Full-process automatic intelligent system for detecting low-pressure turbine shaft by ultrasonic water immersion method

Abstract

The invention provides a full-process automatic intelligent system and method for detecting a low-pressure turbine shaft through an ultrasonic water immersion method, and belongs to the technical field of industrial nondestructive detection. In order to solve the problems that the detection process is tedious and a model cannot be continuously optimized, the system comprises a data standardization and preprocessing module for converting an original waveform signal into a standardized visual image; the artificial intelligence interpretation module is used for analyzing the image to identify defects and automatically generating detection information; the model optimization and self-learning module associates an interpretation result with a subsequent physical verification result and optimizes the model according to the difference; and the intelligent judgment and application interface module schedules the detection process and executes a decision according to the confidence coefficient. According to the invention, full-process automation is realized, the detection efficiency and accuracy are improved, and a closed-loop self-learning mechanism is constructed, so that the system can continuously self-evolve, and effective precipitation and intelligent reuse of expert experience are realized.

Description

Technical Field

[0001] This invention relates to the field of industrial nondestructive testing technology, specifically to a fully automated intelligent system and method for ultrasonic immersion testing of low-pressure turbine shafts. More particularly, it relates to an ultrasonic immersion testing system and method integrating artificial intelligence technology. Background Technology

[0002] In the field of industrial nondestructive testing, especially for quality control of high-value critical components such as low-pressure turbine shafts of aero-engines, ultrasonic testing is a core and indispensable technical means. In recent years, in order to improve testing efficiency and automation, the industry has begun to introduce artificial intelligence technology, such as using deep learning models to analyze C-scan images generated by ultrasonic testing to automatically identify potential defects.

[0003] Patent document CN116818891A discloses an ultrasonic defect detection and classification system using machine learning. The system includes one or more computing devices accessing scan data acquired during an ultrasonic scan of an object. A first input feature map, comprising a two-dimensional (2D) scan image, is constructed based on the scan data and fed into a first deep neural network to generate a first output feature map. A second input feature map, comprising an image of a defect-free object, is fed into a second deep neural network having the same structure and weight values ​​as the first deep neural network to produce a second output feature map. A large distance between the first and second output feature maps indicates that the scanned object contains a defect.

[0004] However, this patent document still has several significant shortcomings in practical applications. First, its automation process is incomplete. The identification results of the artificial intelligence model are usually only the location and preliminary classification of defects. Subsequent tedious tasks such as defect size measurement, detection parameter verification, and test report writing still need to be completed manually by inspection engineers, for example, by recording data through screenshots or copying data. The overall inspection process is still time-consuming and labor-intensive, with limited efficiency improvement. Second, the existing system is an open-loop system. Once the artificial intelligence model is deployed, its performance is usually fixed. In inspection practice, for some suspected defects identified by artificial intelligence, it is often necessary to conduct subsequent physical destructive testing, such as metallographic analysis, to ultimately confirm their nature. The existing technology fails to feed these more authoritative physical verification results back to the artificial intelligence system, causing the model to be unable to learn and evolve from real inspection results. Knowledge and experience cannot be accumulated and iterated in the system, making it difficult to cope with new defect types or pseudo-defect patterns. The long-term performance and reliability of the system cannot be continuously improved.

[0005] Therefore, the market needs a fully automated intelligent system and method for ultrasonic immersion testing of low-pressure turbine shafts that can simplify the retrieval process, improve detection efficiency and accuracy, and enable the system to continuously evolve through closed-loop self-learning. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the present invention aims to provide a fully automated intelligent system and method for ultrasonic testing of low-pressure turbine shafts. This system and method are designed to solve the problems of cumbersome testing processes, low efficiency, highly subjective results, and the inability of artificial intelligence models to utilize subsequent physical verification data for closed-loop feedback and continuous optimization in existing technologies.

[0007] The present invention provides a fully automated intelligent system for ultrasonic immersion testing of low-pressure turbine shafts, comprising: a data standardization and preprocessing module, an AI intelligent interpretation module, a model optimization and self-learning module, and an intelligent judgment and application interface module. The system interacts with an external physicochemical metallographic analysis module to construct a closed-loop learning and optimization path.

