An online deformation monitoring and grading early warning method and system in a hemp felt production process

Abstract

The application discloses a kind of hemp felt production process deformation online monitoring and grading early warning method and system, involve hemp felt quality monitoring technical field, the steps of this method include: through sensor array real-time acquisition hemp felt production process data, and read into factory batch attribute, record batch ID;According to hemp felt production process data, by laying-mixing uneven detection algorithm, obtain hemp felt local error score;Further obtain hemp felt deformation grade;Judge whether S4 is run;Establish defect prediction algorithm, the local error score of the hemp felt and batch information are as input, compensate batch and sensor error, output each grid defect probability and next process time step hemp felt unevenness;Build automated remix strategy, the local error score of the hemp felt and next process time step hemp felt unevenness are as input, output control parameter.The application solves the problems of parameter drift caused by equipment calibration and non-standardization, laying / net strip directionality uneven and uneven mixing.

Description

Technical Field

[0001] This invention relates to the field of hemp felt quality monitoring technology, specifically to a method and system for online monitoring and graded early warning of deformation during hemp felt production. Background Technology

[0002] Hemp felt, a functional material with excellent mechanical properties, thermal insulation, sound absorption, and eco-recyclability, is widely used in construction, home furnishings, industrial vibration isolation, and automotive interiors. As downstream applications demand higher standards of product consistency and reliability, the production process of hemp felt places greater emphasis on controlling thickness uniformity, mechanical isotropy, and appearance integrity. However, the diverse sources of raw materials, significant fluctuations in fiber properties, complex production processes, and the prevalence of customized (non-standard) equipment in hemp felt production make it difficult to achieve effective online monitoring and rapid response throughout the entire production process using traditional quality inspection methods that rely on manual inspections or single sensors.

[0003] First, fibers are prone to directional bias or localized enrichment during web formation, carding, and traction, leading to significant differences in the stiffness, elongation, and resilience of the sheet in the longitudinal and transverse directions. After setting or stretching, this manifests as deformation defects such as bending, edge curling, or corner curling. Existing detection methods mostly involve downstream tensile testing or batch sampling, which makes it difficult to locate and determine the cause of localized orientation anomalies in real time, and also fails to provide timely feedback to the web formation or carding mechanism for closed-loop correction. This results in a decrease in yield and an increase in rework costs.

[0004] Secondly, hemp felt production lines often consist of multiple customized machines, irregularly shaped rollers, and non-uniform control units. Different machines exhibit significant differences in response under the same process settings. Simultaneously, mechanical wear, changes in friction coefficients, and the coupling effect of temperature and humidity caused by equipment use lead to drift in process parameters over time. Traditional methods often rely on manual calibration based on experience or periodic shutdowns for calibration, which not only affects production capacity but also fails to detect and compensate for slow parameter drift in a timely manner. This results in frequent issues of large differences in products at different times under the same process settings, severely hindering mass production and stable operation.

[0005] Furthermore, natural fibers are characterized by wide length distribution, easy entanglement, varying specific gravity, and strong hygroscopicity. Factors such as mixer structure, blade design, feeding method, and silo flow patterns can all induce stratification or agglomeration, manifesting as uneven thickness and clumping on the production surface, and localized breakage or excessive shrinkage during subsequent stretching and shaping stages. Existing online detection technologies mostly focus on single physical quantities, making it difficult to simultaneously and sensitively capture multi-dimensional information such as density, moisture content, thickness, and mechanical vibration, thus making early warning and precise location of stratification formation mechanisms challenging.

[0006] Therefore, the present invention provides a method and system for online monitoring and graded early warning of deformation during the production of hemp felt. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for online monitoring and graded early warning of deformation during the production of hemp felt, so as to solve the existing problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a method for online monitoring and graded early warning of deformation during the production of hemp felt, comprising the following steps: S1. Collect data on the production process of hemp felt in real time through a sensor array, read the batch attribute B upon arrival, and record the batch ID; S2. Based on the data from the hemp felt production process, the local error score of the hemp felt is obtained through a lay-and-mix unevenness detection algorithm. S3. Based on the local error score of the hemp felt and the set local error threshold of the hemp felt, obtain the deformation level of the hemp felt; determine whether to run S4. S4. Establish a defect prediction algorithm, take the local error score of the felt and batch information as input, compensate for batch and sensor errors, and output the defect probability of each grid and the unevenness of the felt in the next process time step. S5. Construct an automated remixing strategy, taking the local error score of the felt and the unevenness of the felt in the next process time step as inputs, and output control parameters.

[0009] A further improvement of this invention lies in the specific implementation steps of the lay-up-mixing unevenness detection algorithm, which include: dividing the hemp felt production surface into several local grids, obtaining a local orientation histogram for each grid based on the structure tensor and Gabor filtering, and quantifying the orientation uniformity and its confidence level to obtain the lay-up unevenness detection score Uni; simultaneously constructing a comprehensive mixing unevenness index Umi through multi-source features, and performing normalization mapping and confidence assignment; obtaining confidence weights based on sensor SNR and orientation coherence, and using confidence-weighted linear fusion to obtain the local error score of the hemp felt, wherein the global weights... Adaptive updates based on historical correlation.

