Remote monitoring and fault prediction system and method for intelligent digital printing equipment

Abstract

The invention relates to a remote monitoring and fault prediction system and method for intelligent digital printing equipment. The method comprises the following steps: acquiring anti-counterfeiting code printing operation data of a spraying device and drying equipment in intelligent digital printing equipment to obtain a comprehensive operation data set; constructing an equipment component relation graph structure; carrying out spraying uniformity and drying efficiency feature extraction, and creating a fusion spatio-temporal feature vector; performing multi-task prediction through a heterogeneous multi-task learning graph neural network to obtain the residual ink amount of the anti-counterfeiting code and the optimal drying parameter; optimization is carried out based on constraint conditions, and a printing process intelligent diagnosis model is obtained; and analyzing a real-time data set acquired in the anti-counterfeiting code printing process through the printing process intelligent diagnosis model, and generating early warning information of spraying quality and drying efficiency. According to the implementation of the invention, multi-source data can be comprehensively utilized, and a new method of intelligent remote monitoring and fault prediction is realized, so that the operation reliability and the production efficiency of digital printing equipment are improved.

Description

Technical Field

[0001] The present invention relates to the field of printing monitoring technology, and in particular to a remote monitoring and fault prediction system and method for intelligent digital printing equipment. Background Art

[0002] With the rapid development of digital printing technology, intelligent digital printing equipment has been widely used in fields such as anti-counterfeiting code printing. However, such equipment faces complex process parameter control and equipment status monitoring problems during operation, especially in terms of spraying uniformity and drying efficiency. Traditional monitoring methods often rely on manual experience, making it difficult to detect and predict potential failure risks in a timely manner, resulting in unstable printing quality and low production efficiency.

[0003] In addition, there are complex interactions between the various components of digital printing equipment, so the failure of a single component may trigger a chain reaction and affect the performance of the entire system. Most existing monitoring systems use isolated single-point monitoring methods, which cannot effectively capture the dynamic relationship between components and make it difficult to achieve comprehensive system status evaluation and fault warning. Summary of the invention

[0004] The main purpose of the present invention is to provide a remote monitoring and fault prediction system and method for intelligent digital printing equipment, so as to realize the comprehensive utilization of multi-source data, realize a new method of intelligent remote monitoring and fault prediction, and improve the operating reliability and production efficiency of digital printing equipment.

[0005] To achieve the above object, the present invention provides a remote monitoring and fault prediction method for intelligent digital printing equipment, comprising the following steps: Collect anti-counterfeiting code printing operation data from the spraying device and drying equipment in the intelligent digital printing equipment to obtain a comprehensive operation data set; Building a dynamic weighted adjacency matrix based on the comprehensive operation data set, taking the spraying device and the drying device as target nodes, and building an equipment component relationship graph structure; Based on the equipment component relationship graph structure, extract the spraying uniformity and drying efficiency features of the comprehensive operation data set to create a fused spatiotemporal feature vector; The fused spatiotemporal feature vector is input into a heterogeneous multi-task learning graph neural network for multi-task prediction to obtain the residual ink amount of the anti-counterfeiting code and the optimal drying parameters; The relationship between ink drying time and drying temperature is converted into constraint conditions, and the heterogeneous multi-task learning graph neural network is optimized based on the residual ink amount of the anti-counterfeiting code and the optimal drying parameters to obtain an intelligent diagnosis model for the printing process; The printing process intelligent diagnosis model is used to analyze the real-time data set collected during the anti-counterfeiting code printing process, and to generate early warning information on the spraying quality and drying efficiency.

[0006] The present invention also provides a remote monitoring and fault prediction system for intelligent digital printing equipment, comprising: The collection module is used to collect the anti-counterfeiting code printing operation data of the spraying device and the drying equipment in the intelligent digital printing equipment to obtain a comprehensive operation data set; A construction module, used to construct a dynamic weighted adjacency matrix based on the comprehensive operation data set, taking the spraying device and the drying device as target nodes, and constructing a device component relationship graph structure; An extraction module, for extracting spraying uniformity and drying efficiency features of the comprehensive operation data set based on the equipment component relationship graph structure, and creating a fused spatiotemporal feature vector; A prediction module, used for inputting the fused spatiotemporal feature vector into a heterogeneous multi-task learning graph neural network for multi-task prediction to obtain the residual ink amount of the anti-counterfeiting code and the optimal drying parameters; An optimization module, used for converting the relationship between ink drying time and drying temperature into constraint conditions, and optimizing the heterogeneous multi-task learning graph neural network based on the residual ink amount of the anti-counterfeiting code and the optimal drying parameters to obtain an intelligent diagnosis model for the printing process; The generation module is used to analyze the real-time data set collected during the anti-counterfeiting code printing process through the printing process intelligent diagnosis model, and generate early warning information on spraying quality and drying efficiency.

[0007] The present invention also provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any one of the above methods when executing the computer program.

[0008] The present invention also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of any of the above-mentioned methods are implemented.

[0009] In summary, the technical solution provided by the present invention collects and fuses multiple sensor data of the spraying device and the drying equipment to construct a comprehensive operation data set, which provides a comprehensive data basis for subsequent analysis and prediction, and helps to improve the accuracy and comprehensiveness of monitoring. By constructing a dynamic weighted adjacency matrix and a graph structure of equipment component relationships, this method can effectively capture the dynamic interactions between the components of the equipment. By combining graph convolution and attention mechanisms, the spatiotemporal feature extraction of spraying uniformity and drying efficiency is realized, and the long-term dependent features in the data are effectively extracted, which improves the richness of feature expression. A heterogeneous multi-task learning graph neural network containing two subnetworks, namely, ink residue estimation and drying parameter optimization, is designed to achieve parallel prediction of multiple related tasks, thereby improving the generalization ability and prediction accuracy of the model. By introducing a bidirectional long short-term memory network and an interactive enhancement module of the self-attention layer, information interaction and feature enhancement between different tasks are achieved, the relationship between ink drying time and drying temperature is converted into a soft constraint condition, and the model is optimized so that the prediction results are more in line with the actual process requirements, thereby improving the practicality of the model. Decision rules and threshold judgment logic are constructed based on the optimized heterogeneous multi-task learning graph neural network, which realizes the intelligent diagnosis of the printing process and can generate early warning information on spraying quality and drying efficiency in time, which helps to improve the controllability and stability of the production process. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 It is a schematic diagram of the steps of a remote monitoring and fault prediction method for an intelligent digital printing device in one embodiment of the present invention; Figure 2 It is a structural block diagram of a remote monitoring and fault prediction system for intelligent digital printing equipment in one embodiment of the present invention; Figure 3 It is a schematic block diagram of the structure of a computer device according to an embodiment of the present invention.

[0011] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0012] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0013] Reference Figure 1 This embodiment provides a remote monitoring and fault prediction method for an intelligent digital printing device, comprising the following steps: S1, collect anti-counterfeiting code printing operation data of the spraying device and drying equipment in the intelligent digital printing equipment to obtain a comprehensive operation data set; Specifically, the ink flow sensor in the spraying device is subjected to real-time data collection to obtain the time series data of the ink flow, which reflects the change of the ink flow of the spraying device at different time points. Through further analysis and calculation, the spraying speed data of the equipment is derived. The spraying speed is one of the important parameters for measuring the working state of the spraying device. The accuracy and consistency of the spraying speed are ensured by analyzing the ink flow time series data. At the same time, the temperature sensor and humidity sensor in the drying device are subjected to data collection to obtain the temperature and humidity data related to the drying process, which reflects the temperature and humidity environment of the drying device during operation. Based on the drying temperature and humidity data, the drying efficiency is estimated to obtain the drying efficiency index reflecting the working state of the drying device. By collecting the spraying pressure data, the pressure change of the spraying device during operation is monitored in real time. The spraying pressure data is combined with the spraying speed data to calculate the uniformity index during the spraying process. The current sensors of the spraying device and the drying device are subjected to data collection to obtain the power consumption data of the equipment to reflect the energy consumption of the equipment. Based on ink flow time series data, spraying speed data, drying temperature and humidity data, drying efficiency index, spraying pressure data, spraying uniformity index and equipment power consumption data, multi-dimensional data fusion is performed to obtain a comprehensive operation data set.

[0014] S2, constructs a dynamic weighted adjacency matrix based on the comprehensive operation data set, takes the spraying device and drying equipment as the target nodes, and constructs the equipment component relationship graph structure; Specifically, the comprehensive operation data set is divided into time windows, and the entire operation time period is divided into multiple time segments to obtain multiple time segment data subsets. Based on the time segment data subsets, the correlation between each device component is calculated to generate a component correlation matrix reflecting the relationship between each component. The component correlation matrix is ​​subjected to threshold filtering to exclude low-correlation noise data to obtain a sparse correlation matrix. The sparse correlation matrix reduces data redundancy and enhances the representativeness of the data by retaining only the component relationships with higher criticality. Based on the sparse correlation matrix, the graph structure of the device components is initialized to construct an initial graph structure. The initial graph structure is topologically optimized, and the spraying device and the drying device are set as target nodes to obtain a hierarchical graph structure, which more clearly reflects the upstream and downstream relationships and mutual dependence between each device component. Based on the hierarchical graph structure, the edge weights in the graph are dynamically updated and calculated to generate a dynamic weighted adjacency matrix. The dynamic weighted adjacency matrix enables the graph structure to reflect the operating status of the device in real time by giving the edge weight value a dynamic adjustment capability. The dynamic weighted adjacency matrix is ​​normalized to obtain a standardized adjacency matrix. Based on the standardized adjacency matrix and preset equipment topology rules, the hierarchical graph structure is adjusted to ensure the rationality and accuracy of the equipment component relationship graph structure, and finally a equipment component relationship graph structure that reflects the complex relationship between the internal components of the intelligent digital printing equipment is generated.

