A real-time printing defect detection method and system based on artificial intelligence

Abstract

The present application provides a real-time 3D printing defect detection method and system based on artificial intelligence, aiming to solve the problem of relying on single visual information, lacking dynamic tracking and insufficient model generalization in the prior art. The method realizes comprehensive perception of the printing process through multi-modal data fusion, and uses a deep learning model to analyze and classify defects in real time. The model output includes defect category, location, abnormal score and severity assessment, which can dynamically track the development trend of defects and provide early warning. By comparing with the printed digital model, the system can accurately identify geometric deviations and implement early intervention through trend analysis. When serious defects are detected, the system automatically triggers response measures such as pausing printing, adjusting parameters or providing correction schemes, forming a complete quality control closed loop. The present application significantly improves the detection accuracy and robustness, and supports incremental learning to adapt to different printing environments.

Description

Technical Field

[0001] This invention belongs to the field of 3D printing, specifically, it relates to a real-time printing defect detection method and system based on artificial intelligence. Background Technology

[0002] Traditional 3D printing defect monitoring relies primarily on manual observation and a limited number of sensors, which has many shortcomings. In traditional solutions, operators often observe the printing process through monitoring cameras or remotely view videos, but the human eye struggles to consistently monitor and capture subtle changes over extended periods. Furthermore, the simple sensors integrated into consumer printers (such as filament depletion switches) can only detect obvious faults and cannot effectively identify numerous, cumulative, and subtly changing quality issues. Early image-processing-based automatic detection methods required manual feature extraction, often only identifying defects of specific shapes, which is both cumbersome and lacks generalization ability.

[0003] To address the aforementioned issues, existing patents and products have incorporated artificial intelligence into 3D printing defect monitoring. Among them, Chinese patent CN202110337787A proposes a real-time online monitoring method based on neural networks. This method utilizes an industrial camera mounted above the printer's worktable to capture top-view images, converts them into grayscale images, and inputs them into a pre-trained convolutional neural network defect segmentation model. This model achieves pixel-level classification, determining each pixel in the image as background, normal, or one of several preset defect types (such as bubbling, collapse, wrinkling, missing filaments, etc.). The defect area is then calculated, and an alarm is triggered once it exceeds a preset threshold. This method automatically extracts features using deep learning, overcoming the limitations of manual feature extraction and achieving a certain level of real-time monitoring effectiveness.

[0004] However, this intelligent monitoring solution lacks dynamic tracking of defect evolution. Currently, it mainly relies on triggering alarms based on the defect area exceeding the limit in a single frame image, and cannot provide early warnings for gradual accumulation and development or early, minute changes.

[0005] Visual information alone is not comprehensive enough. Some faults may not be visually obvious in their early stages, but they may be accompanied by abnormal sounds, vibrations, or temperature changes. Relying solely on image monitoring will reduce the robustness of detection.

[0006] The model has limited generalization ability. The distribution of training data determines the model's adaptability; if the material color, printer model, or camera angle is changed, the model's detection accuracy may decrease. It also lacks automatic calibration and online learning capabilities. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide a real-time printing defect detection method and system based on artificial intelligence, which addresses the problems existing in the prior art.

[0008] The technical solution of this invention is: Firstly, an artificial intelligence-based real-time printing defect detection method is provided, comprising the following steps: Step S1: Acquire a real-time image sequence during the 3D printing process, and simultaneously acquire sound, infrared, and vibration data during the 3D printing process. Combine the real-time image sequence, sound, infrared, and vibration data into multimodal data. Step S2: Preprocess the multimodal data collected in step S2; and align and fuse the multimodal data collected from different sensors in time and space to form a multi-channel feature input synchronized with the printing process; Step S3: Input the preprocessed multimodal data into a pre-trained artificial intelligence defect detection model. The artificial intelligence defect detection model contains at least one deep neural network for real-time analysis of the printing process. The model output of the artificial intelligence defect detection model includes representation information of the current printing state, including defect segmentation results and / or defect category determination of the printing area, as well as anomaly degree evaluation index. Based on the model output, determine whether there is a printing defect and its type. If a defect is detected, obtain its position and area feature information in the current layer or current frame. If no defect is detected, continue to monitor subsequent printing data in a loop. Step S4: For the defects detected in step S3, the severity of the defects is assessed by combining multi-frame timing information and printed digital model data; Step S5: Compare the severity of the defect assessed in step S4 with a preset threshold. When the severity of the defect exceeds the safety threshold, immediately trigger the predetermined response measures. Step S6: After the printing job is completed, save the log of defects detected and handling decisions throughout the printing process. If parameter adjustment and correction measures are taken during the printing process, the relevant data will be fed back to the training module of the artificial intelligence defect detection model to perform incremental training or update of the artificial intelligence defect detection model.

[0009] Specifically, the assessment of the severity of the defect in step S4 includes: calculating the growth rate or cumulative impact based on the development trend of the defect in the continuous layer; determining the functional importance of the corresponding physical part of the defect in the finished product based on the G-Code or digital model provided by the printer; and quantifying the severity score of the defect by combining the defect area ratio and shape characteristics.

