Printed matter quality intelligent detection and data tracing system

Abstract

The invention relates to the field of printing industry automation, and discloses a printed matter quality intelligent detection and data tracing system. Comprising a multi-modal data synchronous acquisition and preprocessing module for establishing working condition parameter and image signal space-time mapping, a physical constraint dynamic reference generation module for generating a controlled reference image by using a conditional generative adversarial network, and a differential abnormal flow analysis module for extracting abnormal signal space-time evolution characteristics. The knowledge-based causal inference module is used for carrying out fault attribution in combination with an equipment physical fingerprint database; the disturbance injection-based adaptive causal calibration module is used for executing active detection in a security envelope; and the knowledge graph self-evolution module is used for updating a model based on verification data. Real-time physical process parameters are fused to a reference image generation process, dynamic characteristics are extracted by using differential abnormal flow, and fault attribution is carried out in combination with an equipment physical fingerprint database, so that accurate tracing of physical root causes of printing defects and closed-loop adaptive learning of a diagnosis model are realized.

Description

Technical Field

[0001] This invention relates to the field of printing industry automation, specifically to an intelligent inspection and data traceability system for printed materials. Background Technology

[0002] As the printing industry moves towards higher speeds and greater precision, automated quality inspection technology based on machine vision has become a key means of ensuring production efficiency. Current mainstream inspection solutions typically use line-scan industrial cameras to capture images of the printed surface. Image processing algorithms then perform pixel-level comparisons between the real-time captured images and pre-stored standard reference templates, identifying appearance defects such as color casts, blemishes, and missing prints by calculating grayscale or color differences. This method can achieve relatively effective quality monitoring in scenarios with relatively stable production environments and fixed process parameters.

[0003] The actual printing process involves the coupling effects of multiple physical fields, including fluid dynamics, mechanical transmission, and material deformation. Most existing visual inspection systems operate independently of the equipment's physical control system, relying primarily on image signals for monitoring, without fully considering the dynamic impact of physical parameters on image quality. When physical conditions such as equipment speed, paper tension, or ink viscosity fluctuate, printed materials are prone to non-defect-related elastic deformation or slight color shifts. Traditional algorithms based on static templates struggle to adapt to such dynamic changes, easily leading to increased false alarm rates. Furthermore, current technologies typically only output the geometric location of defects at the image level, making it difficult to establish a direct mapping between visual anomalies and the physical equipment's operating state, thus hindering automatic tracing from image defects to the root cause in physical components.

[0004] Existing intelligent diagnostic models primarily rely on pre-defined expert rules or supervised learning based on historical labeled data, and their diagnostic coverage is limited by the completeness of the training samples. Faced with complex faults caused by multiple coupled factors or unknown fault modes not previously recorded during production, existing passive detection systems lack the ability to actively verify causal relationships. These systems struggle to observe system responses to determine fault mechanisms through fine-tuning control parameters, and cannot automatically update their internal diagnostic knowledge base based on new operating conditions. Such systems, lacking closed-loop evolution capabilities, may experience a decline in diagnostic accuracy as equipment ages or processes change, failing to meet the requirements of intelligent manufacturing for full lifecycle data traceability. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an intelligent inspection and data traceability system for printed materials, which solves the problems of decoupling physical conditions and image features, difficulty in tracing dynamic root causes through static inspection, and lack of adaptive diagnostic capabilities for unknown faults in existing printing quality inspection technologies.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solution: a smart detection and data traceability system for printed matter quality, comprising: a multimodal data synchronous acquisition and preprocessing module, a physical constraint dynamic reference generation module, a differential anomaly flow analysis module, a knowledge-based causal inference module, an adaptive causal calibration module based on perturbation injection, and a knowledge graph self-evolution module.

[0007] The multimodal data synchronous acquisition and preprocessing module is used to perform synchronous acquisition and standardization processing of physical condition parameters and digital image signals. This module uses a synchronization controller to generate trigger signals, controlling the industrial camera's line scanning action and the sensor's sampling action to maintain temporal synchronization. To address the mismatch between the sampling frequencies of image data and physical parameter data, this module uses the image acquisition time as a time reference and performs linear interpolation calculations on the physical process parameters acquired at adjacent times. In the data normalization stage, this module employs a maximum-minimum-maximum normalization method, subtracting the statistical minimum value from the original physical parameter value and dividing by the difference between the statistical maximum and minimum values. This maps each physical quantity to a unified dimensionless interval, outputting a standardized physical process parameter vector and a preprocessed image.

[0008] The physical constraint dynamic reference generation module is connected to the multimodal data synchronous acquisition and preprocessing module to generate reference images constrained by physical conditions. This module is based on a conditional generative adversarial network (GAN) architecture, which includes a generator and a discriminator. The generator employs an encoder-decoder network architecture, receiving a static standard template image and a normalized physical process parameter vector as input. The physical process parameter vector is then fused into the feature layer of the encoder-decoder network after feature mapping. The discriminator uses a fully convolutional architecture, dividing the input image into multiple local patches and outputting a discrimination result matrix. During model training, the total loss function is defined as the weighted sum of the conditional adversarial loss and the pixel-level reconstruction loss. The conditional adversarial loss is calculated based on the discriminator's expected discrimination probability for the real image and the expected discrimination probability for the generated image; the pixel-level reconstruction loss is calculated based on the expected sum of the absolute values ​​of the pixel differences between the generated image and the real image under the distributions of real image data, physical process parameters, and random noise.