[0008] The data standardization and preprocessing module is used to convert the raw waveform signals acquired by the ultrasound equipment into standardized visual images. The AI ​​intelligent interpretation module is used to analyze the visual image to identify defects, extract and quantify information, and automatically generate structured detection information. The model optimization and self-learning module is used to associate the interpretation results of the AI ​​intelligent interpretation module with the verification results of the physicochemical metallographic analysis module, generate a difference dataset, and drive the deep learning model to incrementally learn and perform iterative optimization of the model. The intelligent judgment and application interface module is used to schedule the entire detection process and execute intelligent decisions based on the confidence level of the judgment results.

[0009] Preferably, the process by which the data standardization and preprocessing module converts the original waveform signal into a standardized C-scan grayscale image includes: Analyze the probe scanning path and step spacing in the raw data to create a two-dimensional mesh corresponding to the physical space; For the A-scan signal at each scanning position, extract the characteristic value of the echo signal within the preset gate; The feature values ​​extracted from all scan points are constructed into a two-dimensional matrix and normalized to generate a standardized C-scan grayscale image with traceable physical coordinate information.

[0010] Preferably, the AI ​​intelligent interpretation module includes a defect detection submodule and an image and text recording submodule; The defect detection submodule adopts a fusion architecture of convolutional neural network and visual Transformer. The convolutional neural network is used to extract local subtle features of defects, while the visual Transformer is used to analyze the global context of the image through a self-attention mechanism, thereby achieving accurate identification of defects and outputting the bounding box, category and confidence score of the defects. The image and text recording submodule integrates an OCR engine, a quantitative feature calculation engine, and a report generator. These are used to identify the scan parameter text embedded in the image and parse it into structured key-value pairs, to perform quantitative feature calculation on the defect areas identified by artificial intelligence, and the report generator to automatically integrate all the extracted and calculated information to generate a structured inspection report that meets industry standards.

[0011] Preferably, the quantification features include the geometric features, intensity features, and location features of the defect; The geometric features include defect area, aspect ratio, perimeter, and roundness; The positional features include principal axis direction angle and centroid coordinates that can be converted into physical coordinates; The intensity features include the average gray value, maximum gray value, and gray standard deviation of the pixels within the defect bounding box.

[0012] Preferably, the model optimization and self-learning module includes a data association unit, a difference analysis unit, and an incremental learning unit; The data association unit uses the unique identifier of the detected component to accurately map and associate the AI ​​prediction record with the real label of the physicochemical metallographic analysis; The difference analysis unit filters out inconsistent cases of missed detections, false detections, and classification errors, and generates a difference dataset. The incremental learning unit employs knowledge distillation or elastic weight consolidation techniques to incrementally learn the deep learning model using differential datasets.

[0013] Preferably, the intelligent decision-making process includes verifying the confidence score for each defect in the report. When the confidence level of all defects is higher than the system's preset confidence level threshold, the automatic judgment result is determined to be reliable, and the report is directly archived or automatically distributed to relevant quality management personnel. If the confidence level of any defect in the report is lower than the preset threshold, the case will be automatically marked as requiring manual review and pushed to the engineer for final decision. At the same time, the case will also be automatically sent to the waiting queue of the model optimization and self-learning module as data to drive model evolution.

[0014] Preferably, the actual label of the physicochemical metallographic analysis module includes relevant data on the nature, size, and location of defects confirmed by physical sectioning and metallographic analysis; The data association unit uses a unique component identifier to associate AI prediction records with real labels in the database.

[0015] Preferably, for the low-pressure turbine shaft cylindrical component, the analyzed scanning path information is converted into an equivalent plane unfolding or polar coordinate representation of the axial position and circumferential angle, and the actual scanning position of the probe is placed at the corresponding position of the two-dimensional virtual space grid by interpolation or direct mapping.