[0010] Specifically, the predetermined response measures in step S5 include: issuing an alarm to the operator to indicate the specific defect type and location, automatically pausing the printing job, and providing a correction plan or implementing a pre-set parameter adjustment strategy based on the defect type. Otherwise, the alarm will not be triggered and the subsequent printing process will continue to be monitored. After the defect response measures are implemented, the system returns to step S1 to continue monitoring until the printing job is completed.

[0011] Preferably, the artificial intelligence defect detection model is a multi-task neural network architecture that simultaneously outputs multiple judgment results related to printing quality: its first output is the category and location of the printing defect, and its second output is a score or probability of whether there is an anomaly in the printing process; wherein the predefined defect categories include at least stringing, warping, layer displacement, under-extrusion, over-extrusion, and uneven surface roughness, and the model provides an anomaly score for unknown types of abnormal situations while identifying any of the above defects.

[0012] Preferably, the defect determination in step S3 includes introducing a reference comparison of the printed digital model, obtaining the target cross-sectional contour data corresponding to the current printed layer, comparing the reference contour with the contour in the actual printed image; if it is found that the actual printed shape deviates from the reference model, it is determined that there is a printing defect, and the geometric error is quantified by calculating the offset distance and overlap between the reference contour and the contour in the actual image, and used as one of the bases for defect identification.

[0013] Preferably, the reference model comparison process includes: extracting the edge features of the printed layer in the printed image and matching them with the theoretical contour of the layer to detect displacement errors; when it is found that the actual printing height or position of a local area is lower or higher than the theoretical model, or the actual boundary exceeds the theoretical range, it is determined to be a corresponding defect and its position is recorded.

[0014] Preferably, the defect severity assessment in step S4 includes trend analysis of defect development over printing time; for defects identified in step S3, their area or feature changes are continuously monitored in each subsequent layer. When the area of ​​a defect at the same location gradually increases across n consecutive layers, or when any defect indicator shows a continuous upward trend and the increase exceeds a preset threshold, even if the current layer does not exceed the absolute alarm threshold, the defect is determined to have a serious trend in advance, thereby triggering an early warning.

[0015] Preferably, the response measures in step S5 include: automatically pausing the printer and locking the current state, sending an alarm message to the user terminal, and suggesting handling measures; when the defect type and cause are identifiable, the system provides a corresponding correction scheme, which includes adjusting the extruder temperature, reducing the printing speed, increasing the cooling fan speed, or modifying the printing path of the current layer; if the defect severity is below a first set threshold and is correctable, the system performs predetermined process parameter adjustments after pausing and then continues printing to attempt to repair the defect or prevent further deterioration; if the defect severity is higher than the set threshold and printing cannot continue, the printing job is terminated.

[0016] Preferably, the target cross-sectional contour data is derived from printer G-Code instructions or original CAD model slice data.

[0017] Secondly, an artificial intelligence-based real-time printing defect detection system is provided, the system comprising: The multimodal data acquisition module is used to acquire real-time image sequences during the 3D printing process, and simultaneously acquire sound, infrared, and vibration data during the 3D printing process to form multimodal data. The multimodal data preprocessing module is used to preprocess the multimodal data collected in step S1, and to align and fuse the multimodal data collected from different sensors in time and space to form a multi-channel feature input synchronized with the printing process. The defect detection module is used to input pre-processed multimodal data into a pre-trained artificial intelligence defect detection model. The artificial intelligence defect detection model contains at least one deep neural network for real-time analysis of the state during the printing process. The model output of the artificial intelligence defect detection model includes representation information of the current printing state, including defect segmentation results and / or defect category determination of the printing area, as well as anomaly degree evaluation indicators. Based on the model output, it is determined whether there are printing defects and their types. If a defect is detected, its position and area feature information in the current layer or current frame are obtained. If no defect is detected, the monitoring of subsequent printing data continues in a loop. The evaluation module is used to assess the severity of defects detected by the defect detection module by integrating multi-frame timing information and printed digital model data. The response module compares the severity of the defect assessed by the evaluation module with a preset threshold. When the severity of the defect exceeds the safety threshold, the predetermined response measures are triggered immediately. The parameter optimization module is used to save the log of defects detected during the entire printing process and the handling decisions after the printing job is completed. If parameter adjustment and correction measures are taken during the printing process, the relevant data will be fed back to the training module of the artificial intelligence defect detection model for incremental training or updating of the artificial intelligence defect detection model.

[0018] Compared to existing technologies, the advantages of this invention are as follows: This invention breaks through the limitations of traditional methods that rely solely on alarms for defect area exceeding limits in a single frame image. It employs continuous time-series data tracking, trend extrapolation, anomaly score accumulation, and multimodal data fusion to dynamically capture defects that accumulate layer by layer and undergo early, minute changes during the printing process. Through this mechanism, the system can not only detect the growth trend of defects when they first appear but also promptly issue warnings to the operator and suggest corresponding corrective measures to prevent further deterioration, thereby greatly improving printing quality and stability. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments or examples of this application, the accompanying drawings used in the embodiments or examples will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other drawings can be obtained according to these drawings without creative effort.