[0009] The differential anomaly flow analysis module connects the multimodal data synchronous acquisition and preprocessing module and the physical constraint dynamic reference generation module to extract dynamic features. This module performs pixel-by-pixel absolute difference operations between the real-time image and the dynamic reference image, and performs morphological opening operations on the results to generate a grayscale net difference image. This module stacks the grayscale net difference images generated at consecutive time points in chronological order to construct a three-dimensional differential anomaly flow. For the one-dimensional signal sequence of the brightness mean of the region of interest in the differential anomaly flow changing over time, this module performs a fast Fourier transform and calculates the spectral complex coefficients, selecting the frequency component with the largest spectral amplitude as the dominant frequency feature. Simultaneously, this module uses optical flow to extract motion trajectory features, establishing an optical flow constraint equation based on the assumption of constant brightness. This equation limits the sum of the product of the horizontal spatial gradient and the horizontal velocity component, the product of the vertical spatial gradient and the vertical velocity component, and the gradient in the time dimension of the grayscale net difference image to zero. This module introduces local smoothing constraints and uses the least squares method to solve the equation to obtain the velocity field vector.

[0010] The knowledge-based causal inference module is connected to the differential anomaly flow analysis module and is used to perform fault attribution. This module has a pre-installed physical fingerprint database for equipment, which stores the theoretical characteristic frequencies of each rotating component. The value of the theoretical characteristic frequency is equal to the real-time machine operating linear velocity divided by the product of pi and the physical diameter of the component, then multiplied by a mechanical coefficient. This module calculates the matching error between the measured dominant frequency characteristic and the theoretical characteristic frequency. When the matching error is less than a preset frequency tolerance threshold, the corresponding candidate root cause node is activated. This module further utilizes a Bayesian inference algorithm, combining prior probabilities and likelihood probabilities from the knowledge graph, to calculate the posterior probability of a specific fault root cause.

[0011] An adaptive causal calibration module based on perturbation injection is used to perform active verification when causal inference is uncertain. This module has a preset safe operating envelope composed of the physical limit parameters of the device actuators. Before generating a perturbation signal, the module calculates the vector sum of the proposed perturbation command and the current state value, and issues the perturbation signal only if the predicted operating state value corresponding to this vector sum falls within the safe operating envelope. During perturbation injection, the module constructs a perturbation response matrix, where the element values ​​are equal to the difference between the dynamic feature vectors before and after the perturbation injection divided by the applied perturbation amplitude, characterizing the sensitivity of changes in physical control parameters to image feature components.

[0012] The knowledge graph self-evolution module updates the causal knowledge base based on the perturbation response matrix. If the analysis results show that a certain physical parameter is related to an image feature and a corresponding directed edge already exists in the knowledge graph, the module adds the original weight value of the directed edge to the product of the response coefficient value obtained in this experiment and the learning rate coefficient to update the weight. If no corresponding directed edge exists in the knowledge graph, the module creates a new root cause node and a directed edge, and uses the standardized response coefficient value as the initial weight. If the measured response coefficient value is lower than a preset invalidation threshold, the knowledge graph self-evolution module performs linear decay on the weight of the corresponding directed edge, and removes the directed edge when the weight is lower than the minimum retention threshold.

[0013] The physical process parameters include the ink roller speed, ink layer thickness setting value and ink viscosity of the ink supply system, the roller pressure and packing compression of the printing and imprinting unit, the hot air temperature and wind speed of the drying system, and the real-time operating speed and paper tension of the substrate transport system.

[0014] This invention provides an intelligent inspection and data traceability system for printed materials. It has the following beneficial effects: 1. This invention constructs a dynamic reference image generation mechanism constrained by physical conditions by fusing real-time physical process parameters into the feature layer of a generative adversarial network. This allows the reference image to adjust in real time according to changes in equipment operating speed and pressure. Combined with spatiotemporal analysis of differential anomaly flow, the system can effectively filter out background fluctuation interference caused by normal mechanical vibration or operating condition switching of the equipment. While retaining the true defect signal, it reduces the false detection rate and achieves high signal-to-noise ratio detection under dynamic and complex operating conditions.

[0015] 2. This invention establishes a physical fingerprint database containing the mechanical geometric parameters and kinematic relationships of equipment, and quantitatively matches the spectral characteristics of abnormal image signals with the theoretical rotational frequencies of mechanical components. This cross-modal causal inference method based on a physical model breaks through the limitation of traditional visual inspection, which can only detect surface defects and cannot identify physical causes. It can directly trace printing quality problems to specific physical components such as ink rollers, gears, or bearings, improving the accuracy of fault diagnosis and maintenance efficiency.

[0016] 3. This invention quantifies the sensitivity relationship between control parameters and quality characteristics by implementing an active exploration mechanism based on perturbation injection within the safe operation envelope, utilizing minute and controllable physical signals to excite the system and calculate the perturbation response matrix. Combined with a closed-loop self-evolutionary update strategy using a knowledge graph, the system can automatically correct its diagnostic model by acquiring definitive causal data through online experiments when historical fault samples are lacking or when facing unknown and complex operating conditions, thus achieving adaptive learning and capability expansion of the detection system. Attached Figure Description

[0017] Figure 1This is a system framework diagram of the present invention; Figure 2 This is a flowchart illustrating the workflow of the multimodal data synchronous acquisition and preprocessing module of the present invention. Figure 3 This is a flowchart illustrating the construction and operation of the physical constraint dynamic reference generation module of the present invention. Figure 4 This is a data processing flowchart of the differential anomaly flow analysis module of the present invention; Figure 5 This is a flowchart illustrating the operation of the knowledge base-based preliminary causal inference module of the present invention. Figure 6 This is a flowchart illustrating the operation of the adaptive causal calibration module based on perturbation injection according to the present invention. Figure 7 This is a flowchart illustrating the operation of the knowledge graph closed-loop self-evolution module of the present invention.

