Artificial intelligence-based high release liner production management system and method

Abstract

This invention relates to the field of intelligent control technology in textile production, specifically to a high-peel-resistance lining production management system and method based on artificial intelligence. The system includes: collecting a set of real-time process parameters and a set of online quality inspection images from the production line; extracting a set of process features and a set of image features including coating uniformity, bubbles, and wrinkles; calculating the correlation strength matrix between process parameters and quality inspection results; identifying key process parameters and their corresponding correlation patterns; constructing a process adjustment decision model; and using the real-time image feature set to deduce recommended values ​​for optimized process parameters. This method can quantify the correspondence between process parameters and various defect features, forming a closed-loop control from quality inspection feedback to process optimization, reducing reliance on manual experience, improving the accuracy and intelligence level of high-peel-resistance lining production control, and optimizing the control effect of lining surface defects.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology in textile production, and in particular to a high-peeling interlining production management system and method based on artificial intelligence. Background Technology

[0002] In the current production management of high-peel lining fabrics, production lines mostly use independent acquisition of process parameters for monitoring. Online inspection systems only make overall pass / fail judgments on the lining quality inspection images, and image recognition only extracts general defect features, without classifying and extracting features such as coating uniformity, bubbles, and wrinkles. Process parameter control and quality inspection image analysis are carried out independently. Process parameter adjustments rely on operator experience and judgment, without establishing a quantitative correspondence between process parameters and quality inspection defect features. Production control is mainly driven by manual experience, and intelligent correlation analysis methods are lacking.

[0003] Existing technologies cannot quantitatively characterize the sequence of each parameter in the process parameter set against the features of quality inspection images one by one. The influence of key process parameters on different defect characteristics cannot be accurately distinguished, and process adjustments lack objective quantitative basis. During production, quality inspection results can only be observed after forward adjustment of process parameters. It is impossible to reverse match the optimal process parameters based on the real-time acquired quality inspection image features. The control efficiency for defects such as insufficient uniformity of lining coating and bubbles and wrinkles is low, and the production process and quality inspection results cannot form a closed loop adaptation.

[0004] It is necessary to establish a quantitative correlation between process parameters and three types of image features: coating uniformity, bubbles, and wrinkles. At the same time, based on real-time image features, the optimized values ​​of process parameters should be derived in reverse, thereby making up for the lack of quantitative correlation and reverse adjustment in the existing production management. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a high-peeling lining production management system and method based on artificial intelligence.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a high-peelability lining production management system and method based on artificial intelligence, comprising:

[0007] Collect the set of process parameters collected in real time from the lining production line, and simultaneously collect the set of quality inspection images generated by the online detection system of the production line;

[0008] The process parameter set is subjected to feature extraction processing to obtain a process feature set, and the quality inspection image set is subjected to image recognition processing to extract an image feature set, wherein the image feature set includes coating uniformity features, bubble features and wrinkle features.

[0009] Based on the process feature set and the image feature set, a correlation strength matrix between process parameters and quality inspection results is calculated. The correlation strength matrix is ​​used to characterize the quantitative relationship between each parameter sequence in the process parameter set and each feature in the image feature set.

[0010] Based on the correlation strength matrix, the set of key process parameters that have the greatest impact on quality inspection results and their correlation patterns with image features are identified.

[0011] Based on the identified set of key process parameters and their correlation patterns with image features, a process adjustment decision model is constructed. This model is used to deduce optimized recommended values ​​for process parameters based on the real-time set of image features.

[0012] As a further aspect of the present invention, feature extraction processing is performed on the set of process parameters to obtain a set of process features, including:

[0013] The set of process parameters includes coating thickness data, curing temperature timing data, and winding tension timing data;

[0014] The coating thickness data is subjected to fluctuation analysis to extract coating thickness trend characteristics and thickness stability indicators;

[0015] Periodic pattern decomposition was performed on the curing temperature time series data to separate steady-state temperature characteristics, temperature rise rate characteristics, and temperature fluctuation frequency characteristics.

[0016] The winding tension time series data is subjected to abnormal peak detection and smoothness calculation to obtain tension anomaly characteristics and tension stability characteristics;

[0017] The extracted coating thickness trend features, thickness stability index, steady-state temperature features, temperature rise rate features, temperature fluctuation frequency features, tension anomaly features, and tension stability features are normalized and vectorized to form the process feature set.

[0018] As a further aspect of the present invention, image recognition processing is performed on the quality inspection image set to extract an image feature set, including:

[0019] The quality inspection image set includes images of the lining surface;

[0020] The surface image of the lining fabric is preprocessed, including image enhancement, noise removal, and region segmentation, to obtain a standardized surface analysis image;

[0021] The standardized surface analysis image is input into a convolutional neural network, and deep image features are extracted through multiple convolutional and pooling layers to obtain a basic image feature map;

[0022] Multi-scale feature fusion is performed on the base image feature map to generate a fused feature map. The multi-scale feature fusion includes extracting features of different scales from the base image feature map and connecting them.

[0023] On the fused feature map, a pre-trained defect detection head is applied to locate and identify uneven coating regions, bubble regions, and wrinkle regions, and to calculate the geometric and texture features of each region, forming the image feature set that includes coating uniformity features, bubble features, and wrinkle features.

[0024] As a further aspect of the present invention, based on the process feature set and the image feature set, a correlation strength matrix between process parameters and quality inspection results is calculated, including:

[0025] The process feature set is mapped to a process feature vector through a fully connected network, and the image feature set is mapped to a quality inspection feature vector through a fully connected network.

[0026] Construct a process-quality inspection association network, taking the process feature vector and the quality inspection feature vector as input, and calculate the attention weight between each feature dimension in the process feature vector and each feature dimension in the quality inspection feature vector;

[0027] Based on the calculated attention weights, the process feature vector and the quality inspection feature vector are subjected to bidirectional weighted interaction to generate process enhancement feature vector and quality inspection enhancement feature vector.

[0028] Calculate the cosine similarity between the process enhancement feature vector and the quality inspection enhancement feature vector, and arrange them into a matrix to obtain the correlation strength matrix. The rows of the correlation strength matrix correspond to the process feature dimension, and the columns correspond to the quality inspection feature dimension.

