AI optimization and defect real-time correction process of ink jet path

Abstract

The invention discloses an AI optimization and defect real-time correction process for an ink jet path, and the process comprises the steps: collecting a powder bed image in real time, inputting the powder bed image into a pre-trained weak supervision semantic segmentation neural network model, and precisely recognizing multiple types of pixel-level defects including wiredrawing, scratching, stacking, spheroidizing, layer shifting and warping. A subsequent ink jet path and process parameters are dynamically optimized according to defect characteristics, a correction instruction is issued in real time, and closed-loop control is formed; according to the method, a weak supervision learning framework based on image-level labels is adopted, so that the problem of refined labeling and recognition of complex defects is solved, meanwhile, an online correction system integrating perception, decision making and execution is constructed, and the precision and stability of the printing process and the forming quality of final parts are greatly improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent quality control technology, and in particular to an AI optimization and real-time defect correction process for inkjet paths. Background Technology

[0002] During the inkjet process, various factors such as nozzle condition, powder uniformity, binder-powder interaction, and process parameter matching can cause diverse forming defects on the powder bed surface. Typical defects include inkjet stringing caused by partial nozzle blockage leading to binder tailing, and scratches in inkjet / non-inkjet areas caused by improper contact between the powder spreading roller or scraper and the cured area. Currently, industry monitoring of these defects relies mainly on periodic offline sampling or online two-dimensional visual inspection based on simple thresholds.

[0003] First, existing technologies lack the ability to accurately and precisely identify numerous small and complex defects of various types and shapes in real time. Because defects vary greatly in texture, scale, contrast, and integration with the background (e.g., minor scratches versus obvious build-up), traditional detection methods based on fixed thresholds or shallow machine vision struggle to achieve precise pixel-level segmentation and classification, easily leading to missed detections or misjudgments, especially for defects with inconspicuous features such as "stringing" or "minor scratches." Second, there is a lack of a closed-loop intelligent decision-making and real-time correction link between defect identification results and the printing execution system. Even when a defect is detected, existing systems typically only issue an alarm or record it, failing to adaptively adjust the inkjet path, jetting parameters, or supplementary process actions based on the type, location, and severity of the defect. This allows defects to accumulate layer by layer once they occur, severely impacting the geometric accuracy and mechanical properties of the final part.

[0004] Therefore, in response to the problems mentioned above, this invention proposes an AI optimization and real-time defect correction process for inkjet paths. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies in accurately identifying complex and multi-category defects, and the disconnect between identification results and printing process control, which prevents real-time online correction, this invention proposes an AI optimization and real-time defect correction process for inkjet paths. This process introduces weakly supervised semantic segmentation technology based on image-level labels to achieve high-precision pixel-level identification and classification of various inkjet printing defects. It also constructs a mapping model between defect features and inkjet control strategies, enabling dynamic optimization of subsequent inkjet paths and process parameters based on real-time identification results, forming an intelligent correction closed loop integrating perception, decision-making, and execution.

