The prohibited set defines parameter combinations or ranges that are not permitted under the current operating conditions. In some implementations, the prohibited set includes a prohibited frequency band, which refers to a frequency range that should be avoided during equipment operation, as adjusting the frequency within the prohibited band may cause resonance or other instability phenomena. In some implementations, the prohibited set includes a prohibited window, which refers to an interval within a specific time period or location segment where parameter adjustments are not permitted. The prohibited window may correspond to a quality-sensitive window, a splicing segment, or a detection-sensitive segment. In some implementations, the prohibited set includes a prohibited combination, which refers to a combination of specific parameter values ​​that are not allowed to be used simultaneously under the current operating conditions. The prohibited combination may be determined based on failure cases in historical execution data or equipment characteristic constraints.

[0083] The parameter package may include an execution mode flag. The execution mode flag specifies the execution method and human-machine interaction mode of the parameter package. In some implementations, the execution mode flag is set to automatic execution mode. In automatic execution mode, the host execution object automatically issues parameters according to control instructions without manual intervention. In some implementations, the execution mode flag is set to manual confirmation mode. In manual confirmation mode, the operator interface displays recommended parameters, expected effects, and risk margins, and parameter adjustments are performed only after operator confirmation. In some implementations, the execution mode flag is set to reminder mode only. In reminder mode, the recommended parameter results are only displayed to the operator as a prompt and do not trigger any automatic execution actions. In some implementations, the execution mode flag is set to prohibited execution mode. In prohibited execution mode, no parameter adjustment operations are allowed under the current operating conditions, and the parameter package is only saved as a record without outputting execution instructions. The execution mode flag can be automatically set or manually specified based on the current risk level, prediction uncertainty, equipment status, and process requirements.

[0084] During parameter optimization, strategy explanation information can also be output. This information provides operators or higher-level control logic with the decision-making basis and background information for parameter recommendations, improving the interpretability of the recommendations. In some implementations, the strategy explanation information includes the reasons for recommending the parameter package, which can be described as selecting the package based on a comprehensive evaluation of the current wrinkle index level, trend prediction results, operating conditions, and historical performance. In some implementations, the information includes the expected wrinkle improvement, which is based on the calculation of wrinkle improvement benefits and expressed numerically or as a percentage of the expected decrease in the wrinkle index after implementing the parameter package. In some implementations, the information includes the current risk margin, which represents the distance between the execution risk of the current parameter package and the safety boundary. A larger risk margin indicates that the current parameter package is further from the safety boundary, resulting in higher execution safety. In some implementations, the information includes the reasons for excluding parameter packages, which can be described as specific parameter packages not being selected due to violations of safety constraints, excessive risk, insufficient improvement benefits, or excessive costs. By providing policy explanations, operators can understand the decision-making logic behind parameter recommendations and make more reasonable judgments in manual confirmation mode.

[0085] After completing the configuration definition of the parameter package, the algorithm parameters in the parameter package need to be converted into control instructions that the actual execution object can recognize and execute. There is a mapping relationship between the algorithm parameters and the execution object, which converts the abstract parameter values ​​output by parameter optimization into control variables acceptable to the specific device.

[0086] The execution objects may include one or more of the following: tension setpoint, roller system adjustment, oscillation mechanism parameters, linear speed correction, execution window control, and threshold switching signal. The tension setpoint specifies the target tension value for the continuous strip material during transport; adjusting the tension setpoint affects the flattening state and wrinkle formation trend of the material. The roller system adjustment specifies the adjustment amount for the roller system position, angle, or pressure; changes in the roller system adjustment can alter the force distribution and motion trajectory of the material. The oscillation mechanism parameters specify the operating frequency, amplitude, and phase configuration of the oscillation mechanism; adjusting these parameters can improve the flattening effect of the material through periodic perturbations. The linear speed correction specifies the correction value of the production line speed relative to the reference speed; adjusting the linear speed correction affects the tension distribution and wrinkle formation conditions of the material. The execution window control specifies the boundary parameters of the time window or position window for parameter adjustment; setting the execution window control controls the effective range of parameter adjustment. The threshold switching signal triggers the switching of threshold parameters in the control logic; the threshold switching signal allows the control strategy to use different judgment thresholds under different operating conditions.