[0016] A fully automated ultrasonic water immersion method for testing low-pressure turbine shafts based on the system includes the following steps: Step S1: The intelligent judgment and application interface module receives the raw test data of the low-pressure turbine shaft collected by the ultrasonic testing equipment, starts the automated processing flow, and transmits the data to the data standardization and preprocessing module; Step S2: The data standardization and preprocessing module analyzes, extracts features, constructs matrices and normalizes the raw detection data to generate standardized C-scan grayscale images and coordinate mapping information, which are then transmitted to the AI ​​intelligent interpretation module. Step S3: The AI ​​intelligent interpretation module performs defect detection, text recognition, and quantitative feature calculation on the standardized C-scan grayscale image, integrates the information to generate a structured inspection report, and feeds it back to the intelligent judgment and application interface module; Step S4: The intelligent judgment and application interface module check the defect confidence in the report. If all of them are higher than the preset threshold, the report is archived and distributed; if there are defects below the threshold, they are marked as difficult cases and pushed to the manual review and model optimization and self-learning module. Step S5: The model optimization and self-learning module associates the AI ​​interpretation results of difficult cases with the verification results of the physicochemical metallographic analysis module, generates a difference dataset, and performs incremental learning on the deep learning model. The updated model is then redeployed to the AI ​​intelligent interpretation module.

[0017] Preferably, the raw detection data in step S1 is a binary format file containing A-scan signals recorded in chronological order, context information, and metadata; The data standardization and preprocessing module converts the original waveform signal into a standardized C-scan grayscale image, including the following steps: Analyze the probe scanning path and step spacing in the raw data to create a two-dimensional mesh corresponding to the physical space; For the A-scan signal at each scanning position, extract the characteristic value of the echo signal within the preset gate; The feature values ​​extracted from all scan points are constructed into a two-dimensional matrix and normalized to generate a standardized C-scan grayscale image with traceable physical coordinate information.

[0018] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention automates the entire process from data conversion and defect interpretation to report generation, replacing tedious and error-prone manual operations and significantly improving detection efficiency and standardization.

[0019] 2. This invention employs a hybrid artificial intelligence architecture that integrates convolutional neural networks and visual transformer technology, which can simultaneously focus on the microscopic details and macroscopic morphology of defects, effectively distinguish between real defects and pseudo-defects, and improve the accuracy and robustness of detection.

[0020] 3. This invention constructs a closed-loop self-learning mechanism based on physical verification, which breaks down the data barrier between artificial intelligence detection and real physical inspection, enabling the system to continuously learn and evolve from the "gold standard" and realize the effective accumulation and intelligent reuse of expert experience.

[0021] 4. By introducing a confidence-based intelligent review process, this invention achieves a balance between efficient automated processing and highly reliable expert adjudication, optimizes the human-machine collaboration mode, and ensures the reliability of the final test results. Attached Figure Description

[0022] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is an overall architecture diagram of a fully automated intelligent system for ultrasonic testing of low-pressure turbine shafts, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the ultrasonic water immersion detection principle in an embodiment of the present invention. Detailed Implementation

[0023] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0024] This invention constructs standard C-scan images from raw data and integrates CNN and ViT models to achieve automatic defect identification. The system simultaneously uses OCR technology to extract textual information from the images, completing defect judgment and recording concurrently. Furthermore, by comparing the AI ​​results with actual physicochemical and metallographic analysis results, the model is continuously optimized, forming a closed-loop intelligent system integrating detection, judgment, recording, and self-evolution.

[0025] Example 1 The present invention provides a fully automated intelligent system for ultrasonic immersion testing of low-pressure turbine shafts, such as... Figure 1 As shown, this system can be implemented based on general-purpose or special-purpose computing devices, such as industrial computers, servers, or embedded systems. Its hardware architecture typically includes a processor, memory, and communication interfaces. The computer program instructions stored in the memory are executed by the processor to implement the system functions described in this application. At the logical level, the system mainly consists of four core modules: a data standardization and preprocessing module, an AI intelligent interpretation module, a model optimization and self-learning module, and an intelligent judgment and application interface module. Furthermore, the system also interacts with an external physicochemical metallographic analysis module to construct a closed-loop learning and optimization path.

[0026] The following will be combined with the appendix Figure 1 and attached Figure 2 The workflow of the system and the specific implementation methods of each module are described.

[0027] Specifically, in the system's workflow, the intelligent judgment and application interface module acts as the central dispatch center, responsible for receiving raw test data from the ultrasonic testing equipment (water immersion ultrasonic C-scan equipment), such as raw data files stored in binary format. These files typically contain all the information collected during water immersion ultrasonic scanning of a specific low-pressure turbine shaft component. Upon receiving the data, this module immediately initiates an automated processing flow and transfers the raw data file to the data standardization and preprocessing module.