[0006] The technical solution of this invention is: an AI optimization and real-time defect correction process for inkjet paths, comprising the following steps: S1, real-time acquisition of powder bed surface image of the current printed layer during binder jet additive manufacturing; S2, input the powder bed surface image into a pre-trained weakly supervised semantic segmentation neural network model, which outputs a pixel-level defect segmentation map, wherein the defects include at least the following: inkjet area - stringing, inkjet area - scratches, inkjet area - accumulation, inkjet area - spheroidization, non-inkjet area - scratches, non-inkjet area - accumulation, layer shift and warping. S3, based on the defect segmentation map, identify and locate the pixel regions, geometric shapes and severity of various defects; S4. Based on the type, location and severity of the defect, optimize the subsequent unexecuted inkjet path planning and generate an optimized inkjet control instruction set. This optimization includes adjusting the inkjet parameters, path sequence or compensation strategy of the defect area and its affected area. For inkjet area stringing defects, the optimization instructions include: pausing the current nozzle operation, triggering the cleaning pulse of the nozzle, and performing local path replanning and re-spraying in the defect area after cleaning; For inkjet areas with accumulation or balling defects, optimization instructions include: reducing the inkjet flow rate or jet frequency in the corresponding area, and adjusting the spacing between adjacent paths; For layer displacement defects, the optimization instructions include: calculating the interlayer misalignment based on the offset vector and performing global coordinate translation compensation on the subsequent inkjet path; For warpage defects, optimization instructions include: reducing inkjet volume at the edge of the warpage area and adjusting toner spreading or compaction parameters in adjacent non-inkjet areas; Furthermore, step S4 also includes establishing a defect impact propagation model, predicting the potential impact of the current defect on subsequent printing layers, and proactively adjusting the printing strategy for several subsequent layers based on the prediction results. S5 sends the optimized inkjet control instruction set to the inkjet printing control system to perform real-time inkjet correction for the current layer or subsequent layers. S6. Repeat steps S1 to S5 until the current layer is printed. Repeat the process when moving to the next printing layer. The pre-trained weakly supervised semantic segmentation neural network model mentioned in step S2 is obtained through training in the following manner: Collect a dataset of powder bed surface images containing various defect types, and label only the defect category at the image level; A weakly supervised learning framework based on image-level labels is adopted. A dual-domain pseudo-label mining module and a self-refinement strategy are used to generate pixel-level pseudo-labels from image-level labels. The dual-domain pseudo-label mining module includes a graph-based global inference unit and a salient region expansion controller. The pseudo-labels are used to train a semantic segmentation network, resulting in a model capable of pixel-level fine segmentation of various defects.

[0007] As a preferred embodiment, the weakly supervised semantic segmentation neural network model in step S2 further integrates a boundary feature transformation mechanism, a cross-pixel dependency module, a visual word encoding module, and a saliency mapping selection mechanism; The boundary feature transformation mechanism enhances the connection between pixels within the same defect region and weakens the connection between pixels in different regions during feature propagation by calculating the boundary confidence of pixels, thereby improving the segmentation accuracy of defect boundaries. The cross-pixel dependency module generates an attention map representing global context dependency by modeling the pairwise relationships of all feature nodes in the channel space, which is used to refine the initial defect response map. The visual word encoding module uses a learnable codebook to encode feature maps into visual words and generates fine-grained visual word labels to supervise the classification network and discover more semantically defective regions. The saliency mapping selection mechanism uses the saliency map provided by the pre-trained saliency detection model as pseudo-pixel level supervision. By calculating the overlap rate between the defect localization map and the saliency map, it adaptively distinguishes between foreground defects and background interference.

[0008] Preferably, the method for identifying the severity of defects in step S3 includes: calculating the area, aspect ratio, contrast and texture features of each segmented defective connected region, inputting the above features into a lightweight evaluation subnetwork, and outputting the severity level of the defect. The severity level is used to determine the radicalness of the corrective measures in step S4.

[0009] As an alternative, an offline reinforcement learning optimization step is also included: collecting defect occurrence data, corrective measures taken, and final printing quality results from multiple printing tasks, constructing a state-action-reward dataset, training a reinforcement learning agent whose state space includes current defect features, printing stage, and device state, and whose action space is a combination of parameters for various corrective strategies, and integrating the trained agent into the path optimization decision in step S4.

[0010] The beneficial effects of this invention are: 1. This invention solves the problems of high cost and insufficient recognition accuracy of fine labeling of multi-category defects in additive manufacturing by introducing weakly supervised semantic segmentation technology based on image-level labels. This invention only requires image-level defect category labels to train a model that can accurately segment and classify eight types of defects with different shapes, scales and contrasts, such as stringing, scratches, accumulation, and spheroidization in inkjet areas and scratches, accumulation, layer displacement, and warping in non-inkjet areas. Through mechanisms such as dual-domain pseudo-label mining, boundary feature transformation and visual word encoding, it enhances the model's ability to perceive and distinguish weak and ambiguous defects, thereby achieving high-precision and full-coverage recognition of complex defects.

