Method for detecting 3D printing precision of oral cavity model based on image recognition

Abstract

The invention relates to a method for detecting 3D printing precision of an oral cavity model based on image recognition. Comprising the following steps that multi-source heterogeneous data of an oral cavity model are obtained, a multi-dimensional data vector is formed, and the multi-source heterogeneous data comprise image information, equipment parameters, process parameters, environment and material and design data; performing precision detection and feature extraction on the oral cavity model based on image recognition, the feature extraction being used for extracting a deviation value of precision detection, and adding the deviation value to the multi-dimensional data vector in S1 to form a multi-dimensional data vector set; performing association feature mining on the multi-dimensional data vector set, and outputting a core association feature vector; performing anomaly diagnosis of improved fuzzy clustering on the core association feature vector to obtain a key parameter causing the occurrence of a deviation value; and generating an intelligent diagnosis report: optimizing the process according to the key parameters.

Description

Technical Field

[0001] This invention relates to the field of 3D printing quality inspection technology, and in particular to a method for detecting the accuracy of 3D printing of oral cavity models using image recognition. Background Technology

[0002] With the widespread application of 3D printing technology in the field of oral healthcare, such as the fabrication of dental crowns, bridges, implant guides, and orthodontic models, the dimensional accuracy of the printed models directly affects the success or failure of subsequent medical procedures. Currently, the accuracy inspection of printed dental models largely relies on manual sampling measurements using tools such as calipers and 3D scanners. Manual measurement methods suffer from inefficiency, susceptibility to subjective errors, inability to achieve full inspection, and difficulty in measuring complex curved surface features. While high-precision 3D scanners offer accurate measurements, their high equipment cost, complex operation, and time-consuming data processing make rapid, batch testing difficult in production lines or clinical settings.

[0003] Therefore, there is an urgent need in this field for a fast, accurate, low-cost, and automated full-inspection solution for the accuracy testing of 3D printed dental models. Existing patent application number CN202411039333.5 discloses a method for detecting the accuracy of 3D printed dental devices based on image recognition, relating to the fields of image processing and image recognition technology. The specific steps of this method are as follows: S100, Device Image Acquisition: Using a high-resolution camera or scanner, images of the 3D printed dental device are acquired from multiple different perspectives (such as front, side, and top), ensuring that the images from each perspective clearly reflect the surface details and structural features of the device. Through multi-scale image analysis, various details from macro to micro can be captured simultaneously, including overall shape, size, surface roughness, and edge sharpness, thereby achieving a comprehensive evaluation of the 3D printing accuracy of dental devices. Furthermore, multi-view image fusion technology can overcome the limitations of a single perspective, ensuring that every detail of the device can be observed from multiple angles, improving the accuracy and reliability of the detection. However, it can only provide a binary judgment of "pass / fail" or dimensional deviation data, but cannot further reveal the underlying causes of the accuracy abnormalities, such as systematic drift of the 3D printing equipment, instability of material batches, or improper setting of slicing parameters. The lack of in-depth correlation analysis of manufacturing process data keeps quality control at the level of post-inspection, making it difficult to achieve predictive maintenance and proactive optimization of process parameters.

[0004] This invention provides an automated detection method for the accuracy of 3D printing of oral models based on image recognition. The method acquires multi-angle two-dimensional images of the oral model under a standard light source, uses advanced image recognition and deep learning algorithms to extract key dimensional features, and compares them with the theoretical dimensions of the original digital model, thereby achieving fast and high-precision non-contact automated detection. At the same time, by integrating printing process data and visual detection results, the method deeply explores the root causes of accuracy abnormalities, achieving a leap from "detection" to "diagnosis". Summary of the Invention

[0005] This invention addresses the technical problems existing in the prior art by providing a method for detecting the accuracy of 3D printing of oral cavity models based on image recognition.

[0006] The technical solution of this invention to solve the above-mentioned technical problems is as follows: a method for detecting the accuracy of 3D printing of oral cavity models based on image recognition; comprising the following steps: S1: Acquire multi-source heterogeneous data of the oral cavity model and form a multi-dimensional data vector. The multi-source heterogeneous data includes: image information, equipment parameters, process parameters, environmental and material data, and design data. S2: Based on image recognition, the oral cavity model is subjected to accuracy detection and feature extraction. The feature extraction is used to extract the deviation value of the accuracy detection and add it to the multidimensional data vector in S1 to form a multidimensional data vector set. S3: Perform correlation feature mining on a multidimensional data vector set and output the core correlation feature vector; S4: Improve the anomaly diagnosis by performing fuzzy clustering on the core associated feature vectors to obtain the key parameters that cause the deviation values; S5: Intelligent diagnostic report generation: Optimizes the process based on key parameters.

[0007] Furthermore, the step of acquiring multi-source heterogeneous data of the oral cavity model and forming a multi-dimensional data vector includes: S11: Standard Image Acquisition: Automated multi-angle image acquisition of oral cavity models; S12: Manufacturing process data extraction: Automatically capture the full process parameters bound to the printing task ID through the API interface or log file of the 3D printer control system. S13: Design Data Synchronization: Export the original design contour data of the oral cavity model from the slicing software or obtain it by parsing G-code, that is, the set of theoretical two-dimensional polygon coordinates for each layer.