[0010] A further improvement of the present invention is that the uneven mesh detection score is obtained by measuring the mesh unevenness in each cell. Within the image, extract several local image patches and calculate the gradient of each local image patch. The structure tensor matrix J is obtained, and then the principal directions of each local image patch are plotted into a histogram. Using directional intensity confidence as weights, uniformity is represented by information entropy normalization, and the calculation formula is expressed as follows: And through maximum entropy Obtain the normalized net unevenness detection score .

[0011] A further improvement of this invention is that the multi-source features in the comprehensive mixing unevenness index Umi include local average density. Local density variance Localized clumping energy Coefficient of variation of thickness Texture roughness characteristics; Define a lattice set average density on , and standard deviation This allows us to obtain the local density parameters of the hemp felt. ; The density and area distribution of local extrema are statistically analyzed and normalized to obtain the local clumping energy. The local average thickness is obtained by short-time Fourier transform, and the peak value detected in the set frequency band is normalized and used as the vibrational spectrum energy. ; Obtain continuous thickness curves from sensor array The thickness variation coefficient is obtained by the ratio of the local average thickness to the local thickness standard deviation. This leads to the comprehensive mixing unevenness index Umi: ; This represents the weighting parameter.

[0012] A further improvement of the present invention is that step S3 further includes recording without alarming when the local error score of the hemp felt is less than the local error threshold of the hemp felt and the uneven web laying detection score Uni and the comprehensive uneven mixing index Umi are both less than the corresponding thresholds; When the local error score of the hemp felt is less than the local error threshold of the hemp felt, and the uneven web laying detection score Uni or the comprehensive uneven mixing index Umi is greater than or equal to the corresponding threshold, a yellow warning is activated and sent to S4. When the local error score of the hemp felt is greater than or equal to the local error threshold of the hemp felt, and the uneven web laying detection score Uni and the comprehensive uneven mixing index Umi are both less than the corresponding threshold, an orange warning is activated and sent to S4. When the local error score of the hemp felt is greater than or equal to the local error threshold of the hemp felt, and the uneven web laying detection score Uni or the comprehensive uneven mixing index Umi is greater than or equal to the corresponding threshold, a red alert is activated, the production line is stopped, and the batch is isolated.

[0013] A further improvement of the present invention is that the defect prediction algorithm includes processing sensor measurements. Perform baseline estimation In addition to bias compensation, the baseline estimation uses an exponentially weighted moving average to obtain a sensor bias estimate, resulting in a compensated signal. ,in It is the reference baseline obtained from the last complete calibration; The compensated multimodal features The data, along with the incoming batch attribute B, is input into an SVM-based conditional prediction model. This model incorporates a batch adaptation strategy to adapt to different batches. When the ground truth is obtained, the conditional prediction model operates at a set small learning rate. Represented as Output the defect probability for each cell. The predicted value of the local error fraction of the felt in the next process time step.

[0014] A further improvement of this invention is that the batch adaptation strategy uses data from the incoming batch attributes as intermediate features for batch vector modulation during data input, represented as follows: ,in, and This represents the scaling and translation coefficients obtained through the convolutional neural network. This represents element-wise multiplication, with batch weighting adjustments applied to the output confidence or threshold. ,in, This represents the batch attribute of the i-th batch. This represents the historical mean and standard deviation.

[0015] A further improvement of the present invention is that the automated remixing strategy includes an index correction compensation layer and a mixing compensation layer; The index correction compensation layer receives output data based on the yellow warning and uses it as input to the defect prediction algorithm, outputting the index-corrected felt unevenness prediction value. ; When the detection score for uneven mesh coverage is greater than or equal to the corresponding threshold Tuni, a simple linear response matrix is ​​established. : ,in, The executor command vector is then used to solve the optimization problem using the following formula: , obtain the executor command vector in , Step size factor; When the overall mixing unevenness index is greater than or equal to the corresponding threshold Tumi, the remixing, pause, reflux, and rate adjustment are used as the action space, and the current state strategy is trained through RL. ,state Includes: Current indicator correction for predicted unevenness of felt. The parameters include MHI, silo vibration spectrum, feeding rate, and historical actions. The reward is defined as the difference between the MHI increase and the time penalty.

[0016] A further improvement of the present invention is that the hybrid compensation layer is used to receive output data based on orange warning, as input to the defect prediction algorithm, and outputs a hybrid compensation felt unevenness prediction value. Local error score based on step S2 Defect probability in step S4 And prediction of unevenness in the next process An automated remixing strategy is used to calculate local perturbation quantities, which are determined by a mapping function. Once determined, the improvement rate is evaluated using a trial-and-error mechanism after the perturbation. ,like Then continue or exit; if it fails and reaches the failure count, it escalates to remix preparation; during the remix process, it is based on the severity index. Quantify severity, by Allocate pause duration and phased return traffic The specific process includes initializing the timing t=0, and running the strong mixing program in the mixer. and send back Pause the reflux level and wait for the material to enter the production line; repeat steps S1 to S4 until the output only records without alarm.