[0018] Among them, 100 is the multimodal data synchronous acquisition and preprocessing module; 200 is the physical constraint dynamic reference generation module; 300 is the differential anomaly flow analysis module; 400 is the knowledge-based causal inference module; 500 is the perturbation injection-based adaptive causal calibration module; and 600 is the knowledge graph self-evolution module. Detailed Implementation

[0019] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] See attached document Figure 1 , Figure 1 This is a system framework diagram of a printed matter quality intelligent inspection and data traceability system according to an embodiment of the present invention. The present invention provides a printed matter quality intelligent inspection and data traceability system, comprising: The multimodal data synchronous acquisition and preprocessing module 100 is used to acquire physical process parameters and printed images synchronously with the printing process, and to perform preprocessing. The physical constraint dynamic reference generation module 200 is used to generate corresponding dynamic reference images based on real-time physical process parameters. The differential anomaly flow analysis module 300 is used to purify anomalous signals by performing differential operations between real-time images and dynamic reference images, and to analyze continuous differential image sequences to extract dynamic features. The knowledge-based causal inference module 400 is used to match the extracted dynamic features with the built-in knowledge base to infer the physical root cause of the anomaly. The perturbation-injection-based adaptive causal calibration module 500 is used to identify the causal relationship of an anomaly online by applying a small perturbation to the device and analyzing the response when encountering an unknown anomaly. The Knowledge Graph Self-Evolution Module 600 is used to add newly identified causal relationships to the knowledge base in order to achieve the system's self-learning and capability evolution.

[0021] The various modules described above in the embodiments of the present invention will be explained in detail below.

[0022] See attached document Figure 2 The following describes the specific implementation method and workflow of the multimodal data synchronous acquisition and preprocessing module 100. As the basic data access point for the entire system, the multimodal data synchronous acquisition and preprocessing module 100 is responsible for establishing a precise spatiotemporal mapping between physical world operating parameters and digital image signals, providing standardized multimodal data input for subsequent physical constraint dynamic reference generation and causal inference. This includes the following steps: Step S110: Perform parallel acquisition of multimodal data.

[0023] The system acquires data through two independent but time-correlated channels. The first channel collects surface image data of the printed material via a line-scan industrial camera connected to a gigabit network or Camera-Link interface. The industrial camera is mounted at the output end of the printing press or between key color groups to acquire high-resolution color or multispectral images. The second channel collects physical process parameters characterizing the operating status of the equipment through sensor arrays deployed at key physical nodes of the printing equipment.

[0024] The physical process parameters specifically encompass mechanical, thermodynamic, and fluid dynamic variables that affect printing quality, including but not limited to: the ink roller speed, ink layer thickness setpoint, and ink viscosity of the ink supply system; the roller pressure and packing compression of the printing press unit; the hot air temperature and wind speed of the drying system; and the real-time operating speed and paper tension of the substrate transport system. All of the above physical quantities are digitized in real time and encapsulated into a physical process parameter vector. The analog-to-digital conversion (A / D) of sensor analog signals and the data transmission configuration based on fieldbus (such as Ether-CAT, Profinet) are well-known technologies in this field and will not be elaborated here.

[0025] Step S120: Perform timestamp synchronization and alignment of multimodal data.

[0026] Because image data and physical parameter data have different sampling frequencies and transmission delays, a unified time base is needed. The system employs a combination of hardware triggering and software alignment. At the hardware level, a synchronous controller connected to the printing press spindle encoder generates a TTL level trigger signal to synchronously trigger the camera's line scanning and the sensor's sampling actions, ensuring that each frame of image... With physical process parameter vector Consistency in physical time. At the data processing level, if there is a mismatch in sampling rates, the time of image acquisition should be considered. Based on this, linear interpolation is performed on the physical process parameters acquired at adjacent time points to map low-frequency physical parameter data to the time points of image acquisition, ensuring data consistency. A strict correspondence.

[0027] Step S130: Perform data cleaning and noise reduction.

[0028] For physical process parameters, the system employs moving average filtering or median filtering algorithms to remove random noise and outliers caused by electromagnetic interference. For image data, it performs region of interest (ROI) extraction, removes image information from background conveyor belts or non-printing areas, and uses pre-calibrated camera intrinsic parameter matrices and distortion coefficients for geometric distortion correction to eliminate the impact of lens distortion on subsequent feature extraction.

[0029] Step S140: Perform data normalization. This is because the physical process parameter vector... The dataset contains physical quantities with varying dimensions and numerical values ​​of different orders of magnitude. Directly inputting these into subsequent models can lead to weight bias. Therefore, all physical parameters need to be mapped to a unified dimensionless interval. A max-min normalization method is used to map the data to the first... physical parameters The processing and calculation formula are as follows: ; in, These are the original physical parameter values. and These are the minimum and maximum values ​​of the parameter within historical statistical data or the rated range of the equipment, respectively. These are the normalized parameter values. After the above processing, a standardized physical process parameter vector is output. With the corresponding preprocessed image The data is stored in a high-speed cache queue for subsequent use by the physical constraint dynamic reference generation module 200.

[0030] For a detailed explanation of the implementation method and working principle of the physical constraint dynamic reference generation module 200, please refer to the appendix. Figure 3As shown. The physical constraint dynamic reference generation module 200 aims to establish a mapping model from the physical process space to the image space. By injecting real-time physical condition data into the image generation process, it generates a dynamic reference image that strictly corresponds to the current production conditions. Specifically, the construction, training, and operation process of the physical constraint dynamic reference generation module 200 includes the following steps: Step S210: Construct a model architecture based on Conditional Generative Adversarial Network (cGAN).

[0031] The model architecture consists of a generator. and discriminator Composition. The generator An encoder-decoder network architecture based on U-Net is adopted. Its input receives two parts of data: First, a pre-stored static standard template image serves as the basic structural skeleton for generation; Second, the normalized physical process parameter vector .

[0032] Within the network, the physical process parameter vector After mapping through fully connected layers, the features are reshaped into feature tensors with the same spatial resolution as the intermediate feature maps of the encoder, and then fused into the bottleneck layer or decoder layer of U-Net through channel-dimensional concatenation. The network includes downsampling and upsampling paths, with skip connections between corresponding layers of the two paths, directly passing the low-level texture features extracted by the encoder to the corresponding layer of the decoder to preserve the high-frequency texture details of the printed material. The discriminator... The PatchGAN architecture is employed. Instead of outputting a single true / false probability for the entire image, this discriminator divides the input image into... The output is a set of overlapping local tiles. A dimensional discrimination result matrix. Each element in the discrimination result matrix represents the probability that the corresponding patch is a true sample. This structure forces the model to focus on the high-frequency structure and consistency of local texture in the image.