[0015] S3, based on the equipment component relationship graph structure, extract the spraying uniformity and drying efficiency features of the comprehensive operation data set to create a fused spatiotemporal feature vector; Specifically, a graph convolution operation is performed on the equipment component relationship graph structure to extract rich feature information from each node in the graph and its neighboring nodes. Through the graph convolution operation, an initial node feature matrix is ​​obtained, which contains the preliminary feature expression of the equipment component. Based on the initial node feature matrix, feature aggregation is performed on the spraying device node, and all feature information related to the spraying device is integrated to generate the spraying device aggregate feature. The spraying device aggregate feature is nonlinearly transformed to extract the initial feature of spraying uniformity. At the same time, feature aggregation operation is performed on the features of the drying device node to obtain the drying device aggregate feature. After the feature is also processed by nonlinear transformation, the initial feature of drying efficiency is obtained. The first and second one-dimensional convolutional neural networks are used to extract the time series features of the initial features of spraying uniformity and the initial features of drying efficiency respectively. Through the time series feature extraction, the time series features of spraying uniformity and drying efficiency are obtained, which reflect the changes in the operating status of the equipment in different time periods. By calculating the attention mechanism of the spraying uniformity time series features and the drying efficiency time series features, the key features that have a greater impact on the equipment status at different time points are identified, and they are given higher weights to generate weighted feature vectors. Based on the weighted feature vector, time dependency modeling is performed to generate long-term dependency features, which reflect the operation rules and trends of the equipment over a long period of time. The long-term dependency features are concatenated with the weighted feature vector to generate a fused spatiotemporal feature vector.

[0016] S4, input the fused spatiotemporal feature vector into the heterogeneous multi-task learning graph neural network for multi-task prediction to obtain the residual ink volume and optimal drying parameters of the anti-counterfeiting code; Specifically, the fused spatiotemporal feature vector is feature decomposed and split into a spraying feature subset and a drying feature subset, which correspond to the feature information of the two key links of spraying and drying in the printing process. The spraying feature subset is input into the ink residual estimation subnetwork in the heterogeneous multi-task learning graph neural network, and the estimated value of the residual ink of the initial anti-counterfeiting code is calculated. The architecture of the ink residual estimation subnetwork includes three layers of graph convolution layers and two layers of fully connected layers. Among them, the graph convolution layer uses Chebyshev polynomials to approximate spectral graph convolution, which can efficiently capture local features in graph structure data, and uses ReLU as the activation function to ensure the nonlinear expression ability of the network. The features processed by the graph convolution layer are extracted and combined through two layers of fully connected layers. The Dropout regularization technology is used in the fully connected layer to prevent the network from overfitting, and the Sigmoid activation function is used in the output layer to ensure that the output estimate is within a reasonable range. At the same time, the drying feature subset is input into the drying parameter optimization subnetwork in the same heterogeneous multi-task learning graph neural network, and the initial optimal drying parameter estimate is calculated through the forward propagation process. The drying parameter optimization subnetwork includes two graph attention layers and three fully connected layers. The graph attention layer uses a multi-head attention mechanism and sets 8 attention heads, which can focus on different aspects of drying features from multiple angles and extract richer feature expressions. In the fully connected layer, BatchNormalization is used to accelerate network convergence and improve training stability. Finally, the Tanh activation function is used in the output layer to ensure that the output drying parameter value is within the expected range. The initial anti-counterfeiting code residual ink volume estimate and the initial optimal drying parameter estimate are input into the interactive enhancement module in the heterogeneous multi-task learning graph neural network. The interactive enhancement module strengthens the interaction and fusion of features through bidirectional information flow and attention mechanism to generate more accurate enhanced feature representation. The interactive enhancement module contains a bidirectional long short-term memory network and a self-attention layer. The bidirectional long short-term memory network processes sequence data in both forward and backward directions and contains 128 hidden units, which enables it to effectively capture long-term dependencies in time series. The self-attention layer improves the understanding of the relationship between features through the scaled dot product attention mechanism to obtain a more accurate feature representation. The enhanced feature representation is input into the multi-task output layer for final prediction. The multi-task output layer consists of two parallel fully connected networks, each of which contains two layers and uses LeakyReLU activation function and L2 regularization to improve the robustness of the model and prevent overfitting. Through parallel calculation, the final predicted value of the residual ink volume and the optimal drying parameters of the anti-counterfeiting code is finally obtained.

[0017] Normalize and standardize the spray feature subset to eliminate the dimensional differences between features, so that each feature is processed on a unified scale to obtain standardized spray features, which is conducive to improving the convergence speed and stability of subsequent models. Based on the standardized spray features, the first layer of graph convolution is performed using Chebyshev polynomial expansion. Chebyshev polynomial expansion is an effective spectral graph convolution method that can perform convolution operations on graph structure data through polynomial approximation. In this layer, the Chebyshev polynomial expansion order used is set to 3, and the size of the convolution kernel is 32, which means that the convolution layer can capture the local features of each node and its neighborhood in the graph and extract the first-order graph convolution features. Apply the ReLU activation function to the first-order graph convolution features to obtain the first-layer activation feature map after nonlinear processing. The introduction of the ReLU activation function helps to introduce nonlinear expression capabilities into the features, so that the model can better learn complex feature relationships. Based on the feature map after the first layer is activated, the second layer of graph convolution is performed. Similar to the first layer, the second layer of graph convolution uses the Chebyshev polynomial expansion method for convolution processing, but the size of the convolution kernel is increased to 64 to capture higher-dimensional feature information. Through the convolution operation of this layer, the second-order graph convolution features are extracted and processed using the ReLU activation function to generate the second-layer activated feature map. The third layer of graph convolution operation is performed on the second-layer activated feature map, using the same Chebyshev polynomial expansion method, but the size of the convolution kernel is further increased to 128. The convolution processing of this layer can capture more complex relationships in the graph structure and generate third-order graph convolution features. Through the third ReLU activation process, the third layer of activated feature map is obtained, which contains deep feature information after three convolutions and nonlinear transformations. The activated feature map of the third layer is input to the first fully connected layer for linear transformation, and the Dropout operation is applied to prevent overfitting. In this layer, the dimension of the feature is reduced to 64, and some neurons are randomly discarded with a Dropout rate of 0.5 to obtain a more robust reduced-dimensional feature vector. Dimensionality reduction not only helps to reduce the complexity of the model, but also retains the most critical feature information. After applying the ReLU activation function to the reduced-dimensional feature vector, the activation output of the first fully connected layer is obtained. The activation output of the first fully connected layer is input to the second fully connected layer, linearly transformed and the Sigmoid activation function is applied to obtain the estimated value of the residual ink amount of the initial anti-counterfeiting code. The Sigmoid activation function ensures that the output value is mapped between 0 and 1. The estimated value reflects the ink residue during the spraying process.

[0018] Data preprocessing is performed on the dried feature subset, including normalization and standardization operations, to eliminate the dimensional differences and distribution deviations between different features, and obtain standardized dried features with uniform distribution and uniform scale. Based on the standardized dried features, the multi-head attention mechanism is used to perform the first-layer graph attention calculation. By setting 8 attention heads, each with an output dimension of 32, the model can pay attention to the various correlations and importance of dried features in parallel from different subspaces, and generate comprehensive first-layer attention weighted features. The first-layer attention weighted features are layer-normalized. This step helps to accelerate the convergence speed of the model and improve the stability of training. It can also prevent common problems such as gradient disappearance and gradient explosion, and ensure that the features maintain effective information expression during the transmission process. On this basis, the multi-head attention mechanism is used to calculate the second-layer graph attention. The parameter settings are the same as the first layer. The more complex and abstract correlations between features are further explored to produce more representative second-layer attention weighted features. The second-layer attention weighted features are also layer-normalized to standardize the feature distribution, improve the generalization ability and robustness of the model, and obtain the second-layer normalized features. The normalized features of the second layer are input into the first fully connected layer for linear transformation and BatchNormalization operations. By setting the output dimension to 128, the model can capture and integrate high-dimensional feature information, and the application of BatchNormalization effectively reduces internal covariate shift, improves training efficiency and model performance. The ReLU activation function is applied to the output of the first layer of full connection to give the model nonlinear expression ability, so that it can better fit complex functional relationships and obtain the features after activation of the first layer. The features after activation of the first layer are input into the second fully connected layer for linear transformation and BatchNormalization operations to reduce the output dimension to 64, gradually condense feature information, and extract more critical feature expressions. The output of the second layer of full connection is processed nonlinearly by the ReLU activation function to generate the features after activation of the second layer. The features after activation of the second layer are input into the third fully connected layer for linear transformation and Tanh activation function processing. The use of the Tanh function limits the output value to between -1 and 1, which is suitable for representing continuous and bounded drying parameters, and finally obtains the estimated value of the initial optimal drying parameter.

[0019] The estimated value of the residual ink volume of the initial anti-counterfeiting code and the estimated value of the initial optimal drying parameter are concatenated to generate an initial joint feature vector. The initial joint feature vector is input into the forward propagation layer of the bidirectional long short-term memory network. Through forward propagation, the temporal information is extracted to generate a forward hidden state sequence. This process helps to capture the sequentially dependent temporal features in the joint feature vector and reflect the dynamic pattern of the features changing over time. At the same time, the same initial joint feature vector is input into the back propagation layer of the bidirectional long short-term memory network to extract the temporal information from the reverse time order to obtain the reverse hidden state sequence. The addition of back propagation enables the model to fully understand the dependency of features on the bidirectional time axis and improve the integrity and accuracy of the temporal information. The forward hidden state sequence and the reverse hidden state sequence are concatenated to obtain a bidirectional fused feature sequence. Based on the bidirectional fused feature sequence, the self-attention weight matrix is ​​calculated to identify important time points and feature relationships in the feature sequence. The self-attention layer uses a scaled dot product attention mechanism and sets the scaling factor to the square root of 8. This mechanism can effectively handle feature associations between different time steps and enable the model to better understand complex feature interactions. The attention scores are Softmax normalized to obtain normalized attention weights. Softmax normalization ensures that the sum of the attention scores between each time step is 1, so that the attention weights are assigned to the most critical features. Based on the normalized attention weights, the bidirectional fusion feature sequence is weighted summed to generate a context vector. The context vector is residually connected to the original bidirectional fusion feature sequence. Residual connections help retain the original feature information while enhancing the expressiveness of feature representations and preventing information from attenuating when it is passed layer by layer in the deep network. The residual features are layer-normalized to standardize the distribution of features and ensure their stability in subsequent processing. The normalized feature representation is input into the feedforward neural network for nonlinear transformation to generate the final enhanced feature representation. The nonlinear transformation of the feedforward neural network gives the features stronger expressiveness, enabling the system to more accurately predict the amount of residual ink and the optimal drying parameters of the anti-counterfeiting code in the scenario of multi-task learning.

