UV printing error detection and correction method based on self-learning system

Abstract

The invention relates to the technical field of UV printing, and discloses a UV printing error detection and correction method based on a self-learning system, and the method comprises the steps: carrying out the normal state representation learning based on comparison self-supervised learning; modeling based on the state transition probability of the normalized flow; performing anomaly detection and early warning based on multi-metric fusion; performing anomaly interpretation and root cause analysis based on a variational auto-encoder; generating and executing a correction scheme based on anti-factual reasoning; according to the method, normal state representation is constructed by comparing self-supervised learning, and data dependence is reduced; a standardized flow modeling state transition probability is adopted, and printing errors are found as soon as possible; understandable abnormal root cause analysis is generated based on a variational auto-encoder; an optimal correction scheme is selected through anti-factual reasoning; the system has good adaptive ability, and can be continuously optimized and adapted to a new environment.

Description

Technical Field

[0001] The present invention relates to the technical field of UV printing, and more particularly to a UV printing error detection and correction method based on a self-learning system. Background Art

[0002] As a new printing process, UV printing technology uses ultraviolet light to cure inks for rapid printing. It offers advantages such as environmental friendliness, high efficiency, and wide applicability to a wide range of materials. It is widely used in packaging printing, advertising production, and personalized customization. With the continuous development of UV printing technology and the expansion of its application scenarios, the requirements for print quality and production efficiency are also increasing.

[0003] During the UV printing process, various printing errors, such as color difference, misregistration, and ink layer cracking, can easily occur due to factors such as material properties, equipment status, and environmental factors. These errors not only result in material waste and reduced production efficiency, but can also cause equipment damage. Therefore, timely and accurate detection and correction of printing errors are crucial to ensuring product quality and reducing production costs.

[0004] Current UV printing error detection relies primarily on manual experience or rule-based automated detection systems. While manual detection is highly reliable, it is inefficient and susceptible to subjective factors. While rule-based automated detection systems can improve efficiency, their detection capabilities are limited by the completeness of their pre-set rules, making them inadequate for the complex and ever-changing production environment. Furthermore, these methods often only detect errors that have already occurred and lack the ability to detect early warning symptoms.

[0005] Existing technologies for correcting printing errors primarily rely on operator experience or simple parameter adjustment rules. This approach struggles to accurately determine the direction and magnitude of parameter adjustments, often requiring multiple attempts to find the right correction solution. This is not only time-consuming and labor-intensive, but also increases material waste. Furthermore, with the continuous emergence of new printing materials and processes, traditional error detection and correction methods have limited adaptability and are unable to meet the needs of industrial development. Summary of the Invention

[0006] The present invention provides a UV printing error detection and correction method based on a self-learning system, which solves the technical problems existing in the printing error detection and correction technology in the related art, such as strong data dependence, delayed error discovery, lack of interpretability, blind correction process and lack of adaptability.

[0007] The present invention provides a UV printing error detection and correction method based on a self-learning system, comprising the following steps: Normal state representation learning based on contrastive self-supervised learning, and constructing a feature distribution model of the normal printing process through a multimodal feature extraction network; Based on the state transition probability modeling of normalized flow, the feature distribution model is used to learn the transition trajectory from normal state to various abnormal states; Perform anomaly detection and early warning based on multi-metric fusion, map the current printing status to the feature space and compare it with the normal state model, and predict the development trend based on the state transition probability; Abnormal interpretation and root cause analysis based on variational autoencoders encodes and decodes detected abnormal states, providing visualization and explainable abnormal root causes; Generate and execute correction plans based on counterfactual reasoning, predict the correction effect based on the root cause of the anomaly, and select the optimal plan for execution.

[0008] In a preferred embodiment, the normal state representation learning based on contrastive self-supervised learning includes: Multimodal data acquisition and preprocessing: acquisition and preprocessing of multimodal data during normal printing; Data enhancement and multi-view sample construction generate different view representations of the same printing state; Multimodal feature extraction network construction, constructing feature extraction networks for different modal data; The contrastive learning objective function is optimized so that the Euclidean distance of the positive sample pairs in the feature space is less than a preset threshold, and the Euclidean distance of the negative sample pairs in the feature space is greater than the preset threshold.

[0009] In a preferred embodiment, the state transition probability modeling based on normalized flow includes: Latent space construction, mapping samples in the normal printing state to the latent space; Normalized flow model construction to learn normal state distribution and potential abnormal state distribution; State transition path simulation, generating transition paths from normal states to various potential abnormal states; The state transition probability graph is constructed to calculate the transition probability between states.

[0010] In a preferred embodiment, the anomaly detection and early warning using multi-metric fusion includes: Real-time state feature extraction, mapping the current printing state into the latent space; Multi-metric anomaly score calculation, which calculates the degree of deviation based on reconstruction error metric, latent space metric, and energy metric; State development trajectory prediction, predicting the possible development trajectory of the current state at a future time point; Adaptive threshold warning: trigger an alert when the anomaly score or predicted state trajectory exceeds the threshold.

[0011] In a preferred embodiment, the anomaly interpretation and root cause analysis based on the variational autoencoder includes: Variational autoencoder construction and training, building a variational autoencoder model including encoder and decoder; Latent space traversal and normal reference generation: exploring and generating reference samples in the latent space whose feature distribution distance from the normal state is less than a preset threshold; Difference feature extraction and visualization, calculating the difference between abnormal samples and normal reference samples and generating a visual representation; Exception type identification and description generation, identifies the exception type and generates a human-understandable description.

[0012] In a preferred embodiment, the generation and execution of the correction scheme based on counterfactual reasoning includes: Model the relationship between printing parameters and abnormalities, and build a causal relationship model between printing parameters and printing status; Candidate generation of correction solutions: generating possible correction solutions based on anomaly root cause analysis and causal relationship models; Counterfactual effect prediction and solution optimization: predict the possible effects of each revised solution and select the optimal solution; Modify the execution and feedback optimization of the plan, execute the selected plan and update the model based on the actual results.

[0013] In a preferred embodiment, the multimodal data includes optical image sequences, temperature distribution data, pressure distribution data and UV light intensity data; the multimodal feature extraction network includes a convolutional neural network for image data, a recurrent neural network or a one-dimensional convolutional network for time series data, and a feature fusion module.

[0014] In a preferred embodiment, the multi-metric anomaly score calculation adopts a weighted combination method, and the three indicators of reconstruction error metric, latent space metric and energy metric are weightedly combined through weight parameters to obtain the final anomaly score.

[0015] In a preferred embodiment, the variational autoencoder includes dedicated encoders, sampling layers, and decoders for different modal data; The encoder outputs a mean vector and a log-variance vector, and the latent variables are sampled from the distribution of the encoder output via a reparameterization trick.

[0016] A computer-readable storage medium is used to store computer-readable instructions. When the computer-readable instructions are read by a computer, a UV printing error detection and correction method based on a self-learning system can be executed.

[0017] The beneficial effects of the present invention are: The UV printing error detection and correction method based on the self-learning system provided by the present invention constructs a normal state representation by comparing self-supervised learning and multimodal feature extraction networks. It can effectively learn the characteristic distribution of the printing process without a large amount of labeled data, significantly reducing data dependence and improving the practicality and promotion value of the system.

