Control valve drawing part intelligent identification system

Abstract

The invention discloses a control valve drawing part intelligent identification system, and belongs to the technical field of intelligent manufacturing and drawing identification, the system obtains a direction control valve drawing sample, combines a QATM algorithm, a YOLO model, EAST and TesseractOCR, and U-Net and a graph theory algorithm to carry out multi-modal feature extraction and joint identification, constructs a three-level classification matching system, and outputs an initial identification result; through label data conversion, manual sampling auditing, initial recognition standard construction, extended label data generation and comprehensive precision label data fusion, multiple times of iterative optimization of an initial model is achieved, and finally a high-precision detection model is output. Automatic and intelligent recognition of part structures, parameters and topological relations in control valve drawings is achieved; the drawing recognition efficiency and accuracy are remarkably improved, and the method is suitable for industrial drawing digitization and intelligent assembly application.

Description

Technical Field

[0001] This invention relates to the field of drawing recognition technology, specifically to an intelligent recognition system for control valve drawing parts. Background Technology

[0002] In the field of industrial manufacturing and assembly, directional control valves are key hydraulic and pneumatic control components. Their drawings contain a large number of complex structural parts, parameter annotations, and assembly relationship information. Traditional drawing recognition mainly relies on manual interpretation, which suffers from low efficiency, error-proneness, and poor consistency, making it difficult to meet the needs of intelligent manufacturing for rapid analysis and reuse of drawing data.

[0003] In existing technologies, recognition methods based on template matching or single visual models are susceptible to interference from factors such as occlusion, noise, and scale variations when processing control valve drawings with similar structures and dense annotations, resulting in limited recognition accuracy. Furthermore, the lack of an effective fusion and verification mechanism for multi-source information such as parameter text and topological relationships makes it difficult to achieve a comprehensive understanding of the structural and functional attributes of parts. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent identification system for control valve drawing parts to solve the problems mentioned in the background art.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: an intelligent identification system for control valve drawing parts, comprising the following steps: S1: Obtain the drawing sample of the directional control valve; S2: Obtain the initial recognition result based on the drawing sample; S3: Transform the initial labeled data based on the initial identification results; S4: Based on the initial annotation data, initial precision annotation data is obtained through manual sampling and review; S5: Generate initial recognition criteria based on the initial precision annotation data using the initial model; S6: Calculate extended annotation data based on the initial identification criteria; S7: Based on the extended annotation data, further extended precision annotation data is obtained through manual sampling and review; S8: Obtain comprehensive precision annotation data based on the initial precision annotation data and the extended precision annotation data; S9: Based on the comprehensive accuracy annotation data, the initial model is iterated and the final detection model is output.

[0006] According to the above technical solution, S2 includes: collecting standard drawing samples of core components of the directional control valve and classifying and archiving them; extracting core features from the component drawings after normalization to construct a tagged component template library; using the QATM algorithm to extract local feature points from the query image and the template image to generate a high-dimensional feature description vector; calculating the matching degree using the cosine similarity algorithm; setting a similarity threshold T1; and verifying and outputting a list of validly matched components and labeling information in conjunction with the IOU threshold T2; wherein T1 is the feature vector matching degree filtering threshold; and T2 is the IOU verification threshold.

[0007] According to the above technical solution, S2 further includes: detecting structural feature information through the YOLO model, identifying parameter annotation information through a combination of the EAST text detector and the TesseractOCR engine, and constructing topological association information through the U-Net image segmentation model and graph theory algorithm; performing cross-validation on the three types of information, setting a position overlap threshold T3, outputting a structured key information set, and constructing a three-level classification matching system based on the information set to output the initial recognition result; wherein T3 is the position overlap threshold.

[0008] According to the above technical solution, S3 includes: extracting data from the initial recognition results to generate an intermediate data table based on preset target category rules, bounding box annotation rules, and feature attribute association rules; performing automated verification on the converted annotation data, setting a verification pass rate threshold T5, storing the data in the form of one annotation file corresponding to one drawing after verification, and performing cropping, normalization, and grayscale processing on the accessory image area; T5 is the automated verification pass rate threshold for annotation data.

[0009] According to the above technical solution, S3 further includes: using OpenCV's SIFT algorithm to extract feature vectors from the accessory images, performing cosine similarity calculation with the feature vectors of templates encoded in the same category in the template library, setting a similarity threshold T6, filtering out samples that meet the threshold, and converting the labeled data into YOLOv8 format; wherein T6 is the matching and filtering threshold between the accessory feature vector and the template feature vector.

[0010] According to the above technical solution, S4 includes: reviewers comparing the original drawings, annotation files, and template library samples to check and correct bounding box offset errors and category misjudgment errors; setting IOU thresholds T7, T8, sampling ratio T9, and annotation accuracy threshold T10; when IOU < T7, it is judged as a serious offset; when IOU > T8, it is judged as qualified; sampling verification is performed on the integrated data, and after meeting the standards, the initial accuracy annotation data is archived according to image ID, annotation file, and review log; T7 is the boundary box serious offset judgment threshold, T8 is the boundary box qualified judgment threshold, T9 is the sampling verification ratio threshold, and T10 is the annotation accuracy compliance threshold.

[0011] According to the above technical solution, S5 includes: visually locating the core distinguishing features of various valves using Grad-CAM; conducting statistical analysis on the core features in the initial precision annotation data; setting feature response value thresholds T11, confidence threshold T12, matching degree threshold T13, accuracy threshold T14, quantity threshold T15, quantity threshold T16, parameter range deviation threshold T17, matching confidence threshold T18, and quantity threshold T19 to quantify the feature response value thresholds; constructing parameter identification standards to classify the initial precision annotation data. The results are classified into first-level and second-level hierarchical categories; T11 is the threshold value of the characteristic response value of the solenoid valve electromagnet component, T12 is the confidence threshold value of the voltage annotation text recognition, T13 is the matching degree threshold value of the one-way valve cone structure, T14 is the accuracy threshold value of the one-way valve arrow annotation recognition, T15 is the threshold value of the number of port positions of the reversing valve interface, T16 is the threshold value of the number of steps of the reversing valve core, T17 is the parameter range deviation threshold value, T18 is the confidence threshold value of the material matching, and T19 is the threshold value of the number of manual reversing valve interfaces.

