Top-level project design method for printing press based on multivariate data

Abstract

The present invention provides a top-level project design method for a printing press based on multivariate data, which relates to the technical field of printing equipment. The method comprises the following steps: obtaining printing press operation data including printing quality parameters, equipment operating parameters and environmental parameters; performing feature analysis and parameter extraction by using an adaptive fuzzy clustering algorithm, dynamically adjusting sample classification membership, identifying key parameter combinations affecting printing quality and constructing a correlation matrix; constructing a dynamic Bayesian probability network model based on a parameter mapping relationship, constructing a core network structure by using information entropy and information gain, and establishing a probabilistic association between printing quality parameters and equipment operating parameters; utilizing a bidirectionally constrained dynamic programming algorithm, combining coupling weights between functional modules and state transition equations, determining the optimal parameter configuration of the functional modules of the printing press, and generating a top-level design scheme.

Description

Technical Field

[0001] The present invention relates to the technical field of printing equipment, and in particular to a top-level project design method for a printing machine based on multivariate data. Background Art

[0002] With the rapid development of the printing industry, market demand for customized printing equipment is growing. The top-level design of a printing press requires considering multiple factors, including the technical specifications of existing equipment and the specific needs of users. Currently, when developing new product projects, printing press manufacturers primarily rely on experience to select parameters and assign weights to requirements, lacking a scientific and systematic decision-making support approach.

[0003] However, the existing technology still has shortcomings. Traditional parameter analysis methods are difficult to effectively handle the complex correlation between multi-source heterogeneous data of printing machines, and cannot accurately identify the key parameter combinations that affect printing quality; most of the existing quality prediction models use static modeling methods, which fail to fully consider the dynamic coupling characteristics between parameters, resulting in low prediction accuracy; the parameter configuration optimization of printing machine functional modules usually adopts a one-way constraint method, without comprehensively considering the mutual influence between modules and quality target requirements, making it difficult to obtain the globally optimal design solution.

[0004] In summary, the present invention aims to solve the above-mentioned technical problems and proposes a top-level project design method for printing presses based on multivariate data analysis. An adaptive fuzzy clustering algorithm is used to analyze the characteristics of printing press operating data and identify key parameters. A dynamic Bayesian probability network is combined to establish a dynamic correlation model between parameters. A bidirectionally constrained dynamic programming algorithm is used to optimize the allocation of parameter weights, achieving a scientific weight configuration for equipment parameters and user needs. This method maintains the ability to deeply mine the multi-source heterogeneous data of printing presses while enhancing the adaptability of the top-level design solution to actual user needs, thereby improving the reliability and practical value of the top-level design of printing presses. Summary of the Invention

[0005] The embodiment of the present invention provides a top-level project design method for a multivariate data printing machine, which can solve the problems in the prior art.

[0006] A top-level project design method for a multivariate data printing machine is provided, including:

[0007] Acquiring printing press operating data, wherein the printing press operating data includes printing quality parameters, equipment operating condition parameters, and environmental parameters;

[0008] Adopting an adaptive fuzzy clustering algorithm to perform feature analysis and parameter extraction on the printing press operation data, by determining the number of clusters, dynamically adjusting the sample classification membership, and identifying the key parameter combination that affects the printing quality;

[0009] Constructing a correlation matrix between the key parameter combinations;

[0010] Based on the parameter mapping relationship in the correlation matrix, a dynamic Bayesian probability network model is constructed, a core network structure is constructed by calculating the information entropy value and information gain, and the conditional probability distribution is optimized using the log-likelihood expectation value. A probabilistic association is established between the printing quality parameters and the equipment operating condition parameters to generate a printing quality prediction result;

[0011] Based on the correlation matrix and the printing quality prediction results, a bidirectional constraint dynamic programming algorithm is used to construct a state transfer equation through the coupling weights between functional modules. Combined with the bidirectional constraints, the optimal parameter configuration of the printing press functional modules is determined to generate a top-level design plan for the printing press.

[0012] In an optional embodiment,

[0013] An adaptive fuzzy clustering algorithm is used to perform feature analysis and parameter extraction on the printing press operation data. By determining the number of clusters and dynamically adjusting the sample classification membership, the key parameter combinations that affect printing quality are identified, including:

[0014] The printing press operation data is segmented by a sliding window method, and a feature vector is determined based on calculations of each window;

[0015] Calculate the minimum description length of the eigenvector to determine the target cluster number interval, calculate the fuzzy partition coefficient value and partition entropy value within the target cluster number interval, select the cluster number corresponding to the maximum fuzzy partition coefficient value and the minimum partition entropy value, and determine the final cluster number;

[0016] Constructing an initial value of the sample classification membership, calculating the distribution parameters of each dimension of the feature vector, and obtaining the sample classification membership value through dynamic adjustment based on the distribution parameters;

[0017] The cluster center coordinates are calculated based on the sample classification membership values, and the spatial position relationship strength between samples is calculated as a penalty factor to construct the objective function;

[0018] Iteratively update the sample classification membership value and cluster center coordinates according to the final number of clusters, and determine the clustering result by minimizing the objective function;

[0019] Analyzing the characteristic patterns of each cluster center based on the clustering results, calculating the contribution of each feature in distinguishing different categories using a statistical discrimination ratio, and selecting a feature combination with a statistical discrimination ratio greater than a preset threshold as a key parameter combination for printing quality;

[0020] The key parameter combination is input into a support vector machine classifier for cross-validation to verify the validity of the key parameter combination.

[0021] In an optional embodiment,

[0022] According to the distribution parameters, obtaining the sample classification membership value through dynamic adjustment includes:

[0023] The local density index is obtained by calculating the mean Euclidean distance between the sample point and the nearest neighboring sample point. The distance deviation value is calculated by combining the standard deviation of the distance sequence from the sample point to the cluster center, and the local reliability value is determined. An adaptive weight factor is generated. Based on whether the local reliability value is lower than the threshold, an exponential decay coefficient is introduced to adjust the sample classification membership value. Specifically, the following steps are performed:

[0024] For each sample point, multiple nearest neighbor sample points are selected, the Euclidean distances between the sample point and the nearest neighbor sample points are calculated, and the mean of the Euclidean distances is used as the local density index of the sample point; the Euclidean distances from each sample point to each cluster center are calculated to obtain a distance sequence, and the standard deviation of the distance sequence is calculated as the distance deviation value of the sample point; the local density index is divided by the distance deviation value to obtain the local reliability value of the sample point;

[0025] generating an adaptive weight factor based on the local reliability value and a preset adjustment coefficient;

[0026] Multiplying the adaptive weight factor by the initial value of the sample classification membership to obtain an adjusted value of the sample classification membership;

[0027] When the local reliability value is lower than the preset reliability threshold, the exponential decay coefficient is multiplied by the sample classification membership adjustment value to obtain the sample classification membership value; when the local reliability value is not lower than the reliability threshold, the sample classification membership adjustment value is used as the sample classification membership value; wherein the exponential decay coefficient specifically refers to:

[0028] ;

[0029] in, D represents the exponential decay coefficient, represents the reliability decay rate, T r represents the reliability threshold, R l Represents the local reliability value.

[0030] In an optional embodiment,

[0031] The contribution of each feature in distinguishing different categories is calculated using the statistical discrimination ratio. Feature combinations with a statistical discrimination ratio greater than a preset threshold are selected as key parameters for printing quality. The combinations include:

[0032] Calculate the inter-class scatter matrix and the intra-class scatter matrix to determine the statistical discrimination ratio sequence of the features; at the same time, calculate the mutual information matrix between the features to determine the redundant features of the labels; through a progressive strategy, starting from the maximum statistical discrimination ratio feature, by iteratively calculating the statistical discrimination ratio gain value, gradually construct a feature combination, specifically including:

[0033] The inter-class dispersion matrix is ​​determined by calculating the difference between the mean values ​​of samples of different categories, and the intra-class dispersion matrix is ​​determined by calculating the difference between the sample in the same category and the category mean value of the corresponding category. Based on the inter-class dispersion matrix and the intra-class dispersion matrix, the statistical discrimination ratio of each feature is calculated to obtain the feature statistical discrimination ratio sequence.