[0087] Mapping relationships can be established through lookup tables, linear transformations, or nonlinear transformations. In some implementations, the mapping relationship is established using a pre-defined parameter lookup table, which records the correspondence between algorithm parameter values ​​and control variables of the execution object. The corresponding control variable values ​​can be directly obtained through a lookup operation. In some implementations, the mapping relationship is established using a linear transformation formula, which converts the algorithm parameter values ​​into control variable values ​​of the execution object through scaling and offset operations. In some implementations, the mapping relationship is established using a nonlinear transformation function, which converts the algorithm parameter values ​​into control variable values ​​of the execution object based on equipment characteristic curves or process response curves. By establishing mapping relationships, the parameter package output by parameter optimization can directly affect the actual execution object, rather than remaining at the abstract numerical level.

[0088] After mapping the parameter package to executable control instructions, different execution methods are adopted according to the setting of the execution mode flag. In automatic execution mode, the upper-level execution logic automatically completes the parameter distribution according to the control instructions. The control instructions are transmitted to the execution object through the communication interface, and the execution object automatically adjusts the operating parameters according to the received control variable values. Automatic execution mode is suitable for scenarios with low risk level, low prediction uncertainty and stable operating conditions. In automatic execution mode, the parameter adjustment process does not require manual intervention, and can achieve rapid response and continuous adjustment.

[0089] In manual confirmation mode, the operator interface displays recommended parameters, expected effects, and risk margins. Parameter adjustments are executed only after operator confirmation. Recommended parameters are displayed numerically or graphically as the specific settings for each adjustment parameter in the parameter package. Expected effects show the expected decrease and trend of the wrinkle index after executing the parameter package. Risk margins show the distance between the execution risk of the current parameter package and the safety boundary, enabling the operator to assess execution safety. After reviewing the above information, the operator can choose to confirm execution, modify parameters and then execute, or cancel execution. When the operator confirms execution, control instructions are issued to the execution object; when the operator modifies parameters and then executes, the modified parameters are verified by a safety filter and then issued to the execution object; when the operator cancels execution, the parameter package is not issued, and the current adjustment cycle ends. Manual confirmation mode is suitable for scenarios with medium risk levels, requiring manual judgment, or where the operating condition is uncertain.

[0090] In high-risk or high-uncertainty modes, only alert messages, degradation strategies, or rollback strategies are output; routine parameter adjustments are not performed. Alert messages inform operators of the current wrinkle status, risk level, and recommended measures. These messages can be output via operator interface, audible and visual alarms, or push notifications. Degradation strategies refer to parameter combinations with smaller adjustment ranges and lower risk, providing limited wrinkle improvement while ensuring safety. Rollback strategies restore parameters to their previous stable or default state, quickly restoring safe operation when current parameter configurations cause anomalies. In some implementations, when the execution risk exceeds a preset high-risk threshold, the system automatically switches to high-risk mode and outputs alert messages. In some implementations, when the prediction uncertainty exceeds a preset high-uncertainty threshold, the system automatically switches to high-uncertainty mode and only allows degradation strategies. In some implementations, when the wrinkle index fails to decrease as expected or exhibits abnormal fluctuations after parameter execution, a rollback strategy is automatically triggered to restore the parameters to their previous stable state. By limiting the execution range in high-risk or high-uncertainty modes, the potential risks of aggressive adjustments under uncertain conditions can be avoided.

[0091] This invention also discloses a visual recognition and trend prediction system for wrinkles in continuous strip materials.

[0092] Reference Figure 5 , Figure 5 A structural block diagram of a visual recognition and trend prediction system 500 is shown. System 500 is used for online visual recognition, short-term trend prediction, and safety self-optimization parameter recommendation of surface wrinkles and defects in continuous strip materials. System 500 includes an image acquisition module 502, a working condition acquisition module 504, a preprocessing module 506, a wrinkle recognition module 508, an index construction module 510, a trend prediction module 512, and a parameter optimization module 514.