[0012] According to the above technical solution, S6 includes: setting a confidence threshold T20 for accessory category identification, a confidence threshold T21 for bounding box identification, and a confidence threshold T22 for core parameter identification as filtering thresholds to filter and validate the labeled data; generating a segmentation mask covering the complete outline of the accessory using the vit-h-SAM model, setting a mask confidence threshold T23, optimizing the initial bounding box based on the segmentation mask, generating extended labeled data and storing it in YOLOv8 format; where T20 is the confidence filtering threshold for accessory category identification, T21 is the confidence filtering threshold for bounding box identification, T22 is the confidence filtering threshold for core parameter identification, and T23 is the confidence threshold for mask identification.

[0013] According to the above technical solution, S8 includes: constructing a three-dimensional feature vector composed of image features, annotation features, and attribute features for each labeled sample; setting a duplicate sample threshold T27, a highly similar sample threshold T28, and a proportion fluctuation threshold T29; performing layered deduplication by using MD5 hash value and comprehensive similarity calculation; merging the deduplicated initial data and extended data; and processing samples with the same image ID according to the principle that the extended data covers the initial data; wherein T27 is the duplicate sample determination threshold, T28 is the highly similar sample determination threshold, and T29 is the category proportion fluctuation threshold.

[0014] According to the above technical solution, S9 includes: the initial model is iterated 3 times; during the iteration process, the convergence threshold and core threshold are verified; a coverage ratio threshold T24, a background ratio threshold T25, an adjacent component ratio threshold T26, and a supplementary quantity T30 are set; if the target is not met, optimization is achieved by supplementing extended data, adjusting regularization weights, or supplementing samples until the core indicators meet the target and the final detection model is output; T24 is the mask coverage core structure ratio threshold, T25 is the mask background ratio threshold, T26 is the mask adjacent component ratio threshold, and T30 is the supplementary parameter annotation sample quantity threshold.

[0015] Compared with the prior art, the beneficial effects achieved by the present invention are: This invention, through multi-model joint recognition and a three-level classification matching mechanism, realizes the automated extraction and verification of the structure, parameters and topological relationships of parts in control valve drawings, greatly reducing manual intervention and improving the efficiency of drawing parsing; This invention utilizes multiple advanced algorithms such as QATM, YOLO, and SAM for feature matching, target detection, and image segmentation, combined with multiple rounds of manual sampling and threshold verification, to effectively address issues such as occlusion, noise, and structural similarity in complex drawings, thereby improving recognition accuracy and system stability. This invention, through dynamic threshold setting, hierarchical classification rules and iterative optimization mechanism, enables the system to adaptively adjust according to different drawing types and recognition requirements, supports rapid access to new part types and continuous model optimization, and has good engineering applicability and scalability. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1, see Figure 1 The present invention proposes the following technical solution: an intelligent identification system for control valve drawing parts, the system comprising the following steps: The process involves: acquiring a directional control valve drawing sample; obtaining initial identification results based on the acquired drawing sample; transforming the initial annotation data based on the initial identification results; obtaining initial precision annotation data through manual sampling review based on the initial annotation data; generating initial identification standards using the initial model based on the initial precision annotation data; calculating extended annotation data based on the initial identification standards; obtaining extended precision annotation data again through manual sampling review based on the extended annotation data; obtaining comprehensive precision annotation data based on the initial precision annotation data and the extended precision annotation data; and performing multiple model iterations on the initial model based on the comprehensive precision annotation data to optimize model accuracy, ultimately outputting the final detection model.

[0019] The steps to obtain the initial recognition results are as follows: First, collect standard drawings and samples of the core components of the directional control valve, and classify and archive them according to component type, specifications, and key features.

[0020] Subsequently, each component drawing is normalized, and its core features such as edge contours, hole diameters, and mounting positioning holes are extracted to construct a tagged component template library. This library enables standardized management of component samples, avoids matching interference caused by inconsistent specifications, and lays the foundation for accurate identification in the future.

[0021] Specifically, core components include, but are not limited to, valve cores, valve sleeves, electromagnets, springs, and interface flanges. Adjust the drawings to a uniform resolution.

[0022] The QATM algorithm is used, with preprocessed drawings as the query image and parts from the template library as the template images. Adaptive convolution kernels in the algorithm are used to extract local feature points from both the query image and the template image, generating high-dimensional feature description vectors. The vector dimension is dynamically adjusted according to the complexity of the parts.

[0023] Next, the matching degree between the feature vector of the query image and the feature vector of each part in the template library is calculated using the cosine similarity algorithm, and a similarity threshold T1 is set.

[0024] For results with a matching degree greater than or equal to T1, further IOU verification is performed. If the IOU is greater than or equal to the IOU threshold T2, it is determined to be a valid match, and false matches caused by local similarity are eliminated.

[0025] The final output lists the matched parts in the drawings, and marks the location coordinates, matching similarity, and corresponding template library tags for each part. A threshold verification mechanism distinguishes parts of different specifications with similar structures, improving the accuracy of the initial matching and reducing the false matching rate.

[0026] Specifically, the similarity threshold T1 can be dynamically fine-tuned according to the clarity of the drawing, while eliminating mismatches of similar contours of valve sleeves of different specifications.

[0027] Then, based on the information characteristics of the directional control valve drawings, the key information is divided into three categories: structural feature information, parameter annotation information, and topological association information.

[0028] For structural feature information, the YOLO model is used for detection to accurately locate the key structural parts of each component in the drawings and output structural feature labels and corresponding location information.

[0029] Among them, structural feature information refers to the geometric structure of the accessories, such as the number of steps in the valve core, the distribution of valve holes, and the installation interface shape of the electromagnet.

[0030] For parameter annotation information, a combined approach of text detection and OCR recognition is adopted. The EAST text detector is used to locate the text areas in the drawing, and then the Tesseract OCR engine is used to recognize the text content, and the recognition results are formatted.

[0031] For topological association information, the connecting lines in the drawings are segmented using the U-Net image segmentation model. Combined with the location coordinates of the parts, a graph theory algorithm is used to construct a topological relationship graph of parts and connecting lines, and the association attributes are labeled.

[0032] Cross-validation is performed on the three types of information to eliminate contradictory data. This enables comprehensive extraction of drawing information and avoids blind spots in the recognition of specific types of information by a single model.

[0033] For example, if the structural features identify the solenoid valve core, but the parameter label does not identify electromagnet-related parameters such as 24V voltage, it is marked as pending verification.

[0034] If the positional overlap between the valve sleeve and the valve core in the topology diagram is greater than or equal to the positional overlap threshold T3, T3 is used to confirm the assembly relationship between the valve sleeve and the valve core. If the structural features are consistent, then the two are confirmed to be in an assembly relationship.