[0033] Step S220: Define the objective function for model training.

[0034] The training process aims to optimize the parameters of the generator and discriminator, ensuring that the images generated by the generator approximate the distribution of real qualified product images and maintain pixel-level consistency. The total loss function... Conditional resistance to loss Compared with pixel-level reconstruction loss It is composed of weighted combinations, and its mathematical expression is: ; in, The hyperparameter for balancing weights, used to adjust the strength of pixel consistency constraints, is described in one embodiment. The value range is from 10 to 100. and These are the generator and the discriminator, respectively.

[0035] Step S230: Calculate the conditional adversarial loss. This loss function reflects the game between the generator and the discriminator. Its formula is defined as: ; In the formula, This represents the input physical process parameter vector. Indicates the relationship with this physical parameter Corresponding real qualified printed image, This represents a random noise vector that follows a standard normal distribution. Indicates that the generator is in condition The generated predicted image This indicates that the discriminator determines the real image. The probability of a result being true is calculated by taking the average of all elements in the discrimination result matrix. It represents the mathematical expectation.

[0036] Step S240, calculate reconstruction loss .

[0037] To reduce the deviation in color and brightness between the generated image and the real image, the following is introduced: Distance, as a constraint term, is defined by the following formula: ; In the formula, Represents generator of Reconstruction losses, Indicating the distribution of real image data Distribution of physical process parameters and random noise distribution Mathematical expectation on Represents a true, qualified printed image. The generator represents the physical process parameter vector. and random noise vector The generated predicted image, Representing a matrix Norm, which is the sum of the absolute values ​​of the differences between image pixels.

[0038] Step S250: Perform offline training of the model.

[0039] Historical datasets provided by the multimodal data synchronous acquisition and preprocessing module 100 The Adam optimizer is used to alternately update the weight parameters of the generator and discriminator. In a specific embodiment, the learning rate is set to 0.0002, and the momentum parameter... Set to 0.5. Set to 0.999. The training termination condition is that the loss function converges or reaches the preset number of iterations.

[0040] Step S260: Perform online dynamic reference generation.

[0041] During the system's online operation phase, a fixed generator is used. The weight parameters. For each time step Received real-time physical process parameter vector The image is input into the generator along with a static standard template image. After one forward propagation calculation, the corresponding dynamic reference image is output. This dynamic reference image It is then fed into the subsequent differential anomaly flow analysis module 300.

[0042] For a detailed implementation method and processing flow of the differential anomaly flow analysis module 300, please refer to the appendix. Figure 4 As shown. The core function of the differential anomaly flow analysis module 300 is to construct a data stream capable of characterizing the spatiotemporal evolution of anomaly signals and extract quantized dynamic features from it, thereby transforming static image defects into dynamic vectors with physical diagnostic significance. Specifically, the operation process of the differential anomaly flow analysis module 300 includes the following steps: Step S310: Generate a high signal-to-noise ratio differential image.

[0043] The differential anomaly flow analysis module 300 receives real-time images from the multimodal data synchronous acquisition and preprocessing module 100. and dynamic reference images from the physical constraint dynamic reference generation module 200. The differential anomaly flow analysis module 300 first performs pixel-by-pixel absolute difference operations to calculate the initial difference map. The calculation formula is as follows: ; In the formula, Represents the coordinates of the initial difference image in two-dimensional space. The pixel grayscale value at that location, Indicates at time Real-time images acquired in spatial coordinates The pixel grayscale value at that location, This indicates the generated dynamic reference image in spatial coordinates. The pixel grayscale value at that location, Represents the x and y coordinates of a pixel in an image. This represents the absolute value operation. To eliminate high-frequency noise caused by sensor thermal noise or minor alignment errors, the module then performs... Perform morphological opening. Morphological opening consists of two sub-steps: erosion followed by dilation. It removes isolated noise points smaller than a preset structuring element while preserving the shape features of the main anomaly regions. Subsequently, the system outputs two types of data: First, it preserves the net grayscale difference image of continuous pixel intensity values. Used for frequency domain analysis and optical flow calculation; Second, a binary mask image generated using Otsu's adaptive threshold segmentation. It is used for geometric morphology analysis.

[0044] Step S320: Construct the differential anomaly stream data cube. The system maintains a fixed-length time sliding window in memory, storing the grayscale net difference image sequence generated at consecutive time points. Stacked chronologically, they form a three-dimensional spatiotemporal data cube, namely, the differential anomaly stream. This represents the frame length of the sliding window. This data structure preserves the spatial location of the anomaly signal. and time dimension Complete evolutionary information on it.

[0045] Step S330: Extract the spatial frequency features of the abnormal signal.

[0046] To address the periodic defects commonly caused by rotating components in the printing process, the differential anomaly flow analysis module 300 performs spectral analysis on the differential anomaly flow. For regions of interest in the differential image where the accumulated energy exceeds a preset threshold, a one-dimensional signal sequence of the mean brightness changing over time is extracted. Subsequently, a Fast Fourier Transform (FFT) is performed on the sequence to obtain its frequency domain features. The formula for calculating the Discrete Fourier Transform is as follows: ; In the formula, Indicates frequency as The complex coefficients of the spectrum at time, Represents the first in a time series The signal value at each sampling point The total length of the sampling sequence. The imaginary unit, This is the base of the natural logarithm. The module calculates the spectral amplitude. The frequency component with the largest amplitude is taken as the dominant frequency characteristic of the abnormal signal. This feature is directly related to the rotation period of the physical components. It is the product of the time index and the frequency index.