[0029] As a further aspect of the present invention, based on the correlation strength matrix, the set of key process parameters that have the greatest impact on quality inspection results and their correlation patterns with image features are identified, including:

[0030] Sum each row of the correlation strength matrix to calculate the total correlation strength of each process feature dimension with all quality inspection feature dimensions;

[0031] The process feature dimensions that exceed a preset threshold are selected, and the process feature dimensions are mapped back to the original set of process parameters to obtain a set of candidate key process parameters.

[0032] For each parameter in the candidate key process parameter set, extract its correlation strength sequence with each quality inspection feature dimension from the correlation strength matrix;

[0033] Cluster analysis is performed on the correlation strength sequences to identify subsets of process parameters with similar correlation patterns, and the correlation pattern of each subset is defined as a correlation pattern.

[0034] A subset of process parameters with high overall correlation strength and clear correlation patterns is identified as the set of key process parameters, and their correlation patterns with specific image features are recorded.

[0035] As a further aspect of the present invention, the step of constructing a process adjustment decision model based on the identified set of key process parameters and their correlation patterns with image features includes:

[0036] A training sample set is constructed using the image feature set as the input variable and the optimization target value of the key process parameter set as the output variable.

[0037] A multi-layer neural network is constructed as the main structure of the decision network, with the number of input layer nodes matching the dimension of the image feature set and the number of output layer nodes matching the dimension of the key process parameter set.

[0038] An association pattern constraint layer is introduced into the decision network. The association pattern constraint layer applies prior constraints to the network weights based on the set of key process parameters and their association patterns with image features.

[0039] The decision network with the introduced correlation pattern constraint layer is trained using the training sample set. The training objective is to minimize the error between the predicted process parameter optimization target value and the actual optimal value.

[0040] After training, the process adjustment decision model is obtained.

[0041] As a further aspect of the present invention, an association pattern constraint layer is introduced into the decision network. This association pattern constraint layer applies prior constraints to the network weights based on the set of key process parameters and their association patterns with image features, including:

[0042] The association pattern between the set of key process parameters and image features is encoded into a connection weight mask matrix, which indicates the relationship between allowed and prohibited connections between the input layer and the output layer.

[0043] During the training process of the decision network, the connection weight mask matrix is ​​multiplied element-wise with the actual connection weight matrix of the network to forcibly prohibit connection relationships that do not exist in the association pattern.

[0044] A regularization term is added to the loss function, which is used to penalize weight changes that are inconsistent with the direction of known association patterns.

[0045] As a further aspect of the present invention, it also includes:

[0046] The process adjustment decision model is used to process the real-time generated set of process parameters and the set of quality inspection images to generate a real-time set of process parameter adjustment instructions. The set of process parameter adjustment instructions includes a sequence of changes in target set values ​​for coating thickness, curing temperature and winding tension.

[0047] The set of process parameter adjustment instructions is sent to the programmable logic controller of the production line to drive the coating machine, curing oven and winding machine to perform parameter adjustments;

[0048] The process adjustment decision model is used to process the real-time generated set of process parameters and the set of quality inspection images to generate a real-time set of process parameter adjustment instructions, including:

[0049] Real-time acquisition of online process parameters and online quality inspection images for the current production batch;

[0050] Feature extraction and image recognition processes are performed on the online process parameter set and the online quality inspection image set, respectively, to obtain the real-time process feature set and the real-time image feature set.

[0051] The real-time image feature set is input into the trained process adjustment decision model, and the model outputs a set of optimized process parameter target values ​​corresponding to the current quality inspection status.

[0052] The set of target values ​​for the optimized process parameters is compared with the current set of online process parameters to calculate the differences in parameters that need to be adjusted.

[0053] Based on the differences in the parameters, and combined with the preset adjustment step size and direction, a specific and executable set of process parameter adjustment instructions is generated.

[0054] As a further aspect of the present invention, the optimized process parameter target value set is compared with the current online process parameter set to calculate the parameter differences that need to be adjusted, including:

[0055] Each target parameter value in the optimized process parameter target value set is aligned and compared with the corresponding real-time parameter value in the current online process parameter set;

[0056] For each parameter, calculate the absolute difference between the target parameter value and the real-time parameter value to obtain the preliminary difference value of the parameter;

[0057] Considering the dynamic response characteristics of the production process, a correction factor is introduced for the initial difference value of each parameter. The correction factor is calculated based on the historical response speed of the parameter adjustment and the current load state of the production line.

[0058] The final parameter difference is obtained by multiplying the initial difference value of each parameter by its corresponding correction factor.

[0059] As a further aspect of the present invention, the present invention also includes a high-peelability lining production management system based on artificial intelligence. The system includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the high-peelability lining production management method based on artificial intelligence as described above.

[0060] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0061] Feature extraction is performed on the set of process parameters to obtain a set of process features. Image recognition is performed on the set of quality inspection images to extract an image feature set containing coating uniformity features, bubble features, and wrinkle features. The correlation strength matrix is ​​calculated through the two types of features, which can completely depict the quantitative correspondence between each parameter sequence in the set of process parameters and each feature in the set of image features. It can clearly distinguish the differences in the impact of different process parameters on the three types of features of coating uniformity, bubbles, and wrinkles, objectively present the degree of correlation between each process parameter and different defect features, abandon the subjective experience-based method of judging the correlation between process parameters and quality inspection features, form a quantitative correspondence system between process parameters and image features, and accurately reflect the degree of effect of changes in each process parameter on different quality inspection features.

[0062] A process adjustment decision model is constructed based on a set of key process parameters and their correlation with image features. Using a set of real-time image features as input, the recommended values ​​of optimized process parameters are directly derived. This establishes a direct derivation logic between quality inspection image features and optimized process parameter values, changing the traditional control mode where process parameters can only be adjusted forward and the results observed. It achieves direct linkage between quality inspection features and process parameter adjustments. Real-time quality inspection image features can be directly converted into quantitative directions for process parameter adjustments. A dynamic matching relationship is formed between process parameter adjustments and real-time quality inspection status during production. Suitable process parameter values ​​can be obtained without repeated parameter trial and error, achieving synchronous adaptation between production process and quality inspection features. This strengthens the pertinence and real-time nature of process adjustments in production management, allowing production control to form a complete closed loop from quality inspection feedback to process optimization. Attached Figure Description

[0063] Figure 1 This is a flowchart of the high-peeling lining production management method based on artificial intelligence as described in this invention;

[0064] Figure 2 A flowchart for image feature extraction;

[0065] Figure 3 A heatmap showing the correlation strength between process parameters and quality inspection characteristics;