[0035] The final output includes a structured key information set containing component names, structural features, key parameters, and topological relationships.

[0036] Based on this key information set, a three-level classification and matching system is constructed, covering valve type classification, core structure determination, and key parameter verification. The hierarchical classification mechanism is used to adapt to the identification needs of valves with different complexities, and verifies them step by step from type to parameters, thereby improving the logic and completeness of the identification results.

[0037] Among them, the valve type is classified as Level 1, the core structure is determined as Level 2, and the key parameters are verified as Level 3.

[0038] Specifically, the primary classification is based on the core control components and topological relationships as the core matching criteria, with the following preset rules: A solenoid valve is defined as one that includes an electromagnet, linear movement of the valve core, and electromagnetic control signals; a check valve is defined as one that includes a one-way valve core, has no control components, and has a one-way flow topology; and a manual directional valve is defined as one that includes a manual operating lever and a rotating valve core. These categories cover the mainstream types of directional control valves.

[0039] The secondary classification is based on the core components and structural features matched by QATM, such as valve core type, control method, number of interfaces, etc., and each structural item corresponds to a specific feature matching threshold.

[0040] For example, a slide valve core needs to be matched with a cylindrical stepped structure and at least two sealing ring grooves.

[0041] The three-level classification is based on the parameters recognized by OCR, and is verified by combining the type-parameter correspondence in the template library.

[0042] For example, for solenoid valves, the rated voltage, rated pressure, and interface diameter need to be verified. If the deviation between the identified parameters and the parameters of similar valves in the template library is less than or equal to the parameter deviation threshold T4, T4 is used to confirm the validity of the parameters. If the parameters are valid, they are confirmed to be valid; otherwise, they are marked as abnormal.

[0043] Finally, the valve type is first determined by primary classification matching, then the corresponding secondary structure matching rule is called based on the type, and finally the tertiary parameters are verified by combining the results of the first two levels to ensure the consistency of the matching, and the initial recognition result is output in the form of a structured dictionary.

[0044] The steps for transforming the initial labeled data are as follows: The conversion is based on preset annotation rules, which include target category rules, bounding box annotation rules, and feature attribute association rules.

[0045] The target category rule is used to construct a classification system for directional control valve accessories, which is divided into three levels of categories according to core function and structural hierarchy.

[0046] Specifically, the first-level category is control components, the second-level category is electromagnets, and the third-level category is DC electromagnets or AC electromagnets. Each category corresponds to a unique numerical code to ensure the uniqueness of the category label.

[0047] The bounding box annotation rules are used to define the bounding box coordinates, which are the coordinates of the accessory positions matched by the QATM algorithm. Specifically, the minimum bounding rectangle principle is used to select the target, and the coordinate format is uniformly pixel coordinates (x1, y1, x2, y2), where (x1, y1) are the coordinates of the top-left corner of the rectangle, and (x2, y2) are the coordinates of the bottom-right corner.

[0048] If an accessory is obstructed, it is labeled with the maximum visible area and the obstruction attribute is marked. Feature attribute association rules are used to bind and label key parameters in the initial recognition results with the accessory.

[0049] Specifically, valve core accessories need to be labeled with the material as steel, the diameter as determined by the valve core specifications, and the number of steps as determined by the valve core structure. When parameters are missing, corresponding labels should be made, and the initial identification matching similarity should be associated with them.

[0050] The storage structure of the unified annotation data is set as image ID-part ID-category code-boundary box coordinates-feature attribute-matching similarity, where the image ID and the original drawing file name have a one-to-one correspondence.

[0051] Specifically, the part ID is generated in the format of image ID-serial number. For example, in the drawing with ID 001, the ID of the first part is 001-01.

[0052] By reading the structured dictionary of the initial recognition results, the image ID, parts list, category information, bounding box coordinates, key parameters, and matching similarity data of each drawing are extracted to generate an intermediate data table. An automated verification and backtracking correction mechanism is used to quickly filter invalid data, reduce manual intervention costs, and ensure the integrity and validity of the labeled data.

[0053] Load the preset annotation rules, match the parsed accessory categories with the category codes in the rules, correct the bounding box coordinates according to the minimum bounding rectangle principle, remove coordinates that exceed the image range due to algorithm errors, and fill the key parameters into the corresponding fields according to the attribute association rules.

[0054] Automated verification is performed on the converted labeled data; Specifically, it checks whether the bounding box coordinates exceed the image resolution range; it also checks whether there are any mismatches in the category coding and whether the key parameter formats conform to the specifications.

[0055] For data that fails the verification, it is marked as abnormal and the initial identification result is returned for re-analysis until the verification pass rate is not lower than the verification pass rate threshold T5, so as to ensure the validity of the labeled data.

[0056] The verified annotation data is stored in the form of one annotation file per drawing. At the same time, a summary list is generated to record the number of annotated accessories, the number of abnormal data entries, and the processing results for each drawing.

[0057] Based on the bounding box coordinates in the initial annotation data, tools are used to batch crop the independent image regions of each component in each drawing. The cropped images are then uniformly normalized and converted to grayscale. Image preprocessing and feature vector comparison are used to further filter high-quality samples. The YOLOv8 format is adapted to mainstream detection models, improving data reusability and model training efficiency.

[0058] Based on the SIFT algorithm of OpenCV, the feature point detection parameters are set as follows: number of Gaussian difference pyramid layers X1, standard deviation of Gaussian kernel X2, and feature point threshold X3. For each cropped part image, a fixed-dimensional feature vector is extracted to describe the part features using SIFT. The vector values ​​are retained to 6 decimal places to ensure the accuracy of feature description, and stored in NumPy array format.

[0059] For the standard samples in the previously constructed accessory template library, perform the same cropping, normalization, grayscale conversion, and SIFT feature extraction process as described above to generate a template feature vector library, which is then classified and stored according to category coding for subsequent similarity comparison.

[0060] The feature vector of the accessory to be calculated is paired with all template feature vectors of the same category in the template library.

[0061] For example, a DC electromagnet coded as 003 is only compared with a template coded as 003 in the template library.

[0062] The similarity between paired vectors is calculated using the cosine similarity formula: cosθ=(A・B) / (|A|×|B|). The similarity is calculated quickly using NumPy's `dot` function and `linalg.norm` function, and the results are rounded to three decimal places to ensure the accuracy of the matching.

[0063] Where A is the feature vector to be calculated, and B is the template feature vector.

[0064] A similarity threshold T6 is set to filter accessory samples that match the template library. Accessory samples with a similarity of not less than T6 are selected, and the highest similarity value and corresponding template ID of each sample are recorded.