[0093] The image acquisition module 502 is used to acquire a continuous sequence of images of the surface of a continuous strip material. The image acquisition module 502 can employ image acquisition devices such as line scan cameras, area scan cameras, or high-speed cameras, selecting the appropriate acquisition device type and configuration parameters based on the production line speed and detection accuracy requirements. In some embodiments, the image acquisition module 502 is also used to acquire synchronous encoder signals, position signals, and operating condition data. The synchronous encoder signal is used to synchronize with the image acquisition time, enabling the acquired image frames to establish a correspondence with the actual position of the material. The position signal is used to identify the specific location of the current acquisition position on the production line, facilitating subsequent quality sensitivity window determination and operating condition correlation analysis. Operating condition data may include real-time measured values ​​such as speed, temperature, and tension; the image acquisition module 502 can acquire this operating condition data through a data interface with the production line control system.

[0094] The operating condition acquisition module 504 is used to synchronously acquire operating condition state vectors. As mentioned earlier, the operating condition state vectors include one or more of the following: speed, position, temperature, material batch, equipment status, and tension status. The operating condition acquisition module 504 can acquire the above operating condition information through a sensor interface, a PLC communication interface, or a database query interface. The operating condition acquisition module 504 maintains time synchronization with the image acquisition module 502, aligning the operating condition state vectors with the corresponding image frames in the time dimension. In some embodiments, the operating condition acquisition module 504 can also acquire historical execution states and historical adjustment records for subsequent operating condition map retrieval and parameter hot-start.

[0095] The preprocessing module 506 is used to preprocess image frames in a continuous image sequence to obtain standardized image frames. As mentioned above, preprocessing includes at least one of illumination normalization, reflection suppression, and denoising. The preprocessing module 506 receives image frames output by the image acquisition module 502, performs processing operations such as illumination normalization, reflection suppression, denoising, detail enhancement, motion compensation, and region cropping on the image frames, and outputs standardized image frames with stable image quality. The preprocessing module 506 can use GPU acceleration or FPGA acceleration to implement real-time image processing to meet the online inspection requirements of high-speed production lines.

[0096] The wrinkle recognition module 508 is used to input standardized image frames into the wrinkle recognition model and output wrinkle region information, wrinkle category information, and corresponding confidence scores. As mentioned above, wrinkle category information includes one or more of edge wrinkles, transverse wrinkles, periodic wrinkles, and pseudo-wrinkles. The wrinkle recognition module 508 receives the standardized image frames output by the preprocessing module 506, analyzes and processes the standardized image frames through the wrinkle recognition model, and outputs wrinkle region information, wrinkle category information, and corresponding confidence scores. The wrinkle region information can be a segmentation mask, bounding box, or contour set. The wrinkle recognition module 508 can also perform confidence calibration, pseudo-wrinkle elimination, and temporal consistency determination on the recognition results, and output reliable recognition results.

[0097] The index construction module 510 is used to construct a wrinkle index vector for the current moment based on wrinkle region information, wrinkle category information, and corresponding confidence levels. As mentioned above, the wrinkle index vector includes at least one of severity index, direction index, density index, area proportion index, and persistence index. The index construction module 510 receives the recognition results output by the wrinkle recognition module 508, calculates the values ​​of each wrinkle index based on the recognition results, and can fuse multiple wrinkle features into a unified wrinkle index. The index construction module 510 can also adaptively adjust the weighting coefficients according to the working condition state vector provided by the working condition acquisition module 504, so that the unified wrinkle index can adapt to the wrinkle assessment needs under different working conditions.