[0020] Figure 1 This is the overall flowchart of Embodiment 1 of the present invention; Figure 2 This is the main flowchart of Embodiment 1 of the present invention; Figure 3 This is a flowchart of step S6 in Embodiment 1 of the present invention; Figure 4 This is a system block diagram of Embodiment 2 of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of this application, but not all embodiments. Generally, the components of the embodiments of this application described and shown in the accompanying drawings can be arranged and designed in various different configurations.

[0022] Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0023] Example 1: refer to Figure 1 and Figure 3 A real-time printing defect detection method based on artificial intelligence includes the following steps: Step S1: Acquire a real-time image sequence during the 3D printing process, and simultaneously acquire sound, infrared, and vibration data during the 3D printing process. Combine the real-time image sequence, sound, infrared, and vibration data into multimodal data.

[0024] Before printing begins, the operator loads the STL or G-Code file of the model to be printed into the monitoring system's software interface. The system preprocesses the model, extracting the contour information of each layer for reference. After printing begins, a regular camera and an infrared camera capture images of the printing area at a frequency of approximately 2 frames per second. Due to the slow printing speed (each layer takes approximately tens of seconds), 2 FPS is sufficient to ensure that at least multiple images are captured for each layer. Simultaneously, the microphone continuously records the audio stream, which the system divides into short audio segments corresponding to the regular image frames and infrared graphic frames. As printing progresses, the data acquisition module continuously provides image frames and audio segments, adding timestamps to align the three.

[0025] This step increases the acquisition and utilization of sound data, expanding the dimensions of defect perception. Many printing faults are accompanied by unique sound characteristics; for example, a rhythmic "clicking" sound is produced when the extruder jams, and loose machine parts cause abnormal noises and vibrations. By integrating the audio collected by the microphone into the model after spectral analysis, the system can detect anomalies that are difficult to detect with images alone. Compared with solutions relying solely on vision, fusing sound signals significantly improves the sensitivity to dynamic faults such as poor feeding and vibration impacts. Once an abnormal sound pattern is detected, the system can issue an early warning before defects form on the surface, enhancing the timeliness and reliability of detection. This multimodal fusion solution fully utilizes the microphone sensing information that is easily obtained from existing equipment, making it highly practical. Infrared data increases the utilization of infrared thermal imaging. Anomalies in the temperature field during printing often indicate quality problems; for example, excessively rapid cooling in localized areas may lead to warping and cracking, while excessively high temperatures may cause material over-melting, collapse, or scorching and discoloration. By introducing infrared images, this method can monitor the temperature distribution of various parts of the printed part in real time and correlate thermal anomalies with geometric anomalies. For example, when an infrared image shows that the temperature of a certain area is significantly lower than the surrounding area and continues to drop, it can be inferred that the area may be warping (because it has detached from the heated bed); similarly, poor interlayer bonding can lead to localized overheating or undercooling spots. This information is difficult to obtain using only visible light images. By fusing thermal imaging and visible light, the system can perceive defects more comprehensively, especially providing predictive evidence for early warping and cracking, thus improving monitoring accuracy. Furthermore, the cost of infrared sensing hardware has decreased in recent years, making it affordable even for small consumer printers, thus making this improvement a practically feasible solution.

[0026] Step S2: Preprocess the multimodal data collected in step S1, and align and fuse the multimodal data collected from different sensors in time and space to form a multi-channel feature input synchronized with the printing process.

[0027] For each new image acquired, the system first performs distortion correction and noise reduction (using pre-calibrated camera parameters to eliminate lens distortion). Then, the image is scaled to the input size required by the neural network (e.g., 512). (512 pixels). For sound data, the system calculates its power spectral density within the corresponding time window and extracts features such as Mel frequency cepstral coefficients to obtain a feature vector representing the sound pattern at that moment. At this point, the system has processed the visual and sound data into a normalized form. Next, the system fuses these two features: specifically, the sound feature is added as an additional channel and concatenated at a certain position in the image feature tensor. That is, a multi-channel input is constructed, where the first three channels are the RGB (or grayscale) information of the image, and the fourth channel contains a matrix mapped to the same size as the image based on the sound features (for example, mapping the main frequency bands of sound energy to uniform values ​​on the image plane, or repeatedly filling in the sound features to form a feature map of the same size as the image). This simple fusion method is efficient in implementation, allowing the subsequent neural network to simultaneously "see" image and sound information. In this embodiment, the fourth channel is infrared intensity, the fifth channel is sound features, and so on. The fused multi-channel tensor is then sent to the defect detection module.

[0028] Step S3: Input the preprocessed multimodal data into a pre-trained artificial intelligence defect detection model. The artificial intelligence defect detection model contains at least one deep neural network for real-time analysis of the state during the printing process. The model output of the artificial intelligence defect detection model includes representation information of the current printing state, including defect segmentation results and / or defect category determination of the printing area, as well as anomaly degree evaluation index. Based on the model output, determine whether there is a printing defect and its type. If a defect is detected, obtain its position and area feature information in the current layer or current frame. If no defect is detected, continue to monitor subsequent printing data in a loop.