[0018] By normalizing the state transition probability of flow modeling and combining it with an anomaly detection mechanism based on multi-metric fusion, the present invention can detect the early symptoms and development trends of printing errors, implement predictive maintenance, effectively reduce material waste and equipment loss, and improve production efficiency.

[0019] The system's anomaly interpretation mechanism, based on a variational autoencoder, enables it to generate intuitive and understandable root cause analysis results, helping operators quickly identify the underlying problem. Furthermore, a counterfactual reasoning framework enables the system to predict the effects of different remediation options and automatically select the optimal one, significantly improving the accuracy and efficiency of the remediation process.

[0020] This invention has excellent adaptive capabilities, enabling it to continuously learn and optimize from production practices, adapting to changes brought about by new printing materials, new equipment, and new environments. The system's self-learning nature enables it to continuously accumulate experience and improve performance, providing strong support for the intelligent upgrade of the UV printing industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a flow chart of the UV printing error detection and correction method based on the self-learning system of the present invention; Figure 2 is a histogram comparing the early warning lead times of different error types of the present invention; Figure 3 It is a line graph showing the trend of the long-term operation effect of the system of the present invention; Figure 4 is a bubble chart of the relationship between the anomaly type and the number of correction attempts of the present invention; Figure 5 is a scatter plot of the feature space dimension importance distribution of the present invention; Figure 6 It is a radar chart comparing the system performance of the present invention. DETAILED DESCRIPTION

[0022] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that these embodiments are discussed solely to enable those skilled in the art to better understand and implement the subject matter described herein, and that the functions and arrangements of the elements discussed may be varied without departing from the scope of this specification. Various examples may omit, substitute, or add various processes or components as needed. Furthermore, features described in some examples may be combined in other examples.

[0023] At least one embodiment of the present invention discloses a UV printing error detection and correction method based on a self-learning system, such as Figure 1 As shown, the following steps are included: Step 1: Based on the normal state representation learning of contrastive self-supervised learning, a feature distribution model of the normal printing process is constructed through a multimodal feature extraction network; It includes the following sub-steps: Step 1.1, multimodal data acquisition and preprocessing; Collect multimodal data from various sensors in the UV printing equipment during the normal printing process, including but not limited to optical image sequences, temperature distribution data, pressure distribution data, UV light intensity data, etc. Perform preprocessing operations such as standardization, denoising, and registration on the collected raw data to align the data of different modalities in time and space.

[0024] In some embodiments, data acquisition can be accomplished using devices such as high-speed industrial cameras, infrared thermal imagers, pressure sensor arrays, and UV intensity detectors. The sampling frequency can be dynamically adjusted based on the printing speed. For example, for high-speed UV roll-to-roll printing equipment, the image acquisition frequency can be set to 100 frames per second, while temperature data can be collected at a rate of 10 frames per second.

[0025] Preprocessing can also include steps such as data outlier detection and removal, time series smoothing, and multimodal data temporal synchronization to improve the accuracy of subsequent analysis. For image data, preprocessing can also include operations such as region segmentation, feature point extraction, and illumination normalization.

[0026] Step 1.2: Data enhancement and multi-view sample construction; A variety of data enhancement techniques are applied to the preprocessed normal printing state data to generate different perspective representations of the same printing state.

[0027] Enhancement methods include but are not limited to: random cropping, rotation, scaling, adding Gaussian noise, occlusion, etc.

[0028] For image data, spatial transformations can be applied; For time series data, techniques such as time window offset can be applied.

[0029] Through these enhancement methods, each original sample is converted into two enhanced views to form a positive sample pair.

[0030] In some embodiments, data enhancement strategies can be dynamically adjusted based on the printing material and process characteristics. For example, for UV printing of transparent materials, a combination of reflected and transmitted light enhancement can be added; for printing on metal substrates, enhancement of surface texture and gloss variation can be added.

[0031] The enhancement strength can be adjusted adaptively. In the initial stage, a larger enhancement strength is used to promote the model to learn robust features, and the enhancement strength is reduced in the later stage to improve feature accuracy.

[0032] In some cases, physical model-guided enhancement methods can also be introduced, such as the simulation of ink droplet morphology changes based on fluid mechanics models.

[0033] Step 1.3, multimodal feature extraction network construction; Construct a feature extraction network suitable for different modal data.

[0034] For image data, convolutional neural networks are used to extract spatial features; For time series data, recurrent neural networks or one-dimensional convolutional networks are used to extract temporal features; For multimodal data, after constructing their own encoders, the features of different modalities are fused through the attention mechanism.

[0035] The feature extraction network outputs a fixed-dimensional feature vector as a low-dimensional representation of the sample.

[0036] The specific implementation of the multimodal feature extraction network is as follows: Network Architecture: A multimodal feature extraction network consists of multiple modality-specific subnetworks and a feature fusion module. For example, a common UV printing monitoring system might include an image subnetwork, a temperature subnetwork, a pressure subnetwork, and a UV light intensity subnetwork.

[0037] Image subnetwork: This uses a deep residual network (ResNet) architecture, consisting of multiple residual blocks. Each residual block consists of two 3×3 convolutional layers with skip connections, effectively mitigating the vanishing gradient problem. It typically includes five convolutional layer groups, each with a varying number of residual blocks, ultimately outputting a 256-dimensional feature vector.

[0038] The temperature subnetwork uses a combination of a one-dimensional convolutional network (1D-CNN) and a long short-term memory (LSTM) network. It first extracts local temperature patterns through three layers of one-dimensional convolutions with kernel sizes of 5, 3, and 3, respectively. It then uses a bidirectional LSTM (containing 128 hidden units) to capture long-term temperature trends, ultimately outputting a 128-dimensional feature vector.

[0039] Pressure sub-network and UV light intensity sub-network: The structure is similar to the temperature sub-network, but the convolution kernel size and the number of hidden units are adjusted according to the data characteristics, and 64-dimensional and 32-dimensional feature vectors are output respectively.

[0040] Feature fusion module: uses the self-attention mechanism to calculate the correlation between different modal features and generate fusion features.

[0041] The specific steps are: Perform linear projection on each modal eigenvector to generate a query vector , key vector Sum value vector ; Calculate the attention weight, which is the similarity between the query vector and the key vector, and then normalize and weight the result with the value vector. Capture information from different representation subspaces through a multi-head attention mechanism; Finally, the final fusion feature vector is generated through the feedforward network with a dimension of 512; Application example: In UV printed signage production, when a missing print is detected, the network can simultaneously analyze: The image data shows uneven distribution of ink drops; The temperature data shows abnormal fluctuations in the nozzle temperature; The pressure data shows that the ink supply system pressure is insufficient; UV light intensity data shows that the local curing is insufficient; Through the feature fusion module, the system identifies the correlation between these modes and determines that insufficient pressure is the root cause, while other anomalies are chain reactions, thus providing accurate root cause analysis for subsequent corrections.

[0042] In some embodiments, the feature extraction network can be initialized using transfer learning methods. For example, the image subnetwork can be initialized using model parameters pre-trained on a large-scale image dataset and then adapted to the UV printing image domain through domain adaptation techniques to accelerate convergence and improve feature quality.