[0047] Step S340: Extract the motion trajectory features of the abnormal signal.

[0048] For anomalies not located at fixed positions, the module uses optical flow to calculate the displacement vector of the anomaly region between consecutive frames. Based on the assumption of constant brightness, i.e., the grayscale value of the same object remains unchanged over a very short time, the optical flow constraint equation is established: ; In the formula, and These represent the spatial gradients of the grayscale net difference image in the horizontal and vertical directions, respectively. This represents the gradient of the grayscale net difference image over time. and These represent the instantaneous velocity components of the abnormal pixel in the horizontal and vertical directions, respectively. To solve the above equation containing two unknowns, the system introduces a local smoothing constraint, assuming a smoothing effect centered on the pixel. (For example Within the neighborhood window, optical flow velocity By keeping the equations constant, an overdetermined system of equations is established, and the velocity field vector of the pixel is obtained by solving it using the least squares method. The system then integrates the velocity field to obtain the motion trajectory features.

[0049] Step S350: Extract the morphological evolution features of the abnormal signal.

[0050] The differential anomaly flow analysis module 300 performs binarized mask images on each frame of the differential anomaly flow. Perform connected component analysis to calculate the geometric and topological properties of the outlier regions. Specific features calculated include the area of ​​the outlier regions. (Total number of pixels), centroid coordinates and firmness ; ; in, Let be the square of the region's perimeter. Then, calculate the derivatives of the aforementioned characteristic quantities over time, such as the area growth rate. This is used to characterize the dynamic severity of the anomaly. All features extracted in steps S330 to S350 above are concatenated into a comprehensive dynamic feature vector. The output is then sent to the subsequent causal inference module.

[0051] For a detailed implementation method and workflow of the knowledge base-based preliminary causal inference module 400, please refer to the appendix. Figure 5As shown. The knowledge-based preliminary causal inference module 400 constitutes the first link of the closed-loop diagnostic system. Its function is to use a pre-built deterministic physical model and expert experience to quickly match and attribute the dynamic feature vectors extracted by the preceding modules, thereby identifying known fault modes. Specifically, the operation process of the knowledge-based preliminary causal inference module 400 includes the following steps: Step S410: Construct the device physical fingerprint database.

[0052] The system establishes a relational database or lookup table containing the kinematic parameters of each key mechanical component of the printing equipment, called the equipment physical fingerprint database. This fingerprint database maps the physical topology of the equipment to a set of characteristic frequencies. For each rotating component, its theoretical characteristic frequency is calculated based on its mechanical dimensions and real-time transmission ratio. The calculation formula is as follows: ; In the formula, The theoretical characteristic frequency of a component is expressed in Hertz (Hz). This indicates the real-time linear velocity of the machine, collected by the encoder or sensor, in meters per second (m / s). This indicates the physical diameter of the component, in meters (m). Pi; This represents the mechanical coefficient for smooth roller components. The value is 1, for gear-type components. The value is taken as the number of teeth on the gear. This step ensures that the vibration or periodic disturbance frequency of each physical component at different vehicle speeds is calculable and known.

[0053] Step S420: Construct an initial causal knowledge graph.

[0054] The system uses a graph database to store the logical relationships between failure modes and root causes. The graph's data structure consists of a series of triples (subject, relation, object). Subjects and objects correspond to nodes in the graph, and relations correspond to edges between nodes. Node types include "evidence nodes" (corresponding to feature components in a dynamic feature vector) representing failure phenomena and "root cause nodes" representing physical components or parameters. Edge attributes include conditional probability weights, which are initialized based on the statistical frequency of historical maintenance records or expert experience to quantify the strength of causal relationships between nodes.

[0055] Step S430: Perform feature matching and candidate node activation.

[0056] The module receives dynamic feature vectors from the differential anomaly flow analysis module 300. Extract the main frequency features. The system traverses the device physical fingerprint database and calculates the measured main frequency. Theoretical characteristic frequencies of each component in the library Frequency matching error between The calculation formula is: ; In the formula, Indicates the first The absolute error of frequency matching of each component The measured dominant frequency extracted from the differential anomaly flow. For the first The theoretical characteristic frequency of each component at the current vehicle speed This represents absolute value operations. The system sets a frequency tolerance threshold. (e.g., 0.5Hz), if Then the corresponding first in the graph The root cause nodes of each component are marked as active and used as candidates for subsequent inference.

[0057] Step S440: Calculate the confidence level of causal inference.

[0058] Based on the activated candidate root cause nodes selected in step S430, the system uses Bayesian inference to calculate the posterior probability of the existence of a specific fault root cause. Posterior probability The calculation formula is as follows: ; In the formula, This indicates that evidence of fault characteristics has been observed. Under the condition of (i.e., successful frequency matching), the first Physical root cause The posterior probability of its validity; This represents the likelihood probability, i.e., when the root cause... Features are generated when they exist The probability, which is directly read from the root cause nodes in the knowledge graph. With evidence nodes The weight of the edge; Indicates root cause The prior probability is set based on the historical failure rate of the component; This represents the total number of currently activated candidate root cause nodes. This indicates evidence of observed fault characteristics. The total marginal probability.

[0059] Step S450: Execute threshold determination and traffic splitting strategy. The system sets a confidence threshold γ (e.g., 0.85). If the calculated highest confidence level... If the causal inference is successful, the system will directly output the root cause. The corresponding diagnostic conclusions and associated operational suggestions in the knowledge graph. If the confidence level of all candidate root causes is below the threshold... If no component meeting the tolerance threshold is found in step S430, it indicates that the current anomaly belongs to a complex or unknown fault mode not covered by the knowledge base, and the system will use the current dynamic feature vector. The candidate root cause list, along with other contextual data, is transmitted to the next-level active physics exploration module for further analysis.