[0011] 2. Based on the defect type, location, and severity assessed in real time through segmentation and quantification, this invention dynamically and adaptively optimizes subsequent inkjet paths, jetting parameters, and process actions (such as nozzle cleaning and local re-jetting for stringing, edge fading strategies for warping, and global coordinate compensation for layer shift). The optimization instructions are then sent to the control system for execution in real time, enabling online real-time correction. This proactively suppresses the generation and accumulation of defects, significantly improving the stability of the printing process and the final part forming quality. Attached Figure Description

[0012] Figure 1 The diagram shown is a schematic representation of the method flow of the present invention. Figure 2 The diagram shown is a schematic of the weakly supervised semantic segmentation neural network model structure of the present invention. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but 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.

[0014] Please see Figure 1 This invention provides an embodiment of an AI optimization and real-time defect correction process for inkjet paths, comprising: (1) First, an image acquisition system is used to acquire inkjet path data. This image acquisition system includes multiple sets of high-resolution industrial cameras integrated on the crossbeam of the toner spreading device or an independent high-precision linear motion module, and a coaxial light source or strip light source with uniform illumination. The camera is a high-speed global shutter CMOS sensor, which ensures that clear images without ghosting can be captured even during the movement of the toner spreading roller or scraper. The acquisition sequence is closely coupled with the printing process, including: At time point 1 (after powder spreading and before inkjet printing), a surface image of the current layer after powder spreading is acquired. This image serves as the "baseline image" to evaluate powder spreading uniformity and as a benchmark for subsequent differential analysis.

[0015] At time point 2 (after inkjet printing and before the next layer of toner), a surface image was captured after the current layer of binder was sprayed. This image directly reflects the effect of the inkjet printing operation and contains most of the defect information in the inkjet printing area.

[0016] Timing point 3 (optional) (specific node in key area): For complex structures or areas known to be prone to warping, image acquisition at a specific angle can be added during or after inkjet printing.

[0017] The acquired raw images are first subjected to standardization preprocessing, including non-uniform illumination correction (using white balance and background fitting subtraction), image denoising (using adaptive median filtering or wavelet thresholding for denoising), and image registration (ensuring pixel-level alignment of images acquired at different time points). The preprocessed image sequence is then fed into a weakly supervised semantic segmentation neural network model.

[0018] (2) Please refer to Figure 2 In actual printing production lines or experimental platforms, tens of thousands of images of powder bed surfaces containing defects of varying degrees and types are collected. Each image is labeled with an image-level defect category (e.g., labeling an image as containing "stringing" and "minor scratches"), eliminating the need for time-consuming and laborious pixel-level contour outlining. A weakly supervised semantic segmentation neural network model is used for training. This model uses a classic classification network as its backbone and integrates three key technologies: For the dual-domain pseudo-label mining module, the classification network first outputs an initial class activation map, but this map usually only highlights the most salient part of the defect (such as the center line of the wire or the highest point of the stack). The graph-based global inference unit maps the features from the spatial domain to the interactive graph domain, and aggregates global context information through graph convolution, enabling the network to notice non-salient defect areas adjacent to the salient region (such as the splashes at the edge of the wire or the faint traces around the scratch). The salient region expansion controller acts like a learnable valve, dynamically deciding how much to expand the initial activation region outward based on the learned feature distribution, so as to cover the complete defect region as much as possible and generate denser initial pseudo-labels. Then, a lightweight self-slimming network (composed of several convolutional layers) is trained with the above pseudo-labels as supervision. Its goal is to repair the missing defect parts in the pseudo-labels (such as incomplete spherical defects due to occlusion) and smooth noise, and finally produce high-quality pixel-level pseudo-labels for training the final segmentation network.

