Green color printing product defect intelligent diagnosis method based on deep learning

Abstract

The invention relates to the technical field of color printing product quality detection, and discloses a green color printing product defect intelligent diagnosis method based on deep learning. Visible light, hyperspectral and infrared thermal imaging data of a green color printing product are collected, multi-scale feature extraction is carried out, defect feature vectors such as surface texture, chromatic aberration distribution and heat conduction abnormity are obtained, and a cross-modal attention mechanism is utilized to fuse the defect feature vectors into a unified defect representation vector. And determining a defect type and a repair scheme through a dynamic classification algorithm in combination with the defect knowledge base. The algorithm comprises the steps of generating an initial candidate defect set, constructing a topological graph, optimizing defect type judgment and the like. A multi-dimensional anomaly detection model is adopted to process sudden defects, and a causal inference engine is utilized to adaptively adjust a repair scheme. And constructing a defect rule knowledge graph to verify the compliance of the repair scheme. The method is accurate in detection, intelligent in diagnosis and efficient in repair, and effectively improves the quality detection and repair level of the green color printing product.

Description

Technical Field

[0001] The present invention relates to the technical field of color print quality detection technology, and specifically to a green color print defect intelligent diagnosis method based on deep learning. Background Art

[0002] In the modern printing industry, green color printing has gained widespread adoption as an environmentally friendly and market-compliant printing method. With increasing consumer demand for color printed product quality and increasingly stringent environmental regulations, accurately and efficiently detecting and diagnosing defects in green color printed products has become a major challenge for printing companies.

[0003] Traditional methods for detecting defects in color printed products rely primarily on manual visual inspection. This approach has numerous drawbacks. For one thing, manual inspection is inefficient and cannot keep up with the pace of production in large-scale environments. For example, some large printing companies process tens of thousands of color printed products daily. Relying on manual inspections of each print is not only labor-intensive but also prone to missed detections. Furthermore, the accuracy of manual inspections is significantly affected by subjective factors such as the inspector's experience and fatigue. Different inspectors may have different criteria for determining defects, and even the same inspector may experience reduced detection accuracy due to fatigue after working for long periods of time.

[0004] While some early automated inspection technologies, such as simple image-based inspection methods based on optical principles, have improved inspection efficiency to a certain extent, they also have limitations. These methods are generally only able to detect obvious surface defects and struggle to effectively identify issues such as internal structural defects, subtle color variations, and thermal conductivity anomalies. With the continuous advancement of color printing technology, the types of defects found in printed products have become increasingly complex and diverse, making traditional optical inspection methods unable to meet actual production needs.

[0005] Furthermore, previous technologies lacked effective integration and intelligent processing in defect diagnosis and repair. Once a defect was detected, it was often impossible to quickly and accurately determine its type, cause, and corresponding repair solution. This often required manual review of extensive data and empirical judgment, which was not only time-consuming and labor-intensive, but also difficult to ensure the optimal repair solution. When faced with multiple defects occurring simultaneously, traditional methods proved insufficient, unable to comprehensively consider various factors and develop a comprehensive and rational repair strategy.

[0006] The widespread application of deep learning technology in areas such as image recognition and data analysis has brought new opportunities for the detection and diagnosis of defects in green color printed products. However, the application of deep learning technology to the detection of defects in green color printed products is still in its developmental stage, and a mature and complete technical system has not yet been formed. How to fully utilize the advantages of deep learning technology to achieve intelligent diagnosis and precise repair of multiple types of defects in green color printed products has become a problem that needs to be solved urgently. It is in this context that the present invention is proposed to address the shortcomings of traditional detection methods and improve the quality detection level and production efficiency of green color printed products. Summary of the Invention

[0007] The purpose of the present invention is to provide an intelligent diagnosis method for green color printed product defects based on deep learning to solve the problems raised in the above background technology.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a deep learning-based intelligent diagnosis method for defects in green color printed products, the method comprising:

[0009] Collecting multi-source image data of green color prints, wherein the multi-source image data includes visible light images, hyperspectral images, and infrared thermal imaging data;

[0010] Performing multi-scale feature extraction on the multi-source image data to obtain a defect feature vector corresponding to each image modality, the defect feature vector including: surface texture features, color difference distribution features, and thermal conduction anomaly features;

[0011] The defect feature vectors are fused into a unified defect representation vector based on a cross-modal attention mechanism;

[0012] Building a defect knowledge base based on a historical defect sample library, wherein the knowledge base stores third data, including: morphological feature templates of various defects, defect cause correlations, and corresponding repair process records;

[0013] Based on the unified defect characterization vector and the third data in the defect knowledge base, a dynamic classification algorithm is used to determine the defect type of the current color printed product and a corresponding repair suggestion solution.