[0066] Figure 4 A graph showing the change in training loss of a neural network over epochs;

[0067] Figure 5 A time-series analysis diagram of the core parameters for the production of high-peel lining fabric. Detailed Implementation

[0068] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0069] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0070] See Figure 1 This invention provides a high-peelability interlining production management method based on artificial intelligence, the specific method including:

[0071] The process parameters are collected in real time from various sensors and monitoring equipment on the lining production line. These sensors continuously monitor the operating status of processes such as coating, curing, and winding. Simultaneously, quality inspection images generated at the end of the production line by online inspection systems, such as high-resolution industrial cameras, are also collected. Feature extraction processing is performed on the collected process parameter sets, for example, using time series analysis algorithms to extract a set of process features that characterize the trends and stability of process changes. Image recognition processing is performed on the quality inspection image sets, using computer vision technology to identify quality information such as coating uniformity, the presence of bubbles, and wrinkles from the images, and quantifying them into a set of image features. Based on the extracted process feature sets and image feature sets, a correlation strength matrix is ​​calculated using a correlation analysis algorithm. This matrix precisely describes the quantitative correlation between the change sequence of each process parameter and each quality inspection image feature in numerical form. Based on the calculated correlation strength matrix, screening and clustering algorithms are used to identify the set of key process parameters that have the greatest impact on the quality inspection results, i.e., the image feature sets, and to clarify the correlation patterns between these key parameters and specific image features. Based on the identified set of key process parameters and their correlation patterns with image features, a process adjustment decision model is constructed. This model, trained by machine learning, can infer and output a set of optimized process parameter recommendations based on a new set of real-time image features, thereby providing a basis for production adjustments.

[0072] In one embodiment of the present invention, feature extraction processing is performed on the process parameter set to obtain a process feature set. The process parameter set includes coating thickness data, curing temperature time series data, and winding tension time series data. Fluctuation analysis is performed on the coating thickness data to extract the trend characteristics of coating thickness changes over time, and its thickness stability index is calculated. Periodic pattern decomposition is performed on the curing temperature time series data to separate the steady-state temperature characteristics representing the isothermal stage, the temperature rise rate characteristics reflecting the rate of temperature increase, and the frequency characteristics reflecting temperature fluctuations from the temperature change curve. Abnormal peak detection and smoothness calculation are performed on the winding tension time series data to identify abnormal characteristics of tension abrupt changes, and the tension stability characteristics of the tension curve are calculated. The extracted coating thickness trend characteristics, thickness stability index, steady-state temperature characteristics, temperature rise rate characteristics, temperature fluctuation frequency characteristics, tension anomaly characteristics, and tension stability characteristics are normalized, and then these feature values ​​are vector-concatenated to form a process feature set representing the comprehensive process state.

[0073] Image recognition processing is performed on the quality inspection image set to extract image feature sets. The quality inspection image set contains images of the lining surface. See also... Figure 2The surface image of the lining fabric is preprocessed, including image enhancement to improve contrast, noise removal to filter out random noise, and region segmentation to obtain a standardized surface analysis image. This standardized surface analysis image is then input into a convolutional neural network. The network abstracts the image layer by layer through multiple alternating convolutional and pooling layers, extracting deep image features to obtain a basic image feature map. Multi-scale feature fusion is then performed on the basic image feature map, extracting features at different scales and concatenating them to generate a feature map that incorporates multi-scale information. A pre-trained defect detection head is applied to the fused feature map. This head can locate and label uneven coating regions, bubble regions, and wrinkle regions in the image, and calculate the area, shape, and other geometric and textural features of each region, ultimately forming an image feature set containing coating uniformity features, bubble features, and wrinkle features.

[0074] In specific implementation, feature extraction processing is performed on the process parameter set to obtain a process feature set. The process parameter set includes coating thickness data, curing temperature time series data, and winding tension time series data. In specific implementation, the coating thickness data comes from time series data points collected by thickness sensors on the production line. Fluctuation analysis is performed on the coating thickness data to extract the trend characteristics of coating thickness changes over time. The trend characteristics are obtained by fitting the time series of coating thickness data through linear regression, and the thickness stability index of the coating thickness data is calculated. The thickness stability index quantifies the dispersion of the coating thickness data. In some embodiments, the thickness stability index is calculated using the following formula:

[0075]

[0076] in: This indicates the thickness stability index. Indicates the first One coating thickness data point, This represents the average value of the coating thickness data. This represents the total number of data points for coating thickness. Periodic mode decomposition is performed on the curing temperature time series data to separate the steady-state temperature characteristics representing the isothermal stage from the temperature change curve. The steady-state temperature characteristics are the average values ​​of the curing temperature time series data during the stable stage. The temperature rise rate characteristic, reflecting the rate of temperature increase, is obtained by calculating the slope of the temperature rise curve. The frequency characteristics reflecting temperature fluctuations are extracted by Fourier transform to obtain the dominant frequency components. The optional periodic mode decomposition method is empirical mode decomposition. Abnormal peak detection and smoothness calculation are performed on the winding tension time series data to identify abnormal characteristics of tension abrupt changes. Abnormal characteristics are detected by setting a threshold to identify peak points in the winding tension time series data that exceed the threshold, and the tension stability characteristics of the tension curve are calculated. The tension stability characteristics are quantified by the first-order difference variance of the winding tension time series data. The optional abnormal peak detection method is a threshold method based on standard deviation. It is understandable that the extracted coating thickness trend features, thickness stability index, steady-state temperature features, temperature rise rate features, temperature fluctuation frequency features, tension anomaly features, and tension stability features are normalized. Then, the coating thickness trend features, thickness stability index, steady-state temperature features, temperature rise rate features, temperature fluctuation frequency features, tension anomaly features, and tension stability features are vectorized and concatenated to form a set of process features that characterize the comprehensive process state.