[0017] On the other hand, the present invention provides an online monitoring and graded early warning system for deformation during the production of hemp felt, comprising: The data acquisition module is used to collect data on the production process of hemp felt in real time through a sensor array, and to read the incoming batch attribute B and record the batch ID; The error monitoring module is used to obtain the local error score of the hemp felt based on the hemp felt production process data and through the uneven spreading-mixing detection algorithm. The early warning level classification module is used to obtain the deformation level of the hemp felt based on the local error score of the hemp felt and the set local error threshold of the hemp felt; and to determine whether to run the defect prediction algorithm establishment module. The defect prediction algorithm establishment module is used to take the local error score of the felt and batch information as input, compensate for batch and sensor errors, and output the defect probability of each grid and the unevenness of the felt in the next process time step. An automated remixing strategy construction module is used to take the local error score of the felt and the unevenness of the felt in the next process time step as inputs and output control parameters.

[0018] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention firstly measures the uniformity of orientation by deploying high-resolution visual measurement on the grid-like surface of the production line, and combines local orientation field gradient with confidence weighting to achieve precise positioning of local orientation deviations; and then uses algebraic response mapping to back-calculate the minimum norm of error into executable fine-tuning commands such as carding roller angle, segment tension and local blowing, so as to gradually eliminate orientation deviations without significant line stop and reduce bending and warping after shaping. 2. Online drift compensation is performed by performing bias estimation on each sensor channel, and a digital twin and recursive calibration process is established for each non-standard machine. The compensated multimodal features and incoming batch information are then input into the predictor via FiLM conditionalization for batch adaptive prediction. This can maintain the stability of detection and control mapping under equipment aging or environmental fluctuations, and reduce misjudgment and product fluctuations caused by parameter drift. 3. By fusing multi-source features, a grid-level mixing unevenness index is calculated and assigned a confidence level. Based on this index and the defect prediction algorithm, a layered remixing strategy is adopted. The pause duration and return flow are driven by the severity and the effect is verified under the MHI closed loop, thereby restoring the mixing uniformity with minimal interruption cost and reducing local fracture and shrinkage defects. Attached Figure Description

[0019] Figure 1 This is a flowchart of a method for online monitoring and graded early warning of deformation during the production of hemp felt according to the present invention. Detailed Implementation

[0020] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0021] The term "and / or" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can represent three cases: A exists alone, A and B exist simultaneously, and B exists alone.

[0022] Example 1 Figure 1 The flowchart of the online monitoring and graded early warning method for deformation during the production of hemp felt disclosed in this embodiment is shown. The steps are as follows: S1. Collect data on the production process of hemp felt in real time through a sensor array, and read the batch attribute B (moisture content, fiber length distribution, impurity rate) of the incoming batch and record the batch ID; The sensor array includes a high-resolution linear / area camera (before and after meshing, with a resolution of 2–5 px / mm depending on the line speed): used for texture spectral density and surface defects; Laser profile / displacement sensor: used to acquire thickness / profile profiles; Multi-point tension sensor (independent measuring points on the left, center, right, and edge): used for tension imbalance detection; Online near-infrared (NIR) moisture content probe (inlet, after mixing, outlet of drying section): used for moisture content / ingredient quality detection; Vibration / acceleration sensors and level sensors for silo: used for detecting material mixing, clumping, or blockage. Environmental (temperature, relative humidity, electrostatic voltage) sensors.

[0023] S2. Based on the data from the hemp felt production process, the local error score of the hemp felt is obtained through a lay-and-mix unevenness detection algorithm. The specific implementation steps of the lay-up-mixing unevenness detection algorithm include: dividing the hemp felt production surface into several local grids, obtaining a local orientation histogram for each grid based on the structure tensor and Gabor filtering, and quantifying the orientation uniformity and its confidence to obtain the lay-up unevenness detection score Uni; simultaneously constructing a comprehensive mixing unevenness index Umi through multi-source features, and performing normalization mapping and confidence assignment; to improve robustness, all sub-scores undergo EWMA / count persistence detection within a time window, Gaussian smoothing in space, and confidence weights are obtained based on sensor SNR and orientation coherence, and a confidence-weighted linear fusion is used to obtain the local error score of the hemp felt, where the global weights... Adaptive updates based on historical correlation.