[0019] The boundary feature transformation mechanism and the cross-pixel dependency module are designed to further improve the accuracy of defect boundary localization, especially for defects with blurred boundaries such as "scratches" and "layer shifts." The boundary feature transformation mechanism operates in the intermediate feature layer of the network. It calculates the probability (boundary confidence) that each pixel belongs to the "boundary" based on features. As features propagate forward, high-confidence boundary pixels suppress the flow of information to different category regions while promoting information propagation within the same category region. This forces the network to learn more discriminative boundary features. The cross-pixel dependency module operates in the channel dimension, generating a global attention weight map by calculating the interrelationships of all feature locations (pixels) in the channel space. This weight map strengthens pixel responses that are consistent with the surrounding context and belong to true defects, while suppressing isolated noise responses, thereby improving the detection rate and segmentation consistency of weak defects (such as shallow scratches).

[0020] The visual word encoding module and saliency mapping selection mechanism address the problems of large intra-class differences (such as stacking of different shapes) and inter-class confusion (such as stacking of non-inkjet areas versus stacking of inkjet areas). The visual word encoding module maps feature map pixels to a learnable visual dictionary, assigning each pixel a "visual word" label. This fine-grained labeling provides the classification network with richer supervision signals than image-level labels, encouraging the network to discover and distinguish different visual patterns under the same category of defects. The saliency mapping selection mechanism introduces an external prior knowledge source: a saliency detection model pre-trained on general scene images. The saliency map generated by this model (highlighting the most salient areas in the image) is used as a reference. This invention uses this map as a benchmark and calculates the overlap rate between the various defect localization maps generated by the network and this saliency map. Defect localization maps with high overlap rates are considered "salient foreground defects" and are enhanced; those with low overlap rates may be classified as background or non-salient interference and are suppressed or merged with the background class, thereby helping the network better distinguish between genuine printing defects and background interference such as powder texture and shadows.

[0021] Using the pre-trained weakly supervised semantic segmentation neural network model, it can receive the pre-processed powder bed image and directly output a pixel-level label image with the same resolution as the input. Each pixel in the image is classified as either background or belongs to one of the aforementioned eight types of defects.

[0022] (3) After obtaining the defect segmentation map, the system performs quantitative analysis on each independent defect connected region. For basic geometric features, it calculates the area, perimeter, minimum bounding rectangle, eccentricity (aspect ratio) and centroid position of the defect region. For morphological and texture features, it extracts the Hu moments (describing shape), gray-level co-occurrence matrix (GLCM) features (describing texture roughness and contrast) of the defect region and the statistics (mean and variance) of pixel gray level / contrast within the region.

[0023] The extracted feature vectors are input into a pre-trained lightweight evaluation sub-network, which outputs a comprehensive "severity score" ("mild", "moderate", and "severe"). For example, for "fragmentation", the severity may be related to its length and width continuity; for "warping", it is related to its curvature and the height of its deviation from the reference plane.

[0024] This invention also establishes a defect impact propagation model, which predicts the potential impact of defects in the current layer on subsequent printed layers based on the type, location, size, and severity of the defects. For example, a severe "piling" defect may create local high points when laying the next layer of powder, leading to scraper collisions or uneven layer thickness; a "warping" defect may cause subsequent layers to have poor adhesion in that area, or even peel off.

[0025] (4) After the defect analysis is completed, the optimization engine immediately performs online replanning for the inkjet commands in the queue that have not yet been executed, specifically including: For inkjet area stringing defects, this invention locates the stringing defect and its associated nozzle number, and immediately sends an instruction to the control system: suspend the operation of the nozzle, execute a cleaning pulse of preset intensity (such as vacuum suction or vibration), and after cleaning, the system replans a local path according to the shape and position of the stringing area, and uses the nozzle or other backup nozzles to perform precise re-spraying, thereby filling the area of ​​material loss caused by stringing.

[0026] For inkjet regions with accumulation / spheroidization defects, if an area of ​​over-deposited material is identified, the optimization engine will reduce the inkjet flow rate in that area in subsequent paths (to reduce droplet volume) and / or increase the jetting dot spacing. At the same time, to avoid new problems caused by poor droplet fusion around the defect, the offset distance between adjacent scan lines may be fine-tuned.

[0027] For layer displacement defects, by comparing the timing point 1 (after the current layer is coated with powder) with the reference image of the previous layer, the overall offset vector (Δx, Δy) is calculated. This invention will perform global translation compensation (-Δx, -Δy) on the coordinates of all subsequent instructions in the inkjet job queue, and prompt the operator to check the mechanical status of the equipment on the control interface.