[0077] In specific implementation, image recognition processing is performed on the quality inspection image set to extract image feature sets. The quality inspection image set includes images of the lining surface. Preprocessing of the lining surface images includes image enhancement to improve image contrast, noise removal by filtering out random noise points, and region segmentation to divide the region of interest, resulting in a standardized surface analysis image. The standardized surface analysis image is input into a convolutional neural network (CNN). The CNN abstracts the image layer by layer through multiple alternating convolutional and pooling layers to extract deep image features, resulting in a basic image feature map. In some embodiments, the CNN contains three convolutional layers, each followed by a max-pooling layer. Multi-scale feature fusion is performed on the basic image feature map, extracting features of different scales from the basic image feature map and performing concatenation operations to generate a feature map that integrates multi-scale information. A pre-trained defect detection head is applied to the fused feature map. The defect detection head can locate and mark the uneven coating regions, bubble regions, and wrinkle regions in the image, and calculate the area, shape, geometric features, and texture features of the uneven coating regions, bubble regions, and wrinkle regions. Finally, an image feature set containing coating uniformity features, bubble features, and wrinkle features is formed. It can be understood that the coating uniformity feature is obtained by calculating the pixel intensity variance of the uneven coating region, the bubble feature is described by the area and roundness of the bubble region, and the wrinkle feature is represented by the texture gradient direction histogram of the wrinkle region.

[0078] In one embodiment of the present invention, a correlation strength matrix between process parameters and quality inspection results is calculated based on a set of process features and a set of image features. The set of process features is mapped to a fixed-dimensional process feature vector through a fully connected network. The set of image features is mapped to a quality inspection feature vector with the same or different dimensions through another fully connected network. A process-quality inspection correlation network is constructed, with both the process feature vector and the quality inspection feature vector serving as inputs. The network calculates the attention weights between each feature dimension of the process feature vector and each feature dimension of the quality inspection feature vector through an internal mechanism. Based on the calculated attention weights, a bidirectional weighted interaction is performed on the process feature vector and the quality inspection feature vector, i.e., the process features are weighted and corrected using quality inspection features, and vice versa, generating process-enhanced feature vectors and quality-enhanced feature vectors. The cosine similarity between the process-enhanced feature vector and the quality-enhanced feature vector is calculated to obtain the similarity values ​​between each pair of feature dimensions. All similarity values ​​are arranged in matrix form with process feature dimensions as rows and quality inspection feature dimensions as columns, ultimately yielding the correlation strength matrix.

[0079] In practical implementation, a correlation strength matrix between process parameters and quality inspection results is calculated based on the process feature set and image feature set. The process feature set is mapped to a fixed-dimensional process feature vector through a fully connected network. The process feature vector is a low-dimensional dense vector representation obtained after linear transformation and nonlinear activation function processing of the process feature set. The image feature set is mapped to a quality inspection feature vector of the same or different dimensions through another fully connected network. The quality inspection feature vector is a vector obtained after mapping the image feature set. A process-quality inspection correlation network is constructed, using both the process feature vector and the quality inspection feature vector as inputs. In some embodiments, the process-quality inspection correlation network is a multi-head attention network. The process-quality inspection correlation network calculates the attention weight between each feature dimension in the process feature vector and each feature dimension in the quality inspection feature vector through an internal mechanism. The attention weight represents the strength of the correlation between the process feature dimension and the quality inspection feature dimension. Based on the calculated attention weights, a bidirectional weighted interaction is performed on the process feature vector and the quality inspection feature vector. This bidirectional weighted interaction involves simultaneously weighting and correcting the process feature vector with the quality inspection feature vector, generating both process-enhanced and quality inspection-enhanced feature vectors. The process-enhanced feature vector is a process feature representation that incorporates quality inspection information. The cosine similarity between the process-enhanced and quality inspection-enhanced feature vectors is calculated. Cosine similarity measures the directional consistency of the process-enhanced and quality inspection-enhanced feature vectors across various dimensions, yielding similarity values ​​between each pair of feature dimensions. All similarity values ​​are arranged into a matrix with process feature dimensions as rows and quality inspection feature dimensions as columns, resulting in the association strength matrix. The elements in the association strength matrix... Indicates the first The first process feature dimension and the first The correlation strength between the various quality inspection feature dimensions can be calculated using the following formula:

[0080]

[0081] in: Represents the first element in the correlation strength matrix. Line number Column elements, The process-enhanced feature vector is represented in the th... Component vectors on each feature The quality inspection enhancement feature vector is represented in the th... The component vectors on each feature, symbol This represents the dot product operation of vectors. Representing vectors The length of the mold, Representing vectors The length of the module.

[0082] In practical implementation, the fully connected network mapping the process feature set contains two layers. The first layer converts the input dimension to 128 dimensions, and the second layer converts the 128 dimensions to 64 dimensions to output the process feature vector. The fully connected network mapping the image feature set has a similar structure, outputting the same 64-dimensional quality inspection feature vector. It can be understood that the parameters of the two fully connected networks are independently trained. The process of calculating the attention weights in the process-quality inspection association network is as follows: the process feature vector and the quality inspection feature vector are concatenated and then input into a feedforward neural network, which outputs an attention score matrix. Based on the attention score matrix, the original feature vectors are weighted and summed to generate process-enhanced feature vectors and quality inspection-enhanced feature vectors. In some embodiments, the attention scores are normalized using the Softmax function during weighted summation. The association strength matrix obtained after calculating the cosine similarity is a 64-row, 64-column real number matrix. Each row of the matrix represents the association pattern between a process feature and all quality inspection features. The optional association strength matrix is ​​persistently stored after generation for subsequent analysis modules to use. It is understandable that the training objective of the process-quality inspection association network is to minimize the reconstruction error, so that the process enhancement feature vector can better predict the original quality inspection feature vector, and the quality inspection enhancement feature vector can better predict the original process feature vector.

[0083] In one embodiment of the present invention, the set of key process parameters that have the greatest impact on quality inspection results and their association patterns with image features are identified based on the association strength matrix. Each row of the association strength matrix is ​​summed to calculate the total association strength of each process feature dimension with all quality inspection feature dimensions. Rows with total association strength exceeding a preset threshold are selected, representing the corresponding process feature dimensions, and these process feature dimensions are mapped back to the original set of process parameters to obtain a candidate set of key process parameters. For each parameter in the candidate set of key process parameters, its corresponding row vector is extracted from the association strength matrix, representing the association strength sequence between the parameter and each quality inspection feature dimension. Cluster analysis is performed on all extracted association strength sequences, grouping process parameters with similar association patterns into the same subset, and defining the common association pattern of each subset as an association pattern. For example, some parameters are strongly correlated with coating uniformity features, while others are strongly correlated with bubble features. The subsets of process parameters with high total association strength and whose association patterns are clear and explicit in the cluster analysis are determined as the final set of key process parameters, and the association pattern between each parameter subset in this set and specific image features is recorded.