[0065] If a component has a similarity of at least T6 with multiple templates of the same category, retain the record with the highest similarity. If a component has a similarity of less than T6 with all templates of the same category, mark it as low similarity and temporarily store it in the backup dataset.

[0066] YOLOv8 uses the txt format with a single annotation file for each image. The annotation file has the same name as the image file and is stored in a subfolder of the same directory. Each annotation information corresponds to a line of data, and its format is: class index - normalized x center - normalized y center - normalized width - normalized height - confidence.

[0067] Among them, the class index is based on the previous class encoding system, converting the three-level class encoding into consecutive integers starting from 0; the normalized coordinates are obtained by converting the original pixel bounding box (x1, y1, x2, y2) into normalized values. The calculation formulas are as follows: x center = (x1 + x2) / (2 × image width); y center = (y1 + y2) / (2 × image height); width = (x2 - x1) / image width; height = (y2 - y1) / image height; The confidence is the highest similarity value calculated by the SIFT feature.

[0068] The steps to obtain the initial precision standard data are as follows: The reviewer checks the original drawing, annotation file, and template library samples to verify the bounding box offset error and class misjudgment error.

[0069] The bounding box offset is determined by checking whether the annotation box completely covers all the key structures of the accessory. Use a tool to calculate the IOU value between the annotation box and the manually redrawn standard box. When IOU < IOU threshold T7, it is determined as a serious offset; When T7 ≤ IOU ≤ IOU threshold T8, it is determined as a slight offset; when IOU > T8, it is determined as qualified.

[0070] For a slight offset, the reviewer manually adjusts the coordinates to ensure that the annotation box covers all the core structures and there is no redundant background.

[0071] For a serious offset, it is necessary to trace back to the initial recognition result, check whether the coordinates matched by the QATM algorithm are incorrect, recalculate the bounding box after correction and perform verification. If it still does not reach IOU < T7 after multiple corrections, it will be marked as invalid data and excluded.

[0072] The class misjudgment is determined by comparing the structural features of the accessory with the similar samples in the template library and checking whether the class encoding is consistent with the actual accessory. If there is a situation where the core features do not match, it is determined as a class misjudgment.

[0073] Check the core features of the accessory against the template library and correct the class encoding and corresponding labels.

[0074] If the misjudgment stems from an error in feature extraction during the initial identification process, the features of that accessory will be added to the template library to improve the accuracy of subsequent identifications.

[0075] For duplicate samples, the one with the highest similarity is retained, and the rest are discarded.

[0076] The approved samples are merged with the unsampled samples to form a preliminary precision-annotated dataset. The unsampled samples need to be batch-verified using a script to ensure format consistency.

[0077] Randomly select a sampling ratio T9 from the integrated data for sampling verification, calculate the labeling accuracy (number of qualified samples / total number of samples), and the accuracy shall not be lower than the labeling accuracy threshold T10. T9 is used to verify the accuracy of the precision annotation data, and T10 is used to ensure the quality of the precision annotation data.

[0078] If the standard is not met, the proportion of unsampled samples to be audited needs to be increased and corrections made until the standard is met.

[0079] The final initial precision annotation data is archived according to image ID, annotation file and review log, and a data traceability table is established: Class A is defined as initial qualified data, Class B is defined as corrected qualified data, and Class C is defined as discarded or invalid data, which provides a basis for data quality for subsequent model training and iteration.

[0080] Sampling verification and traceability archiving mechanisms are used to ensure the traceability of data quality and facilitate problem localization during subsequent model iterations.

[0081] The initial identification criteria acquisition steps are as follows: By using feature heatmap analysis and statistical threshold extraction methods, the core distinguishing features of various valves are located and the thresholds are quantified.

[0082] Specifically, feature localization uses Grad-CAM (Grad Weighted Class Activation Mapping) of the initial model to visualize the key response regions of different valve types.

[0083] For example, the core response area of ​​a solenoid valve is the electromagnet component and the voltage parameter marking area; a one-way valve is a conical valve core structure and a one-way flow arrow marking; and a directional valve is a multi-port orifice distribution and a valve core stepped structure.

[0084] Threshold quantization involves conducting statistical analysis on the core features of various valves in the initial precision-annotated data to extract the threshold value of the feature response.

[0085] For example, the characteristic response value of the electromagnet component of the solenoid valve is not lower than the characteristic response value threshold T11, which is used to determine the validity of the solenoid valve electromagnet component; the confidence level of voltage label text recognition is not lower than the confidence threshold T12, which is used to determine the validity of voltage label text recognition. The one-way valve's conical structure profile matching degree is not lower than the matching degree threshold T13, used to determine the validity of the one-way valve's conical structure; the arrow annotation recognition accuracy is not lower than the accuracy threshold T14, used to determine the validity of the one-way valve's arrow annotation recognition; the directional control valve's number of interface holes is not less than the quantity threshold T15, used to determine the validity of the directional control valve's interface holes; the number of valve core steps is not less than the quantity threshold T16, used to determine the validity of the directional control valve's valve core steps, and the step height difference is not lower than the directional control valve's specification. These thresholds must meet the statistical requirement that more than 95% of positive samples comply and more than 98% of negative samples do not comply.

[0086] By combining the correspondence between OCR recognition results and labeled parameters, a parameter recognition standard is constructed: By statistically analyzing the positional distribution of parameters in the initial precision-annotated data, the high-frequency regions of parameter annotation are determined, and the initial model prioritizes extracting text from these regions as parameter candidates. For parameter text in non-high-frequency areas, correlation verification is performed in conjunction with the accessory type.

[0087] Specifically, the dimensional parameters must conform to the format of numerical value plus unit, and the value must be within the parameter range deviation threshold T17 of the same type of accessory parameters. Model parameters must match a regular expression composed of letters, numbers, and symbols, and include a feature code; material parameters must be matched in a preset material dictionary, with a matching confidence level not lower than the matching confidence threshold T18.

[0088] If the parameter identification result does not meet the above standards, it will be marked as pending review and a second identification will be triggered; if the second identification still fails, the default parameters of similar accessories will be associated.

[0089] Hierarchical classification determination is performed on the classification results of the initial precision-annotated data: The primary classification, or valve category determination, is based on core structural characteristics, with priority from highest to lowest as follows: control component type, valve core structure, and number of interfaces.