[0020] S5, converts the relationship between ink drying time and drying temperature into constraint conditions, and optimizes the heterogeneous multi-task learning graph neural network based on the residual ink amount of the anti-counterfeiting code and the optimal drying parameters to obtain an intelligent diagnosis model for the printing process; Among them, the relationship between ink drying time and drying temperature is modeled. By analyzing the equipment operation data and experimental results, a mathematical model that reflects how the ink drying time changes with the drying temperature is established to obtain the ink drying time function. The ink drying time function is discretized to convert the continuous function curve into discrete corresponding relationship points of ink drying time and drying temperature. Based on the corresponding relationship points, a constraint condition matrix is ​​constructed to define the soft constraints in the drying parameter optimization process. The soft constraints reflect the reasonable range and relationship between the ink drying time and the drying temperature, and guide the model to generate parameters that are more in line with the actual process requirements during the optimization process. The loss function of the heterogeneous multi-task learning graph neural network is modified, and the aforementioned soft constraint terms are introduced into the loss function to form a constrained optimization objective function. Based on the constrained optimization objective function and the residual ink amount of the anti-counterfeiting code, the parameters of the ink residual amount estimation subnetwork are updated. By continuously adjusting the network parameters, the model can more accurately estimate the residual ink amount of the anti-counterfeiting code during the spraying process and generate an optimized ink residual amount estimation subnetwork. Similarly, based on the constrained optimization objective function and the optimal drying parameters, the parameters of the drying parameter optimization subnetwork are updated to ensure that the drying parameter optimization subnetwork can better consider the impact of drying temperature on ink drying time in the actual process, output more reasonable and effective drying parameters, and obtain the optimized drying parameter optimization subnetwork. The optimized ink residue estimation subnetwork is integrated with the drying parameter optimization subnetwork to obtain the optimized heterogeneous multi-task learning graph neural network. Based on the optimized heterogeneous multi-task learning graph neural network, decision rules and threshold judgment logic are constructed to form a complete intelligent diagnosis model for the printing process. The diagnostic model can automatically determine whether the key parameters in the printing process are within the normal range based on real-time data, and provide early warning information in abnormal situations, thereby improving the reliability of equipment operation and printing quality.

[0021] S6, analyzes the real-time data set collected during the anti-counterfeiting code printing process through the intelligent diagnosis model of the printing process, and generates early warning information on spraying quality and drying efficiency.

[0022] Specifically, the real-time data set collected by the intelligent digital printing equipment during operation is input into the constructed intelligent diagnosis model. The real-time data set includes the ink flow, spraying speed, spraying pressure of the spraying device, and key parameters such as the temperature, humidity and equipment power consumption of the drying equipment. The intelligent diagnosis model uses the optimized ink residual estimation subnetwork and drying parameter optimization subnetwork previously trained by heterogeneous multi-task learning graph neural network to make real-time predictions on the spraying quality and drying efficiency. The model estimates the residual ink volume of the anti-counterfeiting code based on the real-time data, and evaluates the efficiency of the drying process in combination with the drying parameters. These prediction results are generated instantly and compared with the pre-set standard values ​​to determine whether the current printing process is running within the ideal range. During the analysis process, the diagnosis model identifies the trend changes in spraying quality and drying efficiency by processing time series data. By comparing the difference between current data and historical data, the model can capture subtle changes in the equipment status and identify potential abnormalities in advance. For example, if the spraying uniformity fluctuates or the drying efficiency drops significantly, the model can immediately identify these anomalies and determine whether these changes will affect the printing quality through internal decision rules and threshold judgment logic. When the diagnostic model detects that the spray quality or drying efficiency exceeds the preset safety range, or there is a potential abnormal trend, the system will generate an early warning message. The early warning message will explain the problem, including possible uneven spraying or insufficient drying, and suggest that the operator take appropriate measures, such as adjusting the spray speed, spray pressure, or adjusting the drying temperature and humidity. These early warning messages will be promptly communicated to the operator through the equipment's monitoring interface to ensure timely intervention before the problem expands, thereby avoiding affecting the printing quality.

[0023] In one example, anti-counterfeiting code printing operation data is collected for the spraying device and drying device in the intelligent digital printing equipment to obtain a comprehensive operation data set, including: The ink flow sensor of the spraying device in the intelligent digital printing equipment is used to collect real-time data to obtain the ink flow time series data, and the spraying speed is calculated based on the ink flow time series data to obtain the spraying speed data; Collect data from the temperature sensor and humidity sensor of the drying equipment to obtain drying temperature and humidity data, and estimate the drying efficiency based on the drying temperature and humidity data to obtain a drying efficiency index; Collecting data from a pressure sensor of a spraying device to obtain spraying pressure data, and calculating spraying uniformity based on the spraying pressure data and spraying speed data to obtain a spraying uniformity index; Collect data from the current sensors of the spraying device and drying equipment to obtain the equipment power consumption data; Based on the ink flow time series data, spraying speed data, drying temperature and humidity data, drying efficiency index, spraying pressure data, spraying uniformity index and equipment power consumption data, multi-dimensional data fusion is performed to obtain a comprehensive operation data set.

[0024] In this example, the ink flow sensor of the spraying device is used for real-time data collection to obtain the ink flow time series data, record the flow changes of the ink flowing through the spray head at each moment during the spraying process, and reflect the dynamic situation of the spraying process. By analyzing the time series data, the spraying speed is calculated. Assuming that in the time period The ink flow rate of the spray head is , the cross-sectional area of ​​the spray head is , then the spraying speed It can be calculated by the formula: ; in, is the ink flow timing data, is the cross-sectional area of ​​the spray head. Through calculation, the speed data of the spraying process is obtained in real time. At the same time, the temperature sensor and humidity sensor of the drying equipment are collected to obtain the temperature and humidity data of the drying process. The efficiency of the drying process is closely related to temperature and humidity. Assume that the drying efficiency can be calculated using the formula: ; in, is the weight of ink before drying, It is the weight of the ink remaining after drying. The temperature and humidity data are used to adjust and control the environmental conditions during the drying process to ensure that the ink can be fully dried within the specified time to achieve the ideal printing effect. At the same time, the pressure sensor of the spraying device is used to collect data to obtain the pressure data during the spraying process. Based on the pressure data and spraying speed data, the spraying uniformity index is calculated. Spraying uniformity It can be expressed as the square of the sum of the deviations between the speed at a certain point in time and the ideal speed. The specific formula is: ; in is the current spraying speed, is the preset ideal spraying speed, It is the number of times or time points of measurement. In this way, the uniformity of the spraying process is evaluated in real time to ensure that the spraying quality meets the requirements. Data is collected from the current sensors of the spraying device and the drying equipment to obtain the power consumption data of the equipment. The power consumption data can not only reflect the energy usage of the equipment, but also serve as an important basis for judging whether the equipment is in normal operation. If the power consumption increases or decreases abnormally, it may indicate that the equipment is faulty or working abnormally. Multi-dimensional data fusion is performed based on ink flow time series data, spraying speed data, drying temperature and humidity data, drying efficiency index, spraying pressure data, spraying uniformity index and equipment power consumption data to generate a comprehensive operation data set. Data from different sources are integrated and correlated to form a global and comprehensive description of the equipment status. For example, the spraying speed and uniformity are cross-analyzed to find out the change law of uniformity at different spraying speeds, or the drying temperature and humidity are correlated with the drying efficiency to optimize the parameter settings in the drying process.

[0025] In one example, a dynamic weighted adjacency matrix is ​​constructed based on a comprehensive operating data set, with spraying equipment and drying equipment as target nodes, and an equipment component relationship graph structure is constructed, including: The comprehensive operation data set is divided into time windows to obtain multiple time segment data subsets, and the correlation between each device component is calculated based on the time segment data subsets to obtain the component correlation matrix; Perform threshold filtering on the correlation matrix between components to obtain a sparse correlation matrix, and initialize the graph structure of the device components based on the sparse correlation matrix to obtain an initial graph structure; The initial graph structure is topologically optimized, and the spraying device and drying equipment are set as target nodes to obtain a hierarchical graph structure. Based on the hierarchical graph structure, the edge weights are dynamically updated and calculated to obtain a dynamic weighted adjacency matrix, and the dynamic weighted adjacency matrix is ​​normalized to obtain a standardized adjacency matrix; Based on the standardized adjacency matrix and preset device topology rules, the hierarchical graph structure is adjusted to obtain the device component relationship graph structure.