[0024] By dividing the production surface into grids and simultaneously calculating orientation and blending sub-fractions at the grid level, anomalies can be accurately located and their local intensity can be reflected, facilitating precise local treatment and reducing overall line downtime or excessive intervention. FOUI, obtained through structural tensor / Gabor and information entropy measurements, provides robust quantification of fiber orientation distribution with added confidence levels, enhancing early sensitivity and explanatory power for orientation inhomogeneity (leading to warping and bending). The fusion of multi-source blending features and confidence weighting enhances the detection capability of hidden defects such as agglomeration, delamination, and uneven moisture content, reducing false alarms. Furthermore, adaptive weight adjustment based on historical correlation improves model stability under different operating conditions.

[0025] The uneven mesh detection score is obtained by measuring the mesh unevenness in each cell. Within the sample, several local image patches are extracted. Since the structure tensor can stably estimate the principal direction even in noisy, low-contrast, or fiber-crossing environments, and can provide a quantitative confidence level of direction intensity (coherence), facilitating subsequent confidence weighting, the gradient of each local image patch is calculated. The structure tensor matrix J is obtained, and then the principal directions of each local image patch are plotted into a histogram. Gabor filtering is highly sensitive to the local frequency and orientation of textures and is often used to emphasize regular fiber orientation stripes. It performs better under certain lighting / material properties, so in this embodiment, it is used as a redundant channel of the structure tensor. Entropy is a classic measure of the unbiasedness of distribution: entropy is maximum when the fiber orientation is uniformly distributed; entropy is low when the orientation is concentrated (single principal direction). Entropy can effectively quantify the diversity of orientations and is robust to various complex interwoven scenarios. Therefore, in this embodiment, the orientation intensity confidence is used as a weight, and uniformity is represented by information entropy normalization. The calculation formula is expressed as follows: And through maximum entropy Obtain the normalized net unevenness detection score .

[0026] The multi-source characteristics of the comprehensive mixing unevenness index Umi include local average density. Using NIR or visible light to correct intensity as a density proxy and local density variance It is used to measure the inhomogeneity within the lattice, and is obtained through blob detection (LoG) / local extremum statistics + silo vibration spectrum (if local clumps pass through, high-frequency impact peaks will be generated). In the web-forming or stretching section, localized areas of insufficient (or excessive) thickness are a significant indicator of uneven mixing. The thickness variation coefficient is obtained by using the cross-sectional laser data to calculate the variance of the grid thickness. Texture roughness features, such as LBP (local binary patterns) energy or contrast and homogeneity in GLCM (gray-level co-occurrence matrix), reflect the fine-grained changes (agglomeration or sparseness) of the surface structure. Define a lattice set Average density on (e.g., a row or several adjacent cells) , and standard deviation This allows us to obtain the local density parameters of the hemp felt. The MHI is limited to [0,1]. The closer the MHI is to 1, the more uniform it is. Local grids are often used... Let MHI be its small neighborhood. MHI.

[0027] For events such as clumping and aggregation, relying solely on density variance is insufficient. Therefore, we statistically analyze the density and area distribution of local extrema and normalize them to obtain the local clumping energy. The local average thickness is analyzed using short-time Fourier transform. Peak values ​​detected within a set frequency band (easily set to 100–2kHz, depending on the material) are normalized and used as the vibrational spectrum energy. ; Obtain continuous thickness curves from sensor array The thickness variation coefficient is obtained by comparing the local average thickness with the local thickness standard deviation; thus, the comprehensive mixing unevenness index Umi is obtained. ; This represents the weighting parameter, which is determined by... Indicates the amount of density change; Capture obvious clumps; Capture thickness-related issues; It can detect material clumping issues in the silo in advance. (Weight) Initially, the values ​​can be adjusted equally or according to historical importance. Finally, a confidence-weighted linear combination is used. , Indicates the feature type; Normalized sub-fractions (the larger the fraction, the more outlier); The confidence level of this sub-score is (0–1); The global weight reflects the historical importance of this feature in defect prediction (the weights can be normalized to 1). Prevent division by zero.

[0028] The proportional method based on historical correlation is used to calculate the mutual information or Pearson correlation coefficient between features and backend quality issues (manual labeling or downstream inspection). ;make Only positive correlations are retained (negative correlations are considered as having no information).

[0029] By employing structural tensors to estimate the principal orientation and using coherence as a confidence weight, orientation information can be stably obtained even under noisy, fiber-crossing, or low-contrast conditions, thereby reducing texture misjudgments and improving the reliability of FOUI. The entropy of the orientation histogram is normalized to a Uni index, making orientation uniformity comparable across different grids, lighting conditions, or batches, facilitating the setting of uniform or batch-adaptive alarm thresholds. Using confidence-weighted entropy as the final orientation score balances orientation strength with suppressing the misleading influence of low-quality samples, improving the system's detection rate of true orientation anomalies and enhancing interpretability.