[0014] Preferably, the method of determining the defect type of the current color printed product and the corresponding repair suggestion by a dynamic classification algorithm includes:

[0015] generating an initial candidate defect set based on the similarity between the unified defect representation vector and the defect templates in the knowledge base;

[0016] Constructing a defect cause topology graph based on the defect cause association relationship, wherein nodes in the topology graph represent defect types and edges represent logical or temporal associations between causes;

[0017] Based on the defect cause topology map and real-time detection data, the initial candidate defect set is optimized through a graph reasoning algorithm, and the final defect type and repair solution priority ranking are output.

[0018] Preferably, the performing multi-scale feature extraction on the multi-source image data includes:

[0019] A multi-branch convolutional network is used to extract local and global texture features from the visible light image, a spatiotemporal attention network is used to extract abnormal band features in the spectral sequence from the hyperspectral image, and an adaptive threshold segmentation algorithm is used to extract abnormal thermal distribution areas from the infrared thermal imaging data.

[0020] Preferably, the fusing of the defect feature vectors into a unified defect representation vector based on a cross-modal attention mechanism includes:

[0021] The importance distribution of each image feature is calculated by the inter-modality correlation weight;

[0022] The weighted features are input into a bidirectional gated recurrent network to capture the temporal dependencies of cross-modal features.

[0023] Preferably, the dynamic classification algorithm also includes: constructing a defect classification optimization model based on a meta-learning framework, the optimization model dynamically adjusts the feature fusion strategy through a multi-task loss function, and the loss function includes category discrimination loss, feature consistency loss and causal correlation loss.

[0024] Preferably, the graph reasoning algorithm is an improved simulated annealing algorithm, comprising:

[0025] Encoding the defect cause topology graph into a state space, wherein the state nodes represent defect cause combinations;

[0026] Generate candidate state transition paths through neighborhood search strategy and introduce random perturbations to escape from local optimal solutions;

[0027] The optimal state path is evaluated according to an energy function, which is calculated based on the defect matching degree and the repair cost weight.

[0028] Preferably, the method further comprises:

[0029] Using a multi-dimensional anomaly detection model on the real-time detection data to extract the characteristics of sudden defects in the production process;

[0030] When a sudden defect is detected, adaptive adjustment of the repair plan is triggered, and the repair priority is updated according to the adjusted plan.

[0031] Preferably, the adaptive adjustment adopts a causal reasoning engine, and the specific method includes:

[0032] Define the defect cause chain, process parameter influencing factors and repair resource constraints;

[0033] The potential root causes are reversely deduced through the causal graph, and the minimum intervention repair strategy is generated based on the gradient optimization algorithm.

[0034] Preferably, the method further includes: constructing a defect rule knowledge graph, in which nodes represent defect entities and process parameters, and edges represent quality inspection standards or material performance constraints, and in the process of generating a repair plan, verifying the compliance of the defect type with the knowledge graph through a graph embedding matching algorithm.

[0035] Preferably, the graph embedding matching algorithm is a contrastive learning algorithm based on a hierarchical graph neural network, including:

[0036] Extract subgraph structural features from the current defect type;

[0037] Neighbor node information is hierarchically aggregated in the knowledge graph, and matching subgraph patterns are screened through similarity comparison.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] The deep learning-based intelligent diagnosis method for green color print defects proposed in this paper significantly improves the ability to detect and repair green color print defects in multiple ways, bringing numerous positive benefits to the printing industry. Regarding defect detection accuracy, this method acquires multi-source information by collecting visible light images, hyperspectral images, and infrared thermal imaging data from green color prints. Image data from different modalities can reflect the characteristics of the color print from different perspectives: visible light images reveal surface texture, hyperspectral images capture spectral anomalies, and infrared thermal imaging data can reveal thermal conductivity problems. Using multi-scale feature extraction technology, corresponding defect feature vectors are extracted for each image modality, including surface texture features, color difference distribution features, and thermal conductivity anomaly features, significantly enriching the defect information. Compared with traditional methods that rely solely on single images or simple feature detection, this method can more comprehensively and accurately identify various potential defects. For example, when detecting subtle color difference defects, the abnormal band feature extraction technology of hyperspectral images can accurately locate the color difference area, while traditional methods may not be able to detect such subtle changes.