[0034] Calculate the mutual information value between the features to obtain a feature mutual information matrix. When the mutual information value is greater than a first preset threshold, mark the feature whose statistical discrimination ratio in the corresponding feature pair is less than a second preset threshold as a redundant feature; select the feature with the largest statistical discrimination ratio and not marked as a redundant feature from the feature statistical discrimination ratio sequence as the initial key feature;

[0035] The features that are not marked as redundant features and not selected as initial key features are sorted from large to small according to the statistical discrimination ratio to obtain a candidate feature sequence; the features in the candidate feature sequence are added to the existing key feature combination one by one, and the statistical discrimination ratio of the combination before and after the addition is calculated to obtain the statistical discrimination ratio gain value;

[0036] When the statistical discrimination ratio gain value is greater than a third preset threshold, adding the corresponding candidate feature to the key feature combination;

[0037] Repeat until all features in the candidate feature sequence are processed and the final key feature combination is obtained.

[0038] In an optional embodiment,

[0039] Calculate the inter-class scatter matrix and the intra-class scatter matrix and determine the statistical discriminant ratio sequence of the features including:

[0040] Calculate the number of samples in each category in the multi-category sample data, perform normalization processing to obtain the category weight vector, and construct the category weight matrix;

[0041] Calculating the category mean vector of each category of samples in the multi-category sample data and the overall mean vector; calculating the difference between the category mean vector and the overall mean vector to construct a category mean difference matrix; multiplying the category mean difference matrix with the category weight matrix to obtain an inter-class dispersion matrix;

[0042] Calculate the difference between each category sample and the corresponding category mean vector in the multi-category sample data to construct a sample difference matrix; multiply the sample difference matrix by the category weight matrix to obtain an intra-category dispersion matrix;

[0043] Extracting the diagonal elements of the inter-class scatter matrix and the diagonal elements of the intra-class scatter matrix to determine the inter-class scatter vector and the intra-class scatter vector respectively;

[0044] The ratio of the inter-class scatter vector to the intra-class scatter vector is calculated to obtain a statistical discriminant ratio sequence of features.

[0045] In an optional embodiment,

[0046] Based on the parameter mapping relationship in the correlation matrix, a dynamic Bayesian probability network model is constructed. The core network structure is constructed by calculating the information entropy value and information gain. The conditional probability distribution is optimized using the log-likelihood expectation value. The printing quality parameters are probabilistically associated with the equipment operating condition parameters to generate the printing quality prediction results, including:

[0047] Based on the dynamic Bayesian probability network, the initial network structure is constructed through the correlation matrix. The information entropy value is used to calculate the information gain, and the parameter nodes are screened to build the core network structure. The parameter prediction is achieved through the optimization of the log-likelihood expectation value and the real-time calculation of the information gain. Specifically,

[0048] Establishing connections between parameter pairs based on correlation coefficients of the parameter pairs in the correlation matrix; determining the direction of the connections based on the temporal characteristics of the parameter pairs, and constructing an initial network structure;

[0049] Calculate the information entropy value of each parameter node in the initial network structure, construct a parameter importance evaluation index based on the information entropy value, and calculate the information gain of each parameter node for the prediction result; sort the parameter nodes according to the information gain, select the parameter nodes with information gain greater than the dynamic threshold to construct a core network structure, obtain the corresponding parent node set for the parameter nodes in the core network structure, and construct a conditional probability distribution based on the parent node set;

[0050] Calculating a log-likelihood expectation value based on the core network structure and the conditional probability distribution, maximizing the log-likelihood expectation value through iterative optimization, and obtaining an optimized conditional probability distribution;

[0051] Calculate the parameter node information gain at the current sampling moment, update the parameter composition of the core network structure according to the change in information gain, and construct a new core network structure; based on the new core network structure and the corresponding conditional probability distribution, calculate the parameter prediction probability at the next sampling moment;

[0052] The parameter value corresponding to the maximum parameter prediction probability is selected as the printing quality prediction result.

[0053] In an optional embodiment,

[0054] Based on the correlation matrix and the printing quality prediction results, a bidirectionally constrained dynamic programming algorithm is used to construct a state transition equation through the coupling weights between functional modules. The optimal parameter configuration of the printing press functional modules is determined in combination with the bidirectional constraints, and a top-level design scheme for the printing press is generated, including:

[0055] Calculating coupling weights between functional modules based on the correlation matrix, wherein the coupling weights represent the degree of mutual influence between different functional modules;

[0056] Constructing a state transfer equation based on the coupling weight, and multiplying and summing the parameter state of the previous function module with the corresponding coupling weight through the state transfer equation to obtain the parameter state of the current function module;

[0057] Establishing bidirectional constraints based on the print quality prediction results, the bidirectional constraints including forward constraints and backward constraints, the forward constraints limiting the adjustment range of parameters corresponding to the current functional module based on the parameter state of the preceding functional module; and the backward constraints ensuring that subsequent functional modules achieve preset quality targets based on the print quality prediction results;

[0058] According to the forward constraints, each functional module is traversed forward, the local optimal solution is calculated and the state transition path is recorded to obtain a state transition sequence; according to the backward constraints, starting from the last functional module, the state transition sequence is traced backward to determine the optimal parameter configuration of the functional module;

[0059] The optimal parameter configuration is input into a printing quality prediction model for verification. When the printing quality prediction result obtained by verification does not meet the requirements, the bidirectional constraint condition is adjusted and the parameter configuration optimization is re-executed.

[0060] In an embodiment of the present invention, by collecting multidimensional data such as printing quality, equipment operating conditions and environment, and adopting an adaptive fuzzy clustering algorithm to realize dynamic sample classification and key parameter combination identification, data feature extraction is made more accurate; cluster analysis is used to determine the core parameters affecting printing quality, and a correlation matrix is ​​constructed, which can reveal the intrinsic mapping relationship between each parameter, help designers intuitively understand the interaction and influence mechanism between multiple factors, and effectively reduce design uncertainty; a dynamic Bayesian probability network model is constructed, and the conditional probability distribution is optimized by information entropy, information gain and log-likelihood expectation value to realize the probabilistic association between printing quality and equipment operating parameters, thereby accurately predicting printing quality and improving the system prediction ability and response speed; based on the correlation matrix and prediction results, a state transfer equation is constructed using a bidirectional constraint dynamic programming algorithm, and the optimal parameter configuration is determined from the perspective of functional module coupling weights, so as to realize the overall coordination and optimization of each functional module of the printing machine, and finally generate a scientific and reasonable top-level design scheme, reducing trial and error costs and improving design efficiency. The present invention integrates multi-dimensional information in a data-driven manner and adopts advanced algorithms to help achieve intelligent decision-making and system optimization under complex working conditions, thereby improving the overall performance and product quality of the printing press, and promoting the research and development and design of printing equipment to a higher level. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is a flowchart of a top-level project design method for a multivariate data printing machine according to an embodiment of the present invention. DETAILED DESCRIPTION

[0062] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0063] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0064] Figure 1 FIG. 1 is a flow chart of a top-level project design method for a multivariate data printer according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0065] Acquiring printing press operating data, wherein the printing press operating data includes printing quality parameters, equipment operating condition parameters, and environmental parameters;

[0066] The sensor collection system installed on the printing press acquires real-time operating data from the printing press. Printing quality parameters include registration accuracy, ink density, dot gain, print contrast, ink layer thickness, and image clarity; equipment operating parameters include printing speed, paper tension, impression pressure, ink fountain opening, feed roller speed, and transmission gear clearance; and environmental parameters include ambient temperature, humidity, dust concentration, and air velocity.

[0067] Specifically, the collection of printing quality parameters is achieved in the following ways: installing a CCD camera at the paper output end of the printing press to capture the printed image, and calculating the registration accuracy through image processing algorithms; using a spectrophotometer to measure the ink density and printing contrast of the printed product; using a microscope to measure the dot size and calculate the dot gain value; using an ink layer thickness gauge to measure the ink layer thickness; and evaluating the clarity of the image and text through image analysis software.

[0068] The equipment operating parameters are collected in the following ways: using a photoelectric encoder to measure the printing speed and the speed of the paper feed wheel; using a tension sensor to detect the paper tension; monitoring the printing pressure through a pressure sensor; using a displacement sensor to measure the ink fountain opening; and using a clearance meter to measure the transmission gear clearance.

[0069] The collection of environmental parameters is achieved through the following methods: installing temperature and humidity sensors around the printing press to monitor the ambient temperature and humidity in real time; using a dust concentration detector to measure the dust content in the air; and using an anemometer to measure the air flow rate in the printing environment.