[0084] In practical implementation, the key process parameter set that has the greatest impact on quality inspection results and its association pattern with image features are identified based on the correlation strength matrix. Each row of the correlation strength matrix is ​​summed to calculate the total correlation strength of each process feature dimension with all quality inspection feature dimensions. In some embodiments, the total correlation strength of the k-th process feature dimension is... Through the formula:

[0085]

[0086] in: This represents the total correlation strength of the k-th process feature dimension. The element in the k-th row and i-th column of the association strength matrix represents the association strength value. This represents the total number of quality inspection feature dimensions. The summation operation iterates through all elements in the k-th row of the correlation strength matrix. Process feature dimensions whose total correlation strength exceeds a preset threshold are selected. The preset threshold, which can be set to 0.85 based on the distribution of total correlation strength in historical data, is then mapped back to the original set of process parameters. This mapping process is completed by looking up the correspondence between process feature dimension numbers and original process parameter names in a table, thus obtaining a set of candidate key process parameters.

[0087] In practical implementation, for each parameter in the candidate key process parameter set, its correlation strength sequence with each quality inspection feature dimension is extracted from the correlation strength matrix. The correlation strength sequence is a vector, and each element in the vector represents the correlation strength value between the process parameter and a specific quality inspection feature dimension. Cluster analysis is performed on all extracted correlation strength sequences. The cluster analysis groups process parameters with similar correlation patterns into the same subset. It can be understood that the cluster analysis uses the K-means clustering algorithm, and the correlation pattern common to each subset is defined as a correlation pattern. For example, if the correlation strength sequences of process parameters in a subset all have high values ​​in the dimension corresponding to the "coating uniformity feature" and low values ​​in other dimensions, then the correlation pattern of this subset is defined as "strongly correlated with the coating uniformity feature". A subset of process parameters with high overall correlation strength and whose correlation patterns are clear and explicit in cluster analysis is identified as the final set of key process parameters. In some embodiments, a clear correlation pattern is defined as the average distance between the correlation strength sequences of all parameters within the subset being less than a set threshold. The correlation pattern between each parameter subset in the key process parameter set and specific image features is recorded. Optional recording methods include storing the parameter subset name, the list of included parameters, and a description of the correlation pattern in a structured list. Refer to Table 1, which illustrates a key process parameter identification and clustering result.

[0088] Table 1: Examples of Correlation Strength and Clustering Results for Key Process Parameters

[0089]

[0090] In practice, cluster analysis clearly identifies different subsets of parameters, each corresponding to a unique association pattern. For example, in Table 1, "Cluster 1" represents one association pattern where parameters are closely related to coating uniformity and bubble characteristics, while "Cluster 2" represents another, where parameters are primarily opposite to changes in wrinkle characteristics. The association pattern descriptions, in text form, summarize the main direction and strength of the association between the parameter subset and one or more quality inspection features. These pattern descriptions, along with the cluster center vectors of the association strength sequence, serve as prior knowledge for subsequent decision model construction.

[0091] See Figure 3This is a heatmap showing the correlation strength between process parameters and quality inspection characteristics. This chart visualizes the correlation strength between process parameters and quality inspection defect characteristics in the production management of high-peel lining fabric. The redder the color, the stronger the correlation; the bluer, the weaker the correlation. Coating thickness trend, tension stability, coating uniformity, and bubble characteristics show a very strong positive correlation, indicating that these two parameters are the core factors affecting surface smoothness and bubble defects. Temperature rise rate, steady-state temperature, and wrinkle characteristics show a very strong positive correlation, indicating that temperature control is crucial for wrinkle defects. Thickness stability, temperature fluctuation frequency, tension anomalies, and the three types of quality inspection characteristics all show moderate or low correlation strengths, belonging to secondary influencing parameters. Coating thickness trend, tension stability, temperature rise rate, and steady-state temperature can be identified as core process control parameters and prioritized for optimization.

[0092] In one embodiment of the present invention, a process adjustment decision model is constructed based on the identified set of key process parameters and their correlation patterns with image features. Using the set of image features from historical production data as input variables and the target values ​​of the optimized and validated set of key process parameters for that batch as output variables, a training sample set is constructed for model training. A multi-layer neural network is constructed as the main structure of the decision network. The number of nodes in the input layer of this network is consistent with the dimension of the image feature set, and the number of nodes in the output layer is consistent with the dimension of the key process parameter set. A correlation pattern constraint layer is introduced into the decision network. This layer applies prior constraints to the connection weights between the network's input layer and hidden or output layer based on the identified set of key process parameters and their correlation patterns with image features. The constructed training sample set is used to train this decision network with the correlation pattern constraint layer. The training process aims to minimize the error between the network's predicted optimization target value of the process parameters and the actual optimal value in the training samples. After training, the resulting network is the process adjustment decision model.

[0093] A correlation pattern constraint layer is introduced into the decision network. This layer imposes prior constraints on the network weights based on the set of key process parameters and their correlation patterns with image features. The correlation patterns between the set of key process parameters and image features are encoded into a connection weight mask matrix with the same dimension as the network weight matrix. The elements in this mask matrix indicate the allowed and prohibited connections between the input and output layers, with allowed connections set to 1 and prohibited connections set to 0. During the training of the decision network, before each weight update, the connection weight mask matrix is ​​multiplied element-wise with the actual connection weight matrix of the network, thereby forcing the weights corresponding to connections that do not exist in the correlation pattern to be reset to zero. A regularization term is added to the model's loss function. This regularization term is used to calculate the difference between the current network weights and the weight directions implied by the known correlation patterns, and penalizes weight changes with excessively large differences.

[0094] In practical implementation, a process adjustment decision model is constructed based on the identified set of key process parameters and their correlation patterns with image features. The image feature set from historical production data serves as the input variable, which is the extracted image feature vector. The target values ​​of the optimized and validated set of key process parameters from historical production batches serve as the output variable, which is the target value vector of the key process parameters. A training sample set is constructed for model training, where each sample is an input-output pair. A multi-layer neural network is constructed as the main structure of the decision network. The multi-layer neural network includes an input layer, at least one hidden layer, and an output layer. The number of nodes in the input layer of the decision network is consistent with the dimension of the image feature set, and the number of nodes in the output layer is consistent with the dimension of the key process parameter set. In some embodiments, the hidden layer is set to two layers, each with 128 nodes. A correlation pattern constraint layer is introduced into the decision network. This layer applies prior constraints to the connection weights between the network's input layer and the hidden or output layer based on the identified set of key process parameters and their correlation patterns with image features. The decision network with an introduced correlation pattern constraint layer is trained using the constructed training sample set. The training process aims to minimize the error between the network's predicted process parameter optimization target value and the actual optimal value in the training sample. The network obtained after training is the process adjustment decision model.