[0026] In this example, the entire data set is divided into multiple time-segment data subsets according to the time dimension, and the operating status of the equipment in different time periods is analyzed in a fine-grained manner to capture the changing trend of the equipment in the time dimension. Assume that the size of the time window is , then each time segment data subset Contains all the running data within the time window. Based on each time segment data subset , calculate the correlation between each device component in the intelligent digital printing device. By calculating the correlation between each component in different time segments, a correlation matrix reflecting the internal structure and operating status of the device is formed. Assumptions and Respectively, in time Internal components and components The state data of the component and The correlation between It can be calculated using the Pearson correlation coefficient formula: ; in, and The components are and components The average state value within the time window, Represents the total number of time points in the time window. The correlation matrix calculated by the above formula , which reflects the degree of correlation between device components in each time segment. Threshold filtering is performed on the correlation matrix to remove component connections with weak correlation and retain more meaningful strong correlation connections to obtain a sparse correlation matrix. . Set a threshold for each correlation matrix , and set the correlation values ​​below this threshold to zero, and retain the correlation values ​​greater than or equal to the threshold. Sparse correlation matrix It can better reflect the main relationship between device components without being disturbed by low-correlation noise. Correlation matrix based on sparseness , initialize the graph structure of the device components. Treat each device component as a node in the graph structure, use the correlation values ​​retained in the sparse matrix as the edges in the graph, connect the related device component nodes, and generate an initial graph structure , where nodes represent the components of the equipment and edges represent the strong associations between components. The spraying device and drying equipment are set as the target nodes in the graph structure, and the entire graph structure is hierarchically processed to obtain a hierarchical graph structure The goal of hierarchical processing is to highlight the core components of the device and clarify the upstream and downstream relationships between the components, so that the graph structure can better reflect the workflow and interdependencies of the device. The edge weights are dynamically updated to reflect the real-time changes in the associations between device components. Dynamic weighting can be achieved by analyzing the real-time data during device operation and recalculating the edge weights. Assume that the edge weights are expressed as , its update formula can be expressed as: ; in, is the previous weight value, is the currently calculated correlation value, is a smoothing factor used to control the balance between historical weights and current relevance values. Through dynamic updates, a dynamic weighted adjacency matrix reflecting the current device operating status is obtained. . For the dynamic weighted adjacency matrix Perform normalization to obtain the standardized adjacency matrix The normalization operation usually uses the graph Laplacian matrix or other normalization methods to ensure that each edge weight in the graph structure is in a comparable range. and preset device topology rules, hierarchical graph structure Adjustments are made to ensure that the graph structure can reflect the real-time status of the device and conform to the topological design principles of the device. The optimized device component relationship graph structure is finally obtained. .

[0027] In one example, based on the equipment component relationship graph structure, the spray uniformity and drying efficiency features of the comprehensive operation data set are extracted to create a fused spatiotemporal feature vector, including: Perform graph convolution operation on the equipment component relationship graph structure to obtain an initial node feature matrix, and based on the initial node feature matrix, perform feature aggregation on the spray device nodes to obtain the spray device aggregation features; Perform nonlinear transformation on the aggregated features of the spraying device to obtain the initial features of the spraying uniformity, and based on the initial node feature matrix, perform feature aggregation on the nodes of the drying equipment to obtain the aggregated features of the drying equipment, and perform nonlinear transformation on the aggregated features of the drying equipment to obtain the initial features of the drying efficiency; Based on the initial features of spraying uniformity, the first one-dimensional convolutional neural network is used to extract time series features to obtain the time series features of spraying uniformity, and based on the initial features of drying efficiency, the second one-dimensional convolutional neural network is used to extract time series features to obtain the time series features of drying efficiency; The attention mechanism is used to calculate the time series features of spraying uniformity and drying efficiency to obtain weighted feature vectors, and time dependency modeling is performed based on the weighted feature vectors to obtain long-term dependency features. The long-term dependency features and weighted feature vectors are concatenated to obtain a fused spatiotemporal feature vector.

[0028] In this example, the device component relationship graph structure is processed to extract key features. Device component relationship graph structure By node collection and edge set Composition, where each node Represents a component of a device, edge It represents the association between the two components. Through the graph convolution operation, the features of each node are extracted from the graph structure, and the initial node feature matrix of the entire graph is obtained through information transmission between nodes. The process of graph convolution can be described as a combination of linear transformation of node features and neighborhood aggregation. Assume that the node The eigenvector of , the set of its neighbor nodes is , then the update formula of graph convolution can be expressed as: ; in, Representation Node In the The feature vector after layer convolution, It is The weight matrix of the layer, is the bias vector, is the activation function, and The nodes are and nodes Through the graph convolution operation, the feature information of each node is aggregated together to obtain the initial node feature matrix of the entire graph structure. Based on the initial node feature matrix , perform feature aggregation on the spray device nodes to obtain the aggregated features of the spray device. The spray device is usually composed of multiple components, and the features of these components need to be aggregated to reflect the overall characteristics of the entire spray device. Feature aggregation can be achieved by averaging or weighted summation. Assume that the set of component nodes of the spray device is , then the aggregation characteristics of the spraying device It can be expressed as: ; in, is the number of components of the spraying device. By aggregating these features, the overall feature representation of the spraying device is obtained. The aggregated features of the spraying device are nonlinearly transformed to obtain the initial features of the spraying uniformity. Nonlinear transformation is usually implemented through activation functions such as ReLU or Tanh, so that the features can capture more complex patterns. Similarly, based on the initial node feature matrix, the features of the drying equipment nodes are aggregated to obtain the aggregated features of the drying equipment Then, the aggregated features of the drying equipment are transformed nonlinearly to generate the initial features of the drying efficiency. The initial features are input into a one-dimensional convolutional neural network (1D CNN) for temporal feature extraction. Assume that the initial features of the spray uniformity are , then the temporal features after 1DCNN

[0029] Conv1D ; Among them, Conv1D represents a one-dimensional convolution operation. This convolution process moves on the time series through a sliding window and extracts the characteristic patterns in the local time period. Similarly, the initial characteristics of the drying efficiency After the second 1D CNN, the time series characteristics of drying efficiency are obtained The attention mechanism is used to calculate the time series features of spray uniformity and drying efficiency to generate weighted feature vectors. The attention mechanism calculates the importance score of each time point and weights the features according to these scores to highlight the feature information of key time points. Assume that the attention score matrix is , then the weighted eigenvector It can be expressed as: ; in, Represents the input time series characteristics (such as It is a weight matrix calculated by the attention mechanism. In this way, the time point features that have an important impact on the spraying quality and drying efficiency are captured. Based on the weighted feature vector, time dependency modeling is performed to obtain long-term dependency features. Time dependency modeling is usually implemented through recursive neural networks (RNN) or long short-term memory networks (LSTM), which can capture long-term dependencies in sequence data. For example, LSTM can store and process important time dependency information through its memory units to generate more representative long-term dependency features. The long-term dependency features are concatenated with the previous weighted features in the same amount to form a fused spatiotemporal feature vector.

[0030] In one example, the fused spatiotemporal feature vector is input into a heterogeneous multi-task learning graph neural network for multi-task prediction to obtain the residual ink volume and optimal drying parameters of the anti-counterfeiting code, including: Perform feature decomposition on the fused spatiotemporal feature vector to obtain a spraying feature subset and a drying feature subset; The spray feature subset is input into the ink residual estimation subnetwork in the heterogeneous multi-task learning graph neural network, and the estimated value of the residual ink of the initial anti-counterfeiting code is obtained through forward propagation calculation; the ink residual estimation subnetwork includes three graph convolution layers and two fully connected layers; the graph convolution layer uses Chebyshev polynomials to approximate the spectral graph convolution, and the activation function is ReLU; the fully connected layer uses Dropout regularization, and the output layer uses Sigmoid activation function; The drying feature subset is input into the drying parameter optimization subnetwork in the heterogeneous multi-task learning graph neural network, and the initial optimal drying parameter estimate is obtained through forward propagation calculation; the drying parameter optimization subnetwork includes two graph attention layers and three fully connected layers; the graph attention layer uses a multi-head attention mechanism, and the number of heads is set to 8; the fully connected layer uses BatchNormalization, and the output layer uses the Tanh activation function; The estimated value of the residual ink volume of the initial anti-counterfeiting code and the estimated value of the initial optimal drying parameter are input into the interactive enhancement module in the heterogeneous multi-task learning graph neural network, and the enhanced feature representation is obtained through bidirectional information flow and attention mechanism; the interactive enhancement module includes a bidirectional long short-term memory network and a self-attention layer; the bidirectional long short-term memory network contains 128 hidden units, and the self-attention layer uses a scaled dot product attention mechanism; The enhanced feature representation is input into the multi-task output layer, and the residual ink amount and the optimal drying parameters of the anti-counterfeiting code are obtained through parallel calculation; the multi-task output layer includes two parallel fully connected networks, each of which contains two layers, using the LeakyReLU activation function and L2 regularization.

[0031] In this example, the fused spatiotemporal feature vector is feature decomposed into a spraying feature subset and a drying feature subset. The spraying feature subset mainly contains features related to the spraying device, such as spraying speed, pressure, etc., while the drying feature subset includes features related to the drying equipment, such as drying temperature, humidity, etc. The spraying feature subset is input into the ink residue estimation subnetwork in the heterogeneous multi-task learning graph neural network. This subnetwork estimates the amount of residual ink in the anti-counterfeiting code through forward propagation calculation. The ink residue estimation subnetwork consists of three graph convolution layers and two fully connected layers. The graph convolution layer uses Chebyshev polynomials to approximate spectral graph convolution. This method can effectively process graph structured data and capture the complex relationship between nodes and their neighbor nodes. Assume that the input node feature matrix is , the calculation formula of Chebyshev polynomial approximate spectral graph convolution is: ; in, Indicates The node feature matrix of the layer, is the normalized Laplacian matrix of the graph, are Chebyshev polynomials, It is The weight matrix of the layer, is the activation function ReLU. Through the graph convolution operation, deeper features are extracted from the graph structure. The features are input into two fully connected layers for processing. The fully connected layer uses Dropout regularization to prevent the model from overfitting. The Dropout operation randomly discards the output of a part of the neurons, so that the model will not rely too much on certain specific features during training, thereby improving the generalization ability of the model. In the final output layer, the Sigmoid activation function is used to limit the output value between 0 and 1 to obtain the estimated value of the residual ink volume of the initial anti-counterfeiting code. At the same time, the drying feature subset is input into the drying parameter optimization subnetwork in the heterogeneous multi-task learning graph neural network. This subnetwork estimates the optimal drying parameters through forward propagation calculation. The drying parameter optimization subnetwork includes two graph attention layers and three fully connected layers. The graph attention layer uses a multi-head attention mechanism, and the number of heads in each layer is set to 8, which enables the model to focus on different aspects of the drying features in parallel from different subspaces and capture richer feature information. The calculation process of the graph attention layer can be expressed as: ; in, is the query matrix, is the key matrix, is the value matrix, is the dimension of the key vector. The scaled dot product attention mechanism can effectively capture the correlation between features. In the graph attention layer, the features are input into three fully connected layers for processing. The fully connected layer uses BatchNormalization operation to accelerate the convergence of the network and improve the stability of training. In the final output layer, the Tanh activation function is used to limit the output value between -1 and 1 to obtain the initial optimal drying parameter estimate. The initial anti-counterfeiting code residual ink volume estimate and the initial optimal drying parameter estimate are input into the interaction enhancement module in the heterogeneous multi-task learning graph neural network. The interaction enhancement module extracts the interaction between features through bidirectional information flow and attention mechanism to obtain enhanced feature representation. The module contains a bidirectional long short-term memory network (BiLSTM) and a self-attention layer. BiLSTM processes sequence data in both forward and reverse directions and contains 128 hidden units to capture long-term dependencies in time series. The self-attention layer improves the understanding of the relationship between features through the scaled dot product attention mechanism. Its calculation formula is similar to the previous attention layer. By weighted summing the features of each time step in the sequence, the feature information of key time points is highlighted to generate enhanced feature representation. The enhanced feature representation is input into the multi-task output layer. The multi-task output layer includes two parallel fully connected networks, each of which contains two layers. The fully connected network uses the LeakyReLU activation function to solve the "neuron death" problem that may be caused by ReLU, and uses L2 regularization to prevent the model from overfitting. Through parallel calculation, the final estimated value of the residual ink amount of the anti-counterfeiting code and the optimal drying parameters are finally obtained.