[0008] Furthermore, the step of performing accuracy detection and feature extraction on the oral cavity model based on image recognition includes: S21: Perform grayscale conversion, Gaussian filtering for noise reduction, and adaptive histogram equalization with contrast limiting on the acquired image to enhance feature edges; calibrate the camera using a standard calibration board of known size to obtain the accurate conversion relationship between pixels and physical size; S22: Use a pre-trained deep learning model to segment the image, automatically identify and label key regions; use an iterative nearest-point algorithm to compare the identified feature contours with the theoretical design contours obtained in S13 to eliminate the overall deviation caused by the placement angle. S23: Based on the registration, the system automatically measures a series of predefined key dimensions and generates dimensional deviation values; S24: The size deviation vector calculated in S23 is used as a new, core data column and associated with the multi-source heterogeneous data collected in S1 to form a multi-dimensional data vector set for subsequent analysis.

[0009] Furthermore, the process of performing correlation feature mining on the multidimensional data vector set, and outputting the core correlation feature vector, includes: S31: Construct a multidimensional data association feature distribution function, and use the autocorrelation function combined with data similarity to process the attribute distribution of the multidimensional data vector set; ; in, Let K represent the distribution function of multidimensional data association features, where K represents the autocorrelation of features, n represents the dimension level of data features, and i represents the number of data nodes. S32: Introducing an adaptive weighted learning mechanism, which dynamically adjusts the weights of each data dimension based on its importance in the feature space, so that the feature distribution function can more accurately fit the true distribution of the data. ; in, Indicates the first The adaptive weighted learning output of the next iteration Indicates data weights, Indicates the number of iterations; S33: Extract association features and combine them with association rules to extract core association features that can characterize the system state from the weighted learning data. ; in, This represents the extracted multidimensional data association features. Represents the data distribution function. This represents the initial weight of the data, and m represents the node connection probability.

[0010] Furthermore, the step of improving fuzzy clustering of the core associated feature vectors for anomaly diagnosis to obtain the key parameters leading to the deviation values ​​includes: in order to accurately locate the abnormal batches and diagnose the root cause, the associated feature vectors extracted in step S3 are... The data is fed into an improved fuzzy clustering algorithm to highlight outliers in multidimensional data. S41: Calculate the inter-class distance and determine the average distance between the cluster centers and the remaining data in the multidimensional dataset; ; in, This represents the average distance between the cluster centers and the remaining data in the multidimensional dataset. and Represents a random point. This represents the total number of clusters. and Represents a set of random numbers; S42: Calculate the average distance within a class, cluster the features of all association rules, and obtain the average distance between all feature data; ; in, This represents the average distance between all feature data. This represents the association rule, where D represents the similarity. S43: Generate feature clustering results, and obtain a quantitative clustering effectiveness index by comparing two average distances; ; Where Y represents the feature clustering result; the better the clustering effect of the data, the closer the value of Y is to 1; conversely, the worse the clustering effect of the data, the closer the value of Y is to -1.

[0011] Furthermore, the intelligent diagnostic report generation in S5 specifically includes: S51: Root cause tracing and visualization. Based on the key parameters diagnosed by S4, the system automatically retrieves the historical best value, process allowable range and current batch value of the parameter from the multi-source heterogeneous database obtained by S1, and presents them in the report in a highlighted contrast form. S52: Optimization suggestion knowledge base matching. The report generation module has an embedded optimization knowledge base containing expert rules and historical cases. Based on the diagnosed abnormal patterns, it automatically matches and outputs one or more specific and actionable process parameter adjustment suggestions. S53: Report integration and output. The final diagnostic report is a structured document that includes at least a summary table of accuracy test results, a heatmap of key dimension deviations, a root cause parameter analysis diagram, and optimization suggestions in text form. It also supports one-click export.

[0012] Furthermore, after generating the feature clustering results, the method further includes: Anomaly pattern matching and diagnosis: The calculated feature clustering result Y is matched with a pre-established anomaly pattern feature library; the anomaly pattern feature library stores the typical Y value range and key parameter features corresponding to different process faults; if the current Y value falls within the typical range of a certain anomaly pattern, it is determined that the current printing task is affected by the anomaly pattern, and the key process parameters that caused the pattern are output as the final diagnostic result.

[0013] Furthermore, the detection method also includes: S6: Process parameter self-calibration and feedback; S61: Parameter adjustment calculation: Based on the optimization suggestions in the S5 diagnostic report, the system automatically calculates the specific adjustment amount required for key process parameters according to the built-in process parameter-accuracy deviation response model. S62: Command Issuance and Verification: Automatically sends the adjusted set of process parameters to the 3D printing control system for the next printing task of the same digital model; S63: Closed-loop verification learning: The system records the printing tasks and corresponding accuracy detection results before and after parameter adjustment, forming a decision-result feedback pair, which is used to continuously update and optimize the response model in S61.

[0014] The beneficial effects of this invention are: Automated 3D scanning enables 100% comprehensive inspection of every printed part. This not only ensures "zero defects" at the factory, but more importantly, it yields massive amounts of complete full-size data, providing a solid data foundation for subsequent root cause analysis and process optimization. The root cause analysis engine, powered by AI and big data, can instantly and intelligently correlate detected dimensional deviations and defects with hundreds of upstream process parameters (such as laser power, scanning speed, and atmosphere concentration) to accurately pinpoint the root cause of the problem. This enables quality issues to be explained, traceable, and attributable, shifting quality control from post-inspection to in-process early warning and pre-inspection prevention.