[0070] All collected data is transmitted to a data acquisition server via industrial Ethernet and stored with unified time tags. The data sampling period is 100ms, and the data is normalized and outliers are removed every hour to generate a standardized data set for subsequent analysis.

[0071] The collected data is stored in XML format. The specific data structure includes fields such as data collection timestamp, parameter type identifier, parameter value, data unit, and collection device number. By establishing a parameter index table, fast retrieval and correlation analysis of different types of parameters can be achieved.

[0072] To ensure data reliability, the system also establishes a data quality assessment mechanism: regular calibration of sensors; setting parameter change range limits to determine abnormal values; adopting a data redundancy backup strategy; and establishing a sensor fault diagnosis and data compensation mechanism.

[0073] In an optional embodiment, an adaptive fuzzy clustering algorithm is used to perform feature analysis and parameter extraction on the printing press operation data. By determining the number of clusters and dynamically adjusting the sample classification membership, the key parameter combinations affecting printing quality are identified, including:

[0074] The printing press operation data is segmented by a sliding window method, and a feature vector is determined based on calculations of each window;

[0075] Calculate the minimum description length of the eigenvector to determine the target cluster number interval, calculate the fuzzy partition coefficient value and partition entropy value within the target cluster number interval, select the cluster number corresponding to the maximum fuzzy partition coefficient value and the minimum partition entropy value, and determine the final cluster number;

[0076] Constructing an initial value of the sample classification membership, calculating the distribution parameters of each dimension of the feature vector, and obtaining the sample classification membership value through dynamic adjustment based on the distribution parameters;

[0077] The cluster center coordinates are calculated based on the sample classification membership values, and the spatial position relationship strength between samples is calculated as a penalty factor to construct the objective function;

[0078] Iteratively update the sample classification membership value and cluster center coordinates according to the final number of clusters, and determine the clustering result by minimizing the objective function;

[0079] Analyzing the characteristic patterns of each cluster center based on the clustering results, calculating the contribution of each feature in distinguishing different categories using a statistical discrimination ratio, and selecting a feature combination with a statistical discrimination ratio greater than a preset threshold as a key parameter combination for printing quality;

[0080] The key parameter combination is input into a support vector machine classifier for cross-validation to verify the validity of the key parameter combination.

[0081] In a specific embodiment,

[0082] Step 1: Data Segmentation and Feature Extraction: The printing press operating data was segmented using a sliding window method. The window length was set to 1 hour, with a sliding step of 10 minutes. Statistical features, including mean, standard deviation, peak value, skewness, and kurtosis, were calculated for the data within each window to form a feature vector.

[0083] Step 2: Determine the optimal number of clusters. First, calculate the minimum description length of the feature vector and preliminarily determine the target number of clusters to be between 2 and 10. Within this range, calculate the fuzzy partition coefficient and partition entropy for each number of clusters. A larger fuzzy partition coefficient indicates a clearer clustering result, while a smaller partition entropy indicates a more stable clustering result. Through comparative analysis, the number of clusters corresponding to the maximum fuzzy partition coefficient and the minimum partition entropy is selected as the final number of clusters.

[0084] Step 3: Initialize and dynamically adjust membership. Initial values ​​for sample classification membership are constructed using a random method, and distribution parameters such as the mean and variance of each dimension of the feature vector are calculated. Local density metrics are calculated based on the distance relationship between the sample point and its nearest neighbor. Dynamically adjust the sample classification membership value based on the distance distribution from the sample point to each cluster center.

[0085] Step 4: Construct and optimize the objective function. The coordinates of the cluster centers for each category are calculated based on the sample classification membership values. The Euclidean distance between sample points is calculated to determine the strength of the spatial relationship between samples, which is used as a penalty factor to construct the objective function. The objective function comprehensively considers the distance between the sample and the cluster center and the spatial relationship between samples.

[0086] Step 5: Iteratively optimize the clustering results. Based on the final number of clusters determined, alternately update the sample classification membership values ​​and cluster center coordinates. By minimizing the objective function, the clustering results are continuously optimized until convergence conditions are reached or the maximum number of iterations is reached.

[0087] Step 6: Feature Contribution Analysis: Analyze the characteristic patterns of each cluster center and calculate the statistical discrimination ratio of each feature in distinguishing different categories. A larger statistical discrimination ratio indicates a greater contribution of the feature to class differentiation. Feature combinations with a statistical discrimination ratio greater than 0.8 are selected as key parameters for print quality.

[0088] Step 7: Validation: Input the identified key parameter combinations into the support vector machine classifier and verify their validity using a 10-fold cross-validation method. The reliability of the key parameter combinations is evaluated by calculating indicators such as classification accuracy, precision, and recall.

[0089] For example, taking the data from a certain run of a printing press as an example, 480 samples were obtained after sliding window segmentation. The optimal number of clusters was calculated to be four. By dynamically adjusting the sample classification membership, the final clustering results showed that the first category mainly represents the normal printing state, the second category reflects the registration deviation state, the third category corresponds to the abnormal ink color state, and the fourth category represents the overall quality abnormal state. Statistical discriminant ratio analysis identified eight key parameters, including ink fountain opening, printing pressure, paper tension, and ambient humidity. These parameters were input into the support vector machine classifier for cross-validation, and the classification accuracy reached 92.5%, verifying the effectiveness of the key parameter combination. These key parameters were then used to construct a correlation matrix, providing a basis for subsequent quality prediction and parameter optimization.

[0090] In this embodiment, the printing press operation data is finely segmented through the sliding window method, and key statistical features (such as mean, standard deviation, peak, etc.) are extracted, effectively capturing the time-varying characteristics of the data and providing high-quality feature data for subsequent clustering and analysis; through the optimization analysis of the fuzzy partition coefficient and the partition entropy value, the optimal number of clusters is accurately determined, making the clustering results clearer and more stable, and improving the accuracy and interpretability of data classification; by dynamically adjusting the sample classification membership, combining the local density and the distance relationship between the sample and the cluster center, the clustering effect is optimized, and the adaptability and classification accuracy of the clustering model to different data types are enhanced; through feature contribution analysis and support vector machine verification, the key parameter combination with the greatest impact on printing quality is successfully identified, and its high accuracy in actual classification is verified, providing a reliable basis for quality prediction and parameter optimization.

[0091] In an optional embodiment, obtaining the sample classification membership value through dynamic adjustment according to the distribution parameter includes:

[0092] The local density index is obtained by calculating the mean Euclidean distance between the sample point and the nearest neighboring sample point. The distance deviation value is calculated by combining the standard deviation of the distance sequence from the sample point to the cluster center, and the local reliability value is determined. An adaptive weight factor is generated. Based on whether the local reliability value is lower than the threshold, an exponential decay coefficient is introduced to adjust the sample classification membership value. Specifically, the following steps are performed:

[0093] For each sample point, multiple nearest neighbor sample points are selected, the Euclidean distances between the sample point and the nearest neighbor sample points are calculated, and the mean of the Euclidean distances is used as the local density index of the sample point; the Euclidean distances from each sample point to each cluster center are calculated to obtain a distance sequence, and the standard deviation of the distance sequence is calculated as the distance deviation value of the sample point; the local density index is divided by the distance deviation value to obtain the local reliability value of the sample point;

[0094] generating an adaptive weight factor based on the local reliability value and a preset adjustment coefficient;

[0095] Multiplying the adaptive weight factor by the initial value of the sample classification membership to obtain an adjusted value of the sample classification membership;

[0096] When the local reliability value is lower than the preset reliability threshold, the exponential decay coefficient is multiplied by the sample classification membership adjustment value to obtain the sample classification membership value; when the local reliability value is not lower than the reliability threshold, the sample classification membership adjustment value is used as the sample classification membership value; wherein the exponential decay coefficient specifically refers to:

[0097] ;

[0098] in,D represents the exponential decay coefficient, represents the reliability decay rate, T r represents the reliability threshold, R l Represents the local reliability value.

[0099] In a specific embodiment,

[0100] Step 1: Calculate the local density index. For each sample point, select the five nearest neighboring sample points as its nearest neighbors. Calculate the Euclidean distance from the current sample point to these five nearest neighbors, and use the average of these distance values ​​as the local density index for that sample point. A smaller local density index indicates a denser sample distribution in the area.