[0029] The defect detection module runs a convolutional neural network model. This model was trained during development using a large amount of printing process data, including normal printing and data from various defect scenarios. Part of the training data came from public datasets, and the other part was obtained by repeatedly printing "defective test pieces" using our experimental printer. These test pieces had their parameters intentionally adjusted to produce typical defects such as stringing, warping, and under-extrusion, thus providing labeled learning samples for the model. The model employs a structure combining U-Net and a multi-head classifier: the encoder extracts deep features from the fused image, the decoder outputs defect segmentation results pixel-by-pixel, and the global classification head outputs the defect category probability. The model also includes a parallel small multilayer perceptron used to predict the parameters of possible current anomalies based on encoder features. When the model in this embodiment receives new multi-channel data, it immediately performs forward inference to obtain the output result. During the first few layers of printing, the system did not detect any anomalies, the model output "no defects," and the anomaly score remained at a low level.

[0030] For the wire-drawing defect, for example: printing continues until the model begins printing two vertical columnar structures and needs to draw out the filament across the gap. At around layer 50, the camera view shows fine wires appearing between the columns. Our model successfully segmented and labeled these wire-drawing areas in this frame, classifying the defect as "wire-drawing". Specifically, the segmentation output shows the non-zero defect class at the location of the filament pixels between the columns, and assigns the category label "wire-drawing"; the probability of the "wire-drawing" category in the global classification head increases to 0.85, while the probabilities of other defect categories remain very low. The anomaly score also increases slightly, but remains at a moderate level. The model's cause prediction output at this point shows "possibly due to insufficient wire retraction distance or high nozzle temperature". Based on this, the system records the currently detected defect 1 – type: wire-drawing, location: between the two columns at the top of the model, cause: possibly insufficient wire retraction.

[0031] For warping defects, for example: when printing reaches around layer 100, the support structures at the four bottom corners are complete, and the main model begins to rapidly accumulate height. At this point, a corner of the model appears to be slightly warped and detached from the print bed in the camera image. In the layer 105 image, a shadow and gap appear at the contact point between the model's left front bottom corner and the print bed, which is identified as an abnormal area in the segmentation output. The model classifies this as a "warping" defect with a probability of approximately 0.9. The corresponding anomaly score also increases significantly, as warping often indicates serious quality problems. The system therefore records defect 2—type: warping, location: lower left front base edge of the model, area: approximately 20 square millimeters (obtained by converting pixel count to actual size). At this point, the global anomaly score is approaching the warning line, as warping could lead to misalignment in subsequent layers or even component detachment. The model's cause output indicates: "Possibly due to insufficient base plate temperature or excessively low ambient temperature." This analysis aligns with the actual situation: assuming that the cooling airflow in the printing environment was too strong, causing the left front area to over-cool and warp.

[0032] For under-extrusion defects, for example, near the completion stage of printing (approximately layer 180), the camera image shows an abnormal surface texture in a section of the top layer of the model: gaps appear between several adjacent lines, indicating incomplete filling. This phenomenon is identified by the model as an "under-extrusion" defect (i.e., insufficient material), with a probability of 0.8. Simultaneously, the microphone records intermittent "clicking" sounds from the extruder while printing this area, indicating gear slippage and poor filament feeding. The model, incorporating the sound signal, therefore has a higher anomaly score and stronger confidence in the under-extrusion judgment. The system records defect 3—Type: Localized under-extrusion, Location: Localized area of ​​the top layer, Area: Approximately 5 square millimeters. The cause analysis module suggests: "Possibly due to slight nozzle blockage or filament slippage."

[0033] This invention employs a multi-task output design, enabling the model to identify specific defect categories while simultaneously providing an overall anomaly score. This allows for both the classification of known defects and the alerting of unknown anomalies. On one hand, the enumerated typical defects (stringing, warping, layer displacement, under-extrusion, over-extrusion, surface defects, etc.) cover common problem types in consumer-grade 3D printing, ensuring the system can monitor issues of general user concern. On the other hand, the introduction of an anomaly scoring mechanism gives the system a degree of sensitivity to uncommon defects (new situations the model has not seen before), preventing complete misses. This is more flexible than existing technologies that rigidly judge based on fixed categories. Furthermore, pre-defining multiple defect output categories facilitates explaining the specific problem to the user in alert prompts, improving the user-friendliness of the human-computer interaction.

[0034] Step S4: For the defects detected in step S3, the severity of the defects is assessed by combining multi-frame timing information and printed digital model data; Our severity assessment strategy involves observing defect growth trends and the importance of their location.