[0015] After diagnosing the root cause, the system can automatically adjust equipment parameters through a closed-loop feedback system for compensation and optimization. This is like an "autopilot" system installed on a 3D printer, which can automatically find the optimal process window, significantly improve the first-piece success rate, shorten the new product development cycle, and ensure high consistency in mass production. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a method for detecting the accuracy of 3D printing of oral cavity models using image recognition. Detailed Implementation

[0017] 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 only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0019] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0020] In one embodiment, a method for detecting the accuracy of 3D printing of an oral cavity model based on image recognition includes the following steps: S1: Acquire multi-source heterogeneous data of the oral cavity model and form a multi-dimensional data vector. The multi-source heterogeneous data includes: image information, equipment parameters, process parameters, environmental and material data, and design data. S2: Based on image recognition, the oral cavity model is subjected to accuracy detection and feature extraction. The feature extraction is used to extract the deviation value of the accuracy detection and add it to the multidimensional data vector in S1 to form a multidimensional data vector set. S3: Perform correlation feature mining on a multidimensional data vector set and output the core correlation feature vector; S4: Improve the anomaly diagnosis by performing fuzzy clustering on the core associated feature vectors to obtain the key parameters that cause the deviation values; S5: Intelligent diagnostic report generation: Optimizes the process based on key parameters.

[0021] In this embodiment, in step S1, Image information: High-resolution multi-angle images of the model acquired through an automated vision system. This is the direct basis for accuracy testing.

[0022] Equipment parameters: The inherent attributes of a 3D printer, such as printer model, laser model, and optical system precision.

[0023] Process parameters: Variable parameters set for this printing task, such as layer thickness, laser power, scanning speed, exposure time, and support structure. This is the main focus of optimization.

[0024] Environment and Materials: Ambient temperature and humidity during printing, and the type and batch of photosensitive resin used. Material properties and environmental stability have a significant impact on molding quality.

[0025] Design data: The original CAD model or the G-code after slicing represents the "theoretically perfect model" and serves as the benchmark for accuracy comparison.

[0026] Multidimensional data vector: After normalizing and standardizing all the above data, a feature vector is constructed for a printing task. For example: [Task ID, Image Feature 1, Image Feature 2, ..., Laser Power, Scanning Speed, Layer Thickness, Resin Batch, Ambient Temperature, ..., Design Profile Data Feature 1, ...]; This step transforms "heterogeneous" data of different dimensions and types into "homogeneous" digital vectors that can be uniformly processed by computers.

[0027] In step S2: the print quality is quantitatively evaluated, and the evaluation results are fed back into the dataset as new features; image preprocessing: the acquired images are subjected to operations such as denoising, contrast enhancement, and scale normalization.

[0028] Feature alignment and registration: Accurately align the captured model image with the theoretical contour from the design data.

[0029] Deviation calculation: Through pixel-level comparison, calculate the errors in key dimensions (such as crown margin width, cusp height, interproximal gap, etc.) and morphological errors (such as whether there is excess resin on the surface or whether there are missing structures).

[0030] The calculated deviation values ​​(such as "edge deviation_0.12mm" and "height deviation_-0.05mm") are used as new features. These deviation values ​​are added to the multidimensional data vector corresponding to this task in S1, forming a richer multidimensional data vector set containing both "cause" (process parameters) and "effect" (accuracy deviation). This dataset becomes a treasure trove containing a large number of "problem cases" and "success cases" for subsequent mining and analysis.

[0031] In step S3: From the massive, mixed data, identify which process parameters, equipment settings, or environmental factors are most relevant to specific accuracy deviations; the core correlation feature vector is a filtered subset containing only those key parameters strongly correlated with accuracy deviations; for example, analysis might find that the four parameters [laser power, scanning speed, layer thickness, and resin batch] explain over 90% of the "edge sharpness deviation," while other parameters (such as ambient temperature) have minimal impact. Therefore, the core feature vector consists of these four parameters.

[0032] In step S4: the problem is finely categorized to precisely pinpoint the specific parameter combinations causing the anomaly. Anomaly diagnosis process: Use the core correlation feature vector (containing only key parameters and corresponding deviation values) output by S3 as input.

[0033] The algorithm automatically clusters the dataset into several typical "failure modes". For example: Cluster1 (Insufficient Power Type): Characterized by low laser power, accompanied by deviations such as missing details and incomplete shape.

[0034] Cluster2 (overexposure type): characterized by high exposure time, accompanied by edge swelling, burrs and deviation.

[0035] Cluster3 (Material Instability Type): Characterized by specific resin batches exhibiting deviations such as surface roughness and poor uniformity.

[0036] When a new model is identified as having a bias, the system calculates its membership degree to each failure mode. The mode with the highest membership degree indicates the key parameter that caused the bias.

[0037] In step S5, the system automatically generates a structured report, which includes: Problem description: Specific deviations detected (e.g., the width of the near-mid surface edge exceeds the standard by 0.15mm).

[0038] Root cause diagnosis: The deviation was determined to be "Cluster2 (overexposure type)", and the key parameter was excessive exposure time.

[0039] Direct suggestion: "It is recommended to adjust the exposure time from the current 12 seconds to 10-11 seconds."