[0043] The feature extraction network can also integrate attention distillation technology to implicitly incorporate expert knowledge into the model. By building a "teacher-student" network architecture, the teacher network incorporates the expert-annotated attention area information and guides the student network to learn more effective feature representations.

[0044] Step 1.4, comparative learning objective function optimization; Based on the extracted feature vectors, a contrastive learning objective function is constructed and optimized. The core idea of contrastive learning is to maximize the mutual information between representations of the same sample from different perspectives, while minimizing the mutual information between representations of different samples.

[0045] Specifically, the system uses different perspectives of the same printing state as positive sample pairs and representations of different printing states as negative sample pairs. By optimizing the loss function, the positive sample pairs are close in the feature space, and the negative sample pairs are far away in the feature space.

[0046] ; in, Represents the loss function of contrastive learning, which is used to measure the quality of the model parameters θ; Indicates the number of samples contained in each training batch; Represents the original feature representation of the i-th sample; Represents the feature representation of the i-th sample after data enhancement; Represents a function that calculates the similarity between two feature vectors, usually using cosine similarity; Represents the temperature parameter, which is used to adjust the contrast intensity of positive and negative sample pairs. A smaller value will make the contrast more "sharp"; Represents the indicator function, which takes the value 1 when the subscript j is not equal to i, otherwise it takes the value 0, and is used to exclude itself as a negative sample; represents the exponential function, which is used to map the similarity value to the positive number space; represents the logarithmic function, which is used to convert the ratio into the difference to make the optimization more stable; Indicates the sum of all samples in the batch; It represents the sum of all possible negative samples, including original features and enhanced features.

[0047] In some implementations, the contrastive learning objective function can employ a hard negative sample mining strategy to enhance training effectiveness. Specifically, the system can maintain a dynamically updated hard example memory library, storing negative samples that are highly similar to anchor samples. These hard examples are then prioritized for contrastive learning in each iteration to improve the model's discriminative capabilities.

[0048] The objective function can also incorporate multi-granularity contrastive learning strategies. In addition to sample-level contrast, contrastive learning can also be performed at the feature map level and local region level to capture feature similarities at different scales. For example, for image data, contrastive learning can be performed simultaneously at multiple levels—global image, mid-scale blocks, and local details—to improve the model's sensitivity to anomalies at different scales.

[0049] The temperature parameters can be adjusted adaptively according to the changes in the distribution of negative samples during training. Higher temperatures are used in the early stages of training to promote uniform distribution of feature space, and lower temperatures are used in the later stages to enhance intra-class aggregation and inter-class separation.

[0050] By optimizing the above objective function, the feature extraction network can learn the essential feature representation of the normal printing state while ignoring irrelevant changes and noise.

[0051] Step 2: Based on the state transition probability modeling of the normalized flow, the feature distribution model is used to learn the transition trajectory from the normal state to various abnormal states; It includes the following sub-steps: Step 2.1, latent space construction; Using the feature extraction network learned in step 1, we map the samples of the normal printing state into the latent space to form the latent representation distribution of the normal state. If any abnormal samples are available, we also map them into the latent space to initialize the state transition model.

[0052] In some implementations, the latent space can be further optimized to enhance its expressiveness and interpretability. For example, a spherical constraint can be introduced to restrict the latent representation to a hypersphere, improving the compactness of the representation; or a decoupling constraint can be introduced to make different dimensions of the latent space correspond to different physical meanings, such as ink droplet size, distribution density, and degree of solidification.

[0053] Manifold learning techniques, such as t-SNE or UMAP, can also be used to visualize and analyze the latent space, helping to understand the distribution characteristics of normal states and potential abnormal areas. Based on this, the boundaries of the normal state can be defined as a reference for anomaly detection.

[0054] Step 2.2, normalized flow model construction; A normalized flow model is constructed to learn the normal state distribution and potential abnormal state distribution. The normalized flow model consists of a series of reversible transformations that can transform a simple base distribution (such as a Gaussian distribution) into a complex target distribution. The model's forward propagation process sequentially processes the input latent representation through multiple consecutive reversible transformation functions to ultimately obtain the transformed representation.

[0055] Each transformation function satisfies the reversible property, that is, there exists an inverse function that can restore the original input from the transformation result. At the same time, the Jacobian determinant of each transformation function is computable and can be used to transform the probability density.

[0056] The state transition probability can be expressed as: ; in, Indicates that from the initial state At time t, the state The transition probability of represents the state at time t; Indicates the initial state; Indicates the state at time t-1; represents the single-step transition probability from the state at time t-1 to the state at time t; Indicates from time Status to time Single-step transition probability of the state; Indicates from time Status to time Single-step transition probability of the state; Indicates the integration of all states from the intermediate time 1 to t-1; Represents the integral symbol, which means the sum of all possible intermediate state paths.

[0057] The specific implementation of the normalized flow model is as follows: Model Architecture: According to the embodiments of this application, Real Non-Volume Preserving Flow (RealNVP) is used as the basic structure of the normalized flow model. The model contains multiple coupling layers, each of which performs the following transformations: Split the input vector into two parts; The first part remains unchanged, and the second part is transformed by an affine transformation, with the transformation parameters determined by the first part; After the transformation, a new vector is obtained; In the next coupling layer, the roles of the two parts are swapped to ensure the expressiveness of the model; Scaling and translation network: The scaling and translation functions in the affine transformation are implemented using a multilayer perceptron, typically consisting of three to four hidden layers. Each layer uses the LeakyReLU activation function, with the final layer using no activation function. To enhance the model's expressiveness, a batch normalization layer is added before the multilayer perceptron.

[0058] Multi-scale Architecture: To handle high-dimensional data, the model adopts a multi-scale architecture that distributes transformations across different resolution levels. The implementation consists of four scale blocks, each with four coupled layers. Between each scale block, a factorization operation maps some variables directly to the output, reducing subsequent computational effort.

[0059] Training optimization: Training is performed by maximizing the log-likelihood function. The training process transforms simple distributions (such as the normal distribution) into complex target distributions while taking into account the volume changes during the transformation.

[0060] Application example: In UV printing inkjet process monitoring, this model can effectively capture the evolution of printing status: The normal state is located in the high-density area of the latent space, forming a relatively concentrated distribution; As the ink viscosity gradually increases (a common UV printing problem), the system status gradually deviates from the normal area along a specific path; By learning this state transition path, the model can predict possible future nozzle clogging problems when there are only slight changes in the ink droplet shape (which are difficult to detect with the naked eye); The system calculates the probability of the current state evolving along each potential fault path, determines the most likely direction of fault development, and issues an early warning 20 to 30 minutes in advance, leaving ample time for maintenance personnel to intervene. In some implementations, in addition to RealNVP, other normalized flow variants, such as Glow, FFJORD, or Flow++, can be selected based on the needs of the specific application scenario. For example, for high-dimensional image data, the Glow model may be more advantageous in handling spatial correlation; while for irregular time series data, FFJORD may perform better in modeling complex dynamic systems.