[0035] For stringing defects, for example, the system calculated the length and number of strings at layer 50 and continued monitoring the same area in each subsequent layer. It was found that fine strings still appeared at this location during layer 51 and 52 printing, with lengths similar to previous layers and no significant increasing trend. Combined with the model's digital reference, there were no design entities between the columns, so these strings did not affect the design geometry. The severity assessment module determined that although stringing existed, its size was within acceptable limits and had not increased. Its impact on the final model was mainly on appearance and aesthetics, with no functional impact; therefore, the severity was classified as "minor." This level is below the warning threshold, so the system does not trigger a pause but only records and marks the stringing defect for later processing. The system highlights the stringing area between the columns on the 3D view of the model in the user interface and indicates that this is a minor defect and does not require interruption. Simultaneously, the model cause analysis suggests checking the string retraction settings. The operator receives a prompt and can choose to adjust parameters (such as increasing the string retraction distance by 1mm) to reduce stringing in subsequent printing. However, this embodiment assumes that the operator does not intervene and printing continues.

[0036] For warpage defects, for example: the system initiates a rigorous evaluation strategy for warpage defects. First, a comparison with the reference model reveals that the corner should be flush with the platform, but the actual height deviates by approximately 2mm. Second, the trend is examined: from layer 100 to layer 105, the images show that the warped area is gradually expanding, and the gap grows from almost invisible initially to clearly visible, indicating that the warpage is rapidly worsening. Without intervention, subsequent layers may experience layer displacement due to suspension, or the nozzle may collide with the warped portion. Based on these factors, the severity is rated as "severe," far exceeding the threshold, requiring immediate action.

[0037] For under-extrusion defects, such as this top-layer under-extrusion defect, the system assesses its impact. Since printing is nearing completion, only a few layers have a small amount of voids, which has little impact on the overall structural strength, but damages surface quality. Model reference comparison shows that this area should be solidly filled, but there are actually about 20% voids. In terms of trends, under-extrusion has only appeared in 1-2 layers and has not yet worsened further; however, judging from the extruder noise, clogging may intensify. If more parts continue to be printed, this clogging could completely stop the material flow. Taking all factors into account, the system rates the severity as "moderate." While it's not enough to scrap a part, measures need to be taken to ensure that subsequent printing jobs do not experience more serious problems.

[0038] This invention utilizes a digital model as a "standard answer" to precisely pinpoint the location and extent of printing errors. For example, if the actual printed outline of a layer is missing a piece compared to the theoretical model, it immediately indicates a material shortage or layer loss; if the actual outline shifts outward, it may indicate over-extruded burrs or filaments. This layer-by-layer comparison not only detects geometric deviations in the model but also identifies the type of deviation. This is more deterministic than relying solely on image feature classification, reducing false positives and false negatives. For instance, in cases involving complex part surface textures, directly examining the image model may make it difficult to determine if details are missing, but with the digital model as a reference, "parts that shouldn't be there or shouldn't be missing" can be clearly identified. Therefore, this improvement significantly enhances the accuracy and scope of defect detection, particularly for monitoring dimensional accuracy and shape integrity, expanding the system's application scenarios (such as quality assurance for printing high-precision functional components).

[0039] Step S5: Compare the severity of the defect assessed in step S4 with a preset threshold. When the severity of the defect exceeds the safety threshold, immediately trigger the predetermined response measures.

[0040] Step S6: After the printing job is completed, save the log of defects detected and handling decisions throughout the printing process. If parameter adjustment and correction measures are taken during the printing process, the relevant data will be fed back to the training module of the artificial intelligence defect detection model to perform incremental training or update of the artificial intelligence defect detection model.

[0041] The system stores the monitoring data from this printing operation (including images, sensor data, three detected defects, and corresponding handling measures) in a log database. Researchers can choose to use this new data for model retraining. For example, the detected "stringing" images and their final processing markers (minor, no need to stop) can be used to enhance the model's ability to distinguish the severity of stringing; "warping" related data can be used to train the model to capture warping signs earlier; and sound and image data from "under-extrusion" can be used for supervised learning to more accurately correlate sound features with material feeding anomalies. As data from more printing tasks accumulates, the model will continuously optimize, and its adaptability to various equipment and materials will improve. This adaptive optimization mechanism reflects the continuous improvement capability of this invention, enabling it to gradually learn to identify unknown new printing faults, significantly outperforming traditional solutions that require manual rule adjustments.

[0042] By employing the above methods, this invention overcomes the limitations of traditional methods that rely solely on alarms for defect area exceeding limits in a single frame image. It utilizes multiple techniques, including continuous time-series data tracking, trend extrapolation, anomaly score accumulation, and multimodal data fusion, to dynamically capture defects that accumulate layer by layer and undergo subtle early changes during the printing process. Through this mechanism, the system can not only detect the growth trend of defects when they first appear but also promptly issue warnings to the operator and suggest corresponding corrective measures to prevent further deterioration, thereby significantly improving print quality and stability.

[0043] Specifically, the assessment of the severity of the defect in step S4 includes: calculating the growth rate or cumulative impact based on the development trend of the defect in the continuous layer; determining the functional importance of the corresponding physical part of the defect in the finished product based on the G-Code or digital model provided by the printer; and quantifying the severity score of the defect by combining the defect area ratio and shape characteristics.