[0040] Parameter linkage suggestion: "If reducing the exposure time results in insufficient adhesion of the substrate, consider increasing the number of exposures for the substrate by 2."

[0041] Specifically, the step of acquiring multi-source heterogeneous data of the oral cavity model and forming a multi-dimensional data vector includes: S11: Standard Image Acquisition: Automated multi-angle image acquisition of oral cavity models; S12: Manufacturing process data extraction: Automatically capture the full process parameters bound to the printing task ID through the API interface or log file of the 3D printer control system. S13: Design Data Synchronization: Export the original design contour data of the oral cavity model from the slicing software or obtain it by parsing G-code, that is, the set of theoretical two-dimensional polygon coordinates for each layer.

[0042] Specifically, the step of performing accuracy detection and feature extraction on the oral cavity model based on image recognition includes: S21: Perform grayscale conversion, Gaussian filtering for noise reduction, and adaptive histogram equalization with contrast limiting on the acquired image to enhance feature edges; calibrate the camera using a standard calibration board of known size to obtain the accurate conversion relationship between pixels and physical size; S22: Use a pre-trained deep learning model to segment the image, automatically identify and label key regions; use an iterative nearest-point algorithm to compare the identified feature contours with the theoretical design contours obtained in S13 to eliminate the overall deviation caused by the placement angle. S23: Based on the registration, the system automatically measures a series of predefined key dimensions and generates dimensional deviation values; S24: The size deviation vector calculated in S23 is used as a new, core data column and associated with the multi-source heterogeneous data collected in S1 to form a multi-dimensional data vector set for subsequent analysis.

[0043] In this embodiment, the pre-trained deep learning model segmentation uses a semantic segmentation model (such as U-Net) trained on a large number of oral cavity model images.

[0044] For example, by inputting the preprocessed image into the model, the model will output a label for each pixel, such as: 0=background, 1=crown, 2=gingiva, 3=supporting structure. In this way, we automatically obtain the precise pixel-level contour of the crown region in the image.

[0045] Iterative Closest Point (ICP) Algorithm Registration: The design profile obtained from S13 is in "theoretical alignment," while the model photographed in S11 is randomly positioned. Direct comparison is meaningless. The ICP algorithm continuously rotates and translates the actual profile to make it coincide with the theoretical profile as much as possible, thereby eliminating the deviation in the overall placement posture.

[0046] For example, a dental crown that should be placed vertically is tilted at 15 degrees in a photograph. The ICP algorithm will automatically calculate a rotation of approximately -15 degrees and a corresponding translation to align the two contours. In this way, the remaining deviation is the actual manufacturing error, not the measurement error.

[0047] In the registered and aligned coordinate system, the specific dimensional error is quantitatively measured: Predefined critical dimensions: Based on the clinical requirements of the oral model, a series of dimensions that need to be measured are defined in advance.

[0048] For example: Mesiodistal diameter: The distance between the mesial and distal points of a tooth crown.

[0049] Buccal-lingual diameter: The distance between the most prominent point on the buccal side of the crown and the most prominent point on the lingual side.

[0050] Crown height: The distance from the gingival margin of the crown to the apex of the tooth.

[0051] Interproximal space: the shortest distance between two adjacent tooth crowns.

[0052] Automatic measurement and deviation calculation: The system automatically locates the measurement points on the registered theoretical and actual contours and performs calculations.

[0053] For example, the theoretical mesial-to-distal diameter is 8.0 mm, while the distance measured by the system on the actual profile is 7.86 mm. Therefore, the dimensional deviation = 7.86 - 8.00 = -0.14 mm. A negative value indicates that the model has shrunk.

[0054] By linking the "results" (accuracy deviations) with the "causes" (process parameters), a dataset that can be used for machine learning is formed.

[0055] Now, take all the deviation values ​​calculated by S23 (such as [-0.14, +0.05, -0.22, ...]) as a new data column and concatenate them into the vector formed by S1.

[0056] Final vector example: [Task ID: Print_12345, Mesial-to-distal diameter deviation: -0.14, Bucto-lingual diameter deviation: +0.05, Crown height deviation: -0.22,..., Slice thickness: 50, Laser power: 300, Scanning speed: 2.0, Resin batch: B20231201A,...] Each row in this vector set is a complete "experimental record," containing process inputs and precision outputs.

[0057] Specifically, the process of performing correlation feature mining on the multidimensional data vector set, and outputting the core correlation feature vector, includes: S31: Construct a multidimensional data association feature distribution function, and use the autocorrelation function combined with data similarity to process the attribute distribution of the multidimensional data vector set; Principle: Use a mathematical method (distribution function) to describe and rank the importance of all process parameters (features). It determines whether a parameter is "critical" by analyzing its own fluctuation pattern (autocorrelation) and its relationship with other parameters.

[0058] ; in, Let K represent the distribution function of multidimensional data association features, where K represents the autocorrelation of features, n represents the dimension level of data features, and i represents the number of data nodes. Work process and examples: Input: Raw data from 100 printing jobs, including 20 parameters: [power, speed, layer thickness, chamber temperature, resin temperature, squeegee speed, ...] and corresponding dimensional deviations.

[0059] Processing: The distribution function f calculates a score for each parameter.