[0061] To enhance the expressive power of the normalized flow model, conditional information can be introduced. For example, printing parameters (such as printing speed, UV light intensity, and ambient temperature) can be used as conditions to construct a conditional normalized flow model and learn the state distribution and transition patterns under different conditions.

[0062] By normalizing the flow model, the system can not only determine whether the current state is abnormal, but also predict its future development trend, achieving a transition from passive response to active prevention.

[0063] Step 2.3, state transition path simulation; Based on the trained normalized flow model, transition paths from the normal state to various potential abnormal states are generated. By interpolating or extrapolating in the latent space, the possible state transitions during the printing process are simulated. The generated paths cover the boundaries of the normal state and potential abnormal regions, providing comprehensive coverage of the state space.

[0064] In some embodiments, state transition path simulation can employ a combination of multiple strategies to achieve more comprehensive coverage. For example, random sampling, directed exploration, and expert guidance can be combined: random sampling uniformly explores the entire state space; directed exploration explores along the main change directions in the latent space; and expert guidance leverages known error patterns to simulate specific types of abnormal development paths in a targeted manner.

[0065] The state transition path generation process incorporates physical constraints to ensure that the generated paths conform to the physical laws of the UV printing process. For example, by considering factors such as ink rheological properties and UV curing dynamics, physically impossible state transition paths are filtered out, improving the accuracy and reliability of the simulation.

[0066] Step 2.4, construct the state transition probability graph; Based on the generated state transition paths, the transition probabilities between states are calculated and a state transition probability graph is constructed. This probability graph describes the transition probability from one state to another and can be calculated using the transformation characteristics of the normalized flow model.

[0067] The state transition probability graph essentially describes the dynamic process of the evolution of the printing state over time and can predict the possible future development direction and probability of the current state.

[0068] In some embodiments, the state transition probability graph may use a sparse representation, retaining only high-probability transition paths to reduce computational and storage overhead. For example, a probability threshold may be set to retain only edges with transition probabilities exceeding the threshold, or for each state, only the top K most likely transition target states may be retained.

[0069] State transition probability graphs can incorporate temporal information to construct time-varying transition probability models. By analyzing state transition characteristics at different time scales, the time course of anomaly development can be more accurately predicted, providing more precise time estimates for early warning systems.

[0070] Through the state transition probability graph, the system can predict the state evolution trend during the printing process and detect potential abnormal states early.

[0071] Step 3: Perform anomaly detection and early warning based on multi-metric fusion, map the current printing state into the feature space and compare it with the normal state model, and predict the development trend based on the state transition probability; It includes the following sub-steps: Step 3.1, real-time state feature extraction; During the UV printing process, multimodal sensor data is collected in real time, and the feature extraction network trained in step 1 is used to map the current printing state into the latent space to obtain its feature representation.

[0072] In some implementations, to improve real-time performance, an incremental feature extraction strategy can be employed. The system maintains a feature cache and only extracts features from newly acquired data, which are then fused with historical features in the cache to avoid repeated computations and improve processing efficiency.

[0073] Multi-resolution feature extraction strategies can also be introduced, using different processing frequencies and precisions for regions or time periods of varying importance. For example, for critical areas (such as print edges) or time periods with a high probability of anomalies, higher sampling frequencies and more refined feature extraction can be used; while for non-critical areas or normal stable phases, the processing frequency can be reduced to conserve computing resources.

[0074] Step 3.2, multi-metric anomaly score calculation; Based on multiple distance measurement methods, the deviation between the current state and the normal state distribution is calculated. Specifically, it includes: Reconstruction error metric: Exploiting the invertibility of the normalized flow model, the current state is mapped back to the data space, and then the reconstruction error with the original input is calculated.

[0075] Latent space metric: Calculate the distance between the current state and the center of the normal state distribution in the latent space. The Mahalanobis distance or other suitable distance metrics can be used.

[0076] Energy metric: Based on the state transition probability graph, the log-likelihood of the current state is calculated, and low-probability states are regarded as abnormal.

[0077] The final anomaly score is obtained by weighted combination of the above three metrics, and the weight parameters can be adjusted through validation data or adaptively learned.

[0078] ; in, is the total anomaly score, which indicates the overall abnormality of the current state x; is the reconstruction error metric, which is used to measure the difference between the current state x after reconstruction by the normalized flow model and the original input; It is a latent space metric used to calculate the distance between the current state x and the center of the normal state distribution in the latent space; It is an energy metric that calculates the log-likelihood value of the current state x based on the state transition probability graph, reflecting the probability of the state occurring. 、 、 Represent the weight parameters of reconstruction error metric, latent space metric, and energy metric, respectively, which are used to adjust the relative importance of the three metric methods in the total anomaly score and satisfy .

[0079] According to one embodiment of the present application, by combining multiple complementary measurement methods, the system can evaluate the degree of abnormality of the state from different perspectives, thereby improving the accuracy and robustness of detection.

[0080] In some implementations, the weighting parameters can be dynamically adjusted based on the characteristics of the print job and historical anomaly detection performance. For example, for high-detail printing jobs, the weight of reconstruction error can be increased; for large-scale batch jobs requiring early warning, the weight of energy metrics can be increased.

[0081] Step 3.3, state development trajectory prediction; Use the state transition probability graph constructed in step 2 to predict the possible development trajectory of the current state at future time points. The specific method is: starting from the current state, based on the state transition probability, simulate multiple possible state sequences and predict the state changes in the future period of time.

[0082] Through trajectory prediction, the system can identify potential abnormal development trends in advance. Even if the current status is still within the normal range, it can foresee that it may deviate from the normal area in the future and achieve advance warning.

[0083] In some implementations, methods such as Monte Carlo Tree Search (MCTS) can be used to efficiently explore trajectories, focusing on high-risk paths. By strategically expanding the search tree in state space, the system can find the development path most likely to lead to serious anomalies within limited computing resources.

[0084] Trajectory prediction can also be combined with printing parameter information to build a conditional prediction model. For example, considering planned changes in printing parameters (such as print speed adjustment and UV light intensity changes), the model can predict the state development trend under these parameter changes and provide more accurate early warning information.

[0085] Step 3.4, adaptive threshold warning; Combining the current anomaly score and the predicted state development trajectory, an adaptive threshold warning is implemented. The system maintains a dynamically updated anomaly score threshold and triggers an alert when any of the following conditions are met: The current anomaly score exceeds the threshold; The maximum anomaly score in the predicted future state trajectory exceeds a threshold; The abnormal score growth rate exceeds the preset value; The threshold can be adaptively adjusted based on the statistical characteristics of historical data to balance detection sensitivity and false alarm rate.

[0086] In some implementations, the threshold can be dynamically adjusted based on the importance and cost of the print job. For example, for high-value, low-tolerance precision print jobs, the threshold can be lowered to increase detection sensitivity; while for routine batch jobs, the threshold can be appropriately increased to reduce false alarms.

[0087] The early warning system can also incorporate a multi-level warning mechanism, issuing different levels of warning signals based on the severity and urgency of the anomaly. For example, a minor anomaly may issue a prompt-level warning, which serves as a reference for operators; a moderate anomaly may issue a warning-level warning, advising operators to pay close attention; and a severe anomaly may issue an emergency-level warning, which may trigger an automatic intervention mechanism.