[0098] The trend prediction module 512 is used to filter and perform time-series modeling on the wrinkle index vectors of the current and historical moments, outputting predicted wrinkle index values, trend labels, and prediction uncertainties for multiple future moments. As mentioned earlier, the trend label includes one of rising, stable, and falling trends. The trend prediction module 512 receives the wrinkle index vector or unified wrinkle index output by the index construction module 510, performs Kalman filtering on the wrinkle index vector to obtain a smoothed state estimate, and uses a time-series model to predict the wrinkle index for multiple future moments. The trend prediction module 512 outputs the future wrinkle index prediction sequence, trend labels, and prediction uncertainties, providing a predictive basis for subsequent risk assessment and parameter optimization.

[0099] The parameter optimization module 514 is used to construct a set of future risk indicators and safety constraints based on the predicted value of the wrinkle index, the prediction uncertainty, and the operating condition state vector. It then solves for the optimal parameter package within the safety constraint set and maps the optimal parameter package into executable control commands. The parameter optimization module 514 receives the prediction results output by the trend prediction module 512 and the operating condition state vector provided by the operating condition acquisition module 504, constructs the set of future risk indicators and safety constraints, and uses an optimizer to solve for the optimal parameter package that maximizes the objective function within the safety constraint set. The parameter optimization module 514 converts the optimal parameter package into executable control commands through equipment parameter mapping relationships and outputs them to the upper control system, actuator, or operator parameter adjustment interface. In some embodiments, the parameter optimization module 514 also outputs strategy explanation information to make the parameter recommendation results interpretable.

[0100] like Figure 5 As shown, the solid arrows represent the main data stream, which is sequentially transmitted through the image acquisition module 502, preprocessing module 506, wrinkle recognition module 508, indicator construction module 510, trend prediction module 512, and parameter optimization module 514. The dashed arrows represent the data transmission of the working condition state vector from the working condition acquisition module 504 to the parameter optimization module 514. The working condition state vector is used by the parameter optimization module 514 for working condition correlation analysis when constructing the risk indicator and safety constraint set. In some embodiments, the working condition state vector can also be transmitted to the indicator construction module 510 for adaptive adjustment of weight coefficients, and to the trend prediction module 512 to improve prediction accuracy.

[0101] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus or device (such as a computer-based system, a system including a processing module or other system that can fetch and execute instructions from, an instruction execution system, apparatus or device).

[0102] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for visual recognition and trend prediction of wrinkles in continuous strip materials, characterized in that, include: A continuous image sequence of the surface of a continuous strip material is acquired, and a working condition state vector is acquired simultaneously. The working condition state vector includes one or more of the following: velocity, position, temperature, material batch, equipment status, and tension status. The image frames in the continuous image sequence are preprocessed to obtain standardized image frames. The preprocessing includes at least one of illumination normalization, reflection suppression, and noise reduction. The standardized image frame is input into the wrinkle recognition model, which outputs wrinkle region information, wrinkle category information and corresponding confidence level, wherein the wrinkle category information includes one or more of edge wrinkles, transverse wrinkles, periodic wrinkles and pseudo-wrinkles. Based on the wrinkle region information, wrinkle category information and corresponding confidence level, a wrinkle index vector is constructed for the current moment. The wrinkle index vector includes at least one of the following: severity index, direction index, density index, area ratio index, and persistence index. The wrinkle index vectors at the current and historical moments are filtered and time-series modeled to output the predicted values ​​of the wrinkle index, trend labels, and prediction uncertainties at multiple future moments, wherein the trend labels include one of rising, stable, and falling. as well as Based on the predicted value of the wrinkle index, the prediction uncertainty, and the working condition vector, a set of future risk indicators and safety constraints is constructed, and the optimal parameter package is solved within the set of safety constraints. The optimal parameter package is then mapped into an executable control command output.

2. The method according to claim 1, characterized in that, Constructing the wrinkle index vector at the current moment also includes: Multiple wrinkle features are fused into a unified wrinkle index, which is calculated using the following formula: ； in Indicators representing the percentage of area Indicates contrast or grayscale difference characteristics. Represents the texture energy index. Indicates a persistent indicator. , , , The weighting coefficient is adaptively adjusted according to the operating condition state vector.