[0028] For warpage defects, the system identifies the warpage area and warpage direction (usually upward). In the warpage area, the system adopts an "edge fading" strategy, which involves gradually and linearly reducing the inkjet volume from the normal area to the warpage edge to reduce the warpage caused by the concentration of shrinkage stress. At the same time, the system can control the toner spreading system to slightly increase the pressure of the toner spreading roller in the non-inkjet area corresponding to the warpage, or trigger the local infrared heating unit to perform online annealing to relieve stress.

[0029] For scratch-type defects (inkjet area / non-inkjet area), the main optimization strategy is prevention. This invention first records the location where the scratch occurs. During the subsequent layer toner application process, the speed of the toner application roller is slightly increased or decreased above the area to change the contact force. At the same time, this location may be marked in the inkjet path planning to avoid the nozzle scanning getting too close.

[0030] Furthermore, based on the results of defect impact propagation prediction, this invention not only adjusts the current layer but may also modify the default parameters of subsequent layers. For example, if it is predicted that the "accumulation" of the current layer will affect the next layer, a thicker powder layer or a lower inkjet saturation can be set for the next layer in this area as compensation.

[0031] This invention also features a reinforcement learning agent that continuously collects printing logs from all networked devices, including original process parameters, defects encountered, corrective actions taken, and final part quality assessment results (such as dimensional accuracy, density, and strength). Through deep reinforcement learning algorithms, the agent learns which actions (parameter combinations of corrective strategies) should be taken to maximize long-term print quality rewards under different "states" (device state, defect combinations, printing stage). The trained policy model is periodically distributed to each terminal device and integrated into an online optimization engine, making its decisions more globally optimal.

[0032] Comparative Example 1 provided by the present invention: This comparative example verifies the present invention through comparative experiments. The experimental platform is a commercial adhesive jetting metal printer, and the printing material is 316L stainless steel powder. The comparative example is a prior art system that only has simple binary threshold detection (detecting obvious defects or large-area contamination).

[0033] (1) In this experiment, several defects were artificially introduced in single-layer printing, including slight nozzle clogging leading to short stringing, improper adjustment of the powder spreading roller causing scratches in non-inkjet areas, excessive local inkjet volume setting leading to accumulation, and slight layer shift caused by uneven powder spreading. The present invention and the comparative example were operated, and the specific results are shown in Table 1.

[0034] Table 1 Comparison of the effects of handling multiple types of defects.

[0035] As shown in the table above, in a scenario where four mixed defects—stringing, scratches, buildup, and layer displacement—are artificially introduced, the comparative model can only identify the most obvious buildup defect, with a defect detection rate as low as 25% and a false alarm rate of 15%. Furthermore, due to the lack of real-time correction capabilities, the surface roughness of this layer reaches as high as 18.7 μm, and the interlayer misalignment reaches 45 μm, requiring lengthy manual intervention. In contrast, this invention achieves a detection rate of up to 98.5% for various defects through a weakly supervised semantic segmentation model, reducing the false alarm rate to 3.2%. Based on the identification results, it automatically triggers targeted cleaning, respraying, parameter adjustment, and coordinate compensation, reducing the surface roughness of this layer by approximately 50% to 9.4 μm and the interlayer misalignment error by approximately 73% to 12 μm. The entire processing is fully automated, taking less than 5 seconds.

[0036] (2) This experiment continuously printed a cylindrical part with a high diameter-to-depth ratio to verify the optimization effect of the present invention on small and weak defects. The structure is prone to progressive warping at the edge, and intermittent micro-fibering and dot-like spherical formation may randomly occur during the process. The present invention and the comparative example were operated, and the specific results are shown in Table 2.