[0028] The data standardization and preprocessing module aims to convert raw, non-intuitive ultrasonic signals into standardized visual images that can be analyzed by subsequent artificial intelligence models. The accuracy and robustness of this conversion process directly determine the accuracy and reliability of subsequent AI intelligent interpretation. Its internal workflow may include the following steps: Step S10: Parse the scan path and metadata. The raw data file not only contains the core echo signal but also embeds key detection metadata. Specifically, the raw data includes a series of A-scan signals recorded in chronological order and their contextual information, as well as rich metadata. Each A-scan signal represents the complete voltage-time sequence of ultrasonic waves received by the probe at a specific physical location, propagating within the material and being reflected back. In this step, the system parses the file header or specific data segments to extract this metadata. This metadata may specifically include: the unique identifier of the detected component, probe model, probe frequency, probe focus, water layer distance, scan speed, scan start and end coordinates, and scan step accuracy. The data parsing unit in the data standardization and preprocessing module first reads the file header information and parses out the two-dimensional physical coordinates of each data point of the probe during the scan. For cylindrical components such as turbine shafts, this typically involves axial position and circumferential angle, which the system converts into equivalent planar unfolding or polar coordinate representation. Then, based on the parsed scan path information and step spacing, a two-dimensional virtual space grid is constructed in memory, where each grid point corresponds to a pixel in the final image. Each actual scanning position of the probe is precisely positioned on the corresponding location of this grid through interpolation or direct mapping. Simultaneously, the time axis of all A-scan signals is corrected using the water layer distance parameter in the metadata, aligning the signal start point with the moment the workpiece surface echo occurs, thereby establishing a unified depth coordinate system with the workpiece surface as the zero depth point.

[0029] Step S20: Extract the amplitude of the A-scan signal. Each A-scan signal is a complex waveform containing various information such as interface waves, defect waves, and bottom waves. For the A-scan signal at each scan position after coordinate mapping and time correction, this step aims to compress its waveform information into a single value to characterize the acoustic properties within that point. To this end, the system allows the operator to preset one or more analysis time windows, which are commonly referred to in the art as "gates." Each gate is used to define an independent signal analysis interval on the time-depth axis. The gate setting depends on the geometry, acoustic characteristics, and specific detection target of the workpiece being inspected. For example, when the primary target is to detect internal defects, a gate (which can be called a "defect gate") is needed to capture the echo of the internal defect. The starting position of this gate should be set outside the near-field interference zone of the ultrasonic probe, and its ending position should be set before the arrival time of the theoretical bottom surface echo of the workpiece. As an exemplary implementation, for a turbine shaft with a wall thickness of 60 mm, the defect gate can be set to a depth range from 2 mm to 55 mm below the workpiece surface. The initial depth of 2 mm is used to avoid near-surface interference, and the final depth of 55 mm is used to ensure that the scanning of the internal region is completed before the bottom surface echo arrives. In this way, the signal analysis within the gate will mainly focus on the internal region of the workpiece body, thereby effectively eliminating interference from surface coupling clutter and expected bottom surface echoes. Within this gate range, the system analyzes the A-scan signal and extracts a key feature value. As a preferred implementation, the maximum peak amplitude of the echo signal within this time window can be extracted. As an alternative implementation, the absolute integral or average value of the echo signal amplitude within this time window can also be calculated, as this integral or average value better reflects the overall energy of the defect. The extracted amplitude is then used as the feature value at this scanning point.

[0030] Step S30: Construct the amplitude matrix. After extracting the amplitude of the A-scan signal from all scanning points, the system fills these feature values ​​into a two-dimensional matrix according to their positions in the two-dimensional spatial grid constructed in step S10. This matrix is ​​the amplitude matrix, whose rows and columns correspond to the axial and circumferential scanning positions of the workpiece, respectively. The value of each element in the matrix represents the acoustic reflection intensity at the corresponding position inside the workpiece.