[0095] In practical implementation, an association pattern constraint layer is introduced into the decision network. This layer imposes prior constraints on the network weights based on the set of key process parameters and their association patterns with image features. The association patterns between the key process parameters and image features are encoded into a connection weight mask matrix with the same dimension as the network weight matrix. Elements in the connection weight mask matrix indicate allowed and prohibited connections between the input and output layers; allowed connections have a value of 1, and prohibited connections have a value of 0. During the training of the decision network, before each weight update, the connection weight mask matrix is ​​multiplied element-wise with the actual connection weight matrix of the network. This forces the weights corresponding to non-existent connections in the association pattern to be reset to zero. This operation ensures that the network learns only through known and meaningful association paths. A regularization term is added to the model's loss function. This term calculates the difference between the current network weights and the weight directions implied by the known association patterns and penalizes weight changes with excessively large differences. In some embodiments, the regularization term... Defined by the following formula:

[0096]

[0097] in: Indicates the weights applied to the weight matrix The regularization term on, It is a hyperparameter that controls the strength of regularization. The connection weight mask matrix is ​​in the th... Line number Column elements, The actual weight matrix of the decision network is at the th... Line number Column elements, These are weight reference values ​​pre-defined based on prior knowledge of association patterns, with symbols... This represents scalar multiplication, summing by traversing all elements of the weight matrix. It is the actual weight of the network. Compared with the weighted reference value preset based on historical knowledge The sum of weighted squared deviations between them. Refer to Table 2, which shows a simplified connection weight mask matrix, where rows correspond to image feature nodes in the input layer and columns correspond to key process parameter nodes in the output layer.

[0098] Table 2: Schematic Table of Connection Weight Mask Matrix

[0099]

[0100] In practical implementation, the construction of the connection weight mask matrix is ​​based on the correlation pattern between the set of key process parameters and image features. For example, as shown in Table 2, it is assumed that the correlation pattern indicates that "coating thickness trend feature" and "tension stability feature" are only related to "coating uniformity feature" and "bubble feature," but not to "wrinkle feature," while "temperature rise rate feature" is only related to "wrinkle feature." This selectable relationship is derived from the "strong correlation" and "irrelevance" information in the correlation pattern description. During training, the actual weight matrix of the decision network is multiplied element-wise with the mask matrix shown in Table 2, thereby ensuring that the weight connection from "wrinkle feature" to "coating thickness trend feature" or "tension stability feature" is forcibly prohibited (remained at 0). The regularization term in the loss function encourages the actual weights of the network. To the preset weight reference value Approaching weights close to each other, but only applying this constraint to weights that are allowed to connect (with a value of 1) within the mask matrix, allows us to understand the preset weight reference values. It can be initialized based on the historical average of the correlation strength.

[0101] See Figure 4This is a graph showing the change in training loss of a neural network over epochs, used to evaluate the model's convergence and training effectiveness. From epochs 0-20, the loss rapidly decreases from approximately 0.88 to approximately 0.38, with a steep slope, indicating that the model is quickly learning data patterns and gradient updates are effective. From epochs 20-70, the loss continues to decrease slowly, accompanied by minor oscillations, indicating that the model is gradually approaching the optimal solution while some gradient fluctuations exist. From epochs 70-100, the loss stabilizes near the convergence threshold, with smaller oscillations, indicating that the model has reached convergence, and further training offers limited gains. After approximately 70 epochs, the loss value stabilizes below the convergence threshold, the model has converged, and the training objective has been achieved. The first 20 epochs account for most of the loss reduction; subsequent epochs are mainly used for fine-tuning. The curve shows a continuous decline followed by a plateau, without exhibiting the overfitting characteristic of a "decline followed by rise."

[0102] In one embodiment of the present invention, a process adjustment decision model is used to process the real-time generated set of process parameters and the set of quality inspection images to generate a real-time set of process parameter adjustment instructions. The set of process parameter adjustment instructions includes a sequence of changes in target set values ​​for coating thickness, curing temperature, and winding tension. The generated set of process parameter adjustment instructions is sent to the programmable logic controller (PLC) of the production line, which drives the actuators such as the doctor blade or pump valve of the coating machine, the heater of the curing oven, and the motor of the winding machine to perform specific parameter adjustment actions.

[0103] A process adjustment decision model is used to process the real-time generated set of process parameters and quality inspection images to generate a real-time set of process parameter adjustment instructions. This includes real-time acquisition of the online process parameter set and online quality inspection image set obtained from the current production batch's online monitoring system. The aforementioned feature extraction and image recognition processes are performed on the online process parameter set and online quality inspection image set respectively to obtain a real-time process feature set and a real-time image feature set. The real-time image feature set is input into the trained process adjustment decision model, which infers and outputs a set of optimized process parameter target values ​​corresponding to the current quality inspection status based on the current quality characteristics. The optimized process parameter target value set is compared with the currently acquired online process parameter set to calculate the numerical difference that each parameter needs to be adjusted. Based on the calculated parameter differences, combined with the preset maximum allowable change (adjustment step size) and adjustment direction for each adjustment, a specific and executable set of process parameter adjustment instructions is generated.

[0104] The optimized set of target process parameters is compared with the current set of online process parameters to calculate the parameter differences that need adjustment. This includes aligning and comparing each target parameter value in the optimized set with the corresponding real-time parameter value in the current set of online process parameters over a time window. For each parameter, the absolute difference between the target parameter value and the real-time parameter value is calculated to obtain the initial difference value for that parameter. Considering the dynamic response characteristics of the production equipment, a correction factor is introduced for the initial difference value of each parameter. This correction factor is calculated based on the historical response speed data after adjustment and the overall load status of the current production line. The initial difference value of each parameter is multiplied by its corresponding correction factor to obtain the parameter difference used to finally generate the adjustment command.