[0030] S3. Based on the local error score of the hemp felt and the set local error threshold of the hemp felt, obtain the deformation level of the hemp felt; determine whether to run S4. When the local error score of the hemp felt is less than the local error threshold of the hemp felt, and the uneven web laying detection score Uni and the comprehensive uneven mixing index Umi are both less than the corresponding threshold, only a record is made and no alarm is triggered. When the local error score of the hemp felt is less than the local error threshold of the hemp felt, and the uneven web laying detection score Uni or the comprehensive uneven mixing index Umi is greater than or equal to the corresponding threshold, a yellow warning is activated and sent to S4. When the local error score of the hemp felt is greater than or equal to the local error threshold of the hemp felt, and the uneven web laying detection score Uni and the comprehensive uneven mixing index Umi are both less than the corresponding threshold, an orange warning is activated and sent to S4. When the local error score of the hemp felt is greater than or equal to the local error threshold of the hemp felt, and the uneven web laying detection score Uni or the comprehensive uneven mixing index Umi is greater than or equal to the corresponding threshold, a red alert is activated, the production line is stopped, and the batch is isolated.

[0031] Based on a hierarchical logic (green / yellow / orange / red) of local error scores and corresponding thresholds, hierarchical alarm management is achieved: it can record or alert only for minor anomalies to avoid accidental line shutdowns, while triggering mandatory isolation or line shutdown for severe anomalies, thereby reducing manual intervention costs while controlling quality risks. By simultaneously referencing combinations of different sub-scores of Uni and Umi, the accuracy of anomaly identification and the ability to distinguish fault types (orientation vs. mixing vs. concurrency) are enhanced, facilitating subsequent targeted remedial measures and improving handling efficiency.

[0032] S4. Establish a defect prediction algorithm, take the local error score of the felt and batch information as input, compensate for batch and sensor errors, and output the defect probability of each grid and the unevenness of the felt in the next process time step. The defect prediction algorithm includes analyzing sensor measurements. Perform baseline estimation In addition to bias compensation, the baseline estimation uses an exponentially weighted moving average to obtain a sensor bias estimate, resulting in a compensated signal. ,in It is the reference baseline obtained from the last complete calibration; The compensated multimodal features The data, along with the incoming batch attribute B, is input into an SVM-based conditional prediction model. This model incorporates a batch adaptation strategy to adapt to different batches. When ground truth is obtained (either manually labeled or from backend quality inspection results), the conditional prediction model operates at a set small learning rate. Represented as It also uses a sample replay / experience buffer to mix historical samples and prevent catastrophic forgetting. A forgetting factor can be set. Weights on older samples are reduced. Output the defect probability for each cell. The predicted values ​​of the local error fraction of the felt in the next process time step are calculated. A confidence level is assigned to each prediction. (Estimated by the output of the prediction model or by ensemble / resampling) When the confidence is low, no action is automatically issued. Instead, a manual confirmation request is sent and the annotations are fed back to update the model.

[0033] EWMA (or Kalman, CUSUM) is used for baseline estimation and bias correction of the sensor channels. This suppresses baseline drift caused by sensor drift, temperature and humidity, or equipment aging online, ensuring the stability of input features and reducing false alarms and model degradation. Compensated multimodal features and batch attributes are used as inputs to the conditional predictor, and FiLM or conditional strategies are employed to enable the model to adapt to the characteristics of different batches of raw materials, thereby improving the generalization ability to predict defect probabilities and unevenness across different batches. Online incremental updates allow the model to smoothly adapt to new batches or new failure modes after obtaining manually labeled or back-end verified true values, avoiding catastrophic forgetting and continuously improving prediction accuracy. Low-confidence judgment and manual confirmation mechanisms ensure the controllability and safety of automated handling.

[0034] The batch adaptation strategy uses data from the incoming batch attributes as intermediate features for batch vector modulation during data input, represented as follows: ,in, and This represents the scaling and translation coefficients obtained through the convolutional neural network. This represents element-wise multiplication, allowing the model to learn feature transformations across different batches, thus improving generalization. Batch-weighted adjustments are made to the output confidence or threshold, denoted as... ,in, This represents the batch attribute of the i-th batch. This represents the historical mean and standard deviation.

[0035] By employing FiLM (or equivalent conditionalization layer) to modulate batch information onto intermediate features, the network can learn the feature transformation patterns corresponding to each batch, thereby significantly improving prediction accuracy and stability under batch diversity. Weighting / correcting the threshold based on batch attributes (τ(B) formula) can automatically relax or tighten alarm sensitivity, avoiding numerous false alarms or missed alarms when raw material characteristics vary greatly, thus improving the robustness and operability of online deployment.

[0036] When uneven mixing or stress drift is detected in a certain area of ​​the hemp felt, the system automatically calculates the necessary adjustments to the mixing ratio, roller tension, or traction speed based on the response matrix G, in order to minimize local errors in the next time step. Different batches of hemp will have different responses, resulting in different G values. When a trend deviation occurs in the production process, it is compensated for by the inverse response of G. Variation is limited by λ to avoid drastic fluctuations.