[0040] In terms of the intelligence level of defect diagnosis, the cross-modal attention mechanism fuses different defect feature vectors into a unified defect representation vector, enabling the model to better understand and analyze comprehensive features. Through a dynamic classification algorithm, combined with the morphological feature templates, defect cause correlations, and repair process records in the defect knowledge base, the defect type is accurately determined and repair suggestions are given. Moreover, the defect classification optimization model constructed based on the meta-learning framework uses a multi-task loss function to dynamically adjust the feature fusion strategy, further improving the accuracy and adaptability of diagnosis. When faced with complex defect situations, this intelligent diagnosis method can quickly and accurately determine the defect type, while traditional manual diagnosis or simple algorithm diagnosis methods require a lot of time for analysis and judgment, and accuracy is difficult to guarantee.

[0041] In the defect repair link, the present invention also shows strong advantages. According to the defect cause topology map and graph reasoning algorithm, such as the improved simulated annealing algorithm, it is not only possible to optimize the judgment of defect type, but also to prioritize the repair schemes to ensure that the best scheme is adopted when multiple repair schemes are available. When a sudden defect is detected, the multi-dimensional anomaly detection model and causal reasoning engine can respond quickly, adaptively adjust the repair scheme and update the repair priority, reducing production delays and product losses caused by defects. For example, if a defect with a large color deviation suddenly appears during the printing process, the system can quickly find the possible cause through causal reasoning, such as ink supply problems or abnormal temperature of the printing equipment, and generate a minimum intervention repair strategy based on this, to resume production in the shortest time and reduce losses.

[0042] The constructed defect rule knowledge graph and the contrastive learning algorithm based on the hierarchical graph neural network verify the compliance of defect types with the knowledge graph during the repair solution generation process, ensuring that the repair solution meets quality inspection standards and material performance constraints, further improving the reliability and effectiveness of the repair solution. This method can also effectively integrate various data and knowledge in the production process, providing comprehensive quality control support for printing companies, helping them optimize production processes, improve production efficiency, reduce production costs, enhance their competitiveness in the market, and promote the intelligent development of the green color printing industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a working principle diagram of the green color printing product defect intelligent diagnosis method based on deep learning according to the present invention;

[0044] Figure 2 Flowchart for feature extraction of multi-source image data;

[0045] Figure 3 Flowchart for improving the simulated annealing graph inference algorithm;

[0046] Figure 4Flowchart for real-time detection and repair solution adjustments. DETAILED DESCRIPTION

[0047] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0048] See also Figure 1-Figure 4 The present invention relates to an intelligent diagnosis method for defects in green color printed products based on deep learning, and its specific implementation method will be elaborated in detail below.

[0049] Collect multi-source image data from green color prints, including visible light images, hyperspectral images, and infrared thermal imaging data. In practice, specialized image acquisition equipment is used, such as a high-resolution visible light camera to capture surface texture and color information, a hyperspectral imager to capture reflectance information in different spectral bands, and an infrared thermal imager to capture thermal distribution information. By properly setting acquisition equipment parameters, such as shooting angle, exposure time, and spectral range, the captured image data accurately reflects the characteristics of the color prints.

[0050] Multi-scale feature extraction is performed on multi-source image data to obtain defect feature vectors corresponding to each image modality. These defect feature vectors include surface texture features, color difference distribution features, and thermal conductivity anomaly features. Different feature extraction methods are used for different types of image data. For example, for visible light images, the multi-branch structure of a convolutional neural network (CNN) can be used to extract local and global texture features. For hyperspectral images, a spatiotemporal attention network is used to mine abnormal band features in the spectral sequence. For infrared thermal imaging data, an adaptive threshold segmentation algorithm is used to identify areas of abnormal thermal distribution and obtain the corresponding feature vectors.