[0105] In practical implementation, a process adjustment decision model is used to process the real-time generated set of process parameters and quality inspection images to generate a real-time set of process parameter adjustment instructions. This set of instructions includes a sequence of changes to target set values ​​for coating thickness, curing temperature, and winding tension. The generated set of process parameter adjustment instructions is then sent to the programmable logic controller (PLC) on the production line. The PLC drives actuators such as the doctor blades or pump valves of the coating machine, the heaters of the curing oven, and the motors of the winding machine to perform specific parameter adjustment actions. In some embodiments, the PLC receives the set of process parameter adjustment instructions encapsulated in JSON format via industrial Ethernet.

[0106] In practical implementation, a process adjustment decision model is used to process the real-time generated set of process parameters and quality inspection images to generate a real-time set of process parameter adjustment instructions. This includes the online process parameter set and online quality inspection image set obtained from the current production batch's online monitoring system. The online process parameter set includes instantaneous values ​​of coating thickness, real-time curing temperature curves, and real-time winding tension curves read from on-site sensors. The aforementioned feature extraction and image recognition processes are performed on the online process parameter set and online quality inspection image set, respectively, to obtain a real-time process feature set and a real-time image feature set. It can be understood that the feature extraction and image recognition processes use the exact same algorithms and parameters as in the model training phase. The real-time image feature set is input into the trained process adjustment decision model. Based on the current quality characteristics, the model infers and outputs a set of optimized process parameter target values ​​corresponding to the current quality inspection status. Each value in the optimized process parameter target value set represents the ideal value that a key process parameter should achieve. The optimized process parameter target value set is compared with the currently acquired online process parameter set to calculate the numerical difference that each parameter needs to be adjusted. Based on the calculated parameter differences, combined with the preset maximum allowable change for each adjustment (i.e., the adjustment step size and adjustment direction), a specific and executable set of process parameter adjustment instructions is generated. The selectable adjustment step size is set according to process safety specifications, and the adjustment direction is determined by the sign of the parameter differences.

[0107] In practical implementation, the optimized set of target process parameters is compared with the current set of online process parameters to calculate the parameter differences that need adjustment. This includes aligning and comparing each target parameter value in the optimized set with the corresponding real-time parameter value in the current set of online process parameters within a time window. This alignment ensures that the target parameter value and the real-time parameter value represent the production state over the same time period. For each parameter, the absolute difference between the target parameter value and the real-time parameter value is calculated to obtain the initial difference value. It can be understood that for time-series parameters such as curing temperature, both the target parameter value and the real-time parameter value are characteristic values, and the comparison is performed on the same characteristic dimension. Considering the dynamic response characteristics of the production process equipment, a correction factor is introduced for the initial difference value of each parameter. The correction factor is calculated based on the historical adjusted response speed data of the parameter and the overall load state of the current production line. In some embodiments, the correction factor... The calculation formula is:

[0108]

[0109] in: Indicates the parameter The correction factor Represents the standard response time constant. Indicates parameters The average response time of historical adjustments This indicates the baseline load status value of the production line. This represents the real-time measured value of the current production line load status. The initial difference value of each parameter is multiplied by its corresponding correction factor to obtain the parameter difference used to finally generate the adjustment command. For example, for the coating thickness parameter, the final parameter difference is calculated using the following formula:

[0110]

[0111] in: The final parameter difference representing the coating thickness parameter, This indicates the initial difference in coating thickness parameters. The final parameter difference can be further limited by adjusting the step size to prevent excessive adjustment in a single step.

[0112] See Figure 5 This is a time-series analysis chart of the core parameters for high-peel lining production, showing the dynamic trends of the three core process parameters—coating thickness, curing temperature, and winding tension—over 100 time points. It intuitively reflects the parameter stability and control logic of the production process. The curing temperature initially decreased slightly and then stabilized, before rebounding after 80 rolls. The winding tension was initially stable, then decreased slightly in the middle stage, and then rebounded in the later stage. The coating thickness was almost a straight line around the zero axis. None of the three parameters experienced drastic fluctuations. The curing temperature and coating thickness were controlled with high precision, and the winding tension was smoothly adjusted, demonstrating strong controllability of the production process. The curing temperature, as the core variable, remained within a reasonable production range throughout the process; the winding tension was fine-tuned over time without negatively impacting coating quality. The slight rebound in parameters in the later stage may be due to the system's dynamic adaptive adjustment to the production status, reflecting the flexibility of closed-loop control. Special attention should be paid to the tension low point near the 60th roll to investigate whether there are changes in the winding roller load or fluctuations in equipment resistance.

[0113] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A high-peelability interlining production management method based on artificial intelligence, characterized in that: The method includes: Collect the set of process parameters collected in real time from the lining production line, and simultaneously collect the set of quality inspection images generated by the online detection system of the production line; The process parameter set is subjected to feature extraction processing to obtain a process feature set, and the quality inspection image set is subjected to image recognition processing to extract an image feature set, wherein the image feature set includes coating uniformity features, bubble features and wrinkle features. Based on the process feature set and the image feature set, a correlation strength matrix between process parameters and quality inspection results is calculated. The correlation strength matrix is ​​used to characterize the quantitative relationship between each parameter sequence in the process parameter set and each feature in the image feature set. Based on the correlation strength matrix, the set of key process parameters that have the greatest impact on quality inspection results and their correlation patterns with image features are identified. Based on the identified set of key process parameters and their correlation patterns with image features, a process adjustment decision model is constructed. This model is used to deduce optimized recommended values ​​for process parameters based on the real-time set of image features.

2. The high-peelability interlining production management method based on artificial intelligence according to claim 1, characterized in that, Perform feature extraction processing on the set of process parameters to obtain a set of process features, including: The set of process parameters includes coating thickness data, curing temperature timing data, and winding tension timing data; The coating thickness data is subjected to fluctuation analysis to extract coating thickness trend characteristics and thickness stability indicators; Periodic pattern decomposition was performed on the curing temperature time series data to separate steady-state temperature characteristics, temperature rise rate characteristics, and temperature fluctuation frequency characteristics. The winding tension time series data is subjected to abnormal peak detection and smoothness calculation to obtain tension anomaly characteristics and tension stability characteristics; The extracted coating thickness trend features, thickness stability index, steady-state temperature features, temperature rise rate features, temperature fluctuation frequency features, tension anomaly features, and tension stability features are normalized and vectorized to form the process feature set.