[0044] Specifically, the predetermined response measures in step S5 include: issuing an alarm to the operator to indicate the specific defect type and location, automatically pausing the printing job, and providing a correction plan or implementing a pre-set parameter adjustment strategy based on the defect type. Otherwise, the alarm will not be triggered and the subsequent printing process will continue to be monitored. After the defect response measures are implemented, the system returns to step S1 to continue monitoring until the printing job is completed.

[0045] Preferably, the artificial intelligence defect detection model is a multi-task neural network architecture that simultaneously outputs multiple judgment results related to printing quality; its first output is the category and location of the printing defect, and its second output is a score or probability of whether there is an anomaly in the printing process; wherein the predefined defect categories include at least stringing, warping, layer displacement, under-extrusion, over-extrusion, and uneven surface roughness, and the model provides an anomaly score for unknown types of abnormal situations while identifying any of the above defects.

[0046] Preferably, the defect determination in step S3 includes introducing a reference comparison of the printed digital model, obtaining the target cross-sectional contour data corresponding to the current printed layer, comparing the reference contour with the contour in the actual printed image; if it is found that the actual printed shape deviates from the reference model, it is determined that there is a printing defect, and the geometric error is quantified by calculating the offset distance and overlap between the reference contour and the contour in the actual image, and used as one of the bases for defect identification.

[0047] This invention proposes a method to accurately measure layer-to-layer positional errors (misalignment) and height errors (missing layers or bulges) by extracting the deviation between image edges and theoretical contours. For example, if the height of a certain area in the actual print is detected to be lower than the expected value, it can be determined that a collapse or missing layer has occurred; if the height exceeds the expected value, it indicates a local bulge or excessive material ejection. Similarly, actual boundaries exceeding design boundaries indicate problems such as filament tailing or wire adhesion. These specific rules make the system's judgment of defects more explicit and reliable, equivalent to establishing identification criteria for each geometric deviation. Such multi-rule fusion detection has stronger interpretability and accuracy compared to purely end-to-end networks, helping to reduce potential misjudgments by deep learning models. Furthermore, the method of using G-Code templates to locate deviations is universal and applicable to parts of different shapes, because the comparison benchmark comes from the digital design of each part itself, thus improving the system's adaptability to diverse printing tasks.

[0048] Preferably, the reference model comparison process includes: extracting the edge features of the printed layer in the printed image and matching them with the theoretical contour of the layer to detect displacement errors; when it is found that the actual printing height or position of a local area is lower or higher than the theoretical model, or the actual boundary exceeds the theoretical range, it is determined to be a corresponding defect and its position is recorded.

[0049] Preferably, the defect severity assessment in step S4 includes trend analysis of defect development over printing time; for defects identified in step S3, their area or feature changes are continuously monitored in each subsequent layer. When the area of ​​a defect at the same location gradually increases across n consecutive layers, or when any defect indicator shows a continuous upward trend and the increase exceeds a preset threshold, even if the current layer does not exceed the absolute alarm threshold, the defect is determined to have a serious trend in advance, thereby triggering an early warning.

[0050] This invention identifies progressively worsening potential problems by recording the area changes of the same defect across consecutive layers and the abnormal score changes in the model output. For example, if a slight warping increases by 10% per layer, the system can determine that it will eventually develop into a serious defect before reaching the absolute alarm threshold, thus intervening early. This trend extrapolation method effectively compensates for the lag in fixed threshold judgments, preventing small defects from accumulating into large problems before being discovered. Compared to the original technology's mechanism of only alarming when a defect exceeds the threshold once, this improvement provides a more sensitive "early warning" function, significantly reducing scrap and material waste. Simultaneously, combining trend information can filter out false positive alarms caused by transient noise—if a small defect is detected in a frame but does not subsequently expand, the system can determine its limited impact and not trigger a stop. This time-series-based analysis reflects the application of systems thinking to quality monitoring, making monitoring decisions more robust and intelligent.

[0051] Preferably, the response measures in step S5 include: automatically pausing the printer and locking the current state, sending an alarm message to the user terminal, and suggesting handling measures; when the defect type and cause are identifiable, the system provides a corresponding correction scheme, which includes adjusting the extruder temperature, reducing the printing speed, increasing the cooling fan speed, or modifying the printing path of the current layer; if the defect severity is below a first set threshold and is correctable, the system performs predetermined process parameter adjustments after pausing and then continues printing to attempt to repair the defect or prevent further deterioration; if the defect severity is higher than the set threshold and printing cannot continue, the printing job is terminated.

[0052] Preferably, the target cross-sectional contour data is derived from printer G-Code instructions or original CAD model slice data.