[0060] It was found that the autocorrelation K of "laser power" showed that the actual power value of 5 batches suddenly decreased by 10%, and the deviation values ​​of these 5 batches were all abnormally high. Therefore, "laser power" would get a very high f value.

[0061] It found that the "resin batch" parameter, after a specific batch (node ​​i) was changed, caused a slight but consistent shift in the mean deviation of all subsequent prints. Therefore, the "resin batch" parameter would also receive a high f value.

[0062] It found that the "scraper speed" varied very little across the entire dataset and had no significant relationship with the bias, so its f-value would be very low.

[0063] Output: A list of parameters sorted from highest to lowest based on their f-values. The top-ranked parameters represent the key features initially selected.

[0064] S32: Introducing an adaptive weighted learning mechanism, dynamically adjusting the weights of each data dimension based on its importance in the feature space, so that the feature distribution function can more accurately fit the true distribution of the data; "deep learning" is applied to the features initially selected in S31, dynamically adjusting the weights of each feature, so that the model increasingly focuses on the truly important parameters and ignores irrelevant noise; ; in, Indicates the first The adaptive weighted learning output of the next iteration Indicates data weights, Indicates the number of iterations; Work process and examples: Initial state (t=0): The weights are initialized to the result of S31, such as power=0.9, speed=0.8, and cabin temperature=0.3.

[0065] Iterative learning (t=1,2,3...): The model uses the current weights to perform clustering or prediction.

[0066] It found that if "power" was given a higher weight, abnormal batches could be distinguished more clearly. Therefore, in the next iteration, the weight of "power" was increased (for example, from 0.9 to 0.95).

[0067] It found that even giving "cabin temperature" a high weight did not significantly help in identifying anomalies. Therefore, the weight of "cabin temperature" was gradually reduced (for example, from 0.3 to 0.05).

[0068] Final output (t=T): After multiple rounds of learning, the weights have stabilized. Power and scan speed have very high weights, while the weights of other parameters are close to 0.

[0069] S33: Perform correlation feature extraction. Combining correlation rules, extract core correlation features that can characterize the system state from the weighted learning data. Based on the results of the first two steps, formally output a set of core parameter combinations that are minimal in number but have the strongest explanatory power. ; in, This represents the extracted multidimensional data association features. Represents the data distribution function. This represents the initial weight of the data, and m represents the node connection probability.

[0070] Work process and examples: Input: After weighted learning, the output power and speed are the most important. At the same time, the association rules found that there is an interaction between layer thickness and power (for example, when the layer thickness is larger, higher power is required to ensure curing).

[0071] Extraction: Function f integrates all information and finally outputs a concise core correlation feature vector containing key information and the relationships between them.

[0072] Output example: Core feature vector = [laser power, scanning speed, layer thickness]; This means that in the subsequent anomaly diagnosis, we only need to focus on these three parameters, which greatly reduces the complexity of the problem.

[0073] Specifically, the step of improving fuzzy clustering of the core associated feature vectors for anomaly diagnosis to obtain the key parameters causing the deviation values ​​includes: in order to accurately locate the abnormal batches and diagnose the root cause, the associated feature vectors extracted in step S3 are... The data is fed into an improved fuzzy clustering algorithm to highlight outliers in multidimensional data. S41: Calculate inter-cluster distances to determine the average distance between cluster centers and the remaining data in a multidimensional dataset; measure the average distance between a cluster center (representing a typical set of process parameters) and all other data points. ; in, This represents the average distance between the cluster centers and the remaining data in the multidimensional dataset. and Represents a random point. This represents the total number of clusters. and Represents a set of random numbers; For example: Assume that the core features of S3 output are [laser power, scanning speed].

[0074] The dataset contains 3 clusters (k=3): Cluster 1 (center) (300mW, 2.0m / s) -> Represents "Standard Parameters" Cluster 2 (center) (280mW, 1.8m / s) -> Represents "conservative parameters" Cluster 3 (center) (350mW, 2.5m / s) -> Represents "aggressive parameters" Now calculate the inter-class distance between cluster centers of the radical parameter cluster: Find all data points that do not belong to cluster 3 (i.e., all points that belong to clusters 1 and 2).

[0075] Calculate the center point (350, 2.5) is the Euclidean distance to these points.

[0076] because The power and speed values ​​at these points are much higher than at other points, and these distances would be very large.

[0077] After averaging, the result is a very large value.

[0078] Conclusion: A larger value indicates that the cluster is more unique and further away from the mainstream group.

[0079] S42: Calculate the average distance within a cluster, cluster all association rules, and obtain the average distance between all feature data; measure the average density between all data points within the same cluster; ; in, This represents the average distance between all feature data. This represents the association rule, where D represents the similarity. For each cluster, find all data points that belong to it and meet the association rule R, and then calculate the average distance between each pair of these points.

[0080] For example: See cluster 1 (standard parameters): This cluster may contain data points from 50 print jobs.

[0081] The association rule R could be: "Power is in the range [295, 305] and speed is in the range [1.95, 2.05]".

[0082] We calculate the Euclidean distance between all pairs of points within these 50 points.

[0083] Because the parameter values ​​of these points are very close, the distance between them will be very small.

[0084] After averaging, the result is a very small value.

[0085] Conclusion: The smaller the value, the more consistent the cluster is and the more stable the process is.

