Search

CN-121981994-A - Workpiece quality detection method based on industrial CT and deep learning

CN121981994ACN 121981994 ACN121981994 ACN 121981994ACN-121981994-A

Abstract

The invention relates to the technical field of industrial nondestructive testing, in particular to a workpiece quality detection method based on industrial CT and deep learning, which comprises the following steps of establishing a workpiece multidimensional information model, adapting high-low energy rays according to material partition, carrying out partition scanning after multi-objective optimized scanning parameters, registering and fusing to obtain enhanced two-dimensional projection data, establishing an end-to-end artifact removal and reconstruction cooperative model, outputting artifact-free three-dimensional data, extracting bimodal features, standardizing, inputting a preset multimodal fusion detection model, outputting relevant defect information, quantifying defect parameters according to the relevant defect information, combining workpiece service conditions, carrying out sequential prediction and finite element simulation coupling assessment, and outputting comprehensive quality grades. The invention improves imaging and detection precision through techniques such as material partition adaptive ray, deep learning and the like, combines simulation and blockchain visualization, and realizes full life cycle control traceability.

Inventors

  • TANG QIBO
  • ZHANG LIKAI
  • HUANG XIAOJIE
  • Bu Tianliang

Assignees

  • 上海恩迪检测控制技术有限公司

Dates

Publication Date
20260505
Application Date
20260126

Claims (10)