[0088] The early warning system also integrates with a historical case library to provide reference information on similar cases. When an anomaly is detected, the system automatically retrieves similar historical anomaly cases from the case library, including their development process, root cause analysis, and successful corrective measures, providing operators with richer decision support.

[0089] Step 4: Based on the abnormality interpretation and root cause analysis of the variational autoencoder, the detected abnormal state is encoded and decoded, and the abnormal root cause is visualized and explainable. It includes the following sub-steps: Step 4.1, construction and training of variational autoencoder; Construct a variational autoencoder model, which consists of two parts: an encoder and a decoder. The encoder maps the input data to the distribution parameters of the latent space, and the decoder reconstructs the latent variables sampled from this distribution into output data.

[0090] The training objective of the variational autoencoder is to minimize the weighted sum of the reconstruction loss and the distribution regularization loss.

[0091] Among them, the reconstruction loss measures the difference between the input and the reconstructed output; Distribution regularization loss forces the latent space distribution to be close to the standard normal distribution; The balance parameter controls the trade-off between reconstruction accuracy and latent space regularization.

[0092] According to one embodiment of the present application, the specific implementation of the variational autoencoder is as follows: Model architecture: The UV printed state variational autoencoder adopts a hierarchical structure that can simultaneously process multimodal data and provide interpretable potential representations.

[0093] Encoder: Build specialized encoders for different modal data. For image data, a convolutional architecture is used, consisting of four convolutional blocks. Each block contains a convolutional layer, a batch normalization layer, and a LeakyReLU activation function, gradually encoding the input image into a feature map. For time series data, the GRU (Gated Recurrent Unit) network is used to extract temporal features.

[0094] After the features of all modalities are merged through the fusion layer, the mean vector and logarithmic variance vector are output respectively through two parallel fully connected layers, and the dimension of the latent space is set to 32.

[0095] Sampling layer: Use reparameterization techniques to sample latent variables from the distribution of the encoder output to ensure that the gradient can be back-propagated normally and achieve end-to-end training.

[0096] The decoder uses the opposite structure of the encoder. First, the latent variables are mapped to higher-dimensional features through fully connected layers, and then decomposed into features of each modality. For image data, transposed convolutional layers are used to gradually reconstruct the image; for time series data, a GRU network and fully connected layers are used to reconstruct the time series.

[0097] Training strategy: The reconstruction loss uses different metrics depending on the data type. For image data, a combination of mean squared error and perceptual loss is used; for time series data, mean squared error or mean absolute error is used.

[0098] The KL divergence loss is calculated analytically and measures the difference between the distribution of the encoder output and the standard normal distribution.

[0099] The balance parameter adopts a cyclic annealing strategy, which sets a small value at the beginning of training to focus on optimizing the reconstruction quality, and gradually increases as the training progresses to enhance the regularization effect.

[0100] Batch stochastic gradient descent was used in the training process, with an initial learning rate of 0.001 and cosine annealing to adjust the learning rate.

[0101] Enhanced interpretability: To improve the interpretability of the model, the following improvements are introduced: Latent space decoupling: By adding a total correlation penalty term, the latent variables are encouraged to be independent of each other, so that each dimension encodes independent printing features.

[0102] Attribute guidance: For some known printing parameters (such as UV light intensity, ink flow, etc.), semi-supervised training is introduced to align specific dimensions of the latent space with these parameters.

[0103] Adversarial regularization: A discriminator network is introduced to distinguish between normal samples and reconstructed samples, further improving the reconstruction quality and feature extraction capabilities.

[0104] Application example: In large-format UV flatbed printing equipment, this variational autoencoder can effectively explain the problem of color unevenness: When a part of a sign print is detected to be lighter in color, the system first inputs the entire sign image into the variational autoencoder; The encoder maps it to the latent space to obtain a 32-dimensional latent representation; By perturbing each dimension of the latent variable and reconstructing the image, the system found that the 7th and 18th dimensions had a significant impact on color saturation; By comparing the differences between abnormal samples and normal samples in these two dimensions, the system infers that such color problems are related to the ink supply system and UV curing energy distribution; Generate visual explanation diagrams that directly display the comparison between the problem area and the reference normal state, providing operators with problem location and cause analysis; In some embodiments, the variational autoencoder may adopt a hierarchical structure to construct a hierarchical latent space.

[0105] The top-level latent variables capture global printing state characteristics, such as overall color balance, uniformity, etc. The mid-level latent variables capture region-level features, such as local texture and edge characteristics; The underlying latent variables capture detail-level features such as ink dot shape and distribution.

[0106] This hierarchical representation helps provide multi-scale anomaly explanations.

[0107] The training process of the variational autoencoder can incorporate an active learning strategy, focusing on boundary samples and samples that are difficult to reconstruct. The system regularly evaluates the reconstruction error of samples and prioritizes samples with high error in training, improving the model's ability to represent abnormal boundary regions.

[0108] Through the generative power and latent space operations of the variational autoencoder, the system is able to not only detect anomalies but also generate intuitive explanations and root cause analysis, making black-box detection a transparent and understandable process.

[0109] Step 4.2, latent space traversal and normal reference generation; When an anomaly is detected, the abnormal sample is input into the encoder of the variational autoencoder to obtain its potential representation. Then, a directed traversal is performed in the latent space to generate a series of latent variables, which are then reconstructed into the corresponding output sample through the decoder.

[0110] By comparing the similarity between these reconstructed samples and normal samples, the reference sample closest to the normal state is found as the "normal reference" for the abnormal samples.

[0111] The core of this step is to explore the relationship between abnormal samples and normal samples in the latent space, identify key difference dimensions, and provide a basis for subsequent root cause analysis.

[0112] In some embodiments, the latent space traversal may employ a gradient-guided strategy to search along directions that reduce abnormal features.

[0113] Specifically, the system first defines an anomaly measurement function, then calculates the gradient of the function with respect to the latent variable, moves in the opposite direction of the gradient, gradually reduces the significance of the abnormal features, and generates samples that are closer to the normal state.

[0114] The normal reference generation process can introduce constraints to ensure that the generated reference samples are as close to normality as possible while preserving key characteristics. For example, an optimization objective can be defined that includes both fidelity and normality. By balancing these two objectives, reference samples can be generated that retain the key characteristics of the original samples while conforming to the normal distribution.

[0115] Step 4.3, differential feature extraction and visualization; Compute the difference between anomaly samples and normal reference samples to identify key features of variation. Differences can be calculated in the original data space or in the feature space, generating difference maps or feature importance scores.

[0116] For image data, a difference heat map can be generated to highlight abnormal areas; for multi-dimensional sensor data, a contribution analysis of each dimension can be generated to identify abnormal key sensor channels.

[0117] According to the embodiments of the present application, through differential analysis, the system can convert complex anomaly detection results into intuitive and understandable visual representations, helping operators to quickly locate problem areas and understand the degree of anomalies.