[0031] Step S40: Normalize to generate grayscale image. To adapt to the processing requirements of the visual model and eliminate gain differences between different detections, the amplitude matrix needs to be normalized. A commonly used method is linear normalization. The most common method is linear normalization, that is, through the linear mapping formula G=255×(V Vmin) / (Vmax The amplitude matrix (Vmin) is converted to a standard 8-bit grayscale range (0 to 255), where Vmin and Vmax are the minimum and maximum values ​​of the matrix, respectively. 0 (black) typically represents no echo or very weak echo (e.g., a perfectly defect-free area), while 255 (white) represents the strongest echo. This step converts the amplitude matrix into a standardized C-scan grayscale image. Crucially, the system appends metadata to this image, recording its pixel equivalent (e.g., each pixel represents a physical dimension of 0.2 mm x 0.2 mm) and coordinate origin information. This ensures that any pixel in the image can be accurately converted back to its actual physical coordinates on the workpiece (i.e., axial position and circumferential angle). For example, if the scan step is 0.1 mm, each pixel in the image represents a physical area of ​​0.1 mm x 0.1 mm. This ensures that the geometry and size of defects in the image can be accurately and without distortion, providing a fundamental guarantee for subsequent accurate quantitative defect analysis (such as calculating crack length and pore diameter).

[0032] After completing the above steps, the data standardization and preprocessing module will pass the generated standardized C-scan image and the accompanying coordinate mapping information to the AI ​​intelligent interpretation module.

[0033] The AI-powered intelligent interpretation module, serving as the system's core intelligence, is responsible for receiving C-scan images and performing automated defect interpretation and information recording. This module includes a defect detection submodule and an image and text recording submodule.

[0034] The defect detection submodule is responsible for identifying and locating defects on the input C-scan image. In this embodiment, this submodule adopts a fusion architecture, combining the advantages of Convolutional Neural Networks (CNNs) and Visual Transformers (ViTs). Specifically, when the C-scan image is input, the CNN scans the image using its multi-layered convolutional kernels. Because convolutional kernels are highly sensitive to local pixel patterns, they can effectively extract local and subtle features of defects, such as the elongated linear bright features of cracks or the approximately circular isolated bright spots of pores. Simultaneously, the Visual Transformer segments the image into a series of non-overlapping image patches and, through its core self-attention mechanism, analyzes the intrinsic relationships between any image patches, regardless of their spatial distance. This allows the model to understand the global context of the defects. For example, for a long crack that appears as multiple discontinuous bright spots on the C-scan image, the Visual Transformer can analyze the linear arrangement trend and spatial distribution pattern of these bright spots to determine that they are associated as a complete crack. Similarly, for a cluster of pores, the visual converter can also identify it as a "clustered pore" based on its overall shape and distribution density. By adopting a technical structure that automatically constructs standardized C-scan images from raw ultrasonic data, integrates CNN and ViT models for intelligent defect identification, and combines OCR technology to automatically extract text information, the problem of cumbersome and inefficient processes caused by manual operation, screenshotting, interpretation, and recording by engineers in traditional inspection procedures, as well as the highly subjective and difficult-to-standardize results caused by inconsistent human interpretation standards, is solved.

[0035] Accordingly, by fusing the local feature extraction capabilities of convolutional neural networks with the global context analysis capabilities of visual Transformers, the defect detection submodule can make more accurate and robust judgments on complex and atypical defect patterns. The final output of this submodule is a set of structured information about the defect, typically including: the bounding box of each defect, i.e. (x_min, y_min, x_max, y_max) in the image coordinate system; the defect category label, such as crack, pore, inclusion, or spurious defect; and a score representing the confidence of the model's judgment, such as 0.98.

[0036] In parallel with the defect detection submodule, the image recording submodule also processes the input C-scan images. This submodule integrates an optical character recognition (OCR) engine and a quantization feature calculation engine. Understandably, in practical applications, images generated by ultrasonic testing equipment often embed key scanning parameters in text form in edge or corner areas. The function of the optical character recognition engine is to automatically scan such areas and recognize text information, such as "X=125.4mm, Y=38.1mm, Amp=-12dB," and parse it into structured key-value pair data.

[0037] The quantization feature calculation engine performs in-depth geometric and morphological feature calculations on each defect bounding box identified by the defect detection submodule 210. The calculated features include geometric features, intensity features, and location features. Geometric features include the defect area (calculated by multiplying the number of pixels by the pixel equivalent), aspect ratio, perimeter, and roundness. Location features include the principal axis direction angle and centroid coordinates (which can be converted to physical coordinates). Intensity features include the average gray value, maximum gray value, and gray standard deviation of pixels within the bounding box. These quantization features provide objective and accurate data support for subsequent engineering evaluations and reports.