  1. 1. The workpiece quality detection method based on industrial CT and deep learning is characterized by comprising the following steps of: Establishing a workpiece multidimensional information model, adapting high-low energy rays according to material partition, performing partition scanning after multi-objective optimization of scanning parameters, and registering and fusing to obtain enhanced two-dimensional projection data; constructing an end-to-end artifact removal and reconstruction cooperative model, inputting enhanced two-dimensional projection data, and outputting artifact-free three-dimensional volume data; Based on artifact-free three-dimensional data, extracting and standardizing bimodal features, inputting a preset multimodal fusion detection model, and outputting defect related information; quantifying defect parameters according to the defect related information, combining the service working condition of the workpiece, and outputting comprehensive quality grade through time sequence prediction and finite element simulation coupling evaluation; and integrating all-link detection data, processing the data by a three-dimensional visualization platform, and storing the data into a blockchain database to realize full life cycle tracing.
  2. 2. The workpiece quality detection method based on industrial CT and deep learning as claimed in claim 1, wherein the specific steps of establishing a workpiece multidimensional information model and adapting high-low energy rays according to material partition are as follows: Basic information of a workpiece to be detected is collected, wherein the basic information comprises material composition, size parameters, preset detection precision and service conditions, various information is classified and sorted, information association relations are defined, and a complete multidimensional information model is constructed; Extracting core characteristics of each material based on material composition information in the multidimensional information model, dividing material areas by adopting an improved K-means clustering algorithm, optimizing a clustering center initialization mode, determining boundary ranges of each material area through iterative calculation, and finishing material partition of a workpiece; Based on the material partition of the workpiece, the corresponding high-low energy ray adaptation range is matched according to the density and attenuation characteristics of different material areas.
  3. 3. The workpiece quality detection method based on industrial CT and deep learning according to claim 2, wherein the specific steps of obtaining the enhanced two-dimensional projection data by registration fusion are as follows: Determining a plurality of targets for optimizing the scanning parameters, wherein the targets are used as constraint conditions, and a particle swarm optimization algorithm is adopted to optimize the scanning parameters, wherein the targets ensure that the detection precision meets the preset requirement, the radiation dose is controlled within a safety range, and the scanning efficiency is improved; According to the optimized scanning parameters and the ray energy corresponding to each material area, controlling the industrial CT equipment to switch high-low energy ray sources, carrying out partitioned scanning on a workpiece, and acquiring high-low energy two-dimensional projection data; registering the high-energy two-dimensional projection data and the low-energy two-dimensional projection data by adopting a mutual information registration algorithm, and eliminating data dislocation caused by scanning angle deviation and detector response difference; the registered projection data is input into a deep learning fusion model based on a U-Net architecture, and the enhanced two-dimensional projection data with high penetrability and high resolution is output through the feature extraction and fusion capability of the model.
  4. 4. The method for detecting the quality of a workpiece based on industrial CT and deep learning as claimed in claim 1, wherein the specific steps of constructing an end-to-end artifact removal and reconstruction cooperative model and outputting artifact-free three-dimensional volume data are as follows: the end-to-end artifact removal and reconstruction collaborative model sequentially comprises a projection data coding module, an attention gating artifact characteristic extraction module and an improved 3D full convolution network reconstruction module, and the connection relation and the data transmission path of each module are defined; constructing a labeling data set required by model training, wherein the labeling data set comprises projection data of a real workpiece and projection data after common artifacts are manually added, and simultaneously labeling corresponding artifact-free standard data; Training an end-to-end artifact removal and reconstruction cooperative model by adopting a mixed loss function, monitoring the performance of the model through a verification set in the training process, and adopting an early-stop strategy to prevent overfitting when the performance of the model tends to be stable, so as to obtain an optimal cooperative model after training; the enhanced two-dimensional projection data is input into an optimal collaborative model, and artifact-free three-dimensional volume data meeting the requirements of ultra-micro defect detection is directly output through the processing flows of model coding, artifact removal, reconstruction and the like.
  5. 5. The method for detecting the quality of a workpiece based on industrial CT and deep learning according to claim 1, wherein the specific steps of outputting the defect-related information are as follows: extracting the gray gradient characteristics related to the defects by adopting a 3DSobel operator based on the artifact-free three-dimensional volume data; based on the original data obtained by dual-energy scanning, calculating the material density related parameters corresponding to each voxel through a material decomposition algorithm, and extracting the material density characteristics; respectively carrying out standardization treatment on the gray gradient characteristics and the material density characteristics, eliminating dimension differences among the characteristics, and forming a standardized bimodal characteristic set; Constructing a hybrid architecture defect detection model based on a transducer-3 DCNN, wherein the hybrid architecture defect detection model comprises a parallel feature coding branch, a cross-mode fusion layer and a defect detection head, functions and cooperation logic of all parts are defined, the parallel feature coding branch extracts a bimodal feature set, an extraction result is transmitted to the cross-mode fusion layer for fusion, and the fused features are transmitted to the defect detection head; Dividing the bimodal feature set into a training set, a verification set and a test set, training the mixed architecture defect detection model by using the training set, adjusting model parameters by using the verification set, verifying model performance by using the test set, and completing training optimization to obtain an optimal defect detection model; And inputting the bimodal feature set into an optimal defect detection model, screening a detection result according to a preset confidence threshold, and outputting the type, confidence and three-dimensional boundary frame coordinate related information of the defect.