[0101] Step 2: Calculate the distance deviation value. Calculate the Euclidean distance from each sample point to all cluster centers to obtain a distance sequence. Calculate the standard deviation of this distance sequence and use it as the distance deviation value for the sample point. A larger distance deviation value indicates a more significant difference in the distance between the sample point and the cluster centers.

[0102] Step 3: Determine the local reliability value. Divide the local density index of the sample point by its distance deviation value to obtain the local reliability value. This value reflects the distribution characteristics of the sample point in the local area and the reliability of the classification. The higher the local reliability value, the more reliable the classification of the sample point.

[0103] Step 4: Generate an adaptive weight factor: Based on the local reliability values ​​of the sample points and a pre-set adjustment coefficient of 0.8, an adaptive weight factor is generated. The adjustment coefficient is used to control the adjustment range of the weight factor to prevent over-adjustment.

[0104] Step 5: Calculate the membership adjustment value: Multiply the adaptive weight factor by the initial membership value of the sample to obtain the membership adjustment value. This step implements the initial adjustment based on local features.

[0105] Step 6: Introduce exponential decay adjustment; set the reliability threshold to 0.6 and the reliability decay rate to 0.5. When the local reliability value of a sample point falls below the threshold of 0.6, calculate the exponential decay coefficient and multiply it by the membership adjustment value to obtain the final membership value. If the local reliability value is not below the threshold, directly use the membership adjustment value as the final membership value.

[0106] For example, consider the data from a sample point in a printing process. This sample point represents the printing state at a specific moment, and its feature vector includes parameters such as ink density and registration accuracy. Calculations are performed to find the five nearest neighboring sample points, and the mean Euclidean distance is calculated to be 0.15, which serves as the local density indicator. The distance sequence from this sample point to the four cluster centers is calculated as [0.8, 0.6, 0.9, 1.2], with a standard deviation of 0.25 as the distance deviation.

[0107] The calculated local reliability value is 0.15 / 0.25 = 0.6. Since this value is equal to the reliability threshold of 0.6, an adjustment coefficient of 0.8 is used to generate an adaptive weighting factor of 0.75. Assuming that the initial membership value of this sample point to a certain category is 0.8, the membership adjustment value is 0.8 × 0.75 = 0.6. Since the local reliability value is equal to the threshold, 0.6 is directly used as the final membership value, indicating that the sample point's degree of belonging to this category is 0.6. This dynamic adjustment ensures that the clustering results better reflect the local distribution characteristics of the data.

[0108] In this embodiment, by calculating the mean Euclidean distance between a sample point and its neighbors and the standard deviation of the distance to the cluster center, the local density and distribution differences of the sample points can be accurately assessed. This provides a basis for dynamically adjusting the sample classification membership, ensuring that cluster analysis can better adapt to the local structure of the data. By combining the local reliability value and the adjustment coefficient, an adaptive weighting factor is generated to dynamically adjust the sample point membership. This adaptive mechanism effectively improves the accuracy of sample classification, avoids excessive or insufficient classification adjustments, and ensures that the clustering results are consistent with the actual data distribution. When the local reliability value falls below a preset threshold, an exponential decay coefficient is introduced to adjust the sample membership, further optimizing the clustering effect. This mechanism effectively handles sample points with high uncertainty and ensures the stability and accuracy of clustering. By combining local reliability, weighting factors, and the exponential decay mechanism, the membership value of each sample point is dynamically adjusted, resulting in more refined and reliable clustering results. This dynamic adjustment ensures that the clustering model accurately reflects the degree of belonging of each sample, improving the flexibility and accuracy of the overall clustering algorithm.

[0109] In an optional embodiment, the contribution of each feature in distinguishing different categories is calculated using a statistical discrimination ratio, and a feature combination having a statistical discrimination ratio greater than a preset threshold is selected as a key parameter combination for printing quality, including:

[0110] Calculate the inter-class scatter matrix and the intra-class scatter matrix to determine the statistical discrimination ratio sequence of the features; at the same time, calculate the mutual information matrix between the features to determine the redundant features of the labels; through a progressive strategy, starting from the maximum statistical discrimination ratio feature, by iteratively calculating the statistical discrimination ratio gain value, gradually construct a feature combination, specifically including:

[0111] The inter-class dispersion matrix is ​​determined by calculating the difference between the mean values ​​of samples of different categories, and the intra-class dispersion matrix is ​​determined by calculating the difference between the sample in the same category and the category mean value of the corresponding category. Based on the inter-class dispersion matrix and the intra-class dispersion matrix, the statistical discrimination ratio of each feature is calculated to obtain the feature statistical discrimination ratio sequence.

[0112] Calculate the mutual information value between the features to obtain a feature mutual information matrix. When the mutual information value is greater than a first preset threshold, mark the feature whose statistical discrimination ratio in the corresponding feature pair is less than a second preset threshold as a redundant feature; select the feature with the largest statistical discrimination ratio and not marked as a redundant feature from the feature statistical discrimination ratio sequence as the initial key feature;

[0113] The features that are not marked as redundant features and not selected as initial key features are sorted from large to small according to the statistical discrimination ratio to obtain a candidate feature sequence; the features in the candidate feature sequence are added to the existing key feature combination one by one, and the statistical discrimination ratio of the combination before and after the addition is calculated to obtain the statistical discrimination ratio gain value;

[0114] When the statistical discrimination ratio gain value is greater than a third preset threshold, adding the corresponding candidate feature to the key feature combination;

[0115] Repeat until all features in the candidate feature sequence are processed and the final key feature combination is obtained.

[0116] In a specific embodiment,

[0117] Step 1: Calculate the inter-class dispersion and intra-class dispersion. First, calculate the inter-class dispersion matrix: for each feature, calculate the mean of samples in different classes, and then calculate the degree of difference between these class means. Specifically, the inter-class dispersion matrix is ​​calculated by calculating the sum of the squares of the deviations between the class means and the overall mean.

[0118] Next, we calculate the intra-class dispersion matrix: for each feature, we calculate the sum of squared deviations between each sample within the same class and the class mean to obtain the intra-class dispersion matrix. Based on these two matrices, we calculate the statistical discriminant ratio for each feature, forming a feature statistical discriminant ratio sequence.

[0119] Step 2: Identify redundant features. Calculate the mutual information between all features and construct a feature mutual information matrix. Set the first preset threshold to 0.8 and the second preset threshold to 0.6. When the mutual information value of a pair of features is greater than 0.8, compare the statistical discrimination ratio of the pair of features. Features with a statistical discrimination ratio less than 0.6 are marked as redundant.

[0120] Step 3: Select the initial key feature: From the feature statistical discriminant ratio sequence, select the feature with the largest statistical discriminant ratio and not marked as redundant as the initial key feature. This feature will serve as the starting point for constructing the feature combination.

[0121] Step 4: Construct a candidate feature sequence; sort all features that are not marked as redundant and not selected as initial key features from large to small according to the statistical discrimination ratio to form a candidate feature sequence.

[0122] Step 5: Iteratively construct feature combinations; each feature in the candidate feature sequence is added to the existing key feature combination. Each time a feature is added, the statistical discrimination ratio of the combination before and after the addition is calculated to obtain the statistical discrimination ratio gain. The third preset threshold is set to 0.1. When the gain value is greater than 0.1, the candidate feature is included in the key feature combination.

[0123] Step 6: Complete feature selection; repeat the process of step 5 until all candidate features are processed and a complete key feature combination is finally obtained.

[0124] For example, taking the data of a certain operation of a printing press as an example, the initial features include 12 characteristic parameters: ink density, registration accuracy, printing speed, paper tension, printing pressure, ink fountain opening, ambient temperature, ambient humidity, etc.

[0125] By calculating the inter-class dispersion and intra-class dispersion, the characteristic statistical discrimination ratio sequence was obtained: ink density (0.85), registration accuracy (0.82), printing pressure (0.78), paper tension (0.75), ink fountain opening (0.72), printing speed (0.68), ambient humidity (0.65), ambient temperature (0.58), etc.

[0126] After calculating the feature mutual information matrix, we found that the mutual information value of ambient temperature and ambient humidity is 0.85, which is greater than the first preset threshold of 0.8. Since the statistical discrimination ratio of ambient temperature (0.58) is less than the second preset threshold of 0.6, it is marked as a redundant feature.

[0127] The ink density with the largest statistical discrimination ratio (0.85) is selected as the initial key feature. The remaining features that are not marked as redundant are arranged in descending order of statistical discrimination ratio to form a candidate feature sequence.