[0090] For example, if the characteristic threshold of the electromagnet component is not lower than T11, it is determined to be a solenoid valve; if there is no control component but the matching degree of the conical valve core is not lower than T13, it is determined to be a check valve; if there is no control component, the valve core is cylindrical and stepped and the number of interfaces is not less than the quantity threshold T19, it is determined to be a manual directional valve.

[0091] Secondary classification, or subdivision type determination: further subdivision based on parameters and structural details.

[0092] For example, solenoid valves can be classified by voltage into DC24V solenoid valves and AC220V solenoid valves; directional valves can be classified by the number of ports into two-position four-way valves and three-position five-way valves.

[0093] When multiple features conflict, a weighted average is applied based on the weights of core structural features, parametric features, and template matching similarity. The category with the highest score is the final determination result.

[0094] The steps for obtaining extended labeled data are as follows: First, set the confidence levels for accessory category recognition (≥ confidence threshold T20), bounding box recognition (≥ confidence threshold T21), and core parameter recognition (≥ confidence threshold T22) as filtering thresholds.

[0095] If all three indicators meet the set standards, the project will be retained. If only one item fails to meet the standard, and the part is a non-core component, it can be temporarily stored. If it is a core component, it will be eliminated directly.

[0096] Then, the initial recognition rule system is invoked to verify the retained annotations.

[0097] Multi-dimensional filtering and rule validation are used to filter out low-quality labels in advance, reduce subsequent processing costs, and ensure the basic quality of expanded data.

[0098] Specifically, samples labeled as solenoid valves must include labels for the electromagnet components, and samples labeled as spool valves must be associated with structural parameters having a step count ≥ T16. Labels that fail verification are marked as needing optimization, while the remaining labels are to be masked.

[0099] Finally, model processing and mask generation are performed: The vit-h-SAM model is used. After loading pre-trained weights, the image regions of the selected parts to be masked are input into the model. The model parameters are set as follows: the point hints are the center points of the bounding boxes, and the mask confidence threshold is set to T23. A segmentation mask covering the complete outline of the parts is generated.

[0100] The initial bounding box optimization steps are as follows: The initial bounding box is optimized based on the segmentation mask. Specifically, the minimum bounding rectangle of the mask is extracted using OpenCV, and the pixel coordinates (x1, y1, x2, y2) of this rectangle are calculated and used to replace the initial bounding box output by the model.

[0101] If the mask shows that the slotted fitting, such as the valve sleeve, has a recessed structure, adjust the bounding box to cover the recessed area, ensuring that the bounding box completely covers all structural features of the fitting.

[0102] If the confidence level of the mask generated by the SAM model is less than T23, then the process should backtrack to the initial model to re-identify the component. If the confidence level of the second identification still does not meet the standard, it should be marked as requiring manual assistance and temporarily stored for further review. The steps for generating extended annotation data are as follows: The optimized labeled data is sorted by category code-confidence, mask paths are added, record fields are optimized to form extended labeled data, and it is uniformly stored in YOLOv8 format.

[0103] The steps for obtaining extended precision standard data are as follows: First, check whether the core features of the accessories in the category label match the initial standard; Secondly, review whether the parameter format and numerical range in the parameter annotations conform to the initial parameter standards; Finally, review whether the connection relationships of accessories in the structural association conform to the initial classification judgment logic.

[0104] Multi-dimensional auditing is used to comprehensively identify potential errors in the extended data and ensure that the extended data is consistent with the initial standards.

[0105] Specifically, the solenoid valve to be audited must contain electromagnet characteristics and have a confidence level ≥ T11; the dimensional parameters must include units and be within the range of ± T17 for similar accessories; the solenoid-valve core drive connection must be marked.

[0106] The deviation determination steps are as follows: If the proportion of the mask covering the core structure of the accessory is less than the coverage ratio threshold T24, it is judged as a serious deviation. If the proportion of non-accessory background area in the mask is greater than the background proportion threshold T25 or the proportion of adjacent accessories is greater than the adjacent accessory threshold T26, it is judged as a deviation. If the confidence level of the mask generated by SAM is less than T23 and the outline is not clear upon manual verification, it is determined to be an invalid segmentation.

[0107] The steps for calculating the indicator are as follows: Labeling accuracy: Labeling accuracy = Number of qualified samples / Total number of samples, refer to T10 standard; Deviation correction rate: Deviation correction rate = Number of correctable deviation samples / Total number of deviation samples; SAM segmentation efficiency: SAM segmentation efficiency = (total number of samples - number of invalid segmentation samples) / total number of samples.

[0108] The steps for correcting deviation annotations are as follows: For minor deviations, the bounding box is regenerated by manually adjusting the mask outline to ensure coverage of the core structure; For severe deviations, if they are caused by parts obstruction, mark the obstruction attribute and adjust the bounding box to the visible area; If the SAM model has poor adaptability, it should be re-annotated by manually drawing masks and bounding boxes. For invalid segmentations, the sample is directly removed.

[0109] To address the issue of misclassification, the categories were re-evaluated by comparing them with the core feature thresholds T11 and T13 of the initial identification standard, and the category codes and labels were corrected accordingly. To address parameter deviation issues, we trace back the OCR recognition records, combine them with the original drawings to complete the parameters, and correct the parameter text and confidence level. For structural association errors, the connection relationships are re-labeled according to the initial classification judgment logic, and the association basis is supplemented.

[0110] The preset threshold system and iterative optimization steps are as follows: Set core thresholds T20, T21, and T22, along with auxiliary thresholds, based on model training requirements; If all three core thresholds are met, and the compliance rate of the auxiliary thresholds is greater than or equal to the set value of the auxiliary thresholds, the sampling review is deemed to have passed; otherwise, iterative optimization is triggered.

[0111] Iterative optimization mechanisms are used to dynamically improve the quality of extended data, ensuring that it can effectively supplement the coverage of the initial data.

[0112] The iterative optimization process is as follows: To pinpoint the root cause of non-compliant indicators, if the labeling accuracy is low, check whether the initial standard does not cover the characteristics of the new scenario. If it does not, it needs to be supplemented. If the SAM segmentation efficiency is low, analyze whether it is a sample quality issue. If it is a sample quality issue, it is necessary to re-screen new samples or adjust the SAM model parameters. Based on the analysis results, optimize the full expanded labeled data: after adjusting the SAM parameters, regenerate the segmentation results of low-confidence masked samples; remove invalid samples that still fail to meet the standards after two consecutive corrections.

[0113] Once the standard is met, the qualified samples and the optimized samples that were not sampled are integrated to generate extended precision labeled data.