  6. 6. The method for detecting the quality of a workpiece based on industrial CT and deep learning as recited in claim 5, wherein the step of quantifying the defect parameters according to the defect-related information comprises the steps of: Accurately positioning each defect area in the artifact-free three-dimensional volume data according to the coordinates of the defect three-dimensional boundary frame, carrying out fine analysis on the defect area by adopting a three-dimensional surface fitting algorithm, and extracting various quantization parameters of the defect; collecting CT detection data of the same type of workpiece in different service periods, extracting quantization parameters of defects in each period, and constructing a time sequence defect data set by combining service working condition parameters of the workpiece; training a bidirectional LSTM network based on the time sequence defect data set, determining structural parameters and training strategies of the network, and learning defect growth rules through historical data in the time sequence defect data set to obtain a defect growth prediction model; Inputting the quantization parameter of the current defect and the service condition parameter of the workpiece into a defect growth prediction model, and obtaining defect growth prediction results of different future service periods according to the growth rule learned by the model; and constructing a mechanical analysis model of the workpiece by combining a finite element simulation technology, inputting quantization parameters of defects and material mechanical properties of the workpiece, calculating critical failure conditions corresponding to the defects, comparing a predicted result with the critical failure conditions, and outputting the residual service life and reliability grade of the workpiece.
  7. 7. The method for detecting the quality of the workpiece based on industrial CT and deep learning as set forth in claim 6, wherein the specific steps of outputting the comprehensive quality level are as follows: referring to quality detection standards of corresponding industries, and carrying out grading judgment on the severity of each defect by combining the specific conditions of defect quantification parameters; establishing a quality assessment matrix, wherein the matrix takes the defect severity level as a transverse dimension, takes the reliability level as a longitudinal dimension, and defines quality level mapping relations corresponding to different dimensional combinations; matching corresponding comprehensive quality grades in a quality assessment matrix according to the defect severity grading result and the reliability grade assessment result; and respectively making subsequent treatment strategies for the workpieces with different comprehensive quality grades, wherein the subsequent treatment strategies comprise passing of qualified workpieces, rechecking plans of the workpieces to be monitored, correction directions of the workpieces to be reworked and treatment modes of unqualified workpieces.
  8. 8. The method for detecting the quality of the workpiece based on industrial CT and deep learning as set forth in claim 1, wherein the specific steps of constructing the three-dimensional visualization platform are as follows: developing a three-dimensional visual platform based on an adaptive technical framework, and defining an overall framework and a core function module of the platform, wherein the overall framework and the core function module comprise a data import module, a three-dimensional display module, an interactive operation module and a data export module; Importing artifact-free three-dimensional data, defect-related information and quality grade assessment results into a three-dimensional visualization platform, and realizing visual presentation of the data through a three-dimensional display module of the platform; Developing a defect position interactive positioning function, supporting a user to accurately view the spatial position of the defect and the relation between the defect and a peripheral structure through various operation modes, and switching different visual angles to perform omnibearing observation; Developing a defect growth trend dynamic simulation function, displaying the change process of defects in different service periods by taking time as a dimension, and supporting flexible control of the simulation process; The detection data query function is developed, the user is supported to quickly query target detection data according to various search conditions, and meanwhile, the exporting function of a detection result is provided, so that the follow-up analysis and archiving are convenient.
  9. 9. The method for detecting the quality of the workpiece based on industrial CT and deep learning as set forth in claim 7, wherein the specific steps of implementing full life cycle tracing in the blockchain database are as follows: The key data of the full link detection is combed, wherein the key data comprise scanning parameters, various projection data, three-dimensional data, feature sets, defect detection results, quantization parameters, prediction results, quality grades, model related parameters and training logs; constructing a blockchain database of a alliance chain architecture, and determining the type and authority of the participating nodes, wherein the type and authority comprise a detection mechanism, a production enterprise and a supervision department, and an identity authentication mechanism is configured for each node; carrying out encryption processing on data to be stored by adopting an encryption algorithm, and defining the format and rule of data storage, wherein each data block comprises the associated information, the time stamp, the data abstract and the node signature of the previous data block; Storing the encrypted full-link key data into a blockchain database according to a preset rule, and realizing the non-falsification of the data through a chained storage structure of the blockchain; and a data tracing mechanism is established, which supports the reverse tracing from the detection result to the original scanning data and the related information of each processing step, and the tracing path and the query mode are defined, so that the whole life cycle of the detection data can be traced.
  10. 10. The method for detecting the quality of the workpiece based on industrial CT and deep learning as set forth in claim 1, further comprising the steps of flow control and optimization adjustment, specifically including the following steps: Formulating operation specifications and quality control standards of each detection step, and defining input requirements, processing flows, output indexes and judgment basis of each step to ensure standardization of the detection process; monitoring the running state and the output result of each step in real time in the detection process, comparing the actual result with a preset standard, finding out deviation in time and analyzing the reason; aiming at the problems of deviation or insufficient performance in the detection process, a corresponding optimization adjustment scheme is formulated, including parameter correction, model iterative optimization and operation flow improvement modes; and periodically collecting detection data and applying feedback information, comprehensively evaluating the performance of the whole detection method, and continuously optimizing the technical scheme of each link according to the evaluation result to improve the stability, accuracy and applicability of the detection method.