[0037] Table 2 Comparison of Progressive Defect Suppression Effects

[0038] As shown in the table above, during the continuous printing of parts prone to warping, existing technology systems cannot detect the initial minute warping and random stringing until defects accumulate and cause serious problems, resulting in a final height deviation of 0.82mm, a maximum warping deformation of 0.95mm, poor sidewall surface quality, and a printing success rate of only 40%. This invention, however, uses temporal differential analysis to capture early signs of warping and continuously applies dynamic edge fading strategies and real-time compensation, successfully controlling the final height deviation to 0.11mm, reducing warping deformation by 84% to 0.15mm, obtaining a smooth sidewall surface, increasing the printing success rate to 95%, and raising the tensile strength to 485±18MPa.

[0039] (3) This experiment ran the system of the present invention for a long time under different equipment and different batches of powder, and used the offline reinforcement learning module to collect data.

[0040] Experimental Process: Initially, the system mainly relied on preset rules for correction. After tens of thousands of virtual iterations and thousands of actual printing events, the reinforcement learning agent generated a new optimization strategy model. For example, the new strategy found that for a certain powder, when "spheroidization" occurs, slightly increasing the spray frequency instead of simply reducing the flow rate, combined with a slight oscillation of the scanning path, can achieve better surface smoothness.

[0041] Then, the old and new strategy models were deployed on the same device, and the same set of complex test cases were printed. The results are shown in Table 3.

[0042] Table 3 Comparison of strategies before and after reinforcement learning optimization

[0043] As shown in the table above, in long-term operation, compared with the initial strategy based on fixed rules, the reinforcement learning optimization strategy trained with massive amounts of data can make more intelligent decisions to correct actions, reducing the average defect rate per layer from 8.5% to 5.1%. At the same time, due to the improved strategy, unnecessary operations are reduced, the overall printing time efficiency is improved by 5%, and the average dimensional accuracy of parts is improved from ±52μm to ±38μm.

[0044] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. An AI optimization and real-time defect correction process for inkjet paths, characterized in that, It includes the following steps: S1, real-time acquisition of powder bed surface image of the current printed layer during binder jet additive manufacturing; S2, input the powder bed surface image into a pre-trained weakly supervised semantic segmentation neural network model, which outputs a pixel-level defect segmentation map, wherein the defects include at least inkjet area-stringing, inkjet area-abrasion, inkjet area-accumulation, inkjet area-spheroidization, non-inkjet area-abrasion, non-inkjet area-accumulation, layer displacement and warping. S3, based on the defect segmentation map, identify and locate the pixel regions, geometric shapes and severity of various defects; S4. Based on the type, location and severity of the defect, optimize the subsequent unexecuted inkjet path planning and generate an optimized inkjet control instruction set. This optimization includes adjusting the inkjet parameters, path sequence or compensation strategy of the defect area and its affected area. S5 sends the optimized inkjet control instruction set to the inkjet printing control system to perform real-time inkjet correction for the current layer or subsequent layers. S6. Repeat steps S1 to S5 until the current layer is printed. Repeat the process when moving to the next printing layer.

2. The AI ​​optimization and real-time defect correction process for inkjet paths according to claim 1, characterized in that, The training methods for the pre-trained weakly supervised semantic segmentation neural network model in step S2 include: A dataset of powder bed surface images containing various defect types was collected, and only image-level defect category labels were annotated. Then, a weakly supervised learning framework based on image-level labels was adopted. A dual-domain pseudo-label mining module and a self-refinement strategy were used to generate pixel-level pseudo-labels from the image-level labels. Finally, the pseudo-labels were used to train a semantic segmentation network to obtain a model that can perform pixel-level fine segmentation of various defects.

3. The AI ​​optimization and real-time defect correction process for inkjet paths according to claim 2, characterized in that, The dual-domain pseudo-label mining module includes: The graph-based global inference unit is used to project classification network features from the coordinate space to the interaction space and capture the relationships between feature nodes through graph convolution, thereby exploring non-significant defect regions. A salient region expansion controller is used to extract the maximum response from the back-projected feature map as a baseline and expand the coverage of salient defect regions through learnable control values ​​to generate a denser and more complete initial defect localization map.