[0053] Through the above implementation examples, the workflow and effectiveness of this invention have been fully verified. Without altering the existing printer structure, this invention successfully monitors and handles various defects during the printing process using simple additional devices such as cameras and microphones, along with intelligent algorithms. From minor stringing defects to severe warping, the system provides timely and accurate identification and responses. Compared to manual visual inspection or single-sensor alarms, the intelligent monitoring system significantly improves the comprehensiveness and timeliness of defect detection; while the automatic pause and parameter adjustment functions automate steps that previously relied on manual intervention, reducing waste and labor costs. This real-time printing defect detection method based on artificial intelligence has advantages such as a wide detection range, fast response speed, low deployment cost, and continuous learning and optimization, making it particularly suitable for large-scale application in the consumer 3D printing field. Through this invention, ordinary users' 3D printers can also possess the ability to "self-monitor" and correct errors, thereby significantly improving printing success rates and product quality, providing strong support for the popularization of additive manufacturing technology.

[0054] This invention truly forms a complete quality control closed loop. When a serious defect is detected, automatic printing pauses to avoid continuing to manufacture defective products, wasting materials, and even preventing equipment damage (e.g., printhead impacting a warped component). It provides users with specific defect type and location information, facilitating rapid problem identification. In existing technologies, operators need to manually adjust parameters such as temperature based on alarms, while this solution pre-maps defect types to remedial measures, automatically or semi-automatically performing adjustments. For example, if under-extrusion is detected, the printing speed can be automatically reduced or the extrusion temperature increased to compensate; if severe stringing is detected, the string retraction distance can be increased or the nozzle temperature reduced. In many cases, this corrects problems without manual intervention, salvaging failed prints. Even if manual intervention is required (e.g., cleaning nozzle blockages), timely shutdown reduces subsequent cleaning difficulty and material waste. In summary, this improvement upgrades the traditional monitoring system to an intelligent control system, maximizing the value of defect detection information, and is particularly practical for unattended or large-scale printing scenarios.

[0055] Implementation Example 2: refer to Figure 4 A real-time printing defect detection system based on artificial intelligence is provided, the system comprising: The multimodal data acquisition module is used to acquire real-time image sequences during the 3D printing process, and simultaneously acquire sound, infrared and vibration data during the 3D printing process, and combine the real-time image sequences, sound, infrared and vibration data into multimodal data; The multimodal data preprocessing module is used to preprocess the multimodal data collected in step S1, and to align and fuse the multimodal data collected from different sensors in time and space to form a multi-channel feature input synchronized with the printing process. The defect detection module is used to input pre-processed multimodal data into a pre-trained artificial intelligence defect detection model. The artificial intelligence defect detection model contains at least one deep neural network for real-time analysis of the state during the printing process. The model output of the artificial intelligence defect detection model includes representation information of the current printing state, including defect segmentation results and / or defect category determination of the printing area, as well as anomaly degree evaluation indicators. Based on the model output, it is determined whether there are printing defects and their types. If a defect is detected, its position and area feature information in the current layer or current frame are obtained. If no defect is detected, the monitoring of subsequent printing data continues in a loop. The evaluation module is used to assess the severity of defects detected by the defect detection module by integrating multi-frame timing information and printed digital model data. The response module compares the severity of the defect assessed by the evaluation module with a preset threshold. When the severity of the defect exceeds the safety threshold, the predetermined response measures are triggered immediately. The parameter optimization module is used to save the log of defects detected during the entire printing process and the handling decisions after the printing job is completed. If parameter adjustment and correction measures are taken during the printing process, the relevant data will be fed back to the training module of the artificial intelligence defect detection model for incremental training or updating of the artificial intelligence defect detection model.

[0056] The embodiments described above are merely preferred embodiments of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications, improvements, and substitutions without departing from the concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A real-time printing defect detection method based on artificial intelligence, characterized in that, Includes the following steps: Step S1: Acquire a real-time image sequence during the 3D printing process, and simultaneously acquire sound, infrared, and vibration data during the 3D printing process. Combine the real-time image sequence, sound, infrared, and vibration data into multimodal data. Step S2: Preprocess the multimodal data collected in step S1, and align and fuse the multimodal data collected from different sensors in time and space to form a multi-channel feature input synchronized with the printing process; Step S3: Input the preprocessed multimodal data into the pre-trained artificial intelligence defect detection model. The artificial intelligence defect detection model contains at least one deep neural network for real-time analysis of the state during the printing process. The model output of the artificial intelligence defect detection model includes the representation information of the current printing state, which includes the defect segmentation results and / or defect category determination of the printing area, as well as the anomaly degree evaluation index. Based on the model output, determine whether there are printing defects and their types. If a defect is detected, obtain its position and area features in the current layer or frame. If no defect is detected, continue monitoring subsequent print data in a loop; Step S4: For the defects detected in step S3, the severity of the defects is assessed by combining multi-frame timing information and printed digital model data; Step S5: Compare the severity of the defect assessed in step S4 with a preset threshold. When the severity of the defect exceeds the safety threshold, immediately trigger the predetermined response measures. Step S6: After the printing job is completed, save the log of defects detected and handling decisions throughout the printing process. If parameter adjustment and correction measures are taken during the printing process, the relevant data will be fed back to the training module of the artificial intelligence defect detection model to perform incremental training or update of the artificial intelligence defect detection model. The assessment of the severity of the defect in step S4 includes: The growth rate or cumulative impact is calculated based on the development trend of defects in continuous layers. The functional importance of the corresponding physical part of the defect in the finished product is determined based on the G-Code or printed digital model provided by the printer. The severity score of the defect is quantified by combining the defect area ratio and shape characteristics.