Description

Workpiece quality detection method based on industrial CT and deep learning Technical Field The invention relates to the technical field of industrial nondestructive testing, in particular to a workpiece quality detection method based on industrial CT and deep learning. Background Industrial CT is a core technology for realizing nondestructive inspection of a workpiece based on an X-ray or gamma-ray penetration imaging principle and combining a computer data processing and three-dimensional reconstruction technology, and is widely applied to fields of precision manufacture, aerospace, high-end equipment and the like, wherein rays are emitted by a ray source to penetrate the workpiece, two-dimensional projection data of different angles are acquired by a detector, and three-dimensional volume data of the workpiece is generated through reconstruction algorithm processing, so that internal structure visualization and defect detection are realized. The conventional process comprises setting single energy scanning parameters based on experience or a traditional algorithm, carrying out integral scanning acquisition projection data on a workpiece, processing by a traditional registration algorithm, completing artifact correction and three-dimensional reconstruction by adopting a step-by-step method, extracting gray features from the reconstructed three-dimensional data, inputting a single-structure deep learning model for defect identification, and finally storing detection data in a traditional database to realize basic tracing. However, the prior art still has a plurality of difficult-to-break limitations that a single energy scanning mode is not combined with the partition characteristics of workpiece materials, the adaptive penetrating capacity and imaging resolution cannot be provided for different areas of heterogeneous material or composite material workpieces, so that the quality of projection data is poor, artifact correction and three-dimensional reconstruction are mutually independent, collaborative optimization is not formed, residual ray hardening artifacts and scattering artifacts are easy to interfere with subsequent defect feature extraction, the ultra-tiny defect detection requirement is difficult to meet, a detection model is built only by relying on gray level single features, key attribute information such as material density is not fused, the detection generalization of different types of workpieces is poor, the detection result can only reflect the current defect state, the prediction capacity of defect service growth trend is lacking, the quality control of the whole life cycle of workpieces cannot be supported, and the traditional database storage mode has data tampering risks, and is difficult to adapt to the requirements of high reliability of detection data in the fields such as aerospace, high-end automobile parts and the like. Disclosure of Invention Aiming at the technical defects in the background technology, the invention provides a workpiece quality detection method based on industrial CT and deep learning, which solves the technical problems and meets the actual demands, and the specific technical scheme is as follows: a workpiece quality detection method based on industrial CT and deep learning comprises the following steps: Establishing a workpiece multidimensional information model, adapting high-low energy rays according to material partition, performing partition scanning after multi-objective optimization of scanning parameters, and registering and fusing to obtain enhanced two-dimensional projection data; constructing an end-to-end artifact removal and reconstruction cooperative model, inputting enhanced two-dimensional projection data, and outputting artifact-free three-dimensional volume data; Based on artifact-free three-dimensional data, extracting and standardizing bimodal features, inputting a preset multimodal fusion detection model, and outputting defect related information; quantifying defect parameters according to the defect related information, combining the service working condition of the workpiece, and outputting comprehensive quality grade through time sequence prediction and finite element simulation coupling evaluation; and integrating all-link detection data, processing the data by a three-dimensional visualization platform, and storing the data into a blockchain database to realize full life cycle tracing. Further, the specific steps of establishing the workpiece multidimensional information model and adapting the high-low energy rays according to the material partition are as follows: Basic information of a workpiece to be detected is collected, wherein the basic information comprises material composition, size parameters, preset detection precision and service conditions, various information is classified and sorted, information association relations are defined, and a complete multidimensional information model is constructed; Extracting core characte