4. The AI ​​optimization and real-time defect correction process for inkjet paths according to claim 1, characterized in that, The dynamic optimization of inkjet path planning in step S4 specifically includes: For inkjet area stringing defects, pause the current nozzle operation, trigger the cleaning pulse of the nozzle, and perform local path replanning and re-spraying in the defect area after cleaning; For inkjet zone deposits or spheroidization defects, reduce the inkjet flow rate or jet frequency in the corresponding area and adjust the spacing between adjacent paths to prevent further excessive material deposition. For layer displacement defects, the interlayer misalignment is calculated based on the offset vector, and global coordinate translation compensation is performed on all subsequent inkjet paths. For warping defects, reduce the inkjet volume at the edge of the warped area and adjust the powder spreading roller parameters or increase the compaction operation in the adjacent non-inkjet area. At the same time, local heating can be triggered to relieve stress.

5. The AI ​​optimization and real-time defect correction process for inkjet paths according to claim 4, characterized in that, Step S4 further includes: establishing a defect impact propagation model to predict the potential impact of the current defect on the structural strength and morphology of subsequent printed layers. Based on the prediction results, not only is the inkjet path of the current layer optimized, but the printing strategies of several subsequent layers are also proactively adjusted, including layer thickness, inkjet saturation, or support structure generation strategies.

6. The AI ​​optimization and real-time defect correction process for inkjet paths according to claim 1, characterized in that: The weakly supervised semantic segmentation neural network model in step S2 also integrates a boundary feature transformation mechanism and a cross-pixel dependency module. The boundary feature transformation mechanism enhances the connection of pixels within the same defect region and weakens the connection of pixels in different regions during feature propagation by calculating the boundary confidence of pixels, thereby improving the segmentation accuracy of defect boundaries. The cross-pixel dependency module generates an attention map representing global context dependency by modeling the pairwise relationship of all feature nodes in the channel space, which is used to refine the initial defect response map, especially improving the detection capability of weak and diffuse defects such as abrasions.

7. The AI ​​optimization and real-time defect correction process for inkjet paths according to claim 1, characterized in that: The weakly supervised semantic segmentation neural network model in step S2 also includes a visual word encoding module and a saliency map selection mechanism. The visual word encoding module uses a learnable codebook to encode feature maps into visual words and generates a fine-grained visual word label supervised classification network to discover more semantically defective regions. The saliency map selection mechanism uses the saliency map provided by the pre-trained saliency detection model as pseudo-pixel-level supervision. By calculating the overlap rate between the defect localization map and the saliency map, it adaptively distinguishes between foreground defects and background interference, effectively solving the visual confusion problem between non-inkjet area-pilling and inkjet area-pilling.

8. The AI ​​optimization and real-time defect correction process for inkjet paths according to claim 1, characterized in that, The method for identifying the severity of defects in step S3 includes: calculating the area, aspect ratio, contrast and texture features of each segmented defect connected region, and then inputting the above features into a lightweight evaluation subnetwork to output the severity level of the defect. The severity level is used to determine the aggressiveness of the corrective measures in step S4, including the amount of additional spray, cleaning intensity or path replanning range.

9. The AI ​​optimization and real-time defect correction process for inkjet paths according to claim 1, characterized in that, The method also collects defect occurrence data, corrective measures taken, and final print quality results from multiple printing tasks to construct a state-action-reward dataset. At the same time, a reinforcement learning agent is trained, whose state space includes the current defect features, printing stage, and device state, and whose action space is a combination of parameters for various correction strategies. Finally, the trained agent is integrated into the path optimization decision in step S4 to achieve an inkjet correction strategy based on historical experience and globally optimal.

10. The AI ​​optimization and real-time defect correction process for inkjet paths according to claim 1, characterized in that: The image acquisition in step S1 is achieved by a high-resolution linear or area array camera integrated on the powder spreading mechanism or an independent scanning arm. The acquisition timing includes at least after powder spreading, before inkjet printing, and after inkjet printing. By performing differential analysis on images at different times, the sensitivity and recognition accuracy of defects related to interlayer states, such as layer displacement and warping, are enhanced.

Images