[0118] In some implementations, differential feature extraction can employ feature attribution methods, such as Integrated Gradients or SHAP (SHapley Additive Ex Planations), to quantitatively assess the contribution of each feature to anomaly determination. These methods can provide more accurate feature importance assessments and help understand the key factors influencing anomalies.

[0119] Difference visualization can be performed at multiple scales, displaying abnormal characteristics in a hierarchical manner, from global to local. For example, for large-format printing, a global difference heat map can be first provided to identify the main abnormal areas; then, key areas can be zoomed in to show the differences at the detailed level; and finally, a quantitative comparison of key indicators such as color deviation values and texture consistency indicators can be provided.

[0120] Step 4.4, abnormality type identification and description generation; Based on the extracted differential features and the system's accumulated experience, the system identifies the type of anomaly and its possible causes. The system maintains a continuously updated anomaly type library containing characteristic patterns and cause descriptions of known anomaly types.

[0121] By matching the characteristic pattern of the current anomaly with the patterns in the type library, the system can identify the anomaly type and generate a human-understandable description, such as "color fading due to insufficient ink flow" or "incomplete curing due to uneven UV light intensity".

[0122] In some implementations, anomaly type identification may utilize a prototypical network approach, maintaining a prototype representation for each known anomaly type. When a new anomaly is detected, its similarity to each prototype is calculated to determine the most closely matching anomaly type. For anomalies with similarity below a threshold, the system marks them as new anomalies and initiates a learning process to create a new anomaly category.

[0123] The description generation process can adopt a method that combines template filling with natural language generation.

[0124] For common exception types, use predefined templates to generate standardized descriptions; For complex or new anomalies, more specific descriptions are automatically generated by analyzing the anomaly features.

[0125] The description may include multiple aspects such as the abnormal area, severity, possible causes and recommended measures.

[0126] Step 5: Generate and execute a correction plan based on counterfactual reasoning, predict the correction effect based on the root cause of the anomaly, and select the optimal plan for execution; It includes the following sub-steps: Step 5.1, modeling the relationship between printing parameters and anomalies; Construct a causal relationship model between printing parameters (such as UV light intensity, printing speed, ink flow rate, etc.) and printing status. The model can be represented as a structured causal graph, where nodes represent variables (parameters or state characteristics) and edges represent causal relationships between variables.

[0127] Causal graphs can be partially predefined based on domain knowledge and gradually refined through structure learning and parameter estimation using observational data. The joint distribution of the model can be represented by the product of the conditional probability distributions of each node, where each conditional probability distribution describes the dependency between a node and its parent node.

[0128] According to the embodiment of the present application, the specific implementation of the causal relationship model is as follows: Model architecture: The Structural Causal Model (SCM) based on Bayesian networks is used, which consists of two parts: Causal structure: A directed acyclic graph (DAG) whose nodes include three types of variables: Printing parameter variables (such as UV light intensity, printing speed, ink flow, nozzle temperature, etc.); Intermediate state variables (such as ink drop shape, ink drop landing accuracy, curing degree, etc.); Final quality variables (e.g., color consistency, adhesion, surface gloss, etc.); Local conditional probability distribution: For each node, define its conditional probability distribution, which can be achieved using parametric forms (such as linear Gaussian models, nonlinear neural networks) or nonparametric forms (such as kernel density estimation).

[0129] Model building process: Initial structure construction: Based on expert knowledge of UV printing processes, an initial causal diagram is constructed to identify key variables and main causal paths, for example, the causal chain of "UV light intensity → curing degree → adhesion."

[0130] Structural optimization: Using structural learning methods such as the Peter-Clark algorithm (PC) or the Greedy Equivalence Search (GES) algorithm, we optimize the causal structure using observational data. Expert knowledge is introduced as structural constraints during the learning process to avoid learning causal relationships that are inconsistent with physical laws.

[0131] Parameter Learning: For each local conditional probability distribution, parameters are learned using maximum likelihood estimation or Bayesian estimation methods. For complex nonlinear relationships, neural network modeling is used and parameters are learned using the back-propagation algorithm.

[0132] Causal effect estimation: The intervention effect is calculated by intervention coefficient, which represents the distribution of other variables after the intervention of the variable. Use the backdoor criterion or frontdoor criterion to estimate the causal effect; For complex models, Monte Carlo simulation methods were used to estimate intervention effects; Application example: Correcting color deviation issues in UV printing: The system detected that a batch of products had a bluish color. Through causal model analysis, possible cause paths are identified: Path 1: Abnormal ink ratio → color imbalance → blue color; Path 2: UV light intensity is too high → over-curing → blue color; Path 3: The printing environment temperature is too low → the ink viscosity increases → the ink drop size is abnormal → the color is bluish; The system calculates the strength of the causal effect of each path and determines that path 2 (UV light intensity is too high) is the most likely; Generate multiple possible correction solutions, such as "reduce UV light intensity", "reduce UV light intensity and increase printing speed", etc. Predict the effects of each plan through counterfactual reasoning and select the best plan for execution; After performing corrections, the system continuously monitors the results and updates the causal model parameters to improve the accuracy of future inferences.

[0133] In some implementations, causal models can integrate knowledge from multiple sources, including physical models, expert experience, and data-driven discovery. For example, for areas with clear physical laws, such as ink rheological behavior, physical equations can be directly introduced to constrain causal relationships. For phenomena that are difficult to accurately model, such as device aging, data-driven statistical learning is primarily relied upon.

[0134] A causal diagram can have a hierarchical structure, containing variables and relationships at different levels of abstraction.

[0135] The top layer describes the relationship between the main printing parameters and quality indicators; The middle level describes the cause-effect relationships within each subsystem; The bottom layer describes the interactions at the specific component level.

[0136] This hierarchical structure facilitates multi-scale causal reasoning and intervention decisions.

[0137] Through causal modeling and counterfactual reasoning, the system can make the leap from correlation to causality, not only answering "what happened", but also explaining "why it happened" and predicting "how to correct it", providing a scientific basis for UV printing error correction.

[0138] Step 5.2, generation of candidate correction solutions; Based on the abnormal root cause analysis results of step 4 and the causal relationship model of step 5.1, a set of possible correction solution candidates is generated. Each correction solution is represented by the adjustment of one or more printing parameters.

[0139] The generation of the correction plan takes into account the following factors: Root cause parameter priority: give priority to adjusting parameters directly related to the root cause of the abnormality; Principle of minimum intervention: Adjust as few parameters as possible while still being able to solve the problem; Safety constraints: Parameter adjustments should not exceed the safety range to avoid equipment damage or material waste.

[0140] In some embodiments, the generation of the correction solution may adopt a combination of rule-based and heuristic search methods.

[0141] The system maintains an expert knowledge base containing correction strategy rules for common anomaly types; For complex or multi-factor anomalies, candidate solutions are generated through intelligent search of parameter space.

[0142] The solution generation process can be combined with historical successful cases and similar case searches to provide references. The system analyzes the similarity between the current anomaly and historical cases, extracts successful correction strategies from similar cases, and makes appropriate adjustments based on the current situation to form new candidate solutions.

[0143] Step 5.3: Counterfactual effect prediction and solution optimization; Using causal models to perform counterfactual reasoning, the system predicts the likely effects of each corrective action. For each action, the system evaluates its potential impact by calculating the expected outcome of the intervention.