[0060] For a detailed implementation and working principle of the perturbation-injection-based adaptive causal calibration module 500, please refer to the appendix. Figure 6 As shown. The perturbation-injection-based adaptive causal calibration module 500 is used to correct diagnostic results when the initial causal inference cannot determine the root cause of a fault, i.e., when facing complex operating conditions of "multiple causes for one effect" or "unknown faults." It actively applies a small probe signal to the physical system, observes the system's dynamic response, and thus uses physical-level causal verification to correct the diagnostic results. Specifically, the operation process of the perturbation-injection-based adaptive causal calibration module 500 includes the following steps: Step S510: Establish a safe and controlled industrial control system interaction interface.

[0061] The module establishes a bidirectional communication connection with the underlying programmable logic controller (PLC) of the printing equipment via industry-standard communication protocols (such as OPC-UA or Modbus-TCP). To ensure that the active exploration process does not affect the normal production safety of the equipment, the system predefines a multi-dimensional safe operating envelope. This envelope consists of the physical limit parameters of each actuator of the equipment, including the maximum allowable speed, upper and lower tension thresholds, and maximum ink layer thickness. Before performing any write operation, the system first reads the current operating state vector of the equipment, calculates the vector sum of the proposed disturbance command and the current state value, and obtains the predicted operating state value. The system compares this predicted operating state value with the safe operating envelope. Only when the predicted value falls completely within the envelope range is the command allowed to be issued; otherwise, it will be intercepted and an alarm will be triggered.

[0062] Step S520: Generate a small and controllable test disturbance signal. Based on the candidate root cause list output in step S450, the system generates a corresponding disturbance signal for each physical control parameter to be verified. This disturbance signal is defined as a short-duration, low-amplitude pulse or step change superimposed on the original setpoint. Specifically, the disturbance amplitude... Set to the current setting value to Furthermore, the output deviation value corresponding to this amplitude is less than the preset defective product judgment threshold for printed matter quality inspection, thereby ensuring that the detection process is an online non-destructive testing process. Disturbance duration It is set to cover at least one complete printing cycle (i.e., the time it takes for the roller to rotate once) to ensure that the sensor can capture the complete physical response.

[0063] Step S530: Collect dynamic response data of the system to disturbances.

[0064] Simultaneously with the injection of the disturbance signal, the module synchronously triggers the differential anomaly flow analysis module 300 to capture changes in image features during the duration of the disturbance. The system calculates the difference in dynamic feature vectors before and after the disturbance injection, denoted as the response vector. This response vector quantifies the specific impact of changes in particular physical parameters on the quality of the printed image.

[0065] Step S540: Construct and calculate the disturbance response matrix.

[0066] The system uses data from multiple injection experiments to construct a matrix that quantifies the sensitivity relationship between input parameters and output features, called the perturbation response matrix. Each element in the matrix Representing the The change of the first physical control parameter affects the... The sensitivity to the influence of each image feature component. Its calculation formula is as follows: ; In the formula, This represents the disturbance response sensitivity coefficient; and They represent the times at time 1 and 2 respectively. (Before the disturbance) and time (After the disturbance) the first The values ​​of image feature components, wherein the image feature components are selected from dynamic feature vectors. ; and They represent the times at time 1 and 2 respectively. and time No. The values ​​of each physical control parameter; denominator That is, the amplitude of the applied disturbance. .

[0067] Step S550: Perform causal attribution and model calibration.

[0068] The disturbance response matrix obtained by system traversal calculation If a certain element If the absolute value exceeds the preset causal association threshold, it indicates that the first... The physical parameter is the cause of the first The system identifies the main causes of various fault characteristics, thereby establishing a clear physical causal relationship between them. Based on this established causal relationship, the system corrects previously ambiguous or low-confidence candidate root causes to definitive conclusions, outputting the final fault diagnosis report. Simultaneously, the system feeds back this newly discovered strong correlation to the knowledge-based preliminary causal inference module 400, specifically by searching for the connection between the first and second fault characteristics in the causal knowledge graph. The root cause node and the first For each feature node, if the edge exists, its weight value is increased; if the edge does not exist, a new edge is created and given an initial high weight, thereby realizing the self-evolutionary learning of the diagnostic model.

[0069] For a detailed implementation method and workflow of the knowledge graph closed-loop self-evolution module 600, please refer to the appendix. Figure 7 As shown. The knowledge graph closed-loop self-evolution module 600, as the system's dynamic database management component, is responsible for transforming the verification results obtained by the perturbation-injection-based adaptive causal calibration module 500 during dynamic exploration into persistent causal knowledge, thereby achieving continuous accumulation and self-iteration of the system's diagnostic capabilities. Specifically, the operation process of the knowledge graph closed-loop self-evolution module 600 includes the following steps: Step S610: Receive and parse causal verification data. The module receives the perturbation response matrix analysis results from the adaptive causal calibration module via the internal data bus. The system parses these results and extracts several physically verified strongly correlated data pairs. Each data pair contains three core elements: the physical control parameter identifier of the perturbation object, the image feature component identifier of the response object, and the response coefficient value quantifying the sensitivity relationship between the two. The system marks these data pairs as causal knowledge entries to be updated, preparing them for graph operations.

[0070] Step S620: Perform weight reinforcement update of existing causal relationships.

[0071] For each causal knowledge entry to be updated, the system queries the current causal knowledge graph to determine if a directed edge connecting the physical control parameter node and the image feature component node exists. If the directed edge exists, it indicates that the causal relationship belongs to known knowledge. At this time, the system performs a weight enhancement operation. Specifically, the system reads the previously stored association probability weight value of the directed edge and multiplies the response coefficient value obtained in this experiment by a preset learning rate coefficient to obtain a weight increment value. The system adds this weight increment value to the original weight value to obtain the updated weight value. If the result of the sum is greater than a preset weight upper limit value (e.g., 1.0), the updated weight value is truncated to this upper limit value, and the result is written back to the graph database. This process, through a mathematical accumulation mechanism, enables the system to have a higher diagnostic confidence for repeatedly verified root causes of faults.