[0114] The steps for obtaining comprehensive precision annotation data are as follows: Hierarchical deduplication is performed using multi-dimensional feature matching. First, a three-dimensional feature vector is constructed for each labeled sample: Image features: Extract SIFT features from the image regions of the labeled accessories and perform normalization processing; Labeling features: Extracting category codes, bounding box IOU baseline values, and parameter numerical hash values; Attribute features: encoded values ​​of key fields such as scene label, data source, and recognition confidence level.

[0115] Based on the model training requirements, two levels of thresholds are set: a threshold for repeated samples (T27) and a threshold for highly similar samples (T28). The thresholds are validated using a test set of known similar samples.

[0116] Perform layered deduplication: The first layer of deduplication involves generating a unique identifier string by combining three-dimensional feature vectors and encrypting it into a 32-bit hash value using MD5. The hash values ​​of all samples are traversed, and samples with the same hash value are identified as duplicate samples. The sample with the best combination of data source priority and confidence priority is retained, and the remaining samples are removed and the deduplication log is recorded.

[0117] The second layer of deduplication involves grouping samples by category encoding and scene label. Feature vector matching is performed only on samples within the same group, with samples ordered in the order of initial data first, followed by extended data. A comprehensive similarity is calculated using the formula: Image feature cosine similarity × 0.6 + Label feature matching degree × 0.4. A comprehensive similarity ≥ T28 is considered highly similar. For similar sample groups, the sample with the highest confidence, highest bounding box coverage, and highest parameter completeness is retained, while the rest are discarded. If a similar sample group includes extended data, the extended samples are retained first.

[0118] Finally, quantitative indicators are calculated: Calculate the amount of data after deduplication; Calculate the deduplication rate: number of deduplicated samples / total number of samples; Calculate the variance of the category distribution; Considering the fluctuation of the proportion of each category before and after deduplication, the fluctuation range is ≤ the proportion fluctuation threshold T29 and the scene coverage.

[0119] The initial data and the extended data after deduplication are merged. For samples with the same image ID, the extended data is processed and merged according to the principle of covering the initial data, and finally the comprehensive accuracy annotation data is output.

[0120] The model iteration steps are as follows: When the model training reaches the preset maximum number of rounds, or when the core metric reaches the main threshold after a certain round of training, a full threshold verification is triggered: Confirm that the total amount of data and the sample size of each category meet the standards; Verify the convergence threshold and analyze the deviation between the iteration curve and the training and test sets; Verify the core thresholds T10 and T24 to evaluate the model's performance metrics.

[0121] If the threshold is not reached, the following steps will be taken: If the basic threshold is not met, supplement the extended data and re-execute the data merging process; If the convergence threshold is not met, advance the round and increase the regularization weight; If the core threshold is not met, for the metrics that fail to meet the target, supplement the sample size by T30, add parameter-labeled samples, and re-execute fine-tuning training. T30 is used to increase the number of supplementary parameter-labeled samples for the model's core metrics.

[0122] Example 2, S1: Select 100 directional control valve drawings produced by a certain machinery factory as samples, covering 8 mainstream types such as DC24V solenoid valve, AC220V solenoid valve, straight-through check valve, and two-position four-way manual directional valve. The resolution of each drawing is uniformly adjusted to 1920×1080 pixels, named according to equipment number-drawing type-serial number (e.g., M01-solenoid valve-001), and archived in a local folder according to valve type.

[0123] S2: Acquisition of initial recognition results S21: Template Library Construction Collect standard drawings of core components such as valve cores (slide valve type, conical), valve sleeves, and electromagnets (DC / AC), extract features such as edge contours and orifice dimensions, and construct a tagged component template library. Among them, the slide valve core template is marked with cylindrical steps + ≥2 sealing ring grooves, and the DC electromagnet template is associated with 24V voltage labeling attributes.

[0124] S22: Feature Matching and Verification The QATM algorithm is employed, using sample drawings as query images and a template library as template images, dynamically adjusting the feature vector dimensions (512 dimensions for complex components such as electromagnets, and 256 dimensions for simple components such as springs). A similarity threshold T1=0.85 and an IOU threshold T2=0.7 are set to perform double verification on the matching results, eliminating false matches caused by similar valve sleeve specifications.

[0125] S23: Multi-model joint recognition Structural features: Use the YOLO model to locate the number of valve core steps, valve hole distribution, etc., and output structural feature labels and location information.

[0126] Parameter labeling: The text area is located by the EAST text detector, and the TesseractOCR engine recognizes parameters such as rated voltage and interface diameter, and the format is standardized as numerical value + unit (such as 24V, 15mm).

[0127] Topological association: U-Net segments the connecting lines, and graph theory algorithms are used to construct a topological relationship graph based on the coordinates of the components, marking the driving connection relationship between the electromagnet and the valve core.

[0128] S24: Three-level classification matching Primary classification: If electromagnet characteristics are detected (response value ≥ 0.8), it is determined to be a solenoid valve; if there are no control components and the conical valve core matching degree is ≥ 0.85, it is determined to be a check valve.

[0129] Secondary classification: Solenoid valves are classified into DC24V and AC220V types according to voltage. The spool valve core needs to be matched with a cylindrical step + 3 sealing ring grooves.

[0130] Level 3 classification: Verified parameter deviation ≤ ±10% (T4=10%), such as DC24V solenoid valve rated voltage identified as 23.5V, which is considered valid; interface diameter identified as 18mm (similar templates in the library are 15mm), which marks the parameter as abnormal.

[0131] S25 outputs the initial recognition results in the form of a structured dictionary. S3 Initial Annotation Data Conversion S31 Preset Annotation Rules Application Category code: Control component (01) → Electromagnet (0101) → DC electromagnet (010101).

[0132] Bounding box: Based on the principle of minimum bounding rectangle, the bounding box coordinates of the slide valve core are (380, 100, 520, 250), with no obstruction markers.

[0133] Feature attribute binding: DC electromagnet associated material: copper, rated voltage: 24V, matching similarity: 0.92; valve core associated material: steel, diameter: 25mm, number of steps: 3.

[0134] S32: Data Parsing and Verification Extract image IDs, parts list and other data to generate intermediate tables, verify that the bounding boxes do not exceed the image range and that there are no unmatched items in the category codes, with a verification pass rate of ≥98% (T5=98%), generate annotation files corresponding to single drawings, and summarize the list records M01-solenoid valve-001 annotation for 2 parts, with no abnormal data.