[0128] Through iterative calculations, when register accuracy was added, the statistical discrimination ratio gain for the combination was 0.15; when impression pressure was added, the gain was 0.12; when paper tension was added, the gain was 0.11; and when ink fountain opening was added, the gain was 0.08. Therefore, the final key feature combination included ink density, register accuracy, impression pressure, and paper tension. These parameters will be used for subsequent printing quality control and optimization.

[0129] In this embodiment, the contribution of each feature in distinguishing between different classes is effectively measured by calculating inter-class and intra-class dispersion, combined with the statistical discriminant ratio. This ensures that the most powerful features for discriminating between different classes are prioritized during feature selection, thereby improving the accuracy of subsequent analysis and prediction. By calculating the mutual information matrix between features and setting a threshold, redundant features can be automatically identified and removed. This reduces inter-feature correlation, improves the independence of feature combinations, and ensures that the selected key features are not interfered with by redundant information, thereby enhancing the reliability and performance of the model. A progressive strategy is employed, iteratively calculating the statistical discriminant ratio gain to gradually select the features that contribute most to print quality control. This approach ensures that the final selected feature combination has the highest discriminative power in distinguishing between different classes, thereby improving quality prediction. The key feature combinations (such as ink density and registration accuracy) ultimately obtained through iterative screening effectively reflect the core parameters affecting print quality. These parameters provide a scientific basis for subsequent quality control and optimization, improving the stability of the printing process and product quality.

[0130] In an optional embodiment, calculating the inter-class dispersion matrix and the intra-class dispersion matrix and determining the statistical discriminant ratio sequence of the features includes:

[0131] Calculate the number of samples in each category in the multi-category sample data, perform normalization processing to obtain the category weight vector, and construct the category weight matrix;

[0132] Calculating the category mean vector of each category of samples in the multi-category sample data and the overall mean vector; calculating the difference between the category mean vector and the overall mean vector to construct a category mean difference matrix; multiplying the category mean difference matrix with the category weight matrix to obtain an inter-class dispersion matrix;

[0133] Calculate the difference between each category sample and the corresponding category mean vector in the multi-category sample data to construct a sample difference matrix; multiply the sample difference matrix by the category weight matrix to obtain an intra-category dispersion matrix;

[0134] Extracting the diagonal elements of the inter-class scatter matrix and the diagonal elements of the intra-class scatter matrix to determine the inter-class scatter vector and the intra-class scatter vector respectively;

[0135] The ratio of the inter-class scatter vector to the intra-class scatter vector is calculated to obtain a statistical discriminant ratio sequence of features.

[0136] In a specific embodiment,

[0137] Step 1: Construct a class weight matrix. Count the number of samples in each class and divide the number of samples in each class by the total number of samples to obtain the class weights. For example, for sample data from four classes, calculate the proportion of samples in each class to form a class weight vector. Then, expand the class weight vector into a diagonal matrix to construct the class weight matrix.

[0138] Step 2: Calculate the inter-class dispersion matrix. First, calculate the class mean vector for each class sample and the overall mean vector for all samples. For each class, calculate the difference between its class mean vector and the overall mean vector to form a class mean difference matrix. Multiply the class mean difference matrix by the class weight matrix to obtain the inter-class dispersion matrix.

[0139] Step 3: Calculate the intra-class dispersion matrix. For each class, calculate the difference between the sample and the corresponding class mean vector to construct a sample difference matrix. Multiply the sample difference matrix with the class weight matrix to obtain the intra-class dispersion matrix.

[0140] Step 4: Extract the dispersion vectors. Extract the diagonal elements from the inter-class dispersion matrix to form the inter-class dispersion vector. Extract the diagonal elements from the intra-class dispersion matrix to form the intra-class dispersion vector. These two vectors reflect the performance of each feature in class differentiation and intra-class aggregation, respectively.

[0141] Step 5: Calculate the statistical discriminant ratio: Divide the inter-class scatter vector by the intra-class scatter vector to obtain the feature statistical discriminant ratio sequence. This sequence reflects the contribution of each feature in distinguishing different classes.

[0142] For example, taking the operation data of a printing press as an example, it includes 4 categories (normal state, registration deviation state, ink color abnormal state, and comprehensive quality abnormal state), with a total of 12 characteristic parameters.

[0143] Statistics show that the number of samples in each category are 200, 150, 100, and 50 respectively, and the total number of samples is 500.

[0144] The calculated category weight vector is [0.4, 0.3, 0.2, 0.1].

[0145] Construct a 4×4 diagonal matrix of class weights.

[0146] Taking the ink density feature as an example: the mean of category 1 is 1.5, the mean of category 2 is 1.3, the mean of category 3 is 1.8, and the mean of category 4 is 1.2; the overall mean is 1.45; the difference between the mean of each category and the overall mean is calculated as [0.05, -0.15, 0.35, -0.25]; after matrix operation, the inter-class dispersion value of this feature is 0.068.

[0147] The difference between the samples within each category and the category mean is calculated: the mean square sum of the sample differences of category 1 is 0.02, the mean square sum of the sample differences of category 2 is 0.03, the mean square sum of the sample differences of category 3 is 0.04, and the mean square sum of the sample differences of category 4 is 0.02. After matrix operation, the intra-category dispersion value of this feature is 0.028.

[0148] For the ink density feature, the inter-class dispersion is 0.068, the intra-class dispersion is 0.028, and the calculated statistical discrimination ratio is 0.068 / 0.028=2.43.

[0149] Using this method to calculate the statistical discriminant ratios of all features, we ultimately obtained the following sequence: ink density (2.43), registration accuracy (2.15), embossing pressure (1.89), paper tension (1.76), etc. This sequence intuitively demonstrates the importance of each feature for class differentiation.

[0150] In this embodiment, a category weight matrix is ​​constructed by counting and normalizing the number of samples in each category, providing a basis for subsequent calculations of inter-class and intra-class dispersion. This method effectively quantifies the relative importance of different categories and accurately calculates inter- and intra-class dispersion through matrix operations, ensuring the accuracy and effectiveness of data processing. By calculating the statistical discriminant ratio of features, the contribution of each feature in distinguishing different categories can be quantified. The statistical discriminant ratio sequence of features provides a clear ranking, helping to identify the features most influential in class differentiation, thereby optimizing feature selection and improving the effectiveness of the classification model. The ratio of inter- and intra-class dispersion can clearly identify which features have strong discriminative power for class differentiation and which have weaker power. Combined with mutual information analysis, redundant features can be further removed, improving the independence and discriminative power of feature combinations, and ensuring that the selected features maximize classification accuracy. By calculating the statistical discriminant ratio of each feature, the most discriminative features can be selected based on the ratio ranking, forming an optimized feature combination. This process improves the accuracy of print quality prediction and control, providing a strong feature foundation for subsequent model optimization and quality assessment.

[0151] Constructing a correlation matrix between the key parameter combinations;

[0152] In one embodiment, the data of the identified key parameter combinations are preprocessed. First, data normalization is performed to convert parameters of different dimensions to a unified scale. Outliers are then detected and processed, and missing values ​​are filled using a moving median method. Finally, the data is tested for stationarity and, if necessary, differencing is performed to eliminate trend effects.

[0153] The Pearson correlation coefficient method was used to calculate the linear correlation between parameters. For each pair of parameters, their covariance was calculated and divided by the product of their standard deviations to obtain the correlation coefficient. The correlation coefficient ranges from -1 to 1, where 1 indicates a perfect positive correlation, -1 indicates a perfect negative correlation, and 0 indicates no linear correlation.

[0154] The Spearman rank correlation coefficient method is used to assess the degree of nonlinear correlation between parameters. First, the parameter values ​​are converted to a rank sequence, and then the correlation coefficient between the rank sequences is calculated. This can capture monotonic nonlinear relationships between parameters.

[0155] By sliding the time window, we calculated the correlation coefficients between the parameters at different time delays. We set the maximum time delay to 10 sampling periods, gradually calculated the correlation coefficients at different time delays, and selected the time delay value with the strongest correlation as the final result.

[0156] Calculate the partial correlation coefficient between any two parameters while keeping other parameters constant. This can eliminate the influence of other parameters and reflect the true degree of correlation between the two parameters.