[0051] The defect feature vectors are fused into a unified defect representation vector based on a cross-modal attention mechanism. This process first calculates the importance distribution of each image feature, then determines the weights by analyzing the correlations between features from different modalities. These weighted features are then fed into a bidirectional gated recurrent network (Bi-GRU) to capture the temporal dependencies of cross-modal features. This results in a unified defect representation vector that comprehensively reflects the defect information of color printed products.

[0052] A defect knowledge base is constructed based on a historical defect sample library. This library contains a large amount of sample data for various defect types. Based on this sample data, morphological feature templates for each defect type are extracted, correlations between defect causes are analyzed, and corresponding repair processes are recorded. This information is organized and stored in the defect knowledge base, serving as an important basis for subsequent defect diagnosis and repair plan development.

[0053] Based on the unified defect representation vector and the third data in the defect knowledge base, a dynamic classification algorithm is used to determine the defect type and corresponding repair solution for the current color printed product. The unified defect representation vector is matched with the defect templates in the defect knowledge base for similarity. In combination with the rules and logic of the dynamic classification algorithm, the knowledge base is used to select defect types that match the characteristics of the current color printed product. Based on the defect cause and repair process records, a corresponding repair solution is provided.

[0054] The technical solution of the present invention is further described in detail below with reference to specific embodiments.

[0055] Example 1:

[0056] In this embodiment, the specific process of determining the defect type of the current color printed product and the corresponding repair suggestion plan through the dynamic classification algorithm is further described in detail.

[0057] Generate the initial candidate defect set based on the similarity between the unified defect representation vector and the defect templates in the knowledge base. Calculate the similarity between the unified defect representation vector and the morphological feature templates of various defects pre-stored in the defect knowledge base. Assume that the unified defect representation vector is V and the defect template set is {T1, T2, ..., T n}, the similarity calculation function is Sim(V,T i ), where i = 1, 2, ..., n. A series of similarity values Sim(V, T1), Sim(V, T2), ..., Sim(V, T n ), set a similarity threshold θ, and include the defect types corresponding to the defect templates with similarity greater than θ into the initial candidate defect set C.

[0058] A defect cause topology graph is constructed based on the defect cause associations. In this graph, nodes represent defect types, and edges represent the logical or temporal relationships between causes. For example, defect A may be caused by causes a and b, and defect B may be caused by causes b and c. In this graph, nodes A and B may be connected through cause b, and the direction of the edge indicates the order of the causal relationship.

[0059] Based on the defect cause topology map and real-time detection data, a graph reasoning algorithm is used to optimize the initial candidate defect set, outputting the final defect type and prioritized repair solutions. The graph reasoning algorithm can employ a modified simulated annealing algorithm. The defect cause topology map is encoded as a state space, with state nodes representing defect cause combinations. A neighborhood search strategy is used to generate candidate state transition paths, and random perturbations are introduced to escape local optima. The optimal state path is evaluated using an energy function. The energy function E is calculated based on the defect matching degree and the weighted repair cost. The formula is E = w1 × M + w2 × C, where w1 and w2 are the weights of the defect matching degree and repair cost, respectively, and w1 + w2 = 1. M represents the degree of match between the current defect cause combination and the actual detected defect signature, ranging from 0 to 1, with larger values indicating higher match. C represents the cost required to repair the current defect cause combination, which can be quantified based on factors such as the complexity of the repair process, the required materials, and the time required. Through an iterative search for the optimal state path, the final defect type is determined, and repair solutions are prioritized based on repair cost and other factors.

[0060] Example 2:

[0061] This embodiment focuses on the specific implementation process of multi-scale feature extraction from multi-source image data.

[0062] For visible light images, a multi-branch convolutional network is used to extract local and global texture features. A convolutional neural network with multiple convolution branches is constructed, where one branch uses a smaller convolution kernel, such as a 3×3 convolution kernel, to extract local texture details of the image, and the other branch uses a larger convolution kernel, such as a 7×7 convolution kernel, to obtain global texture features of the image. Assume that the input visible light image is I vis After the convolution operation of the first branch convolution layer Conv1, the local texture feature map F is obtained local , the calculation formula is F local =Conv1(I vis ); After the convolution operation of the second branch convolution layer Conv2, the global texture feature map F is obtained global , the calculation formula is F global =Conv2(I vis ). Then the local texture feature map and the global texture feature map are fused to obtain the texture feature vector V of the visible light image vis , you can use splicing or weighted fusion, such as V vis =Concat(F local ,F global ).