3. The high-peelability interlining production management method based on artificial intelligence according to claim 2, characterized in that, Image recognition processing is performed on the quality inspection image set to extract image feature sets, including: The quality inspection image set includes images of the lining surface; The surface image of the lining fabric is preprocessed, including image enhancement, noise removal, and region segmentation, to obtain a standardized surface analysis image; The standardized surface analysis image is input into a convolutional neural network, and deep image features are extracted through multiple convolutional and pooling layers to obtain a basic image feature map; Multi-scale feature fusion is performed on the base image feature map to generate a fused feature map. The multi-scale feature fusion includes extracting features of different scales from the base image feature map and connecting them. On the fused feature map, a pre-trained defect detection head is applied to locate and identify uneven coating regions, bubble regions, and wrinkle regions, and to calculate the geometric and texture features of each region, forming the image feature set that includes coating uniformity features, bubble features, and wrinkle features.

4. The high-peelability interlining production management method based on artificial intelligence according to claim 3, characterized in that, Based on the aforementioned process feature set and image feature set, the correlation strength matrix between process parameters and quality inspection results is calculated, including: The process feature set is mapped to a process feature vector through a fully connected network, and the image feature set is mapped to a quality inspection feature vector through a fully connected network. Construct a process-quality inspection association network, taking the process feature vector and the quality inspection feature vector as input, and calculate the attention weight between each feature dimension in the process feature vector and each feature dimension in the quality inspection feature vector; Based on the calculated attention weights, the process feature vector and the quality inspection feature vector are subjected to bidirectional weighted interaction to generate process enhancement feature vector and quality inspection enhancement feature vector. Calculate the cosine similarity between the process enhancement feature vector and the quality inspection enhancement feature vector, and arrange them into a matrix to obtain the correlation strength matrix. The rows of the correlation strength matrix correspond to the process feature dimension, and the columns correspond to the quality inspection feature dimension.

5. The high-peelability interlining production management method based on artificial intelligence according to claim 4, characterized in that, Based on the correlation strength matrix, the set of key process parameters that have the greatest impact on quality inspection results and their correlation patterns with image features are identified, including: Sum each row of the correlation strength matrix to calculate the total correlation strength of each process feature dimension with all quality inspection feature dimensions; The process feature dimensions that exceed a preset threshold are selected, and the process feature dimensions are mapped back to the original set of process parameters to obtain a set of candidate key process parameters. For each parameter in the candidate key process parameter set, extract its correlation strength sequence with each quality inspection feature dimension from the correlation strength matrix; Cluster analysis is performed on the correlation strength sequences to identify subsets of process parameters with similar correlation patterns, and the correlation pattern of each subset is defined as a correlation pattern. A subset of process parameters with high overall correlation strength and clear correlation patterns is identified as the set of key process parameters, and their correlation patterns with specific image features are recorded.

6. The high-peelability interlining production management method based on artificial intelligence according to claim 5, characterized in that, The process adjustment decision model is constructed based on the identified set of key process parameters and their correlation patterns with image features, including: A training sample set is constructed using the image feature set as the input variable and the optimization target value of the key process parameter set as the output variable. A multi-layer neural network is constructed as the main structure of the decision network, with the number of input layer nodes matching the dimension of the image feature set and the number of output layer nodes matching the dimension of the key process parameter set. An association pattern constraint layer is introduced into the decision network. The association pattern constraint layer applies prior constraints to the network weights based on the set of key process parameters and their association patterns with image features. The decision network with the introduced correlation pattern constraint layer is trained using the training sample set. The training objective is to minimize the error between the predicted process parameter optimization target value and the actual optimal value. After training, the process adjustment decision model is obtained.

7. The high-peelability interlining production management method based on artificial intelligence according to claim 6, characterized in that, A correlation pattern constraint layer is introduced into the decision network. This layer imposes prior constraints on the network weights based on the set of key process parameters and their correlation patterns with image features, including: The association pattern between the set of key process parameters and image features is encoded into a connection weight mask matrix, which indicates the relationship between allowed and prohibited connections between the input layer and the output layer. During the training process of the decision network, the connection weight mask matrix is ​​multiplied element-wise with the actual connection weight matrix of the network to forcibly prohibit connection relationships that do not exist in the association pattern. A regularization term is added to the loss function, which is used to penalize weight changes that are inconsistent with the direction of known association patterns.

8. The high-peelability interlining production management method based on artificial intelligence according to claim 7, characterized in that, Also includes: The process adjustment decision model is used to process the real-time generated set of process parameters and the set of quality inspection images to generate a real-time set of process parameter adjustment instructions. The set of process parameter adjustment instructions includes a sequence of changes in target set values ​​for coating thickness, curing temperature and winding tension. The set of process parameter adjustment instructions is sent to the programmable logic controller of the production line to drive the coating machine, curing oven and winding machine to perform parameter adjustments; The process adjustment decision model is used to process the real-time generated set of process parameters and the set of quality inspection images to generate a real-time set of process parameter adjustment instructions, including: Real-time acquisition of online process parameters and online quality inspection images for the current production batch; Feature extraction and image recognition processes are performed on the online process parameter set and the online quality inspection image set, respectively, to obtain the real-time process feature set and the real-time image feature set. The real-time image feature set is input into the trained process adjustment decision model, and the model outputs a set of optimized process parameter target values ​​corresponding to the current quality inspection status. The set of target values ​​for the optimized process parameters is compared with the current set of online process parameters to calculate the differences in parameters that need to be adjusted. Based on the differences in the parameters, and combined with the preset adjustment step size and direction, a specific and executable set of process parameter adjustment instructions is generated.

9. The high-peelability interlining production management method based on artificial intelligence according to claim 8, characterized in that, The optimized set of process parameter target values ​​is compared with the current set of online process parameters to calculate the parameter differences that need to be adjusted, including: Each target parameter value in the optimized process parameter target value set is aligned and compared with the corresponding real-time parameter value in the current online process parameter set; For each parameter, calculate the absolute difference between the target parameter value and the real-time parameter value to obtain the preliminary difference value of the parameter; Considering the dynamic response characteristics of the production process, a correction factor is introduced for the initial difference value of each parameter. The correction factor is calculated based on the historical response speed of the parameter adjustment and the current load state of the production line. The final parameter difference is obtained by multiplying the initial difference value of each parameter by its corresponding correction factor.

10. A high-peelability interlining production management system based on artificial intelligence, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the artificial intelligence-based high-peeling lining production management method according to any one of claims 1 to 9.

Images