[0032] In one example, the spray feature subset is input into the ink residual estimation subnetwork in the heterogeneous multi-task learning graph neural network, and the estimated value of the residual ink amount of the initial anti-counterfeiting code is obtained by forward propagation calculation, including: Normalizing and standardizing the spray feature subset to obtain standardized spray features; Based on the standardized spraying features, the first-layer graph convolution operation is performed using Chebyshev polynomial expansion to obtain the first-order graph convolution features; the Chebyshev polynomial expansion order is 3, and the convolution kernel size is 32; Perform ReLU activation function processing on the first-order graph convolution feature to obtain the feature map after the first layer activation; Based on the feature map after activation in the first layer, the second layer of graph convolution operation is performed to obtain the second-order graph convolution feature; the second layer of graph convolution uses the same Chebyshev polynomial expansion method as the first layer, and the convolution kernel size is 64; The second-order graph convolution features are processed by the ReLU activation function to obtain the feature map after the second layer activation; Based on the feature map after activation in the second layer, the third layer of graph convolution operation is performed to obtain the third-order graph convolution features; the third layer of graph convolution uses the same Chebyshev polynomial expansion method as the first two layers, and the convolution kernel size is 128; The third-order graph convolution features are processed by the ReLU activation function to obtain the feature map after the third layer activation; The activated feature map of the third layer is input into the first fully connected layer, and linear transformation and Dropout operation are performed to obtain a reduced-dimensional feature vector; the output dimension of the first fully connected layer is 64, and the Dropout rate is 0.5; The reduced-dimensional feature vector is processed by the ReLU activation function to obtain the activation output of the first fully connected layer; The activation output of the first fully connected layer is input into the second fully connected layer, and linear transformation and Sigmoid activation function processing are performed to obtain the estimated value of the residual ink amount of the initial anti-counterfeiting code.

[0033] In this example, the spray feature subset is normalized and standardized to obtain the standardized spray feature. Normalization adjusts features of different scales to the same range, usually scaling all feature values ​​to between 0 and 1, thereby avoiding the imbalance between features due to different dimensions. Standardization further adjusts the distribution of features so that they have the same mean and variance, and adjusts the feature data to a standard normal distribution with a mean of 0 and a standard deviation of 1. Based on the standardized spray features, the first layer of graph convolution is performed using Chebyshev polynomial expansion. Chebyshev polynomial expansion is an efficient spectral graph convolution method that performs convolution operations by approximating the eigendecomposition of the graph Laplacian operator. In this layer, the order of the Chebyshev polynomial expansion is set to 3, and the size of the convolution kernel is 32. The calculation formula for graph convolution is: ; in, It is The input feature matrix of the layer, is the normalized Laplacian matrix of the graph, are Chebyshev polynomials, It is The weight matrix of the layer, is the activation function. In this process, Chebyshev polynomial expansion is used to approximate spectral graph convolution, and complex processing of features is achieved through the superposition of polynomials. After this layer of processing, the first-order graph convolution feature is obtained. In order to introduce nonlinearity and enhance the expressiveness of the model, the first-order graph convolution feature is processed by the ReLU activation function. The formula of the ReLU activation function is: ; Its characteristic is that it directly outputs the original value when the input is greater than 0, otherwise it outputs 0. This characteristic enables the model to better capture the nonlinear relationship in the data. After ReLU activation, the feature map of the first layer is obtained. Based on the feature map of the first layer, the second layer of graph convolution is performed. In this layer, the same Chebyshev polynomial expansion method as the first layer is used, but the size of the convolution kernel is increased to 64 to capture higher-dimensional feature information. Similar to the first layer, the second layer of graph convolution also processes the input feature map by approximating the graph Laplacian operator, and captures more complex features through polynomial expansion. After this layer of convolution processing, the second-order graph convolution feature is obtained. Similarly, the second-order graph convolution feature is processed by the ReLU activation function to obtain the feature map of the second layer. Based on the feature map of the second layer, the third layer of graph convolution is performed. The third layer of graph convolution still uses the same Chebyshev polynomial expansion method as the first two layers, and the size of the convolution kernel is further increased to 128. By increasing the size of the convolution kernel layer by layer, more complex and abstract feature patterns in the data are captured. After the third layer of graph convolution processing, the third-order graph convolution features are obtained. The third-order graph convolution features are processed by the ReLU activation function to obtain the feature map after the activation of the third layer. The feature map after the activation of the third layer is input into the first fully connected layer for linear transformation and Dropout operation. The function of the fully connected layer is to map the information in the feature map to a low-dimensional space and regularize it through the Dropout operation to prevent overfitting. The output dimension of the first fully connected layer is set to 64, which means that the high-dimensional feature vector is compressed into a 64-dimensional space. The Dropout rate is set to 0.5, which means that half of the neuron outputs will be randomly discarded during the training process, thereby improving the generalization ability of the model. After linear transformation and Dropout operation, the reduced-dimensional feature vector is obtained. The reduced-dimensional feature vector is processed by the ReLU activation function to obtain the activation output of the first fully connected layer. The activation output of the first fully connected layer is input into the second fully connected layer for linear transformation and Sigmoid activation function processing. The formula of the Sigmoid activation function is: ; The output value is limited to between 0 and 1, which is suitable for probability estimation in binary classification or regression tasks. Through this operation, an estimated value of the amount of residual ink in the initial anti-counterfeiting code is obtained, which reflects the amount of ink that may remain during the spraying process under the current conditions.

[0034] In one example, a drying feature subset is input into a drying parameter optimization subnetwork in a heterogeneous multi-task learning graph neural network, and an initial optimal drying parameter estimate is obtained through forward propagation calculation, including: Perform data preprocessing on the drying feature subset, including normalization and standardization operations, to obtain standardized drying features; Based on the standardized drying features, the multi-head attention mechanism is used to perform the first-layer graph attention calculation to obtain the first-layer attention weighted features; the number of heads of the multi-head attention mechanism is 8, and the output dimension of each head is 32; Perform layer normalization on the first layer of attention weighted features to obtain the first layer of normalized features; Based on the normalized features of the first layer, the multi-head attention mechanism is used to perform the second-layer graph attention calculation to obtain the second-layer attention weighted features; the parameter setting of the second-layer multi-head attention mechanism is the same as that of the first layer; Perform layer normalization on the second-layer attention weighted features to obtain the second-layer normalized features; The second layer normalized features are input into the first fully connected layer, and linear transformation and BatchNormalization operations are performed to obtain the first layer fully connected output; the output dimension of the first fully connected layer is 128; Perform ReLU activation function on the fully connected output of the first layer to obtain the features after activation of the first layer; The activated features of the first layer are input into the second fully connected layer, and linear transformation and BatchNormalization operations are performed to obtain the second fully connected output; the output dimension of the second fully connected layer is 64; The fully connected output of the second layer is processed by ReLU activation function to obtain the features after activation of the second layer; The activated features of the second layer are input into the third fully connected layer, and linear transformation and Tanh activation function processing are performed to obtain the initial optimal drying parameter estimation value.