[0063] For hyperspectral images, a spatiotemporal attention network is used to extract abnormal band features in the spectral sequence. The spatiotemporal attention network can be divided into a spatial attention module and a temporal attention module. First, the spatial dimension of the hyperspectral image is processed by the spatial attention module to calculate the spatial attention weight. Assume that the image of the hyperspectral image at a certain moment is I hs (t), the calculation formula of the spatial attention weight matrix S is S = Softmax(f s (I hs (t))), where f s It is a spatial feature extraction function, such as a network composed of convolutional layers and fully connected layers. Then, the spatial attention weight is multiplied with the hyperspectral image to obtain the spatial attention enhanced image I hs-s (t) = S × I hs (t). Then, the spectral sequence is processed by the time attention module to calculate the time attention weight. Assume that the hyperspectral image sequence after T time steps is {I hs (1),I hs (2),…,I hs (T)}, the calculation formula of the time attention weight vector T is where f t Is a temporal feature extraction function. Finally, the temporal attention weight is multiplied by the image sequence enhanced by spatial attention to obtain the abnormal band feature vector V hs ,Right now

[0064] For infrared thermal imaging data, the adaptive threshold segmentation algorithm is used to extract the abnormal heat distribution area. The adaptive threshold segmentation algorithm calculates the threshold of different areas according to the local features of the image. Assume that the infrared thermal imaging image is I ir , divide the image into multiple sub-regions R1, R2, ..., R m For each subregion R i , calculate its grayscale mean μ i and grayscale standard deviation σ i , threshold T i The calculation formula is T i =μ i +k×σ i , where k is an empirical coefficient that is adjusted according to the actual situation. The pixel value of each sub-region is compared with the corresponding threshold value, and the pixels greater than the threshold value are marked as abnormal thermal distribution areas, thereby obtaining the characteristic vector V of the abnormal thermal distribution area. ir .

[0065] Example 3

[0066] In the actual process of diagnosing defects in green color printed products, fusing defect feature vectors based on a cross-modal attention mechanism plays a key role in accurately identifying defects. In this example, the specific operation process of this mechanism is demonstrated using a batch of beautifully printed packaging boxes as an example.

[0067] After acquiring visible light image, hyperspectral image and infrared thermal imaging data, the corresponding defect feature vectors will be obtained, which are denoted as V vis 、V hs 、V ir When calculating the correlation weight between modalities, first find the correlation matrix C. vis and V hs For example, the correlation is calculated as This formula represents the cosine similarity between two vectors, which is used to measure their similarity. Similarly, the other elements in the C matrix are calculated to obtain the complete correlation matrix.

[0068] Calculate the importance weight of each eigenvector according to the correlation matrix, such as w vis Calculation of the numerator ∑ j∈{hs,ir} C vis,j is the sum of the correlations between the visible light eigenvector and the other two modal eigenvectors, and the denominator ∑ i∈{vis,hs,ir} ∑ j∈{hs,ir} C i,j is the sum of all inter-modal correlations. Through such calculations, the importance of each modal eigenvector can be determined and then weighted.

[0069] The weighted eigenvector w vis V vis 、w hs V hs 、w ir V ir are concatenated into a new vector V concat , and then input into the bidirectional gated recurrent network (Bi-GRU). In Bi-GRU, the update gate z t The formula z t =σ(W z x t +U z h t-1 ), where W z and U z is the weight matrix used to control the influence of the input and the previous hidden layer state on the update gate, σ is the Sigmoid function, which maps the result to the [0,1] interval and determines the proportion of the hidden layer state to be retained at the previous moment. t Formula r t =σ(W r x t +U r ht-1 ) Similarly, control the degree of forgetting the information of the previous moment. Candidate hidden layer state Formula The tanh function maps the result to the interval [-1,1] by resetting the gate and input information. The final hidden layer state h t =(1-z t )⊙h t-1 +z t ⊙ It is generated by combining the previous hidden layer state and the candidate hidden layer state.