[0144] By comparing the expected effects of different plans, the optimal plan is selected, that is, the plan that can maximize the improvement of the target quality indicators.

[0145] According to the embodiments of the present application, counterfactual reasoning allows the system to "simulate" the effects of different correction schemes before actual execution, significantly reducing trial and error costs and resource waste.

[0146] In some implementations, the prediction of results can consider multiple objective indicators, such as print quality, material consumption, and production efficiency, to form a multi-objective optimization problem. The system can employ the Pareto optimality principle to identify a set of correction solutions that strikes a good balance between these objectives and select the most appropriate solution based on the priority of the current task.

[0147] The solution optimization process uses a Bayesian optimization algorithm to efficiently search within the parameter space. Based on the performance of the evaluated solutions, the system constructs a probabilistic model of parameters and effects, guiding the generation and evaluation of subsequent solutions, gradually approaching the global optimal solution.

[0148] Step 5.4, implementation of revised plan and feedback optimization; Execute the selected optimal correction solution and adjust the corresponding printing parameters. After the correction is executed, continue to monitor the printing status to verify the correction effect.

[0149] Based on the actual correction results, the causal model and solution generation strategy are updated to form a closed-loop feedback optimization. By continuously accumulating correction experience, the system improves the accuracy and efficiency of future corrections.

[0150] Feedback optimization is one of the core features of the self-learning system of the present invention, enabling the system to continuously improve itself and adapt to the ever-changing printing environment and material characteristics.

[0151] In some embodiments, corrections can be performed using an incremental strategy, especially for large parameter adjustments that pose a high risk.

[0152] The system first implements small adjustments and observes the initial effects; After confirming that the direction is correct, gradually increase the adjustment range until the desired effect is achieved.

[0153] This gradual strategy reduces the risk of overcorrection.

[0154] The feedback optimization process can incorporate online learning techniques to update model parameters and strategies in real time. The system maintains an experience replay buffer, storing recent correction examples (including problem characteristics, measures taken, and actual results). These examples are regularly used to update the causal model and decision-making strategy, continuously improving the system's correction capabilities.

[0155] For successful correction cases, the system automatically generates knowledge items and adds them to the company's knowledge base for easy reference and learning. These items contain problem descriptions, diagnostic procedures, corrective measures, and effectiveness evaluations, becoming valuable process knowledge assets for the company.

[0156] Application examples of this embodiment Application scenario description: This implementation has been put into practical use on a large-scale UV flatbed printing production line. This line primarily produces high-end advertising signs and display materials, processing approximately 100 print batches daily on a variety of materials, including acrylic, aluminum composite panels, and PVC. Due to the high quality requirements, error detection and timely correction are crucial for production efficiency and cost control.

[0157] The production line is equipped with the following equipment and sensor systems: 8 industrial-grade UV flatbed printers with a maximum printing size of 2.5m×1.2m; Each printer is equipped with four high-resolution industrial cameras to cover different angles of the printing area; Each print head is equipped with a temperature sensor and a pressure sensor; The UV lamp group is equipped with a light intensity sensor and a temperature sensor; Environmental monitoring system records environmental parameters such as temperature, humidity, and dust concentration.

[0158] Before implementing the method of the present invention, the production line experienced an average of 35 to 40 abnormal situations per week, with a scrap rate of approximately 8.5%. It took an average of 2.7 attempts to correct a problem, resulting in high material waste and production delays.

[0159] Implementation process example: Data collection and model training: During the initial implementation phase, the system collected two weeks of normal printing data, including: Image data: approximately 28,000 high-resolution printing process images; Temperature data: about 760 hours of temperature time series; Pressure data: approximately 760 hours of pressure time series; UV light intensity data: about 760 hours of light intensity time series; Environmental parameters: approximately 760 hours of environmental parameter records.

[0160] For the acrylic printing process, the system first performed data preprocessing, including image calibration, noise filtering, and time series alignment. It then applied data augmentation techniques to generate 10 enhanced versions of each original sample, expanding the training dataset.

[0161] In constructing the feature extraction network, the system uses a pre-trained ResNet-34 as the basis for image feature extraction and adjusts it through transfer learning. Time series data is processed using a bidirectional LSTM network with 128 hidden units. Feature fusion uses an 8-head self-attention mechanism to fuse features from each modality into a 512-dimensional feature vector.

[0162] Contrastive learning training used a batch size of 128, an initial temperature parameter of 0.07, and a hard negative mining strategy. Training was performed on a GPU and took approximately 12 hours to converge.

[0163] Anomaly detection and warning examples: During a UV printing process on acrylic sheets, the system successfully warned of an ink blockage. The specific process is as follows: After printing began, the system collected multimodal data and extracted features in real time. At approximately 25% of the print run, the system detected a slight fluctuation in the ink pressure parameter using a normalized flow model. The anomaly score was 0.32, which fell below the direct warning threshold of 0.65.

[0164] The system predicted future state trends based on the state transition probability graph, indicating that if the current state continues, the anomaly score will reach 0.78 in approximately 15 minutes, exceeding the warning threshold. The system identified this development pattern as consistent with an early sign of nozzle clogging.

[0165] The system triggered the early warning mechanism, issuing an orange (moderate) warning signal to the operator, while also providing a visual predictive trend chart showing the areas and time points where problems might occur.

[0166] After receiving the early warning, the operator paused the printing task and inspected the designated printhead, where they discovered that tiny ink residues were gradually accumulating, which would later lead to complete blockage.

[0167] Before the problem could cause any substantial damage, the operator cleaned the print head, eliminated the potential risk, and then continued the printing task. The entire process was interrupted for only about 5 minutes.

[0168] Root cause analysis and correction examples: In another case, the system dealt with a color unevenness issue: The system detected a slight color fade in the right area of the printed product, with an abnormality score of 0.72, triggering an early warning.

[0169] The variational autoencoder analyzes the abnormal samples, maps them into a latent space, and performs feature traversal. The system generates multiple normal reference samples and, through comparison, determines that the latent space dimensions most relevant to the abnormalities are the 8th and 23rd dimensions, which are associated with ink jet volume and UV curing intensity, respectively.

[0170] Differential feature extraction generated a heat map that clearly showed the faded area on the right. The system matched this pattern with a library of historical cases and identified that this type of problem is often associated with an uneven ink supply system or an aging UV light source.

[0171] Through causal analysis, the system constructs the possible causal chain of the problem: Path 1: Insufficient pressure in the right ink supply line → ink drop volume decreases → color density decreases; Path 2: The UV lamp on the right is aging → insufficient curing energy → incomplete color fixation; The system calculated the strength of the causal effect of each path and determined that the probability of path 1 was 78% and that of path 2 was 22%.

[0172] Based on this analysis, the system generated three candidate corrections: Solution A: Increase the pressure of the right ink supply system by 5%; Solution B: Increase the overall ink flow by 3% and increase the print head's dwell time in the right area by 2%. Option C: Increase UV light intensity by 7%; Counterfactual reasoning showed that Option B had the best expected improvement (89%) and the least impact on the equipment. The system recommended Option B as the optimal choice.