[0157] The linear correlation coefficient, nonlinear correlation coefficient, time-lag correlation, and conditional correlation are weighted together. The linear correlation weight is set to 0.4, the nonlinear correlation weight is set to 0.3, the time-lag correlation weight is set to 0.2, and the conditional correlation weight is set to 0.1. The comprehensive correlation coefficient is obtained through weighted summation.

[0158] The calculated comprehensive correlation coefficients are organized into a matrix. The rows and columns of the matrix correspond to the key parameters, and the matrix elements represent the comprehensive correlation between the parameters. A diagonal element of 1 indicates that the parameter is completely correlated with itself.

[0159] Perform a significance test on each element in the correlation matrix. Set the significance level to 0.05 and calculate the P value for each correlation coefficient. Set insignificant correlation coefficients to 0 and retain those with significant correlations.

[0160] Apply sparse processing to the correlation matrix. Set the correlation coefficient threshold to 0.3 and set the correlation coefficients below the threshold to 0 to highlight important correlations. At the same time, ensure the symmetry of the matrix, that is, the correlation of parameter A on parameter B should be equal to the correlation of parameter B on parameter A.

[0161] The final correlation matrix is ​​converted into a heatmap, using different shades of color to represent the degree of correlation. Positive correlations are represented by red, negative correlations by blue, and the stronger the correlation, the darker the color. This visualization intuitively demonstrates the strength of the relationship between parameters.

[0162] In this embodiment, data normalization, outlier detection and processing, missing value filling, and stationarity testing ensure the quality and consistency of input data, laying a solid foundation for subsequent correlation analysis. These steps eliminate interfering factors in the data and improve the accuracy of the analysis. A variety of methods, including the Pearson correlation coefficient, Spearman rank correlation coefficient, time-lagged correlation, and conditional correlation coefficient, are used to comprehensively assess the linear and nonlinear relationships between parameters. Sliding time windows and partial correlation analysis consider the true correlations between parameters under different time delays and conditions, providing a more comprehensive and refined analysis of parameter relationships. A comprehensive correlation coefficient matrix is ​​generated by weightedly combining different types of correlations (linear, nonlinear, time-lagged, and conditional). This weighted combination allows for the rational integration of different types of relationships, more accurately reflecting the overall degree of correlation between parameters and ensuring the comprehensiveness and effectiveness of the analysis. Significance testing, sparsification processing, and final heatmap visualization of the correlation matrix provide intuitive and easy-to-understand parameter correlation analysis results. The heatmap effectively displays the strength of correlations between parameters, helping decision makers quickly identify important relationships and providing a scientific basis for printing quality control.

[0163] In an optional embodiment, a dynamic Bayesian probability network model is constructed based on the parameter mapping relationship in the correlation matrix, a core network structure is constructed by calculating the information entropy value and information gain, the conditional probability distribution is optimized using the log-likelihood expectation value, and a probabilistic association is established between the printing quality parameters and the equipment operating condition parameters. Generating a printing quality prediction result includes:

[0164] Based on the dynamic Bayesian probability network, the initial network structure is constructed through the correlation matrix. The information entropy value is used to calculate the information gain, and the parameter nodes are screened to build the core network structure. The parameter prediction is achieved through the optimization of the log-likelihood expectation value and the real-time calculation of the information gain. Specifically,

[0165] Establishing connections between parameter pairs based on correlation coefficients of the parameter pairs in the correlation matrix; determining the direction of the connections based on the temporal characteristics of the parameter pairs, and constructing an initial network structure;

[0166] Calculate the information entropy value of each parameter node in the initial network structure, construct a parameter importance evaluation index based on the information entropy value, and calculate the information gain of each parameter node for the prediction result; sort the parameter nodes according to the information gain, select the parameter nodes with information gain greater than the dynamic threshold to construct a core network structure, obtain the corresponding parent node set for the parameter nodes in the core network structure, and construct a conditional probability distribution based on the parent node set;

[0167] Calculating a log-likelihood expectation value based on the core network structure and the conditional probability distribution, maximizing the log-likelihood expectation value through iterative optimization, and obtaining an optimized conditional probability distribution;

[0168] Calculate the parameter node information gain at the current sampling moment, update the parameter composition of the core network structure according to the change in information gain, and construct a new core network structure; based on the new core network structure and the corresponding conditional probability distribution, calculate the parameter prediction probability at the next sampling moment;

[0169] The parameter value corresponding to the maximum parameter prediction probability is selected as the printing quality prediction result.

[0170] In a specific embodiment,

[0171] Based on the correlation coefficients in the correlation matrix, we determine the connection relationships between parameters. We set the correlation coefficient threshold to 0.5 and establish a connection when the correlation coefficient of a parameter pair exceeds the threshold. We analyze the time series data of the parameters to determine the order in which the parameters change, and based on this, we determine the direction of the connection and construct a directed network structure.

[0172] For each parameter node in the network, the probability distribution of its historical data is calculated and the information entropy value is calculated. A parameter importance index is constructed based on the information entropy value. The information gain of each parameter node with respect to the predicted target is calculated to assess its contribution to the prediction result. A dynamic threshold is set at 1.2 times the mean information gain value, and parameter nodes with information gain greater than this threshold are screened.

[0173] The selected nodes with high information gain parameters are combined into the core network structure. For each node, its parent node set is determined, that is, the upstream nodes that have a direct impact on the node. Based on historical data statistical analysis, a conditional probability distribution table for each node relative to its parent node set is established.

[0174] Use the expectation-maximization algorithm to optimize the conditional probability distribution. Calculate the expected log-likelihood for the current network structure and iteratively adjust the conditional probability parameters to maximize the expected log-likelihood. Set the maximum number of iterations to 100 and the convergence threshold to 0.001.

[0175] At each sampling moment, the information gain of the parameter nodes is recalculated. When the change in information gain exceeds a preset threshold (e.g., 20%), the parameters of the core network structure are updated. The conditional probability distribution is calculated for the newly added nodes, and the related connections of the removed nodes are pruned.

[0176] Based on the updated core network structure and the optimized conditional probability distribution, the probability of each parameter value at the next sampling time is calculated. The parameter value combination with the highest probability is selected as the prediction result.

[0177] For example, take the printing press registration system as an example:

[0178] The correlation matrix shows a correlation coefficient of 0.75 between paper tension and registration accuracy, 0.68 between printing speed and registration accuracy, and 0.55 between impression pressure and registration accuracy. Time series analysis shows that changes in paper tension and printing speed precede changes in registration accuracy, establishing a connection between these two parameters and registration accuracy.

[0179] The information entropy values ​​of each parameter node were calculated as follows: paper tension: 0.85, printing speed: 0.78, impression pressure: 0.62, and registration accuracy: 0.92;

[0180] Calculate the information gain for the prediction of registration accuracy: paper tension: 0.45, printing speed: 0.38, impression pressure: 0.25, and the dynamic threshold is set to 0.35;

[0181] Select paper tension and printing speed as core network nodes; the parent node set of registration accuracy is {paper tension, printing speed}; establish a conditional probability table, such as:

[0182] P(Registration Accuracy=High|Tension=Normal, Speed=Moderate)=0.85;

[0183] P(registration accuracy = medium | tension = high, speed = fast) = 0.65.

[0184] The initial log-likelihood expectation value is -856.3. After 87 iterations of optimization, the log-likelihood expectation value is increased to -782.5. The optimized conditional probability is as follows:

[0185] P(Registration Accuracy=High|Tension=Normal, Speed=Moderate)=0.92.

[0186] The information gain calculation results at the new sampling moment are: paper tension: 0.48 (increase of 6.7%), printing speed: 0.42 (increase of 10.5%), and stamping pressure: 0.36 (increase of 44%). Since the information gain change of stamping pressure exceeds the threshold, it is added to the core network structure.

[0187] Predict the probability distribution of the registration accuracy at the next moment: high accuracy: 0.85, medium accuracy: 0.12, low accuracy: 0.03; select "high accuracy" with the highest probability as the prediction result.