[0070] After Bi-GRU processing, the output unified defect representation vector V unified This system integrates defect features and their temporal dependencies across different modal images. In practical applications, the fused vectors generated by this cross-modal attention mechanism can more accurately reflect the comprehensive characteristics of packaging defects, such as color deviation, blurred patterns, and localized temperature anomalies, providing strong support for subsequent defect diagnosis and repair plan development.

[0071] Example 4:

[0072] This embodiment focuses on building a defect classification optimization model based on a meta-learning framework in a dynamic classification algorithm, and is explained based on the production data of a large color printing factory.

[0073] Category discrimination loss L class It is used to measure the degree of distinguishability between different defect categories. Assume that the common defect categories of the color printing factory are color deviation, registration error, uneven ink layer thickness, etc., with a total of n categories. The predicted defect category probability distribution is P(y), and the actual defect category label is y. Taking an actual inspection as an example, if the predicted probability of color deviation of a color printed product is P(y)1=0.8, the color deviation category in the actual label is y1=1, and the probabilities of other defect categories are all 0, then according to the cross entropy loss function The calculated loss is -log(0.8). This loss reflects the difference in classification between the model's predictions and the actual results. By minimizing this loss, the model can more accurately distinguish different defect categories.

[0074] Feature consistency loss L consistency Ensure the consistency of different modal features after fusion. Assume that the fusion feature vectors of different modalities extracted from multi-source image data are F, and perform random rotation and other transformations on F to obtain F ′ , the feature consistency loss calculation formula is In actual calculation, F and F ′Is a multidimensional vector, calculate the sum of the squares of the differences between their corresponding elements, that is, the square of the L2 norm. If F=[f1,f2,…,f m ], F ′ =[f1 ′ ,f2 ′ ,…,f ′ m ],but By minimizing this loss, we ensure that the features of different modalities remain stable and consistent after fusion, avoiding contradictory or unstable feature representations.

[0075] Cause-related loss L cause Consider the correlation between defect causes. Assume that the defect causes summarized in the color printing factory include ink quality, printing equipment parameters, ambient temperature and humidity, etc., a total of m, the correlation matrix between the causes is A, and the predicted defect cause vector is C. If the correlation strength A between ink quality and printing equipment parameters is ij is 0.6, and the predicted probabilities of the two causes are C i =0.7, C j =0.4, then when calculating the causal correlation loss, this term is 0.6×(0.7-0.4) 2 The sum of these association losses for all cause pairs yields the total cause association loss. By minimizing this loss, the model learns the intrinsic connections between defect causes and improves diagnostic accuracy.

[0076] Total loss function L = α × L class +β×L consistency +γ×L cause , where α, β, and γ are weight coefficients, and α + β + γ = 1. In practical applications, through multiple experiments and adjustments, appropriate weights such as α = 0.5, β = 0.3, and γ = 0.2 were determined. By minimizing the total loss function and dynamically adjusting the feature fusion strategy, the model can more accurately classify and diagnose defects in different production environments and defect types.

[0077] Example 5:

[0078] This embodiment provides a detailed description of real-time detection data processing and repair solution adjustment during the production process, taking a factory specializing in the production of high-end color posters as an example.

[0079] When performing anomaly detection on real-time detection data, an autoencoder model based on deep learning is used. Assume that the multi-source image data of a poster collected in real time is represented as X after preprocessing and is input into the autoencoder AE. The autoencoder consists of an encoder E and a decoder D. The encoder E encodes the input data X into a low-dimensional feature vector z = E(X). This process extracts the key features of the data and reduces the data dimension through a series of operations such as convolution and pooling. The decoder D then decodes the low-dimensional feature vector into reconstructed data Try to restore the original data as much as possible.

[0080] Calculate the reconstruction error This involves calculating the sum of the squares of the differences between the original and reconstructed data elements. A reconstruction error threshold ∈ is set. If the reconstruction error R of the poster image data detected at a certain moment is greater than ∈, a sudden defect is detected. For example, if an abnormality such as a sudden color change occurs on the poster, the autoencoder will have difficulty accurately reconstructing the image, and the reconstruction error will increase and exceed the threshold.