[0037] S5. Construct an automated remixing strategy, taking the local error score of the felt and the unevenness of the felt in the next process time step as inputs, and output control parameters.

[0038] The automated remixing strategy includes an index correction compensation layer and a mixing compensation layer; The index correction compensation layer receives output data based on the yellow warning and uses it as input to the defect prediction algorithm, outputting the index-corrected felt unevenness prediction value. ; When the detection score for uneven mesh coverage is greater than or equal to the corresponding threshold Tuni, a simple linear response matrix is ​​established. (Obtained from digital twin or offline calibration): ,in, The executor command vector is then used to solve the optimization problem using the following formula: , obtain the executor command vector in , This is the step size factor.

[0039] When the overall mixing unevenness index is greater than or equal to the corresponding threshold Tumi, the remixing, pause, reflux, and rate adjustment are used as the action space, and the current state strategy is trained through RL. ,state Includes: Current indicator correction for predicted unevenness of felt. The data includes MHI (Mean Hierarchical Index), silo vibration spectrum, feeding rate, and historical actions. The reward is defined as the difference between the MHI rise and the time penalty. Training data can be obtained from simulators (digital twins) or historical online simulations.

[0040] The index correction and compensation layer takes the output of the yellow warning level as input. It first models the actuator-product surface response relationship using the linear response matrix G and then directly calculates the actuator commands using damped minimum norm solving (ridge regression form). This achieves rapid, controllable, and numerically stable local correction of the netting orientation problem, avoiding secondary disturbances caused by blind, large-scale adjustments. Using G obtained from digital twins or offline calibration, the control mapping is adapted to different non-standard machines. Through a trial-and-verification strategy with finite-step small adjustments, local orientation anomalies can be gradually corrected and the effect verified while ensuring safety, significantly improving the success rate of correction and reducing the need for manual parameter tuning.

[0041] The hybrid compensation layer is used to receive output data based on orange warning, which serves as the input to the defect prediction algorithm, and outputs the hybrid compensation felt unevenness prediction value. Local error score based on step S2 Defect probability in step S4 And prediction of unevenness in the next process An automated remixing strategy is used to calculate local perturbation quantities, which are determined by a mapping function. The determination (mapped to perturbation strength and subject to maximum amplitude limiting) is followed by evaluation of the improvement rate using a trial-and-error mechanism after perturbation. ,like Then continue or exit; if it fails and reaches the failure count, it escalates to remix preparation; during the remix process, it is based on the severity index. Quantify severity, by Allocate pause duration and phased return traffic The specific process includes initializing the timing t=0, and running the strong mixing program in the mixer. (e.g., 10–30 s), and send back Pause the reflux level and wait for materials to enter the production line. Repeat steps S1 to S4 until the output only records no alarms.

[0042] When the overall mixing is severely uneven, actions such as remixing / pause / reflow / rate adjustment are used to form an action space and employ an RL learning strategy. This allows for the learning of near-optimal action sequences under complex entanglement and multivariate coupling conditions, improving the ability to handle refractory stratification problems, reducing repetitive and ineffective operations, and enhancing MHI recovery efficiency (while retaining engineering constraints from simulation / data training). The mixing compensation layer performs closed-loop control on the orange warning output (based on local disturbance mapping and trial-and-error rules), and quantifies pause and reflow volumes through the severity index during remixing. This achieves controllable execution of "allocating remixing resources on demand and gradually restoring the production line," reducing downtime while ensuring quality recovery. A closed-loop mechanism of step-by-step reflow—waiting for settling—real-time evaluation is adopted, which can determine the effectiveness of remixing using real-time indicators such as MHI / E during operation. This avoids prolonged blind remixing or excessive reflow, improving resource utilization and repair efficiency.

[0043] The strategy involves a closed-loop process of step-by-step refluxing in the mixer—waiting for settling—real-time measurement of MHI / to determine whether to end or continue. The entire remixing process includes rate / amplitude limiting, actuator feedback verification, and safety rollback (if multiple consecutive failures occur, a rollback is initiated, triggering manual intervention or stronger mechanical intervention). Simultaneously, the cause of each trigger, action sequence, feedback, and improvement rate are fully logged for traceability and subsequent strategy optimization. This strategy follows the principles of starting with lighter interventions and gradually increasing them, using trial and error verification and quantified input. It combines predictive results to avoid excessive intervention, restoring mixing uniformity and thickness consistency with minimal interruption, reducing rework rates, and improving steady-state production capacity and automation reliability.

[0044] The standard execution process is to test, verify, and rollback, with failure counts and rollback trigger conditions set up. Combined with a manual takeover mechanism, this ensures the system can safely converge and be manually handled when automated actions fail or sensors malfunction, reducing the risk of automated errors. Detailed log records (trigger reasons, action sequences, improvement rates, and whether rollback occurred) provide high-quality data for subsequent offline analysis, strategy optimization, and model retraining, helping to continuously improve system performance and meet quality traceability requirements.