[0135] S33: Feature Extraction and Template Alignment The SIFT algorithm of OpenCV (X1=6, X2=1.6, X3=0.03) is used to extract the feature vector of the accessory. The cosine similarity is calculated with the vector of the same category in the template library (such as DC electromagnet 010101). T6=0.8 is set, and samples with similarity ≥0.8 are selected. The highest similarity and template ID are recorded.

[0136] S34: Format Conversion Generate annotation files in YOLOv8 format. Example: 00.310.210.130.150.92 (Category index 0 corresponds to a DC electromagnet, normalized coordinates and confidence level).

[0137] S4: Acquisition of initial precision annotation data S41: Manual Sampling Audit Twenty drawings were randomly selected at a sampling rate of T9=20% for review of the boundary frames and parameters. Bounding box: The IOU of the electromagnet annotation box of a certain solenoid valve is 0.75 (T7=0.6, T8=0.8), which is determined to be slightly off. The coordinates are manually adjusted to (115,75,355,225).

[0138] Category misclassification: One drawing misclassified an AC electromagnet as a DC electromagnet. The corrected category code is 010102, and the feature is added to the template library.

[0139] Manual sampling and review compensate for the algorithm's deviations in bounding box localization and category determination, improve the accuracy of labeled data, and provide high-quality data support for model training.

[0140] S42: Accuracy Verification The accuracy rate of the annotation is calculated as 19 / 20 = 95% (T10 = 95%). The archived data includes 68 items in category A, 27 items in category B, and 5 items in category C. A data traceability table is established.

[0141] S5: Acquisition of Initial Identification Standards S51: Core Feature Localization and Threshold Quantization Feature localization: Visualized using Grad-CAM, the core response area of ​​the solenoid valve is the electromagnet (T11=0.8) and the voltage label area (T12=0.9), while the one-way valve is the conical valve core (T13=0.85) and the arrow label (T14=95%).

[0142] Threshold settings: Number of reversing valve interface holes T15=4, Number of valve core steps T16=2, Parameter range deviation T17=±10%, Material matching confidence level T18=0.85.

[0143] S52: Construction of Parameter Identification Standards Size parameters: must conform to the numerical value + unit, such as 15mm (within the range of 12-18mm for similar accessories).

[0144] Model parameters: Matching letter + number regular expression (e.g., DFB-25), including the series feature code DFB.

[0145] S53: Hierarchical Classification Determination Primary classification: Electromagnet present (T11≥0.8) → Solenoid valve; No control component + conical valve core (T13≥0.85) → Check valve.

[0146] Secondary classification: Solenoid valves are further subdivided by voltage (DC24V, AC220V), and directional valves are classified by the number of ports (two-position four-way, three-position five-way).

[0147] Conflict handling: The core structure weight is 0.5, the parameter weight is 0.3, and the matching similarity weight is 0.2. The category with the highest weighted score is the final category.

[0148] S6: Extended Annotation Data Acquisition S61: Setting the Filter Threshold Set T20=T21=T22=0.85. Of the 30 drawings, 25 drawings that meet all three indicators are retained, 3 drawings of non-core components that do not meet one indicator are temporarily stored, and 2 drawings of core components that do not meet the standard are removed.

[0149] S62: Rule Validation All solenoid valve samples were labeled with electromagnets. The number of steps in the spool valve core was ≥2 (T16=2). Two samples failed and were marked as needing optimization. The remaining 23 samples were marked as needing masking.

[0150] S63: SAM Model Processing Load the vit-h-SAM pre-trained weights, use the point hints as the center point of the bounding box, set the mask confidence T23=0.8, generate a complete valve sleeve contour mask, and optimize the bounding box coordinates (e.g., adjust the original valve sleeve coordinates (90,60,400,260) to (85,55,410,270) to cover the concave structure).

[0151] S64: Generate extended annotation data Sort by category code-confidence, supplement the mask path, and store in YOLOv8 format. Example: 10.280.320.210.230.89 (category index 1 corresponds to valve sleeve).

[0152] S7: Acquisition of Extended Precision Annotated Data S71: Three rounds of review Category labeling: All solenoid valves contain electromagnet characteristics (confidence level ≥ 0.8), and there are no misclassifications.

[0153] Parameter annotation: All dimensional parameters include units and are within ±10%.

[0154] Structural correlation: Both the electromagnet and the valve core are marked with drive connection, which is logical.

[0155] S72: Deviation judgment Severe bias: One sample mask covers 75% of the core structure (T24=80%) → severe bias.

[0156] Common bias: 18% of the background in 2 samples (T25=15%) → bias.

[0157] Invalid segmentation: No sample mask confidence < 0.8, SAM segmentation success rate 92%.

[0158] S73: Correction and Indicator Calculation Minor deviations: Manually adjust the mask and regenerate the bounding box; severe deviations: Mark the occlusion attributes and adjust the bounding box to the visible area.

[0159] Indicators: 96% labeling accuracy, 88% deviation correction rate, and 92% SAM segmentation effectiveness. Once these indicators are met, extended precision labeling data will be generated.

[0160] S8: Acquisition of Comprehensive Precision Annotation Data S81: Layered deduplication First layer: MD5 hash value deduplication, removing 3 completely duplicate samples.

[0161] Second layer: Calculate the comprehensive similarity within the same category and scene (image features 0.6 + labeled features 0.4), T28=0.9, remove 5 highly similar samples, and prioritize the retention of extended data.

[0162] S82: Quantitative Indicators and Merging After deduplication, there are 92 samples, with a deduplication rate of 8%. The variance of the category distribution is ≤0.05, and the fluctuation range is ≤5% (T29=5%).

[0163] The initial and extended data are merged, and images with the same ID are processed according to the principle that the extended data overwrites the initial data, and the comprehensive accuracy annotation data is output.

[0164] S9: Model Iteration and Output S91: Three-round iterative optimization First iteration: The accuracy of the core metrics was 88%, which did not reach the 95% main threshold. Ten additional parameter annotation samples were added (T30=10).

[0165] Second iteration: Accuracy 93%, regularization weights adjusted, model convergence optimized.

[0166] Three iterations: accuracy of 96.5%, core thresholds T10=95% and T24=80% both met, and the convergence curve was stable.

[0167] S92: Final Model Output The system outputs an intelligent recognition model for the directional control valve drawings of the machinery plant. It can accurately identify the accessory list, structural features, key parameters and topological relationships of 8 types of valves, with a recognition accuracy of 96.5%, a bounding box positioning error of ≤5 pixels, and a parameter recognition error of ≤±10%.