3. The method according to claim 2, characterized in that, Filtering and time-series modeling of the wrinkle index vectors at current and historical moments include: Kalman filtering is applied to the unified wrinkle index to obtain a smoothed state estimate; and The wrinkle index at the next K time points is predicted using ARIMA, LSTM, TCN, or Transformer models to obtain the future wrinkle index prediction sequence.

4. The method according to claim 1, characterized in that, The future risk indicator incorporates at least one of the following factors: Risk of continued deterioration of wrinkles; Equipment disturbance risk; Quality-sensitive window risk; and Model uncertainty risk.

5. The method according to claim 4, characterized in that, The set of security constraints includes at least one of the following constraints: Quality-sensitive window constraints prohibit the execution of certain parameter combinations or only allow the execution of degradation strategies within a preset sensitive window; Safety boundary constraints limit frequency, amplitude, rate of change, duration, or energy budget; Uncertainty constraints allow only conservative strategies to be chosen when the prediction uncertainty exceeds a threshold; and Equipment status constraints restrict high-risk parameter packages when the equipment is in an abnormal, derated, or maintenance state.

6. The method according to claim 5, characterized in that, Solving for the optimal parameter package within the set of security constraints includes: Construct the objective function ,in Indicates the benefits of wrinkle improvement. Indicates execution risk, Indicates execution cost, and For weighting coefficients; and The optimal parameter set that maximizes the objective function is solved within the set of security constraints using a Bayesian optimizer, a secure reinforcement learning policy network, or a hybrid optimizer that combines rules and learning.

7. The method according to claim 6, characterized in that, The wrinkle improvement benefits Calculated using the following formula: ； in The uniform wrinkle index at the current moment, To execute parameter package The expected wrinkle index after the event. Indicates the first The expected improvement in the amount of wrinkle characteristics. For the corresponding weights.

8. The method according to claim 6, characterized in that, The execution risks Calculated using the following formula: ； in Indicates the execution parameter package Time The probability of occurrence of risk events. This indicates the quantified value of the loss from the risk event, which includes resonance risk, band breakage risk, quality-sensitive section disturbance risk, and / or equipment overshoot risk.

9. The method according to claim 6, characterized in that, The execution cost Calculated using the following formula: ； in Indicates the tension adjustment range. Indicates the speed adjustment range. Indicates the number of adjustments. Indicates the execution time. , , , This represents the corresponding cost coefficient.

10. A visual recognition and trend prediction system for wrinkles in continuous strip materials, characterized in that, include: The image acquisition module is used to acquire a continuous sequence of images of the surface of a continuous strip material. The working condition acquisition module is used to synchronously acquire the working condition status vector, which includes one or more of the following: speed, position, temperature, material batch, equipment status, and tension status. The preprocessing module is used to preprocess the image frames in the continuous image sequence to obtain standardized image frames. The preprocessing includes at least one of illumination normalization, reflection suppression, and noise reduction. The wrinkle recognition module is used to input the standardized image frame into the wrinkle recognition model and output wrinkle region information, wrinkle category information and corresponding confidence level, wherein the wrinkle category information includes one or more of edge wrinkles, transverse wrinkles, periodic wrinkles and pseudo wrinkles. The indicator construction module is used to construct a wrinkle indicator vector at the current moment based on the wrinkle region information, wrinkle category information and corresponding confidence level. The wrinkle indicator vector includes at least one of severity indicator, direction indicator, density indicator, area ratio indicator and persistence indicator. The trend prediction module is used to filter and time-series model the wrinkle index vectors at the current and historical moments, and output the predicted values ​​of the wrinkle index, trend labels, and prediction uncertainties at multiple future moments, wherein the trend labels include one of rising, stable, and falling. as well as The parameter optimization module is used to construct a set of future risk indicators and safety constraints based on the predicted value of the wrinkle index, the prediction uncertainty and the working condition vector, solve for the optimal parameter package within the set of safety constraints, and map the optimal parameter package into an executable control command output.

Images