[0188] In this embodiment, a directional parameter network was successfully constructed by calculating correlation coefficients and analyzing time series data between parameters. This network structure reflects the relationships between different parameters and their order of change, providing more accurate parameter dependency analysis and a clear framework for subsequent prediction and optimization. By calculating information entropy and information gain, core parameter nodes that contribute most to the prediction target were successfully selected. Dynamic adjustment of the core network structure ensures that network nodes always contain the most predictive parameters, improving the flexibility and prediction accuracy of the network model. The expectation-maximization algorithm is used to optimize the conditional probability distribution. By iteratively maximizing the expected log-likelihood value, the model's conditional probability is optimized, resulting in more accurate predictions for each node. This optimization process improves the model's convergence speed and stability, enhancing its adaptability to data changes. By dynamically updating the core network structure through real-time calculation of information gain changes, the network model is able to adapt to parameter changes at any time. This mechanism enables the model to make the most accurate predictions at each sampling moment, providing reliable prediction results and effectively improving the accuracy of parameter adjustment and quality control in practical applications.

[0189] In an optional embodiment, based on the correlation matrix and the printing quality prediction results, a bidirectionally constrained dynamic programming algorithm is used to construct a state transition equation using coupling weights between functional modules. The optimal parameter configuration of the functional modules of the printing press is determined in combination with the bidirectional constraints, and a top-level design scheme for the printing press is generated, including:

[0190] Calculating coupling weights between functional modules based on the correlation matrix, wherein the coupling weights represent the degree of mutual influence between different functional modules;

[0191] Constructing a state transfer equation based on the coupling weight, and multiplying and summing the parameter state of the previous function module with the corresponding coupling weight through the state transfer equation to obtain the parameter state of the current function module;

[0192] Establishing bidirectional constraints based on the print quality prediction results, the bidirectional constraints including forward constraints and backward constraints, the forward constraints limiting the adjustment range of parameters corresponding to the current functional module based on the parameter state of the preceding functional module; and the backward constraints ensuring that subsequent functional modules achieve preset quality targets based on the print quality prediction results;

[0193] According to the forward constraints, each functional module is traversed forward, the local optimal solution is calculated and the state transition path is recorded to obtain a state transition sequence; according to the backward constraints, starting from the last functional module, the state transition sequence is traced backward to determine the optimal parameter configuration of the functional module;

[0194] The optimal parameter configuration is input into a printing quality prediction model for verification. When the printing quality prediction result obtained by verification does not meet the requirements, the bidirectional constraint condition is adjusted and the parameter configuration optimization is re-executed.

[0195] In one embodiment, the degree of association between functional modules is analyzed based on a correlation matrix. For each pair of functional modules, the correlation coefficients between the parameter pairs they contain are calculated, and the average value is taken as the coupling weight between the modules. The larger the coupling weight, the stronger the mutual influence between the two functional modules.

[0196] The parameter state of each functional module is represented as a state vector. Based on the coupling weights, a relationship is established between the state of the current functional module and the state of the previous functional module. The state vector of the current module is obtained by multiplying the state vector of the previous module by the corresponding coupling weight and summing them.

[0197] Set forward constraints: Determine the adjustable range of the current module parameters based on the actual parameter status of the preceding functional module to prevent parameter adjustments from exceeding the physical limitations of the device.

[0198] Set backward constraints: Based on the print quality prediction results, set quality target thresholds for subsequent functional modules to ensure that the final print quality meets the requirements.

[0199] Starting from the first functional module, traverse backwards. For each module, enumerate the possible parameter states within the forward constraints, calculate the minimum cost to reach that state, and record the optimal state transition path. For each possible state, save the path to reach it.

[0200] Starting from the last functional module, according to the backward constraints, select the final state that meets the quality requirements. Follow the recorded state transition path forward and back to determine the optimal parameter configuration for each functional module.

[0201] The optimal parameter configuration is input into the printing quality prediction model to verify the expected effect. If the prediction result does not meet the requirements, the parameter range of the bidirectional constraint condition is adjusted and the optimization process is repeated.

[0202] For example, take the paper feeding system, impression system, ink supply system and paper transport system of a four-color printing press as an example:

[0203] Calculate the coupling weights between functional modules: paper feeding-imprinting, 0.75; paper feeding-ink supply, 0.45; imprinting-ink supply, 0.65; imprinting-paper feeding, 0.70; ink supply-paper feeding, 0.55.

[0204] The state vector of the paper feeding system includes: paper tension and paper feeding speed; the state vector of the imprinting system includes: imprinting pressure and printing speed; an example of the state transition equation is: imprinting system state = 0.75 × paper feeding system state.

[0205] Forward constraints: paper feed speed range 20-60 m / min; paper tension range 150-300 N; impression pressure range 2000-4000 N;

[0206] Backward constraints: registration accuracy error ≤ 0.1 mm; color difference value ΔE ≤ 3; printing density error ≤ 0.05;

[0207] Possible states of the paper feeding system: State 1, speed 30 m / min, tension 200 N; State 2, speed 40 m / min, tension 250 N; State 3, speed 50 m / min, tension 280 N;

[0208] The corresponding optimal states of the stamping system are: state 1 corresponds to a pressure of 2500 N and a speed of 28 m / min; state 2 corresponds to a pressure of 3000 N and a speed of 38 m / min; state 3 corresponds to a pressure of 3500 N and a speed of 48 m / min.

[0209] Starting from the paper feeding system, reverse tracing is performed and the state that meets the registration accuracy requirements is selected: paper feeding system speed 38 m / min, ink supply system ink fountain opening 45%, imprinting system pressure 3000 N, speed 38 m / min, paper feeding system speed 40 m / min, tension 250 N.

[0210] The above parameter configurations are input into the prediction model: the predicted registration accuracy error is 0.08 mm, the predicted color difference value ΔE: 2.8; and the predicted printing density error is 0.04.

[0211] The verification results meet the quality requirements and are confirmed as the final parameter configuration plan.

[0212] In this embodiment, by calculating the coupling weights between functional modules, the strength of the mutual influence between different functional modules can be quantified. This analysis helps identify the interdependencies of key modules, provides a reliable basis for subsequent parameter adjustments, and optimizes the overall performance of the system. Forward and backward constraints are set to ensure that the parameter adjustments of each functional module are balanced between physical limitations and quality requirements. Forward constraints control the adjustable range of parameters, avoiding adjustments that exceed the capabilities of the equipment, while backward constraints ensure that the final printing quality meets the requirements, improving the controllability and stability of the system. By recording the optimal state transition path of each module and backtracking, the optimal parameter configuration of each functional module can be determined. This method takes into account the minimum cost path during the optimization process, ensuring that the parameter configuration of each module in the system is optimal and meets the quality goals. By inputting the optimal parameter configuration obtained by optimization into the printing quality prediction model for verification, it is ensured that the prediction results meet the quality requirements. This process ensures the accuracy of the final parameter configuration and improves the quality control capability of the overall printing process.

[0213] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0214] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. The top-level project design method of a multivariate data printing machine is characterized by: include: Acquiring printing press operating data, wherein the printing press operating data includes printing quality parameters, equipment operating condition parameters, and environmental parameters; Adopting an adaptive fuzzy clustering algorithm to perform feature analysis and parameter extraction on the printing press operation data, by determining the number of clusters, dynamically adjusting the sample classification membership, and identifying the key parameter combination that affects the printing quality; Constructing a correlation matrix between the key parameter combinations; Based on the parameter mapping relationship in the correlation matrix, a dynamic Bayesian probability network model is constructed, a core network structure is constructed by calculating the information entropy value and information gain, and the conditional probability distribution is optimized using the log-likelihood expectation value. A probabilistic association is established between the printing quality parameters and the equipment operating condition parameters to generate a printing quality prediction result; Based on the correlation matrix and the printing quality prediction results, a bidirectional constraint dynamic programming algorithm is used to construct a state transfer equation through the coupling weights between functional modules. Combined with the bidirectional constraints, the optimal parameter configuration of the printing press functional modules is determined to generate a top-level design plan for the printing press.

2. The method according to claim 1, characterized in that An adaptive fuzzy clustering algorithm is used to perform feature analysis and parameter extraction on the printing press operation data. By determining the number of clusters and dynamically adjusting the sample classification membership, the key parameter combinations that affect printing quality are identified, including: The printing press operation data is segmented by a sliding window method, and a feature vector is determined based on calculations of each window; Calculate the minimum description length of the eigenvector to determine the target cluster number interval, calculate the fuzzy partition coefficient value and partition entropy value within the target cluster number interval, select the cluster number corresponding to the maximum fuzzy partition coefficient value and the minimum partition entropy value, and determine the final cluster number; Constructing an initial value of the sample classification membership, calculating the distribution parameters of each dimension of the feature vector, and obtaining the sample classification membership value through dynamic adjustment based on the distribution parameters; The cluster center coordinates are calculated based on the sample classification membership values, and the spatial position relationship strength between samples is calculated as a penalty factor to construct the objective function; Iteratively update the sample classification membership value and cluster center coordinates according to the final number of clusters, and determine the clustering result by minimizing the objective function; Analyzing the characteristic patterns of each cluster center based on the clustering results, calculating the contribution of each feature in distinguishing different categories using a statistical discrimination ratio, and selecting a feature combination with a statistical discrimination ratio greater than a preset threshold as a key parameter combination for printing quality; The key parameter combination is input into a support vector machine classifier for cross-validation to verify the validity of the key parameter combination.