[0035] In this example, data preprocessing is performed on the dried feature subset, including normalization and standardization operations. Features of different dimensions are adjusted to the same range and have the same mean and variance. Normalization is usually used to scale the feature values ​​to between 0 and 1, and standardization is used to adjust them to a standard normal distribution with a mean of 0 and a standard deviation of 1 to obtain standardized dried features. The first-level graph attention calculation is performed using a multi-head attention mechanism. The multi-head attention mechanism processes input features in parallel through multiple attention heads to capture the correlation between features from different angles. The output dimension of each attention head is set to 32, and the number of heads is 8, which means that in the first-level graph attention calculation, 8 groups of attention scores will be calculated in parallel, and a 32-dimensional weighted feature representation will be generated for each input feature. The calculation process of the multi-head attention mechanism can be expressed as: ; in, is the query matrix, is the key matrix, is the value matrix, is the dimension of the key vector. By scaling the dot product attention mechanism, the relationship between different features in the dried feature subset is effectively captured, and the first layer of attention weighted features are obtained. In order to ensure the stability and uniform distribution of the features, the first layer of attention weighted features are layer normalized. Layer normalization normalizes the features of each input sample at the level so that the output of each layer remains stable and consistent during training. This helps to accelerate the convergence of the model and prevent the problem of gradient disappearance or explosion. After layer normalization, the first layer of normalized features is obtained. Based on the first layer of normalized features, the second layer of graph attention calculation is performed using the multi-head attention mechanism. The multi-head attention mechanism parameter settings of the second layer are the same as those of the first layer, which is also 8 heads, and the output dimension of each head is 32. Through another attention calculation, the deeper correlation between features is further captured and the model's understanding of the dried features is enhanced. Through the second layer of graph attention calculation, the second layer of attention weighted features are generated. As in the first layer, the second layer of attention weighted features are layer normalized to ensure the consistency of feature distribution and the stability of the model. After this processing, the second layer of normalized features is obtained. The second layer of normalized features is input into the first fully connected layer for linear transformation. The fully connected layer maps the input high-dimensional features to a lower-dimensional space. The output dimension of the first fully connected layer is set to 128, which means that the features are mapped to a 128-dimensional vector space. The BatchNormalization operation is performed on the output of the first fully connected layer, and each small batch of data is normalized through standardization, thereby accelerating the training and stabilizing the learning process of the model. After this step, the first layer of fully connected output is generated. The ReLU activation function is processed on the first layer of fully connected output. The formula of the ReLU activation function is: ; This function sets negative values ​​directly to 0 and leaves positive values ​​unchanged, allowing the model to effectively learn nonlinear patterns in the data. After ReLU processing, the features after activation of the first layer are obtained. The features after activation of the first layer are input into the second fully connected layer for linear transformation and BatchNormalization operation. The output dimension of the second fully connected layer is set to 64 to further compress the dimension of the feature vector. By gradually reducing the dimension, the key features in the data are extracted and redundant information is removed. The introduction of the BatchNormalization operation continues to maintain the consistency of the features at the level, ensuring the stability of the model training process. Similarly, the ReLU activation function is applied to the output of the second fully connected layer to generate the features after activation of the second layer. The features after activation of the second layer are input into the third fully connected layer for linear transformation and processed using the Tanh activation function. The formula of the Tanh activation function is: ; The output value is limited to between -1 and 1. This activation function is often used for regression problems, especially when dealing with continuous values. After Tanh activation, the estimated value of the initial optimal drying parameter is obtained.

[0036] In one example, the estimated value of the residual ink volume of the initial anti-counterfeiting code and the estimated value of the initial optimal drying parameter are input into the interactive enhancement module in the heterogeneous multi-task learning graph neural network. Through the bidirectional information flow and attention mechanism, the enhanced feature representation is obtained, including: Perform feature concatenation on the initial estimated value of the residual ink volume of the anti-counterfeiting code and the initial estimated value of the optimal drying parameter to obtain an initial joint feature vector; The initial joint feature vector is input into the forward propagation layer of the bidirectional long short-term memory network to extract the temporal information and obtain the forward hidden state sequence; The initial joint feature vector is input into the back propagation layer of the bidirectional long short-term memory network to extract the time sequence information and obtain the reverse hidden state sequence; Concatenate the forward hidden state sequence and the reverse hidden state sequence to obtain a bidirectional fusion feature sequence; Based on the bidirectional fusion feature sequence, the self-attention weight matrix is ​​calculated to obtain the attention score; the self-attention layer uses the scaled dot product attention mechanism with a scaling factor of the square root of 8; Perform Softmax normalization on the attention score to obtain the normalized attention weight; Based on the normalized attention weights and the bidirectional fusion feature sequence, a weighted sum operation is performed to obtain the context vector; Perform residual connection between the context vector and the bidirectional fusion feature sequence to obtain the residual feature; The residual features are layer-normalized to obtain normalized feature representation, and the normalized feature representation is input into a feedforward neural network for nonlinear transformation to obtain enhanced feature representation.

[0037] In this example, the initial anti-counterfeiting code residual ink volume estimate and the initial optimal drying parameter estimate are concatenated to obtain a comprehensive initial joint feature vector. The initial joint feature vector is input into the forward propagation layer of the bidirectional long short-term memory network (BiLSTM). BiLSTM is a recurrent neural network that can capture forward and backward dependencies in time series. The forward propagation layer extracts information in the time dimension by sequentially processing the features in the time series and generates a forward hidden state sequence. Assume that at time step When , the initial joint eigenvector is , the forward hidden state The calculation can be expressed as: ; in, is the forward LSTM unit, is the hidden state of the previous time step. This operation gradually extracts the forward dependencies of each time step in the sequence and generates a complete sequence of forward hidden states. At the same time, the same initial joint feature vector is input into the back propagation layer of BiLSTM. In contrast to the forward propagation layer, the back propagation layer starts from the end of the time series, processes the features in reverse order, and extracts the backward dependencies in the time series. Reverse hidden state The calculation formula is: ; in, is the hidden state of the next time step. Through this operation, a complete reverse hidden state sequence is obtained. The forward hidden state sequence and the reverse hidden state sequence are concatenated to obtain a bidirectional fusion feature sequence , and its calculation formula is: ; The splicing method combines the forward and backward information of the time series, so that the feature sequence can not only reflect the current state, but also take into account the past and future contextual information at the same time, enhancing the time dependency of the feature. Based on the bidirectional fusion feature sequence, the self-attention weight matrix is ​​calculated to extract the key time point information. The self-attention mechanism can calculate the importance weight of each time point based on the information in the feature sequence. The calculation process of self-attention can be implemented by the scaled dot product attention mechanism, and its formula is: ; in, is the query matrix, is the key matrix, is the value matrix, is the dimension of the key vector, and the scaling factor is the square root of 8. This mechanism can effectively calculate the similarity between each time step and other time steps to obtain the attention score, thereby determining which time points are most important for feature representation. In order to ensure the comparability and sum of the attention scores to 1, the attention scores are normalized by Softmax to obtain the normalized attention weights. The calculation formula of the Softmax function is: ; in, Indicates The role of Softmax is to convert the scores into probability distributions so that the attention weights at each time point can be compared in the range of [0, 1]. A weighted sum operation is performed based on the normalized attention weights and the bidirectional fusion feature sequence to generate a context vector The calculation formula of the context vector is: ; Among them, Attention(·) is a weighted operation based on attention weight. The context vector integrates the important information in the entire time series and can better reflect the global features. In order to enhance the stability of feature expression and retain more original information, the context vector is residually connected with the bidirectional fusion feature sequence to obtain the residual feature The residual connection alleviates the problem of information loss in deep networks by adding skip connections between features. The calculation formula is: ; After layer normalization of the residual features, a normalized feature representation is obtained. The purpose of layer normalization is to standardize the features at each layer to ensure that information propagation in the deep network is more stable and reduce the risk of gradient vanishing or exploding. The normalized feature representation is input into the feedforward neural network for nonlinear transformation to obtain an enhanced feature representation. The role of the feedforward neural network is to extract and combine complex patterns in the features and enhance the representation ability of the features through nonlinear transformation. Through the above operations, an enhanced feature representation is obtained.

[0038] In one example, the relationship between ink drying time and drying temperature is converted into constraint conditions, and the heterogeneous multi-task learning graph neural network is optimized based on the amount of residual ink in the anti-counterfeiting code and the optimal drying parameters to obtain an intelligent diagnosis model for the printing process, including: Model the relationship between ink drying time and drying temperature to obtain the ink drying time function; Discretize the ink drying time function to obtain the corresponding relationship points between the ink drying time and the drying temperature; Based on the corresponding relationship points, a constraint matrix is ​​constructed to obtain the soft constraints for drying parameter optimization; The loss function of the heterogeneous multi-task learning graph neural network is modified, and soft constraints are added to obtain the constrained optimization objective function; Based on the constrained optimization objective function and the amount of ink remaining in the anti-counterfeiting code, the parameters of the ink remaining estimation subnetwork of the heterogeneous multi-task learning graph neural network are updated to obtain the optimized ink remaining estimation subnetwork; Based on the constrained optimization objective function and the optimal drying parameters, the parameters of the drying parameter optimization subnetwork of the heterogeneous multi-task learning graph neural network are updated to obtain the optimized drying parameter optimization subnetwork; The optimized ink residue estimation subnetwork and the optimized drying parameter optimization subnetwork are integrated to obtain an optimized heterogeneous multi-task learning graph neural network. Based on the optimized heterogeneous multi-task learning graph neural network, decision rules and threshold judgment logic are constructed to obtain an intelligent diagnosis model of the printing process.

[0039] In this example, the relationship between ink drying time and drying temperature is modeled to describe how the drying time of ink changes at different drying temperatures. Assuming that the drying time has a nonlinear relationship with temperature, an ink drying time function is fitted through experimental data. ,in Indicates the drying time, Indicates the drying temperature. This function can be fitted by regression analysis or machine learning methods, and can be in the form of polynomial regression, exponential model, etc. For example: ; in, , and is the fitted model parameter. This function reflects the effect of temperature on ink drying time, where The higher the drying time The shorter the ink drying time, the more discrete the function is. The continuous function is converted into a set of discrete corresponding points. For example, the temperature range is divided into several equally spaced points. , and then calculate the drying time corresponding to each temperature , get the corresponding relationship point set between ink drying time and drying temperature Based on the discrete corresponding relationship points, a constraint matrix is ​​constructed to obtain the soft constraints for drying parameter optimization. The construction of the matrix needs to take into account the reasonable relationship between temperature and drying time, that is, between each pair of temperature and time points, the drying time should not exceed a certain threshold range to ensure that the drying efficiency is guaranteed in actual operation. Elements It can be expressed as: ; The threshold can be determined according to the specific process requirements. The soft constraints constructed in this way can guide the model to avoid unreasonable drying parameters during the optimization process. The loss function of the heterogeneous multi-task learning graph neural network is modified, and the above soft constraints are added to obtain the constrained optimization objective function. The original loss function It may be based on mean square error or cross entropy, and the constrained optimization objective function after adding soft constraints It can be expressed as: ; in, is the weight coefficient used to balance the impact between the original loss and the soft constraint, is the expected drying time. The optimization objective function not only focuses on the prediction accuracy of the model, but also introduces the consideration of process requirements. Based on the constrained optimization objective function, the parameters of the ink residual estimation subnetwork of the heterogeneous multi-task learning graph neural network are updated. Through the back-propagation algorithm, the parameters of the network will be adjusted according to the loss gradient in each iteration, so as to gradually reduce the residual ink estimation error and meet the soft constraints at the same time. Finally, the optimized ink residual estimation subnetwork is obtained, which can more accurately estimate the amount of ink remaining during the spraying process under the optimized model parameters. At the same time, based on the same constrained optimization objective function and the optimal drying parameters, the parameters of the drying parameter optimization subnetwork of the heterogeneous multi-task learning graph neural network are updated. By adjusting the network parameters, the drying parameters are finally optimized while meeting the constraints. The optimized drying parameter optimization subnetwork will be able to provide reasonable time recommendations for different drying temperatures, thereby ensuring the efficiency and consistency of the drying process. The optimized ink residual estimation subnetwork and the optimized drying parameter optimization subnetwork are integrated to obtain the optimized heterogeneous multi-task learning graph neural network. Based on the optimized heterogeneous multi-task learning graph neural network, decision rules and threshold judgment logic are constructed to obtain an intelligent diagnosis model for the printing process. The intelligent diagnosis model can monitor the key parameters in the anti-counterfeiting code printing process in real time, and determine whether adjustments are needed based on the ink residue and drying parameters output by the model. For example, if the ink residue exceeds the set threshold, the system can issue an alarm to prompt the operator to check the spraying device; if the drying time is too long or the temperature is insufficient, it is recommended to adjust the drying parameters to ensure that the ink is completely dry.