[0081] When a sudden defect is detected, adaptive adjustments to the repair plan are triggered. This adjustment is made using a causal inference engine, first defining the defect cause chain, process parameter influencing factors, and repair resource constraints. In this poster production scenario, the defect cause chain might include causal relationships such as unstable ink supply leading to color deviation and clogged printheads leading to blurred patterns; process parameter influencing factors such as the degree to which printing speed affects pattern clarity; and repair resource constraints such as the available ink types and quantities and the number of spare printheads.

[0082] Then, the potential root cause is reversely deduced through the causal graph. Assume that the causal graph is G = (V, E), where V is the node set (defect cause) and E is the edge set (causal relationship). Starting from the detected sudden defect node, such as the color deviation node, search along the reverse direction of the edge to find the possible potential root cause, such as unstable ink supply. Generate a minimum intervention repair strategy based on the gradient optimization algorithm. Assume that the repair strategy is S and the objective function is J(S), which comprehensively considers factors such as repair cost and repair effect. For example, the repair cost can include the cost of replacing ink and the time required to adjust equipment. The repair effect can be measured by the quality indicators of the poster after repair. The repair strategy S is continuously updated through the gradient descent algorithm to minimize the objective function J(S).

[0083] Finally, the repair priority is updated based on the adjusted plan. The repair priority is recalculated and ranked based on factors such as repair cost, repair time, and impact on production. For example, if replacing the ink can quickly resolve a color deviation problem at a low cost and with minimal impact on production, then this repair option will be given a higher priority, ensuring that unexpected defects can be handled promptly and effectively during the production process, thereby ensuring the quality of color printed products.

[0084] Example 6:

[0085] This embodiment focuses on building a defect rule knowledge graph and using a graph embedding matching algorithm to verify the compliance of defect types with the knowledge graph. It is illustrated using a factory that produces color labels as an example.

[0086] When constructing a defect rule knowledge graph, nodes represent defect entities and process parameters, while edges represent quality inspection standards or material performance constraints. In color label production, defect entities may include fading label color or unclear pattern edges; process parameters include printing temperature and pressure. For example, label color fading is linked to the light resistance of printing ink through the quality inspection standard "Standard for the degree of fading of label color after a certain period of light exposure"; unclear pattern edges are linked to printing pressure through the "Standard for the relationship between printing pressure and pattern edge clarity."

[0087] In the process of repair solution generation, a contrastive learning algorithm based on hierarchical graph neural network is used for graph embedding matching. The subgraph structure features are extracted from the current defect type. Assuming that a defect with unclear pattern edges is detected in a certain color label, the corresponding graph structure is G defect . Extract subgraph structure features F through graph neural network GNN1 defect In GNN1, the feature extraction of graph structure data is performed through operations such as convolution to obtain the vector F that can represent the feature of the defect subgraph. defect .

[0088] Then in the knowledge graph G knowledge The knowledge graph is composed of multiple nodes and edges. Through the hierarchical graph neural network GNN2, starting from the bottom layer, the information of neighbor nodes is gradually aggregated. For example, in the first layer, the node only aggregates the information of the directly adjacent nodes; in the second layer, the aggregation includes the information of the directly adjacent nodes and their adjacent nodes, and so on. In each layer, the current defect subgraph feature F is calculated. defect Similarity Sim(F defect ,F knowledge ), calculation methods such as cosine similarity can be used.

[0089] A similarity threshold is set. If the calculated similarity is greater than the threshold, it means that there is a matching sub-graph pattern between the current defect type and the knowledge graph. For example, when the calculated similarity shows that the sub-graph feature of the pattern edge is unclear and the sub-graph feature of the pattern edge is unclear due to printing pressure problems in the knowledge graph is highly similar, the match is considered successful. At this time, a repair plan is generated based on the repair suggestions in the matching sub-graph pattern, such as adjusting the printing pressure. If there is no matching sub-graph pattern, it is necessary to further analyze the cause of the defect, adjust the knowledge graph or repair plan, and ensure that the repair plan meets the requirements of quality inspection standards and material performance constraints, so as to ensure the printing quality of color labels.