[0045] The threshold and weight settings can be set by default according to the present invention, or they can be set by those skilled in the art.

[0046] Example 2 Based on the same inventive concept as Example 1, the present invention provides an online monitoring and graded early warning system for deformation during the production of hemp felt, comprising: The data acquisition module is used to collect data on the production process of hemp felt in real time through a sensor array, read the attributes of incoming batches, and record the batch ID; The error monitoring module is used to obtain the local error score of the hemp felt based on the hemp felt production process data and through the uneven spreading-mixing detection algorithm. The early warning level classification module is used to obtain the deformation level of the hemp felt based on the local error score of the hemp felt and the set local error threshold of the hemp felt; and to determine whether to run the defect prediction algorithm establishment module. The defect prediction algorithm establishment module is used to take the local error score of the felt and batch information as input, compensate for batch and sensor errors, and output the defect probability of each grid and the unevenness of the felt in the next process time step. An automated remixing strategy construction module is used to take the local error score of the felt and the unevenness of the felt in the next process time step as inputs and output control parameters.

[0047] This invention organically combines multimodal sensing, lattice-based feature analysis, batch-conditional prediction, and layered remixing compensation to construct an end-to-end solution capable of highly sensitive identification of the unique web-laying orientation and material layering issues of hemp felt, while also adaptively compensating for differences in non-standard equipment and batches. This solution significantly reduces false alarm rates, rework rates, and unnecessary line stoppages while improving the accuracy of early defect detection, positioning precision, and correction success rate. It is beneficial for enhancing production line automation, product consistency, and production efficiency, and provides a traceable, adjustable, and engineering-feasible technical path for industrial deployment.

[0048] 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.

[0049] 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.

[0050] 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.

[0051] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus 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.

[0052] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for online monitoring and graded early warning of deformation during the production of hemp felt, characterized in that: Includes the following steps: S1. Collect data on the production process of hemp felt in real time through a sensor array, read the batch attribute B upon arrival, and record the batch ID; S2. Based on the data from the hemp felt production process, the local error score of the hemp felt is obtained through a lay-and-mix unevenness detection algorithm. S3. Based on the local error score of the hemp felt and the set local error threshold of the hemp felt, obtain the deformation level of the hemp felt; determine whether to run S4. S4. Establish a defect prediction algorithm, take the local error score of the felt and batch information as input, compensate for batch and sensor errors, and output the defect probability of each grid and the unevenness of the felt in the next process time step. S5. Construct an automated remixing strategy, taking the local error score of the felt and the unevenness of the felt in the next process time step as inputs, and output control parameters.

2. The method for online monitoring and graded early warning of deformation during the production process of hemp felt according to claim 1, characterized in that: The specific implementation steps of the lay-up-mixing unevenness detection algorithm include: dividing the hemp felt production surface into several local grids, obtaining a local orientation histogram for each grid based on the structure tensor and Gabor filtering, and quantifying the orientation uniformity and its confidence level to obtain the lay-up unevenness detection score Uni; simultaneously constructing a comprehensive mixing unevenness index Umi through multi-source features, and performing normalization mapping and confidence assignment; obtaining confidence weights based on sensor SNR and orientation coherence, and using confidence-weighted linear fusion to obtain the local error score of the hemp felt, where the global weights... Adaptive updates based on historical correlation.

3. The method for online monitoring and graded early warning of deformation during the production process of hemp felt according to claim 2, characterized in that: The uneven mesh detection score is obtained by measuring the mesh unevenness in each cell. Within the image, extract several local image patches and calculate the gradient of each local image patch. The structure tensor matrix J is obtained, and then the principal directions of each local image patch are plotted into a histogram. Using directional intensity confidence as weights, uniformity is represented by information entropy normalization, and the calculation formula is expressed as follows: and through maximum entropy Obtain the normalized net unevenness detection score .

4. The method for online monitoring and graded early warning of deformation during the production of hemp felt according to claim 3, characterized in that: The multi-source characteristics of the comprehensive mixing unevenness index Umi include local average density. Local density variance Localized clumping energy Coefficient of variation of thickness Texture roughness characteristics; Define a lattice set average density on , and standard deviation This allows us to obtain the local density parameters of the hemp felt. ; The density and area distribution of local extrema are statistically analyzed and normalized to obtain the local clumping energy. The local average thickness is obtained by short-time Fourier transform, and the peak value detected in the set frequency band is normalized and used as the vibrational spectrum energy. ; Obtain continuous thickness curves from sensor array The thickness variation coefficient is obtained by the ratio of the local average thickness to the local thickness standard deviation. This leads to the comprehensive mixing unevenness index Umi: ; This represents the weighting parameter.