[0086] S43: Generate feature clustering results. By comparing two average distances, a quantitative clustering effectiveness index is obtained. A simple ratio is used to comprehensively evaluate inter-cluster separation and intra-cluster compactness, thus quantifying the "quality" of a cluster. This index Y is specifically used to highlight anomalies. ; Where Y represents the feature clustering result; the better the clustering effect of the data, the closer the value of Y is to 1; conversely, the worse the clustering effect of the data, the closer the value of Y is to -1.

[0087] Anomaly pattern matching and diagnosis: The calculated feature clustering result Y is matched with a pre-established anomaly pattern feature library; the anomaly pattern feature library stores the typical Y value range and key parameter features corresponding to different process faults; if the current Y value falls within the typical range of a certain anomaly pattern, it is determined that the current printing task is affected by the anomaly pattern, and the key process parameters that caused the pattern are output as the final diagnostic result.

[0088] The structure of the database (each record contains the following information): Anomaly Mode ID / Name: Give this type of fault a unique identifier. For example: Mode A001: Power Attenuation Anomaly.

[0089] Typical Y value range: In which range does the clustering validity index Y calculated in step S43 typically fall for this type of failure?

[0090] For example: Y∈[0.85,0.98]. This range is obtained by analyzing the statistical values ​​of Y calculated from multiple historical occurrences of the same type of failure.

[0091] Key parameter feature vector: This describes the core process parameter behavior that most critically affects this type of failure. It is typically a specific instance of the core correlation feature vector output by S3.

[0092] For example: [Laser power: significantly low, scanning speed: normal, slice thickness: normal]. More specifically, this can be quantified as: power < rated value. 0.8.

[0093] Root cause description: The underlying reasons that lead to this key parameter characteristic.

[0094] For example: laser aging, optical lens contamination, power calibration errors.

[0095] Solution / Maintenance Recommendations: Recommended actions to address this root cause.

[0096] For example: clean the F-Theta lens, check and calibrate the laser power meter, and contact the equipment manufacturer to replace the laser module.

[0097] The workflow is as follows: enter: Current Y value: A clustering validity index obtained from step S43, such as Y_current=0.92.

[0098] The current core feature vector is a simplified combination of parameters and their specific values ​​obtained from step S33, such as F_out=[laser power:190mW, scanning speed:2.1m / s] (the set value should be power:300mW, speed:2.0m / s).

[0099] Matching loop: The system iterates through each record in the abnormal pattern feature library.

[0100] Step 1: Y-value matching. Determine whether Y_current falls within the "typical Y-value range" of a certain record.

[0101] Step 2: Feature Verification. Based on the Y-value matching, further verify whether the current F_out matches the "key parameter features" of this record.

[0102] Output diagnostic results: Scenario 1: Exact match (finding the root cause) Scenario: Y_current=0.92, and the current core feature is [Power:190mW,...].

[0103] Matching process: The system found that 0.92 ∈ [0.85, 0.98], which matched the pattern A001.

[0104] Next, verify the characteristic: 190mW < 300mW 0.8 (i.e., 240mW), the condition is met.

[0105] Final diagnostic output: Diagnostic conclusion: Abnormal pattern identified: A001 - Power attenuation abnormality.

[0106] Key process parameters: The laser power is abnormally low (actual value 190mW, far below the set value of 300mW).

[0107] Potential root cause: laser aging or optical lens contamination.

[0108] Recommended action: Clean the F-Theta lens immediately and calibrate the laser output using a power meter.

[0109] Scenario 2: Y-value matches but features do not (new or compound anomalies are found)

[0110] Scenario: Y_current=0.88, but the current core feature is [scanning speed: 2.8m / s,...] (the speed is set to 2.0m / s).

[0111] Matching process: 0.88∈[0.85,0.98], matched pattern A001.

[0112] However, during the verification of the characteristics, it was found that the power was a normal 300mW, which did not meet the characteristics of "power attenuation".

[0113] Final diagnostic output: Diagnostic conclusion: A highly anomalous cluster was found (Y=0.88), but it does not match the characteristics of the known anomalous pattern A001. Suspected novel anomaly or a composite pattern of A002 and A003.

[0114] Key process parameters: Scanning speed is abnormally high (actual value 2.8m / s).

[0115] Follow-up actions: It is recommended that experts conduct in-depth analysis and add this new pattern to the abnormal pattern library.

[0116] Scenario 3: No match (significant fluctuations for unknown reasons).

[0117] Scenario: Y_current=0.50.

[0118] Matching process: After traversing the library, it was found that 0.50 did not fall within the typical Y value range of any known pattern.

[0119] Final diagnostic output: Diagnostic conclusion: A significant process deviation was detected (Y=0.50), but no known anomalous patterns were matched.

[0120] Key process parameters: The core feature vectors obtained from output S33, such as [power, speed, layer thickness] and their values.

[0121] Specifically, the intelligent diagnostic report generation in S5 includes: S51: Root cause tracing and visualization. Based on the key parameters diagnosed by S4, the system automatically retrieves the historical best value, process allowable range and current batch value of the parameter from the multi-source heterogeneous database obtained by S1, and presents them in the report in a highlighted contrast form. Specific execution process: 1. Input trigger: Receive "critical abnormal parameters" (e.g., "laser power") from the output of the S4 / S5 pattern matching module.