[0040] Reference Figure 2 This embodiment provides a remote monitoring and fault prediction system for intelligent digital printing equipment, including: The collection module 1 is used to collect the anti-counterfeiting code printing operation data of the spraying device and the drying device in the intelligent digital printing equipment to obtain a comprehensive operation data set; Construction module 2 is used to construct a dynamic weighted adjacency matrix based on the comprehensive operation data set, taking the spraying device and the drying equipment as target nodes, and constructing the equipment component relationship graph structure; Extraction module 3, used to extract spray uniformity and drying efficiency features of the comprehensive operation data set based on the equipment component relationship graph structure, and create a fused spatiotemporal feature vector; Prediction module 4 is used to input the fused spatiotemporal feature vector into a heterogeneous multi-task learning graph neural network for multi-task prediction to obtain the residual ink amount of the anti-counterfeiting code and the optimal drying parameters; Optimization module 5 is used to convert the relationship between ink drying time and drying temperature into constraint conditions, and optimize the heterogeneous multi-task learning graph neural network based on the residual ink amount of the anti-counterfeiting code and the optimal drying parameters to obtain an intelligent diagnosis model for the printing process; The generation module 6 is used to analyze the real-time data set collected during the anti-counterfeiting code printing process through the intelligent diagnosis model of the printing process, and generate early warning information on the spraying quality and drying efficiency.

[0041] In this embodiment, for the specific implementation of each unit in the above system embodiment, please refer to the description in the above method embodiment, which will not be repeated here.

[0042] Reference Figure 3 In an embodiment of the present invention, a computer device is also provided. The computer device may be a server, and its internal structure may be as follows: Figure 3 As shown. The computer device includes a processor, a memory, a display screen, an input system, a network interface and a database connected through a system bus. Among them, the processor designed by the computer is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the above method is implemented.

[0043] Those skilled in the art will understand that Figure 3 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied.

[0044] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the above method is implemented. It can be understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0045] In summary, a remote monitoring and fault prediction method for intelligent digital printing equipment is proposed in the embodiment of the present invention. By collecting and fusing multiple sensor data of the spraying device and the drying device, a comprehensive operation data set is constructed, which provides a comprehensive data basis for subsequent analysis and prediction, and helps to improve the accuracy and comprehensiveness of monitoring. By constructing a dynamic weighted adjacency matrix and a device component relationship graph structure, the method can effectively capture the dynamic interaction between the components of the device. By combining graph convolution and attention mechanism, the spatiotemporal feature extraction of spraying uniformity and drying efficiency is realized, and the long-term dependent features in the data are effectively extracted, which improves the richness of feature expression. A heterogeneous multi-task learning graph neural network containing two sub-networks of ink residue estimation and drying parameter optimization is designed to realize parallel prediction of multiple related tasks, and improve the generalization ability and prediction accuracy of the model. By introducing a bidirectional long short-term memory network and an interactive enhancement module of the self-attention layer, information interaction and feature enhancement between different tasks are realized, and the relationship between ink drying time and drying temperature is converted into a soft constraint condition. The model is optimized so that the prediction results are more in line with the actual process requirements, and the practicality of the model is improved. Decision rules and threshold judgment logic are constructed based on the optimized heterogeneous multi-task learning graph neural network, which realizes the intelligent diagnosis of the printing process and can generate early warning information on spraying quality and drying efficiency in time, which helps to improve the controllability and stability of the production process.

[0046] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided by the present invention and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double-speed data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM.

[0047] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, system, article or method including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, system, article or method. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the presence of other identical elements in the process, system, article or method including the element.

[0048] The above description is only a preferred embodiment of the present invention, and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A remote monitoring and fault prediction method for intelligent digital printing equipment, characterized in that: The following steps are involved: Collect anti-counterfeiting code printing operation data from the spraying device and drying equipment in the intelligent digital printing equipment to obtain a comprehensive operation data set; Building a dynamic weighted adjacency matrix based on the comprehensive operation data set, taking the spraying device and the drying device as target nodes, and building an equipment component relationship graph structure; Based on the equipment component relationship graph structure, extract the spraying uniformity and drying efficiency features of the comprehensive operation data set to create a fused spatiotemporal feature vector; The fused spatiotemporal feature vector is input into a heterogeneous multi-task learning graph neural network for multi-task prediction to obtain the residual ink amount of the anti-counterfeiting code and the optimal drying parameters; The relationship between ink drying time and drying temperature is converted into constraint conditions, and the heterogeneous multi-task learning graph neural network is optimized based on the residual ink amount of the anti-counterfeiting code and the optimal drying parameters to obtain an intelligent diagnosis model for the printing process; The printing process intelligent diagnosis model is used to analyze the real-time data set collected during the anti-counterfeiting code printing process, and to generate early warning information on the spraying quality and drying efficiency.

2. The remote monitoring and fault prediction method of intelligent digital printing equipment according to claim 1 is characterized in that: The anti-counterfeiting code printing operation data collection of the spraying device and the drying device in the intelligent digital printing equipment is obtained to obtain a comprehensive operation data set, including: Performing real-time data collection on an ink flow sensor of a spraying device in an intelligent digital printing device to obtain ink flow time series data, and calculating a spraying speed based on the ink flow time series data to obtain spraying speed data; Collecting data from the temperature sensor and humidity sensor of the drying equipment to obtain drying temperature and humidity data, and estimating the drying efficiency based on the drying temperature and humidity data to obtain a drying efficiency index; Collecting data from a pressure sensor of the spraying device to obtain spraying pressure data, and calculating spraying uniformity based on the spraying pressure data and the spraying speed data to obtain a spraying uniformity index; Collecting data from current sensors of the spraying device and the drying device to obtain device power consumption data; Multi-dimensional data fusion is performed based on the ink flow timing data, the spraying speed data, the drying temperature and humidity data, the drying efficiency index, the spraying pressure data, the spraying uniformity index and the equipment power consumption data to obtain a comprehensive operation data set.

3. The remote monitoring and fault prediction method of intelligent digital printing equipment according to claim 1, characterized in that: The method of constructing a dynamic weighted adjacency matrix based on the comprehensive operation data set, taking the spraying device and the drying device as target nodes, and constructing a device component relationship graph structure includes: Dividing the comprehensive operation data set into time windows to obtain a plurality of time segment data subsets, and calculating the correlation between each device component based on the time segment data subsets to obtain an inter-component correlation matrix; Performing threshold filtering on the inter-component correlation matrix to obtain a sparse correlation matrix, and initializing the graph structure of the device components based on the sparse correlation matrix to obtain an initial graph structure; Performing topological optimization on the initial graph structure, setting the spraying device and the drying device as target nodes, and obtaining a hierarchical graph structure; Based on the hierarchical graph structure, dynamically update and calculate the edge weights to obtain a dynamic weighted adjacency matrix, and normalize the dynamic weighted adjacency matrix to obtain a standardized adjacency matrix; Based on the standardized adjacency matrix and preset device topology rules, the hierarchical graph structure is adjusted to obtain a device component relationship graph structure.

4. The remote monitoring and fault prediction method of intelligent digital printing equipment according to claim 1, characterized in that: Based on the equipment component relationship graph structure, the spraying uniformity and drying efficiency feature extraction is performed on the comprehensive operation data set to create a fused spatiotemporal feature vector, including: Performing a graph convolution operation on the equipment component relationship graph structure to obtain an initial node feature matrix, and based on the initial node feature matrix, performing feature aggregation on the spray device node to obtain spray device aggregation features; Performing a nonlinear transformation on the aggregation features of the spraying device to obtain the initial features of the spraying uniformity, and based on the initial node feature matrix, performing feature aggregation on the nodes of the drying equipment to obtain the aggregation features of the drying equipment, and performing a nonlinear transformation on the aggregation features of the drying equipment to obtain the initial features of the drying efficiency; Based on the initial characteristics of the spraying uniformity, a first one-dimensional convolutional neural network is used to extract time series characteristics to obtain the time series characteristics of the spraying uniformity, and based on the initial characteristics of the drying efficiency, a second one-dimensional convolutional neural network is used to extract time series characteristics to obtain the time series characteristics of the drying efficiency; Performing attention mechanism calculation on the spraying uniformity time series feature and the drying efficiency time series feature to obtain a weighted feature vector, and performing time dependency modeling based on the weighted feature vector to obtain a long-term dependency feature; The long-term dependency feature and the weighted feature vector are concatenated to obtain a fused spatiotemporal feature vector.