[0168] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0169] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A smart identification system for control valve drawing parts, characterized in that, Includes the following steps: S1: Obtain the drawing sample of the directional control valve; S2: Obtain the initial recognition result based on the drawing sample; S3: Transform the initial labeled data based on the initial identification results; S4: Based on the initial annotation data, initial precision annotation data is obtained through manual sampling and review; S5: Generate initial recognition criteria based on the initial precision annotation data using the initial model; S6: Calculate extended annotation data based on the initial identification criteria; S7: Based on the extended annotation data, further extended precision annotation data is obtained through manual sampling and review; S8: Obtain comprehensive precision annotation data based on the initial precision annotation data and the extended precision annotation data; S9: Based on the comprehensive accuracy annotation data, the initial model is iterated and the final detection model is output.

2. The intelligent identification system for control valve drawing parts according to claim 1, characterized in that, S2 includes: Standard drawings of core components of directional control valves are collected, classified, and archived. After normalization of the component drawings, core features are extracted to construct a tagged component template library. The QATM algorithm is used to extract local feature points from the query image and template image to generate high-dimensional feature description vectors. The matching degree is calculated using the cosine similarity algorithm. A similarity threshold T1 is set, and the IOU threshold T2 is used to verify and output a list of validly matched components and their annotation information. T1 is the feature vector matching degree screening threshold; T2 is the IOU verification threshold.

3. The intelligent identification system for control valve drawing parts according to claim 2, characterized in that, S2 further includes: Structural feature information is detected using the YOLO model, parameter annotation information is identified using a combination of the EAST text detector and the TesseractOCR engine, and topological association information is constructed using the U-Net image segmentation model and graph theory algorithm. Cross-validation is performed on the three types of information, a positional overlap threshold T3 is set, and a set of structured key information is output. Based on this information set, a three-level classification and matching system is constructed to output the initial recognition results. T3 is the positional overlap threshold.

4. The intelligent identification system for control valve drawing parts according to claim 1, characterized in that, S3 includes: Based on preset target category rules, bounding box annotation rules, and feature attribute association rules, data from the initial recognition results are extracted to generate an intermediate data table; the converted annotation data is automatically verified, a verification pass rate threshold T5 is set, and after verification, it is stored in the form of one annotation file corresponding to one drawing, and the accessory image area is cropped, normalized, and grayscale processed. T5 is the threshold for the automatic verification pass rate of labeled data.

5. The intelligent identification system for control valve drawing parts according to claim 4, characterized in that, S3 further includes: The SIFT algorithm of OpenCV is used to extract the feature vectors of the accessory images. The cosine similarity is calculated with the feature vectors of templates encoded in the same category in the template library. A similarity threshold T6 is set, and samples that meet the threshold are selected. The labeled data is then converted into YOLOv8 format. T6 is the matching and filtering threshold between the accessory feature vector and the template feature vector.

6. The intelligent identification system for control valve drawing parts according to claim 1, characterized in that, S4 includes: Auditors compared the original drawings, annotation files, and template library samples to check and correct bounding box offset errors and category misjudgments. They set IOU thresholds T7 and T8, sampling ratio T9, and annotation accuracy threshold T10. IOU < T7 was considered a serious offset, and IOU > T8 was considered acceptable. The integrated data was sampled for verification. After meeting the standards, the initial accuracy annotation data was archived according to image ID, annotation file, and audit log. T7 is the threshold for judging severe boundary box offset, T8 is the threshold for judging qualified boundary box, T9 is the threshold for sampling verification ratio, and T10 is the threshold for achieving the standard of annotation accuracy.

7. The intelligent identification system for control valve drawing parts according to claim 1, characterized in that, S5 includes: Grad-CAM is used to visualize and locate the core distinguishing features of various valves. Statistical analysis is carried out on the core features in the initial precision annotation data. Thresholds T11, T12, T13, T14, T15, T16, T17, T18, and T19 are set to quantify the feature response value thresholds. Parameter identification standards are constructed to perform primary and secondary hierarchical classification judgment on the classification results of the initial precision annotation data. T11 is the threshold value of the characteristic response value of the solenoid valve electromagnet component; T12 is the confidence threshold value of voltage annotation text recognition; T13 is the matching degree threshold value of the one-way valve cone structure; T14 is the accuracy threshold value of the one-way valve arrow annotation recognition; T15 is the threshold value of the number of port positions of the reversing valve interface; T16 is the threshold value of the number of steps of the reversing valve core; T17 is the parameter range deviation threshold value; T18 is the confidence threshold value of material matching; and T19 is the threshold value of the number of manual reversing valve interfaces.

8. The intelligent identification system for control valve drawing parts according to claim 1, characterized in that, S6 includes: Set confidence thresholds T20 for accessory category identification, T21 for bounding box identification, and T22 for core parameter identification as filtering thresholds to filter and validate the labeled data; generate a segmentation mask covering the complete outline of the accessory using the vit-h-SAM model, set a mask confidence threshold T23, optimize the initial bounding box based on the segmentation mask, generate extended labeled data and store it in YOLOv8 format; T20 is the confidence threshold for accessory category identification, T21 is the confidence threshold for bounding box identification, T22 is the confidence threshold for core parameter identification, and T23 is the confidence threshold for mask identification.

9. The intelligent identification system for control valve drawing parts according to claim 1, characterized in that, S8 includes: For each labeled sample, a three-dimensional feature vector composed of image features, label features, and attribute features is constructed; a threshold of duplicate samples T27, a threshold of highly similar samples T28, and a threshold of proportion fluctuation T29 are set; deduplication is performed by using MD5 hash value and comprehensive similarity calculation; the deduplicated initial data and extended data are merged; and samples with the same image ID are processed according to the principle that the extended data covers the initial data. T27 is the threshold for determining duplicate samples, T28 is the threshold for determining highly similar samples, and T29 is the threshold for class proportion fluctuation.

10. The intelligent identification system for control valve drawing parts according to claim 1, characterized in that, S9 includes: The initial model is iterated 3 times. During the iteration, the convergence threshold and core threshold are verified. The coverage ratio threshold T24, background ratio threshold T25, adjacent component ratio threshold T26, and supplementary quantity T30 are set. If the target is not met, the model is optimized by supplementing extended data, adjusting regularization weights, or supplementing samples until the core indicators meet the target and the final detection model is output. T24 is the threshold for the proportion of the mask covering the core structure, T25 is the threshold for the proportion of the mask background, T26 is the threshold for the proportion of the mask adjacent components, and T30 is the threshold for the number of supplementary parameter annotation samples.

Images