[0122] 2. Data Query: The system automatically backtracks to the central database established in S1 to query the following three types of data: Historical best value: Calculated from the average or mode of the parameter among the batches with the highest historical success and accuracy. For example: Historical best power = 300W.

[0123] Process allowable range: Read the upper and lower limits of this parameter from the process file. For example: Process range = [280W, 320W].

[0124] Current batch value: The actual value of this parameter in the current abnormal batch. For example: Current power = 190W.

[0125] Visual presentation: Generate a "parameter status comparison chart" in the report.

[0126] Format: Usually a horizontal bar chart or dashboard.

[0127] content: A baseline marked with "lower limit of process", "historical best" and "upper limit of process".

[0128] A bar or pointer representing the "current value".

[0129] Highlighting logic: If the current value exceeds the process range, it will be highlighted in red and marked "Seriously Exceeding Standards".

[0130] If the current value is within the process range but deviates from the historical best value by more than a certain threshold (such as 10%), it will be highlighted in yellow and marked "significant deviation".

[0131] S52: Optimization suggestion knowledge base matching. The report generation module has an embedded optimization knowledge base containing expert rules and historical cases. Based on the diagnosed abnormal patterns, it automatically matches and outputs one or more specific and actionable process parameter adjustment suggestions. Matching and output process: Input: A specific "abnormal mode ID" (e.g., A001).

[0132] Search: The system searches the optimization suggestion knowledge base using "abnormal pattern ID" as the keyword to find all related expert rules and historical cases.

[0133] Integration and Generation: Extract all matching "suggested actions".

[0134] Remove duplicates and sort them by the urgency of the measures or the order of operations (e.g., "clean" before "calibrate").

[0135] Citing relevant and successful historical cases as supporting evidence increases the credibility of the recommendations.

[0136] Output: Generate one or more optimization suggestions in text format.

[0137] Final output example: [Optimization Suggestions] 1. Immediate action: Turn off the device and clean the F-Theta lens with anhydrous ethanol and a lint-free cloth.

[0138] 2. Equipment maintenance: Use an external power meter to calibrate the laser output power to ensure it matches the set value.

[0139] 3. Reference Case: This problem was successfully resolved on October 1, 2023 by cleaning the lens (Case C-20231001).

[0140] S53: Report integration and output. The final diagnostic report is a structured document that includes at least a summary table of accuracy test results, a heatmap of key dimension deviations, a root cause parameter analysis diagram, and optimization suggestions in text form. It also supports one-click export.

[0141] Report content structure: Summary table of accuracy test results: List the “design value”, “measured value”, “deviation value” and “qualified status” of all critical dimensions in tabular form.

[0142] Heat map of critical dimension deviations: The part model is rendered in color, with different colors representing the magnitude of dimensional deviation in different areas (e.g., blue represents negative deviation / smaller size, and red represents positive deviation / larger size). This visually displays the trend of deformation or shrinkage.

[0143] Root cause parameter analysis diagram: The "parameter status comparison chart" generated in S51 visually displays abnormal situations of key parameters.

[0144] Optimization suggestions for text format: That is, the detailed operation steps generated in S52.

[0145] In one embodiment, the detection method further includes: S6: Process parameter self-calibration and feedback; S61: Parameter adjustment calculation: Based on the optimization suggestions in the S5 diagnostic report, the system automatically calculates the specific adjustment amount required for key process parameters according to the built-in process parameter-accuracy deviation response model. S62: Command Issuance and Verification: Automatically sends the adjusted set of process parameters to the 3D printing control system for the next printing task of the same digital model; S63: Closed-loop verification learning: The system records the printing tasks and corresponding accuracy detection results before and after parameter adjustment, forming a decision-result feedback pair, which is used to continuously update and optimize the response model in S61.

[0146] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0147] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for detecting the 3D printing precision of an image-recognized oral model, characterized in that, The method comprises the following steps: S1: obtaining multi-source heterogeneous data of the oral model and forming a multi-dimensional data vector, wherein the multi-source heterogeneous data comprises image information, device parameters, process parameters, environmental and material data, and design data; S2: performing precision detection and feature extraction on the oral model based on image recognition, wherein the feature extraction is used to extract deviation values for precision detection and is added to the multi-dimensional data vector in S1 to form a multi-dimensional data vector set; S3: performing correlation feature mining on the multi-dimensional data vector set to output a core correlation feature vector; S4: performing abnormal diagnosis on the core correlation feature vector by improved fuzzy clustering to obtain key parameters causing the deviation values; S5: generating an intelligent diagnosis report by optimizing the process according to the key parameters.

2. The method of claim 1, wherein the method is used for detecting the 3D printing accuracy of an image-recognized dental model. The step of obtaining multi-source heterogeneous data of the oral model and forming a multi-dimensional data vector comprises: S11: standard image acquisition: performing multi-angle automatic image acquisition on the oral model; S12: manufacturing process data extraction: automatically capturing whole-process process parameters bound to the printing task ID through an API interface or a log file of a 3D printer control system; S13: design data synchronization: obtaining original design contour data of the oral model, i.e., a theoretical two-dimensional polygon coordinate set of each layer, from slice software or by analyzing G code.