[0173] The operator adjusted the parameters for Option B, and the corrected print results showed a 92% improvement in color uniformity, very close to the prediction. The system recorded this successful correction and updated the causal model and case library.

[0174] Technical effect verification: After 6 months of actual operation, this implementation method has achieved significant technical results on the UV printing production line: Early warning capability: The system can predict potential anomalies before problems fully manifest, significantly improving warning time: Ink system related issues: average early warning of 31 minutes, 63% earlier than traditional methods; Curing system-related issues: 18 minutes of early warning on average, 57% earlier than traditional methods; Material-related issues: 25 minutes of early warning on average, 61% earlier than traditional methods; Average warning lead time: increased from 12 minutes before implementation to 26 minutes after implementation, an increase of 117%.

[0175] The ability to provide early warning significantly reduces losses due to developing problems. Material waste has been reduced from an average of 32 high-value pieces of material per month to 7 pieces, a 78% reduction. Equipment maintenance costs have been reduced by 42%, primarily due to reduced equipment damage caused by errors.

[0176] Improved correction efficiency: The system significantly improves problem-solving efficiency through precise root cause analysis and correction solution optimization: First-time correction success rate: increased from 36% before implementation to 87% after implementation, a 141% improvement; The average number of correction attempts decreased by 56%, from 2.7 before implementation to 1.2 after implementation. Revision completion time: reduced from an average of 27 minutes to 11 minutes, a 59% time saving; The adoption rate of correction suggestions by operators reached 93%, indicating that the correction solutions provided by the system are highly credible.

[0177] In addition, overall production efficiency has also been significantly improved: Average production line downtime: reduced from 465 minutes per week to 178 minutes, a 62% decrease; Scrap rate: reduced from 8.5% to 2.3%, a 73% reduction; Production efficiency: increased by an average of 21%, primarily due to reduced interruptions and rework; Quality consistency: Batch-to-batch color deviation was reduced by 68% and edge accuracy was improved by 43%.

[0178] As the system continues to operate, its self-learning capabilities continue to improve. In the most recent month of operational data, the system achieved an 85% anomaly detection accuracy rate for the newly introduced transparent PET material using only five days of normal data, demonstrating its ability to quickly adapt to new materials and processes.

[0179] like Figures 2 to 6 As shown, there are comparisons of early warning lead times for different error types; trends in the long-term operation of the system; the relationship between anomaly types and the number of correction attempts; the importance distribution of feature space dimensions; and system performance comparisons.

[0180] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Ordinary technicians in this field can also make more forms of equivalent embodiments based on the inspiration of this embodiment, all of which are protected by this embodiment.

Claims

1. A UV printing error detection and correction method based on a self-learning system, characterized in that: The following steps are involved: Normal state representation learning based on contrastive self-supervised learning, and constructing a feature distribution model of the normal printing process through a multimodal feature extraction network; Based on the state transition probability modeling of normalized flow, the feature distribution model is used to learn the transition trajectory from normal state to various abnormal states; Perform anomaly detection and early warning based on multi-metric fusion, map the current printing status to the feature space and compare it with the normal state model, and predict the development trend based on the state transition probability; Abnormal interpretation and root cause analysis based on variational autoencoders encodes and decodes detected abnormal states, providing visualization and explainable abnormal root causes; Generate and execute correction plans based on counterfactual reasoning, predict the correction effect based on the root cause of the anomaly, and select the optimal plan for execution.

2. The UV printing error detection and correction method based on the self-learning system according to claim 1 is characterized in that: The normal state representation learning based on contrastive self-supervised learning includes: Multimodal data acquisition and preprocessing: acquisition and preprocessing of multimodal data during normal printing; Data enhancement and multi-view sample construction generate different view representations of the same printing state; Multimodal feature extraction network construction, constructing feature extraction networks for different modal data; The contrastive learning objective function is optimized so that the Euclidean distance of the positive sample pairs in the feature space is less than a preset threshold, and the Euclidean distance of the negative sample pairs in the feature space is greater than the preset threshold.

3. The UV printing error detection and correction method based on the self-learning system according to claim 1 is characterized in that: The state transition probability modeling based on normalized flow includes: Latent space construction, mapping samples in the normal printing state to the latent space; Normalized flow model construction to learn normal state distribution and potential abnormal state distribution; State transition path simulation, generating transition paths from normal states to various potential abnormal states; The state transition probability graph is constructed to calculate the transition probability between states.

4. The UV printing error detection and correction method based on the self-learning system according to claim 1 is characterized in that: The anomaly detection and early warning of multi-metric fusion includes: Real-time state feature extraction, mapping the current printing state into the latent space; Multi-metric anomaly score calculation, which calculates the degree of deviation based on reconstruction error metric, latent space metric, and energy metric; State development trajectory prediction, predicting the possible development trajectory of the current state at a future time point; Adaptive threshold warning: trigger an alert when the anomaly score or predicted state trajectory exceeds the threshold.

5. The UV printing error detection and correction method based on the self-learning system according to claim 1 is characterized in that: The anomaly explanation and root cause analysis based on the variational autoencoder includes: Variational autoencoder construction and training, building a variational autoencoder model including encoder and decoder; Latent space traversal and normal reference generation: exploring and generating reference samples in the latent space whose feature distribution distance from the normal state is less than a preset threshold; Difference feature extraction and visualization, calculating the difference between abnormal samples and normal reference samples and generating a visual representation; Exception type identification and description generation, identifies the exception type and generates a human-understandable description.

6. The UV printing error detection and correction method based on the self-learning system according to claim 1 is characterized in that: The generation and execution of the correction plan based on counterfactual reasoning includes: Model the relationship between printing parameters and abnormalities, and build a causal relationship model between printing parameters and printing status; Candidate generation of correction solutions: generating possible correction solutions based on anomaly root cause analysis and causal relationship models; Counterfactual effect prediction and solution optimization: predict the possible effects of each revised solution and select the optimal solution; Modify the execution and feedback optimization of the plan, execute the selected plan and update the model based on the actual results.

7. The UV printing error detection and correction method based on the self-learning system according to claim 2, characterized in that: The multimodal data includes optical image sequences, temperature distribution data, pressure distribution data and UV light intensity data; the multimodal feature extraction network includes a convolutional neural network for image data, a recurrent neural network or a one-dimensional convolutional network for time series data and a feature fusion module.

8. The UV printing error detection and correction method based on the self-learning system according to claim 4 is characterized in that: The multi-metric anomaly score calculation adopts a weighted combination method, which combines the three indicators of reconstruction error metric, latent space metric and energy metric through weight parameters to obtain the final anomaly score.

9. The UV printing error detection and correction method based on the self-learning system according to claim 5, characterized in that: The variational autoencoder includes specialized encoders, sampling layers, and decoders for different modal data; The encoder outputs a mean vector and a log-variance vector, and the latent variables are sampled from the distribution of the encoder output via a reparameterization trick.

10. A computer-readable storage medium, characterized in that It is used to store computer-readable instructions, and when the computer-readable instructions are read by a computer, it can run the UV printing error detection and correction method based on the self-learning system as described in any one of claims 1 to 9.

Images