5. The remote monitoring and fault prediction method of intelligent digital printing equipment according to claim 1, characterized in that: The fused spatiotemporal feature vector is input into a heterogeneous multi-task learning graph neural network for multi-task prediction to obtain the amount of residual ink and the optimal drying parameters of the anti-counterfeiting code, including: Performing feature decomposition on the fused spatiotemporal feature vector to obtain a spraying feature subset and a drying feature subset; The spray feature subset is input into the ink residual estimation subnetwork in the heterogeneous multi-task learning graph neural network, and the estimated value of the residual ink of the initial anti-counterfeiting code is obtained by forward propagation calculation; the ink residual estimation subnetwork includes three graph convolution layers and two fully connected layers; the graph convolution layer uses Chebyshev polynomials to approximate the spectral graph convolution, and the activation function is ReLU; the fully connected layer uses Dropout regularization, and the output layer uses Sigmoid activation function; The drying feature subset is input into the drying parameter optimization subnetwork in the heterogeneous multi-task learning graph neural network, and the initial optimal drying parameter estimation value is obtained by forward propagation calculation; the drying parameter optimization subnetwork includes two graph attention layers and three fully connected layers; the graph attention layer uses a multi-head attention mechanism, and the number of heads is set to 8; the fully connected layer uses BatchNormalization, and the output layer uses a Tanh activation function; The estimated value of the residual ink volume of the initial anti-counterfeiting code and the estimated value of the initial optimal drying parameter are input into the interactive enhancement module in the heterogeneous multi-task learning graph neural network, and an enhanced feature representation is obtained through bidirectional information flow and attention mechanism; the interactive enhancement module includes a bidirectional long short-term memory network and a self-attention layer; the bidirectional long short-term memory network includes 128 hidden units, and the self-attention layer uses a scaled dot product attention mechanism; The enhanced feature representation is input into the multi-task output layer, and the residual ink amount and the optimal drying parameters of the anti-counterfeiting code are obtained through parallel calculation; the multi-task output layer includes two parallel fully connected networks, each fully connected network includes two layers, and uses LeakyReLU activation function and L2 regularization.

6. The remote monitoring and fault prediction method of intelligent digital printing equipment according to claim 5, characterized in that: The step of inputting the spray feature subset into the ink residual amount estimation subnetwork in the heterogeneous multi-task learning graph neural network, and obtaining the estimated value of the residual ink amount of the initial anti-counterfeiting code through forward propagation calculation, comprises: Normalizing and standardizing the spray feature subset to obtain standardized spray features; Based on the standardized spraying features, the first layer of graph convolution operation is performed using Chebyshev polynomial expansion to obtain first-order graph convolution features; the Chebyshev polynomial expansion order is 3, and the convolution kernel size is 32; Performing ReLU activation function processing on the first-order graph convolution feature to obtain a feature map after the first layer is activated; Based on the feature map after activation of the first layer, a second-layer graph convolution operation is performed to obtain the second-order graph convolution feature; the second-layer graph convolution uses the same Chebyshev polynomial expansion method as the first layer, and the convolution kernel size is 64; Performing ReLU activation function processing on the second-order graph convolution feature to obtain a feature map after the second layer activation; Based on the feature map after activation of the second layer, a third-layer graph convolution operation is performed to obtain a third-order graph convolution feature; the third-layer graph convolution uses the same Chebyshev polynomial expansion method as the first two layers, and the convolution kernel size is 128; Performing ReLU activation function processing on the third-order graph convolution feature to obtain a feature map after activation of the third layer; Input the activated feature map of the third layer into the first fully connected layer, perform linear transformation and Dropout operation, and obtain a reduced-dimensional feature vector; the output dimension of the first fully connected layer is 64, and the Dropout rate is 0.5; Performing ReLU activation function processing on the dimension-reduced feature vector to obtain an activation output of the first fully connected layer; The activation output of the first fully connected layer is input into the second fully connected layer, and linear transformation and Sigmoid activation function processing are performed to obtain an estimated value of the residual ink amount of the initial anti-counterfeiting code.

7. The remote monitoring and fault prediction method of intelligent digital printing equipment according to claim 5, characterized in that: The step of inputting the drying feature subset into the drying parameter optimization subnetwork in the heterogeneous multi-task learning graph neural network and obtaining an initial optimal drying parameter estimation value through forward propagation calculation includes: Performing data preprocessing on the drying feature subset, including normalization and standardization operations, to obtain standardized drying features; Based on the standardized drying features, a multi-head attention mechanism is used to perform first-layer graph attention calculation to obtain first-layer attention weighted features; the number of heads of the multi-head attention mechanism is 8, and the output dimension of each head is 32; Performing layer normalization processing on the first layer of attention weighted features to obtain first layer normalized features; Based on the normalized features of the first layer, a multi-head attention mechanism is used to perform the second-layer graph attention calculation to obtain the second-layer attention weighted features; the parameter setting of the second-layer multi-head attention mechanism is the same as that of the first layer; Performing layer normalization processing on the second-layer attention weighted features to obtain second-layer normalized features; The normalized features of the second layer are input into the first fully connected layer, and linear transformation and BatchNormalization operations are performed to obtain the first layer fully connected output; the output dimension of the first fully connected layer is 128; Perform ReLU activation function processing on the fully connected output of the first layer to obtain the features after activation of the first layer; The activated features of the first layer are input into the second fully connected layer, and linear transformation and BatchNormalization operations are performed to obtain the second fully connected output; the output dimension of the second fully connected layer is 64; Perform ReLU activation function processing on the fully connected output of the second layer to obtain the features after activation of the second layer; The activated features of the second layer are input into the third fully connected layer, and linear transformation and Tanh activation function processing are performed to obtain the initial optimal drying parameter estimation value.

8. The remote monitoring and fault prediction method of intelligent digital printing equipment according to claim 5, characterized in that: The step of inputting the estimated value of the residual ink volume of the initial anti-counterfeiting code and the estimated value of the initial optimal drying parameter into the interactive enhancement module in the heterogeneous multi-task learning graph neural network, and obtaining enhanced feature representation through bidirectional information flow and attention mechanism, includes: Performing feature concatenation on the initial estimated value of the amount of residual ink in the anti-counterfeiting code and the initial estimated value of the optimal drying parameter to obtain an initial joint feature vector; Inputting the initial joint feature vector into the forward propagation layer of the bidirectional long short-term memory network to extract the time sequence information and obtain a forward hidden state sequence; Inputting the initial joint feature vector into the back propagation layer of the bidirectional long short-term memory network to extract time sequence information and obtain a reverse hidden state sequence; Performing a splicing operation on the forward hidden state sequence and the reverse hidden state sequence to obtain a bidirectional fusion feature sequence; Based on the bidirectional fusion feature sequence, a self-attention weight matrix is ​​calculated to obtain an attention score; the self-attention layer uses a scaled dot product attention mechanism, and the scaling factor is the square root of 8; Performing Softmax normalization on the attention score to obtain a normalized attention weight; Based on the normalized attention weight and the bidirectional fusion feature sequence, a weighted sum operation is performed to obtain a context vector; Performing a residual connection between the context vector and the bidirectional fusion feature sequence to obtain a residual feature; The residual features are subjected to layer normalization processing to obtain a normalized feature representation, and the normalized feature representation is input into a feedforward neural network for nonlinear transformation to obtain an enhanced feature representation.

9. The remote monitoring and fault prediction method of intelligent digital printing equipment according to claim 1, characterized in that: The relationship between the ink drying time and the drying temperature is converted into a constraint condition, and the heterogeneous multi-task learning graph neural network is optimized based on the residual ink amount of the anti-counterfeiting code and the optimal drying parameters to obtain an intelligent diagnosis model for the printing process, including: Model the relationship between ink drying time and drying temperature to obtain the ink drying time function; Discretizing the ink drying time function to obtain corresponding relationship points between the ink drying time and the drying temperature; Based on the corresponding relationship points, a constraint condition matrix is ​​constructed to obtain soft constraints for drying parameter optimization; Modifying the loss function of the heterogeneous multi-task learning graph neural network, adding the soft constraint term, and obtaining a constrained optimization objective function; Based on the constrained optimization objective function and the amount of residual ink in the anti-counterfeiting code, updating the parameters of the residual ink estimation subnetwork of the heterogeneous multi-task learning graph neural network to obtain an optimized residual ink estimation subnetwork; Based on the constrained optimization objective function and the optimal drying parameters, updating the parameters of the drying parameter optimization subnetwork of the heterogeneous multi-task learning graph neural network to obtain an optimized drying parameter optimization subnetwork; Integrating the optimized ink residue estimation subnetwork and the optimized drying parameter optimization subnetwork to obtain an optimized heterogeneous multi-task learning graph neural network; Based on the optimized heterogeneous multi-task learning graph neural network, decision rules and threshold judgment logic are constructed to obtain an intelligent diagnosis model for the printing process.

10. A remote monitoring and fault prediction system for intelligent digital printing equipment, characterized in that: A system for executing a remote monitoring and fault prediction method for an intelligent digital printing device according to any one of claims 1 to 9, the system comprising: The collection module is used to collect the anti-counterfeiting code printing operation data of the spraying device and the drying equipment in the intelligent digital printing equipment to obtain a comprehensive operation data set; A construction module, used to construct a dynamic weighted adjacency matrix based on the comprehensive operation data set, taking the spraying device and the drying device as target nodes, and constructing a device component relationship graph structure; An extraction module, for extracting spraying uniformity and drying efficiency features of the comprehensive operation data set based on the equipment component relationship graph structure, and creating a fused spatiotemporal feature vector; A prediction module, used for inputting the fused spatiotemporal feature vector into a heterogeneous multi-task learning graph neural network for multi-task prediction to obtain the residual ink amount of the anti-counterfeiting code and the optimal drying parameters; An optimization module, used for converting the relationship between ink drying time and drying temperature into constraint conditions, and optimizing the heterogeneous multi-task learning graph neural network based on the residual ink amount of the anti-counterfeiting code and the optimal drying parameters to obtain an intelligent diagnosis model for the printing process; The generation module is used to analyze the real-time data set collected during the anti-counterfeiting code printing process through the printing process intelligent diagnosis model, and generate early warning information on spraying quality and drying efficiency.

Images