[0090] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "includes," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0091] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A green color printing defect intelligent diagnosis method based on deep learning, characterized in that: include: Collecting multi-source image data of green color prints, wherein the multi-source image data includes visible light images, hyperspectral images, and infrared thermal imaging data; Performing multi-scale feature extraction on the multi-source image data to obtain a defect feature vector corresponding to each image modality, the defect feature vector including: surface texture features, color difference distribution features, and thermal conduction anomaly features; The defect feature vectors are fused into a unified defect representation vector based on a cross-modal attention mechanism; Building a defect knowledge base based on a historical defect sample library, wherein the knowledge base stores third data, including: morphological feature templates of various defects, defect cause correlations, and corresponding repair process records; Based on the unified defect characterization vector and the third data in the defect knowledge base, a dynamic classification algorithm is used to determine the defect type of the current color printed product and a corresponding repair suggestion solution.

2. The intelligent diagnosis method for defects in green color printed products according to claim 1, characterized in that: The method of determining the defect type of the current color printed product and the corresponding repair suggestion plan by using a dynamic classification algorithm includes: generating an initial candidate defect set based on the similarity between the unified defect representation vector and the defect templates in the knowledge base; Constructing a defect cause topology graph based on the defect cause association relationship, wherein nodes in the topology graph represent defect types and edges represent logical or temporal associations between causes; Based on the defect cause topology map and real-time detection data, the initial candidate defect set is optimized through a graph reasoning algorithm, and the final defect type and repair solution priority ranking are output.

3. The intelligent diagnosis method for defects in green color printed products according to claim 1, characterized in that: The performing multi-scale feature extraction on the multi-source image data includes: A multi-branch convolutional network is used to extract local and global texture features from the visible light image, a spatiotemporal attention network is used to extract abnormal band features in the spectral sequence from the hyperspectral image, and an adaptive threshold segmentation algorithm is used to extract abnormal thermal distribution areas from the infrared thermal imaging data.

4. The intelligent diagnosis method for defects in green color printed products according to claim 3, characterized in that: The method of fusing the defect feature vectors into a unified defect representation vector based on a cross-modal attention mechanism includes: The importance distribution of each image feature is calculated by the inter-modality correlation weight; The weighted features are input into a bidirectional gated recurrent network to capture the temporal dependencies of cross-modal features.

5. The intelligent diagnosis method for defects in green color printed products according to claim 1, characterized in that: The dynamic classification algorithm also includes: constructing a defect classification optimization model based on a meta-learning framework, and the optimization model dynamically adjusts the feature fusion strategy through a multi-task loss function, and the loss function includes category discrimination loss, feature consistency loss and cause correlation loss.

6. The intelligent diagnosis method for defects in green color printed products according to claim 5, characterized in that: The graph reasoning algorithm is an improved simulated annealing algorithm, including: Encoding the defect cause topology graph into a state space, wherein the state nodes represent defect cause combinations; Generate candidate state transition paths through neighborhood search strategy and introduce random perturbations to escape from local optimal solutions; The optimal state path is evaluated according to an energy function, which is calculated based on the defect matching degree and the repair cost weight.

7. The intelligent diagnosis method for defects in green color printed products according to claim 1, characterized in that: The method further comprises: Using a multi-dimensional anomaly detection model on the real-time detection data to extract the characteristics of sudden defects in the production process; When a sudden defect is detected, adaptive adjustment of the repair plan is triggered, and the repair priority is updated according to the adjusted plan.

8. The intelligent diagnosis method for defects in green color printed products according to claim 7, characterized in that: The adaptive adjustment adopts a causal reasoning engine, and the specific method includes: Define the defect cause chain, process parameter influencing factors and repair resource constraints; The potential root causes are reversely deduced through the causal graph, and the minimum intervention repair strategy is generated based on the gradient optimization algorithm.

9. The intelligent diagnosis method for defects in green color printed products according to claim 1, characterized in that: The method also includes: constructing a defect rule knowledge graph, in which nodes represent defect entities and process parameters, and edges represent quality inspection standards or material performance constraints. During the repair plan generation process, the compliance of the defect type with the knowledge graph is verified through a graph embedding matching algorithm.

10. The intelligent diagnosis method for defects in green color printed products according to claim 9, characterized in that: The graph embedding matching algorithm is a contrastive learning algorithm based on a hierarchical graph neural network, including: Extract subgraph structural features from the current defect type; Neighbor node information is hierarchically aggregated in the knowledge graph, and matching subgraph patterns are screened through similarity comparison.

Images