3. The method according to claim 2, characterized in that According to the distribution parameters, obtaining the sample classification membership value through dynamic adjustment includes: The local density index is obtained by calculating the mean Euclidean distance between the sample point and the nearest neighboring sample point. The distance deviation value is calculated by combining the standard deviation of the distance sequence from the sample point to the cluster center, and the local reliability value is determined. An adaptive weight factor is generated. Based on whether the local reliability value is lower than the threshold, an exponential decay coefficient is introduced to adjust the sample classification membership value. Specifically, the following steps are performed: For each sample point, multiple nearest neighbor sample points are selected, the Euclidean distances between the sample point and the nearest neighbor sample points are calculated, and the mean of the Euclidean distances is used as the local density index of the sample point; the Euclidean distances from each sample point to each cluster center are calculated to obtain a distance sequence, and the standard deviation of the distance sequence is calculated as the distance deviation value of the sample point; the local density index is divided by the distance deviation value to obtain the local reliability value of the sample point; generating an adaptive weight factor based on the local reliability value and a preset adjustment coefficient; Multiplying the adaptive weight factor by the initial value of the sample classification membership to obtain an adjusted value of the sample classification membership; When the local reliability value is lower than the preset reliability threshold, the exponential decay coefficient is multiplied by the sample classification membership adjustment value to obtain the sample classification membership value; when the local reliability value is not lower than the reliability threshold, the sample classification membership adjustment value is used as the sample classification membership value; wherein the exponential decay coefficient specifically refers to: ； in, D represents the exponential decay coefficient, represents the reliability decay rate, T r represents the reliability threshold, R l Represents the local reliability value.

4. The method according to claim 2, characterized in that The contribution of each feature in distinguishing different categories is calculated using the statistical discrimination ratio. Feature combinations with a statistical discrimination ratio greater than a preset threshold are selected as key parameters for printing quality. The combinations include: Calculate the inter-class scatter matrix and the intra-class scatter matrix to determine the statistical discrimination ratio sequence of the features; at the same time, calculate the mutual information matrix between the features to determine the redundant features of the labels; through a progressive strategy, starting from the maximum statistical discrimination ratio feature, by iteratively calculating the statistical discrimination ratio gain value, gradually construct a feature combination, specifically including: The inter-class dispersion matrix is ​​determined by calculating the difference between the mean values ​​of samples of different categories, and the intra-class dispersion matrix is ​​determined by calculating the difference between the sample in the same category and the category mean value of the corresponding category. Based on the inter-class dispersion matrix and the intra-class dispersion matrix, the statistical discrimination ratio of each feature is calculated to obtain the feature statistical discrimination ratio sequence. Calculate the mutual information value between the features to obtain a feature mutual information matrix. When the mutual information value is greater than a first preset threshold, mark the feature whose statistical discrimination ratio in the corresponding feature pair is less than a second preset threshold as a redundant feature; select the feature with the largest statistical discrimination ratio and not marked as a redundant feature from the feature statistical discrimination ratio sequence as the initial key feature; The features that are not marked as redundant features and not selected as initial key features are sorted from large to small according to the statistical discrimination ratio to obtain a candidate feature sequence; the features in the candidate feature sequence are added to the existing key feature combination one by one, and the statistical discrimination ratio of the combination before and after the addition is calculated to obtain the statistical discrimination ratio gain value; When the statistical discrimination ratio gain value is greater than a third preset threshold, adding the corresponding candidate feature to the key feature combination; Repeat until all features in the candidate feature sequence are processed and the final key feature combination is obtained.

5. The method according to claim 4, characterized in that Calculate the inter-class scatter matrix and the intra-class scatter matrix and determine the statistical discriminant ratio sequence of the features including: Calculate the number of samples in each category in the multi-category sample data, perform normalization processing to obtain the category weight vector, and construct the category weight matrix; Calculating the category mean vector of each category of samples in the multi-category sample data and the overall mean vector; calculating the difference between the category mean vector and the overall mean vector to construct a category mean difference matrix; multiplying the category mean difference matrix with the category weight matrix to obtain an inter-class dispersion matrix; Calculate the difference between each category sample and the corresponding category mean vector in the multi-category sample data to construct a sample difference matrix; multiply the sample difference matrix by the category weight matrix to obtain an intra-category dispersion matrix; Extracting the diagonal elements of the inter-class scatter matrix and the diagonal elements of the intra-class scatter matrix to determine the inter-class scatter vector and the intra-class scatter vector respectively; The ratio of the inter-class scatter vector to the intra-class scatter vector is calculated to obtain a statistical discriminant ratio sequence of features.

6. The method according to claim 1, characterized in that Based on the parameter mapping relationship in the correlation matrix, a dynamic Bayesian probability network model is constructed. The core network structure is constructed by calculating the information entropy value and information gain. The conditional probability distribution is optimized using the log-likelihood expectation value. The printing quality parameters are probabilistically associated with the equipment operating condition parameters to generate the printing quality prediction results, including: Based on the dynamic Bayesian probability network, the initial network structure is constructed through the correlation matrix. The information entropy value is used to calculate the information gain, and the parameter nodes are screened to build the core network structure. The parameter prediction is achieved through the optimization of the log-likelihood expectation value and the real-time calculation of the information gain. Specifically, Establishing connections between parameter pairs based on correlation coefficients of the parameter pairs in the correlation matrix; determining the direction of the connections based on the temporal characteristics of the parameter pairs, and constructing an initial network structure; Calculate the information entropy value of each parameter node in the initial network structure, construct a parameter importance evaluation index based on the information entropy value, and calculate the information gain of each parameter node for the prediction result; sort the parameter nodes according to the information gain, select the parameter nodes with information gain greater than the dynamic threshold to construct a core network structure, obtain the corresponding parent node set for the parameter nodes in the core network structure, and construct a conditional probability distribution based on the parent node set; Calculating a log-likelihood expectation value based on the core network structure and the conditional probability distribution, maximizing the log-likelihood expectation value through iterative optimization, and obtaining an optimized conditional probability distribution; Calculate the parameter node information gain at the current sampling moment, update the parameter composition of the core network structure according to the change in information gain, and construct a new core network structure; based on the new core network structure and the corresponding conditional probability distribution, calculate the parameter prediction probability at the next sampling moment; The parameter value corresponding to the maximum parameter prediction probability is selected as the printing quality prediction result.

7. The method according to claim 1, characterized in that Based on the correlation matrix and the printing quality prediction results, a bidirectionally constrained dynamic programming algorithm is used to construct a state transition equation through the coupling weights between functional modules. The optimal parameter configuration of the printing press functional modules is determined in combination with the bidirectional constraints, and a top-level design scheme for the printing press is generated, including: Calculating coupling weights between functional modules based on the correlation matrix, wherein the coupling weights represent the degree of mutual influence between different functional modules; Constructing a state transfer equation based on the coupling weight, and multiplying and summing the parameter state of the previous function module with the corresponding coupling weight through the state transfer equation to obtain the parameter state of the current function module; Establishing bidirectional constraints based on the print quality prediction results, the bidirectional constraints including forward constraints and backward constraints, the forward constraints limiting the adjustment range of parameters corresponding to the current functional module based on the parameter state of the preceding functional module; and the backward constraints ensuring that subsequent functional modules achieve preset quality targets based on the print quality prediction results; According to the forward constraints, each functional module is traversed forward, the local optimal solution is calculated and the state transition path is recorded to obtain a state transition sequence; according to the backward constraints, starting from the last functional module, the state transition sequence is traced backward to determine the optimal parameter configuration of the functional module; The optimal parameter configuration is input into a printing quality prediction model for verification. When the printing quality prediction result obtained by verification does not meet the requirements, the bidirectional constraint condition is adjusted and the parameter configuration optimization is re-executed.

Images