3. The method of claim 2, wherein the method further comprises: determining the 3D printing precision of the dental model based on the image recognition result. The step of performing precision detection and feature extraction on the oral model based on image recognition comprises: S21: performing grayscale, Gaussian filter denoising, and contrast-limited adaptive histogram equalization on the collected images to enhance feature edges; using a standard calibration board with a known size to calibrate the camera to obtain an accurate conversion relationship between pixels and physical sizes; S22: using a pre-trained deep learning model to segment the images to automatically identify and label key regions; using iterative nearest point algorithm to eliminate overall deviation caused by the angle of placement by comparing the identified feature contour with the theoretical design contour obtained in S13; S23: on the basis of registration, the system automatically measures a series of predefined key dimensions; and generating dimension deviation values; S24: taking the dimension deviation vector calculated in S23 as a new core data column, and associating the dimension deviation vector with the multi-source heterogeneous data collected in S1 to jointly form a multi-dimensional data vector set for subsequent analysis.

4. The method of claim 1, wherein the method is a method of detecting the 3D printing accuracy of an image-recognized dental model, characterized by, The step of performing correlation feature mining on the multi-dimensional data vector set to output a core correlation feature vector comprises: S31: constructing a multi-dimensional data correlation feature distribution function, and performing attribute distribution arrangement processing on the multi-dimensional data vector set by using an autocorrelation function in combination with data similarity; ； wherein, wherein, represents the multi-dimensional data correlation feature distribution function, K represents the feature autocorrelation, n represents the data feature dimension hierarchy, and i represents the data node quantity. S32: introducing an adaptive weighted learning mechanism, dynamically adjusting the weight of each data dimension in the feature space according to the importance of the data dimension, so that the feature distribution function can more accurately fit the real distribution of the data; ； wherein, denotes the adaptive weighted learning output of the denotes the data weight, denotes the number of iterations;​ S33: performing correlation feature extraction, and extracting a core correlation feature capable of representing the system state from the data subjected to the weighted learning in combination with the correlation rule; ； wherein, represents the extracted multi-dimensional data association feature, represents a data distribution function, represents a data initialization weight, and m represents a node connection probability.

5. The method of claim 4, wherein the method further comprises: The step of improving the core correlation feature vector by fuzzy clustering to obtain the key parameters causing the deviation value includes: for accurate positioning of the abnormal batch and diagnosing the root cause, the correlation feature vector extracted in step S3 is input into an improved fuzzy clustering algorithm to highlight the abnormal points in the multi-dimensional data: S41: calculating the distance between classes, and measuring the average distance between the clustering center and the remaining data in the multi-dimensional data set; ； wherein, denotes the average distance of the cluster center to the remaining data in the multi-dimensional data set, and denotes a random point, denotes the total number of clusters, and denotes a set of random numbers; S42: Calculate the average distance within the class, cluster the features of all association rules, and obtain the average distance between all feature data; ； wherein, represents the average distance between all feature data, represents the association rule, D represents the similarity; S43: Generate feature clustering results, and obtain a quantitative clustering effectiveness index by comparing the two average distances; ； Wherein, Y represents the feature clustering result; the better the clustering effect of the data, the closer the value of Y is to 1; otherwise, the worse the clustering effect of the data, the closer the value of Y is to -1.

6. The method of claim 1, wherein the method is a method of detecting the 3D printing accuracy of an image-recognized dental model, characterized by, The intelligent diagnosis report generation in S5 specifically includes: S51: Root cause tracing and visualization. The system automatically retrieves the historical optimal value, process allowable range and current batch value of the key parameter from the multi-source heterogeneous database obtained in S1 according to the key parameter diagnosed in S4, and presents them in the report in the form of highlighted comparison; S52: Optimization suggestion knowledge base matching. The report generation module is embedded with an optimization knowledge base containing expert rules and historical cases, which automatically matches and outputs one or more specific and operable process parameter adjustment suggestions according to the diagnosed abnormal pattern; S53: Report integration and output. The final diagnosis report is a structured document, which at least includes the precision detection result summary table, the key dimension deviation heat map, the root cause parameter analysis graph and the optimization suggestion in text form, and supports one-key export.

7. The method of claim 5, wherein the method further comprises: determining the 3D printing precision of the dental model based on the image recognition result. After generating the feature clustering results, it further includes: Abnormal pattern library matching and diagnosis: match the calculated feature clustering result Y with the pre-established abnormal pattern feature library; the abnormal pattern feature library stores the typical Y value range and its key parameter characteristics corresponding to different process faults; if the current Y value falls within the typical range of a certain abnormal pattern, it is determined that the current printing task is affected by the abnormal pattern, and the key process parameters that cause the pattern are output as the final diagnosis result.

8. The method of claim 1, wherein the method is a method of detecting the 3D printing accuracy of an image-recognized dental model, characterized by, The detection method further includes: S6: Process parameter self-correction and feedback; S61: Parameter adjustment amount calculation: based on the optimization suggestions in the S5 diagnosis report, the system automatically calculates the specific adjustment amount of the key process parameters according to the built-in process parameter-precision deviation response model; S62: Instruction issuing and verification: automatically issue the adjusted process parameter set to the 3D printing control system for the next printing task of the same digital model; S63: Closed-loop verification learning: the system records the printing tasks before and after parameter adjustment and the corresponding precision detection results, forms a feedback pair of decision-result, and is used for continuous updating and optimization of the response model in S61.

Images