[0072] Step S630: Perform topological expansion of the new causal relationship.

[0073] If the query result in step S620 shows that there is no directed edge connecting the physical control parameter node and the image feature component node in the causal knowledge graph, or that there is no corresponding physical control parameter node, it indicates that the system has discovered a completely new fault mode that is not covered by pre-defined knowledge. At this point, the system performs a topology expansion operation. First, the system creates a root cause node representing the new physical control parameter in the graph database and assigns it a unique device address encoding attribute. Next, the system creates a directed edge from the root cause node to the corresponding image feature evidence node. Finally, the system standardizes the response coefficient values ​​obtained in this experiment, mapping them to the range of 0 to 1, and assigns the standardized values ​​as initial weights to the newly created directed edge. Through this step, the system's knowledge boundary is substantially expanded, achieving adaptive learning for unknown operating conditions.

[0074] Step S640: Perform decay and pruning of invalid relations.

[0075] To prevent the knowledge graph from containing outdated information over time, the system introduces a forgetting mechanism. When the system performs active perturbation exploration, if it finds that certain causal relationships recorded as high-weight in the graph do not produce the expected response in actual physical testing (i.e., the measured response coefficient value is lower than a preset invalidation threshold), the system will perform a linear decay operation on the weight of that directed edge, subtracting a preset forgetting factor value from the current weight. The system monitors the edge weight values ​​in real time. When the weight value of an edge is reduced to below a preset minimum retention threshold (e.g., 0.05), the system determines that the causal relationship is no longer valid or has become invalid, and then physically deletes the directed edge from the graph database.

[0076] Step S650: Full spectrum normalization and consistency check.

[0077] After completing the incremental update, the system performs a consistency check on the entire causal knowledge graph. The system traverses each evidence node, finds all incoming edges pointing to that node, and calculates the sum of the weights of these incoming edges. If the sum of weights is not equal to 1, the system performs a normalization calculation, that is, divides the current weight value of each edge by the sum of those weights to obtain the normalized new weight value. This step ensures the mathematical rigor of the probability values ​​in subsequent Bayesian inference calculations. The updated causal knowledge graph is persistently stored in non-volatile memory and synchronized in real-time with the preliminary causal inference module based on the knowledge base for the next fault diagnosis task.

[0078] In this embodiment, the system is deployed on a high-speed flexible packaging gravure printing production line to monitor and diagnose hidden quality defects in the printing process.

[0079] First, the multimodal data synchronous acquisition and preprocessing module 100 acquires image data of the printed image in real time. During production, the differential anomaly flow analysis module 300 detects a periodic transverse stripe defect that is difficult to detect with the naked eye in the generated net differential image by comparing the real-time image with the dynamic reference image from the physical constraint dynamic reference generation module 200. The differential anomaly flow analysis module 300 performs frequency domain analysis on the differential anomaly flow and extracts the dominant frequency characteristic of the defect signal as a specific value.

[0080] Next, the preliminary causal inference module 400, based on the knowledge base, matches the main frequency feature input to the device's physical fingerprint database. The query results show that the specific frequency falls within both the meshing frequency range of the printing plate cylinder drive gear and the rotational frequency range of the paper guide roller. Since the theoretical characteristic frequencies of these two components at the current speed are extremely close and both are within the system's frequency tolerance range, passive observation alone cannot distinguish whether the stripes are caused by gear wear or paper guide roller eccentricity. At this point, the system determines that a fuzzy diagnostic state of multiple causes and one effect has occurred.

[0081] Subsequently, the system automatically triggers the perturbation-injection-based adaptive causal calibration module 500. The perturbation-injection-based adaptive causal calibration module 500 first reads the current operating status of the equipment. After confirming the safe operating envelope, it sends a tiny speed step command to the servo motor of the guide roller. The tension fluctuation caused by the speed step command is controlled within a range that does not produce waste. At the instant of perturbation injection, the system detects that the intensity of the transverse stripes in the differential image does not change significantly, and the calculated perturbation response coefficient is extremely low. Immediately afterward, the perturbation-injection-based adaptive causal calibration module 500 injects a tiny phase lead command into the phase control unit of the printing plate cylinder. Monitoring results show that the amplitude of the transverse stripes increases significantly instantaneously, exhibiting extremely high perturbation response sensitivity.

[0082] Finally, based on the above physical-level verification results, the system ruled out the possibility of a guide roller malfunction and diagnosed the problem as an abnormal meshing clearance in the printing plate cylinder drive gear. The system output a diagnostic report, recommending that maintenance personnel check the gear clearance. Simultaneously, the knowledge graph closed-loop self-evolution module 600 automatically updated the background database, increasing the connection weight between the printing plate cylinder drive gear node and the specific frequency stripe feature node, and decreasing the association weight between the guide roller and this feature. Through this diagnosis, the system completed adaptive learning and knowledge solidification of the fuzzy fault under this specific operating condition.

Claims

1. A smart inspection and data traceability system for printed materials, characterized in that, The system includes: The multimodal data synchronous acquisition and preprocessing module is used to acquire physical process parameters and printed images synchronously with the printing process, and to preprocess them to output standardized physical process parameter vectors and preprocessed images. The physical constraint dynamic reference generation module is connected to the multimodal data synchronous acquisition and preprocessing module, and is used to receive the physical process parameter vector and generate the corresponding dynamic reference image based on the real-time physical process parameters. The differential anomaly flow analysis module is connected to the multimodal data synchronous acquisition and preprocessing module and the physical constraint dynamic reference generation module. It is used to purify the anomaly signal by performing differential operations on the real-time image and the dynamic reference image, and to analyze the continuous differential image sequence to extract dynamic features. The knowledge-based causal inference module is connected to the differential anomaly flow analysis module and is used to match the extracted dynamic features with the built-in knowledge base to infer the physical root cause of the anomaly. The perturbation-injection-based adaptive causal calibration module, connected to the knowledge-based causal inference module, is used to identify the causal relationship of the anomaly online by applying a small perturbation to the device and analyzing the response when encountering an unknown anomaly. The knowledge graph self-evolution module connects the perturbation-injection-based adaptive causal calibration module and the knowledge-based causal inference module, and is used to supplement the newly identified causal relationships into the knowledge base to realize the system's self-learning and capability evolution.