5. The method for online monitoring and graded early warning of deformation during the production process of hemp felt according to claim 4, characterized in that: Step S3 further includes recording without alarming when the local error score of the hemp felt is less than the local error threshold of the hemp felt and the uneven web laying detection score Uni and the comprehensive uneven mixing index Umi are both less than the corresponding thresholds; When the local error score of the hemp felt is less than the local error threshold of the hemp felt, and the uneven web laying detection score Uni or the comprehensive uneven mixing index Umi is greater than or equal to the corresponding threshold, a yellow warning is activated and sent to S4. When the local error score of the hemp felt is greater than or equal to the local error threshold of the hemp felt, and the uneven web laying detection score Uni and the comprehensive uneven mixing index Umi are both less than the corresponding threshold, an orange warning is activated and sent to S4. When the local error score of the hemp felt is greater than or equal to the local error threshold of the hemp felt, and the uneven web laying detection score Uni or the comprehensive uneven mixing index Umi is greater than or equal to the corresponding threshold, a red alert is activated, the production line is stopped, and the batch is isolated.

6. The method for online monitoring and graded early warning of deformation during the production process of hemp felt according to claim 5, characterized in that: The defect prediction algorithm includes analyzing sensor measurements. Perform baseline estimation In addition to bias compensation, the baseline estimation uses an exponentially weighted moving average to obtain a sensor bias estimate, resulting in a compensated signal. ,in It is the reference baseline obtained from the last complete calibration; The compensated multimodal features The data, along with the incoming batch attribute B, is input into an SVM-based conditional prediction model. This model incorporates a batch adaptation strategy to adapt to different batches. When the ground truth is obtained, the conditional prediction model operates at a set small learning rate. Represented as Output the defect probability for each cell. The predicted value of the local error fraction of the felt in the next process time step.

7. The method for online monitoring and graded early warning of deformation during the production process of hemp felt according to claim 6, characterized in that: The batch adaptation strategy uses data from the incoming batch attributes as intermediate features for batch vector modulation during data input, represented as follows: ,in, and This represents the scaling and translation coefficients obtained through the convolutional neural network. This represents element-wise multiplication, with batch weighting adjustments applied to the output confidence or threshold. ,in, This represents the batch attribute of the i-th batch. This represents the historical mean and standard deviation.

8. The method for online monitoring and graded early warning of deformation during the production of hemp felt according to claim 7, characterized in that: The automated remixing strategy includes an index correction compensation layer and a mixing compensation layer; The index correction compensation layer receives output data based on the yellow warning and uses it as input to the defect prediction algorithm, outputting the index-corrected felt unevenness prediction value. ; When the detection score for uneven mesh coverage is greater than or equal to the corresponding threshold Tuni, a simple linear response matrix is ​​established. : ,in, The executor command vector is then used to solve the optimization problem using the following formula: , obtain the executor command vector in , Step size factor; When the overall mixing unevenness index is greater than or equal to the corresponding threshold Tumi, the remixing, pause, reflux, and rate adjustment are used as the action space, and the current state strategy is trained through RL. ,state Includes: Current indicator correction for predicted unevenness of felt. The parameters include MHI, silo vibration spectrum, feeding rate, and historical actions. The reward is defined as the difference between the MHI increase and the time penalty.

9. The method for online monitoring and graded early warning of deformation during the production of hemp felt according to claim 8, characterized in that: The hybrid compensation layer is used to receive output data based on orange warning, which serves as the input to the defect prediction algorithm, and outputs the hybrid compensation felt unevenness prediction value. Local error score based on step S2 Defect probability in step S4 And prediction of unevenness in the next process An automated remixing strategy is used to calculate local perturbation quantities, which are determined by a mapping function. Once determined, the improvement rate is evaluated using a trial-and-error mechanism after the perturbation. ,like Then continue or exit; if it fails and reaches the failure count, it escalates to remix preparation; during the remix process, it is based on the severity index. Quantify severity, by Allocate pause duration and phased return traffic The specific process includes initializing the timing t=0, and running the strong mixing program in the mixer. and send back Pause the reflux level and wait for the material to enter the production line; repeat steps S1 to S4 until the output only records without alarm.

10. A system for online monitoring and graded early warning of deformation during the production of hemp felt, used to execute the method for online monitoring and graded early warning of deformation during the production of hemp felt as described in any one of claims 1-9, characterized in that: include: The data acquisition module is used to collect data on the production process of hemp felt in real time through a sensor array, and to read the incoming batch attribute B and record the batch ID; The error monitoring module is used to obtain the local error score of the hemp felt based on the hemp felt production process data and through the uneven spreading-mixing detection algorithm. The early warning level classification module is used to obtain the deformation level of the hemp felt based on the local error score of the hemp felt and the set local error threshold of the hemp felt; and to determine whether to run the defect prediction algorithm establishment module. The defect prediction algorithm establishment module is used to take the local error score of the felt and batch information as input, compensate for batch and sensor errors, and output the defect probability of each grid and the unevenness of the felt in the next process time step. An automated remixing strategy construction module is used to take the local error score of the felt and the unevenness of the felt in the next process time step as inputs and output control parameters.

Images