2. The real-time printing defect detection method based on artificial intelligence according to claim 1, characterized in that, The predetermined response measures in step S5 include: The system sends an alarm to the operator indicating the specific type and location of the defect, automatically pauses the printing job, and provides a corrective solution or executes a pre-set parameter adjustment strategy based on the defect type. Otherwise, it does not trigger an alarm but continues to monitor the subsequent printing process. After the defect response measures are implemented, the system returns to step S1 to continue monitoring until the printing job is completed.

3. The real-time printing defect detection method based on artificial intelligence according to claim 1, characterized in that, The artificial intelligence defect detection model is a multi-task neural network architecture that simultaneously outputs multiple judgment results related to printing quality. Its first output is the type and location of the printing defect, and the second output is the score or probability of whether there is an anomaly in the printing process. The predefined defect categories include at least stringing, warping, layer displacement, under-extrusion, over-extrusion, and uneven surface roughness. When the artificial intelligence defect detection model identifies any of the above defects, it gives an anomaly score for unknown types of abnormal situations.

4. The real-time printing defect detection method based on artificial intelligence according to claim 3, characterized in that, The defect category determination in step S3 includes introducing a reference comparison of the printed digital model, obtaining the target cross-sectional contour data corresponding to the current printed layer, and comparing the reference contour with the contour in the actual printed image. If the actual printed shape is found to deviate from the reference model, a printing defect is determined. The geometric error is quantified by calculating the offset distance and overlap between the reference contour and the contour in the actual image, and this is used as one of the bases for defect identification.

5. The real-time printing defect detection method based on artificial intelligence according to claim 4, characterized in that, The comparison process of the reference model includes: Extract the edge features of the printed layer in the printed image and match them with the theoretical contour of the layer to detect displacement error; When it is found that the actual printed height or position of a local area is lower or higher than the theoretical model, or that the actual boundary exceeds the theoretical range, it is judged as a corresponding defect and its position is recorded.

6. The real-time printing defect detection method based on artificial intelligence according to claim 4, characterized in that, The assessment of the severity of the defect in step S4 includes: Analysis of the trend of defects over printing time; For the defects identified in step S3, their area or feature changes are continuously monitored in each subsequent layer; When the area of ​​a defect at the same location gradually increases across n consecutive layers, or when any defect indicator shows a continuous upward trend and the increase exceeds a preset threshold, even if the current layer does not exceed the absolute alarm threshold, the defect is determined to have a serious trend in advance, thereby triggering an early warning.

7. The real-time printing defect detection method based on artificial intelligence according to claim 4, characterized in that, The response measures in step S5 include: Automatically pause the printer and lock the current status, send an alarm message to the user terminal and suggest appropriate actions; When the defect type and cause are identifiable, the system provides a corresponding correction scheme, which includes adjusting the extruder temperature, reducing the printing speed, increasing the cooling fan speed, or modifying the printing path of the current layer. If the defect severity is below a first set threshold and is correctable, the system will perform a predetermined process parameter adjustment after pausing and then continue printing in an attempt to repair the defect or prevent it from worsening. If the severity of the defect exceeds the set threshold and cannot continue, the printing job will be terminated.

8. A real-time printing defect detection system based on artificial intelligence, characterized in that, The system is used to perform the method as described in claim 1, the system comprising: The multimodal data acquisition module is used to acquire real-time image sequences during the 3D printing process, and simultaneously acquire sound, infrared and vibration data during the 3D printing process, and combine the real-time image sequences, sound, infrared and vibration data into multimodal data; The multimodal data preprocessing module is used to preprocess the multimodal data collected in step S1, and to align and fuse the multimodal data collected from different sensors in time and space to form a multi-channel feature input synchronized with the printing process. The defect detection module is used to input pre-processed multimodal data into a pre-trained artificial intelligence defect detection model. The artificial intelligence defect detection model contains at least one deep neural network for real-time analysis of the state during the printing process. The model output of the artificial intelligence defect detection model includes representation information of the current printing state, including defect segmentation results and / or defect category determination of the printing area, as well as anomaly degree evaluation indicators. Based on the model output, it is determined whether there are printing defects and their types. If a defect is detected, its position and area feature information in the current layer or current frame are obtained. If no defect is detected, the monitoring of subsequent printing data continues in a loop. The evaluation module is used to assess the severity of defects detected by the defect detection module by integrating multi-frame timing information and printed digital model data. The response module compares the severity of the defect assessed by the evaluation module with a preset threshold. When the severity of the defect exceeds the safety threshold, the predetermined response measures are triggered immediately. The parameter optimization module is used to save the log of defects detected during the entire printing process and the handling decisions after the printing job is completed. If parameter adjustment and correction measures are taken during the printing process, the relevant data will be fed back to the training module of the artificial intelligence defect detection model for incremental training or updating of the artificial intelligence defect detection model.

Images