2. The intelligent inspection and data traceability system for printed matter quality according to claim 1, characterized in that, The multimodal data synchronous acquisition and preprocessing module is configured as follows: A trigger signal is generated by a synchronous controller connected to the main encoder of the printing press, which synchronously triggers the line scanning of the camera and the sampling action of the sensor; When there is a sampling rate mismatch, the physical process parameters acquired at adjacent time points are calculated by linear interpolation based on the image acquisition time. The maxima-minus normalization method is used to map the physical process parameters to a unified dimensionless interval.

3. The intelligent inspection and data traceability system for printed matter quality according to claim 1, characterized in that, The physical constraint dynamic reference generation module is built with a model architecture based on conditional generative adversarial networks, which consists of a generator and a discriminator. The generator adopts an encoder-decoder network architecture based on a U-shaped network. Its input receives a pre-stored static standard template image and a normalized physical process parameter vector. The physical process parameter vector is mapped by a fully connected layer and then fused into the bottleneck layer or decoder layer of the U-shaped network. The discriminator adopts the PatchGAN architecture, which divides the input image into multiple overlapping local patches and outputs a discrimination result matrix. All elements in the discrimination result matrix represent the probability that the corresponding patch is a real sample.

4. The intelligent inspection and data traceability system for printed matter quality according to claim 3, characterized in that, When training the conditional generative adversarial network, the physical constraint dynamic reference generation module defines the total loss function as a weighted combination of the conditional adversarial loss and the pixel-level reconstruction loss. The conditional adversarial loss reflects the game between the generator and the discriminator, and its calculation includes the probability that the discriminator determines the real image to be real and the probability that the discriminator determines the predicted image generated by the generator to be fake. The pixel-level reconstruction loss uses L1 distance as a constraint term to calculate the sum of the absolute values ​​of the pixel differences between the real image and the predicted image generated by the generator.

5. The intelligent inspection and data traceability system for printed matter quality according to claim 1, characterized in that, The differential anomaly flow analysis module is configured as follows: Perform a real-time pixel-by-pixel absolute difference operation between the printed image and the dynamic reference image, and perform a morphological opening operation on the result to generate a grayscale net difference image; The grayscale net difference image sequence generated at consecutive time points is stacked in chronological order to form a three-dimensional difference anomaly stream; A fast Fourier transform is performed on the one-dimensional signal sequence of the brightness mean of the region of interest in the differential anomaly stream as it changes over time, and the frequency component with the largest spectral amplitude is taken as the dominant frequency feature of the anomaly signal.

6. The intelligent inspection and data traceability system for printed matter quality according to claim 5, characterized in that, The differential anomaly flow analysis module is also configured to extract motion trajectory features using optical flow, establish optical flow constraint equations based on the assumption of constant brightness, including the spatial gradient of the grayscale net difference image in the horizontal direction, the spatial gradient in the vertical direction, and the gradient in the time dimension, and introduce local smoothing constraints to solve the velocity field vector of the abnormal pixel points using the least squares method.

7. The intelligent inspection and data traceability system for printed matter quality according to claim 1, characterized in that, The knowledge-based causal inference module constructs a device physical fingerprint database, which contains the theoretical characteristic frequencies of each rotating component. The theoretical characteristic frequencies are calculated based on the product of the real-time machine running linear velocity divided by pi, the physical diameter of the component, and the mechanical coefficient. The knowledge-based causal inference module calculates the frequency matching error between the measured main frequency feature and the theoretical feature frequency. When the frequency matching error is less than the frequency tolerance threshold, the corresponding candidate root cause node is activated, and Bayesian inference is used to calculate the posterior probability of the existence of a specific fault root cause based on prior probability and likelihood probability.

8. The intelligent inspection and data traceability system for printed matter quality according to claim 7, characterized in that, The perturbation-injection-based adaptive causal calibration module is configured as follows: A safe operating envelope consisting of the physical limit parameters of each actuator of the equipment is predefined, and the small disturbance is allowed to be issued only when the predicted operating state value falls completely within the safe operating envelope. A perturbation response matrix is ​​constructed, wherein the elements in the perturbation response matrix represent the sensitivity of the changes in physical control parameters to the image feature components. The perturbation response matrix is ​​calculated by dividing the difference between the dynamic feature vectors before and after the perturbation by the applied perturbation amplitude.

9. The intelligent inspection and data traceability system for printed matter quality according to claim 8, characterized in that, The knowledge graph self-evolution module is configured as follows: Receive the analysis results of the disturbance response matrix; If a directed edge already exists in the causal knowledge graph that connects the physical control parameter node and the image feature component node, then the original weight value of the directed edge is read and added to the product of the response coefficient value and the learning rate coefficient obtained in this experiment to perform weight reinforcement update. If the directed edge does not exist in the causal knowledge graph, a new root cause node and a directed edge are created, and the standardized response coefficient value is assigned as the initial weight to the newly created directed edge. If the measured response coefficient value is lower than the preset invalidation threshold, then the weight of the directed edge is subjected to a linear decay operation.

10. The intelligent inspection and data traceability system for printed matter quality according to claim 1, characterized in that, The physical process parameters include the ink roller speed, ink layer thickness setting value and ink viscosity of the ink supply system, the roller pressure and packing compression of the printing and imprinting unit, the hot air temperature and wind speed of the drying system, and the real-time operating speed and paper tension of the substrate transport system.

Images