[0038] Finally, the report generator in the AI ​​intelligent interpretation module summarizes and integrates all information from the two sub-modules. It automatically fills in the scanning parameters extracted by the optical character recognition engine, the defect location, category, and confidence level output by the defect detection sub-module, and the detailed features calculated by the quantification feature calculation engine into a predefined, industry-standard inspection report template (such as a structured format conforming to NAS410 or ISO standards). Thus, the system can automatically generate a complete, formatted, and structured inspection report, achieving full automation of report generation. This report is directly geared towards engineering applications, clearly answering several core questions that quality engineers care about most: "What is the defect?" (e.g., cracks, porosity, or inclusions and their specific classification), "Where is it?" (precise three-dimensional physical coordinates on the component), "How big is it?" (equivalent size calculated based on echo amplitude or image segmentation), and "How certain is it?" (intuitively displaying the AI's judgment with a confidence percentage). This makes the report not just a data compilation, but a direct action guide for decision-making. An example output paragraph is as follows: Three types of defects were found on the component [component serial number]: Crack: Located in the 0-70mm range of the stamped end, with a length of 70mm, a maximum amplitude of φ1.2-8dB, and a burial depth of approximately 127mm.

[0039] Clustered pores: located in the 620-660mm range of the stamped end, with an equivalent diameter of φ1.2-8dB and a burial depth of about 113mm.

[0040] Single inclusion: Located in the 2030-2255mm range of the stamped end, with an equivalent diameter of φ1.2+15dB and a burial depth of approximately 118mm. Overall judgment: Unqualified. The generated structured inspection report is then sent to the intelligent judgment and application interface module. This module checks the confidence score of each defect in the report. If the confidence scores of all defects are higher than the system's preset confidence threshold (e.g., 95%), the system determines that the automatic interpretation result is highly reliable and directly archives the report or automatically distributes it to relevant quality management personnel via email, enterprise messaging system, etc. Alternatively, if the confidence score of any defect in the report is lower than the preset threshold (e.g., the confidence score of a suspected crack is 85%), the system classifies the case as a "difficult case". In this case, the intelligent judgment and application interface module will perform parallel operations: first, mark the report as requiring manual review and push it to the work queue of senior non-destructive testing engineers for manual decision-making; second, automatically send the C-scan image containing the low-confidence defect, the AI ​​preliminary interpretation result, and all relevant data as a high-value difficult sample to the waiting queue of the model optimization and self-learning module.

[0041] The working process of the model optimization and self-learning module is described below. This module endows the system with the ability to continuously self-optimize by constructing a closed-loop feedback mechanism of "prediction-validation-learning".

[0042] The system's self-learning process begins with subsequent physical verification. For turbine shaft components inspected by AI, especially those with questionable or critical defects, subsequent destructive testing procedures may be implemented, such as physical sectioning and metallographic analysis, to ultimately confirm the true nature of the defect. The physicochemical metallographic analysis module represents the verification process that provides the "gold standard" answer. After the metallographic analysis is completed, the inspector enters the final authoritative analysis result (e.g., "Metallographic sectioning confirms a fatigue crack with a depth of 0.5 mm" or "Metallographic verification shows that this is actually normal microstructure reflection at the material grain boundary, a false defect") into the database connected to the system, including the component's unique identifier.

[0043] The model optimization and self-learning module uses the unique identifier of each component to accurately map and associate the AI ​​prediction records (S50) stored in the system with the real labels (S60) from the physicochemical metallographic analysis module. The system periodically or when it receives new metallographic results triggers a comparison and difference analysis procedure to automatically compare the AI's interpretation results with the "real labels" of the metallographic analysis.

[0044] In step S70, all inconsistent cases are filtered out to form a "difference dataset" specifically for model optimization. These inconsistent cases include missed detections (metallographic findings of defects that the AI ​​fails to detect), false positives (the AI ​​reports defects that metallographic findings confirm the absence of defects), and classification errors (e.g., the AI ​​misclassifies a crack as an inclusion). This difference dataset accurately reflects the shortcomings in the current AI model's knowledge system and therefore has extremely high optimization value.

[0045] In step S80, the system uses the newly generated difference dataset to incrementally learn or retrain the deep learning model deployed in the AI ​​intelligent interpretation module. To avoid the "catastrophic forgetting" problem when learning new knowledge—that is, forgetting previously acquired knowledge—various incremental learning techniques can be employed, such as knowledge distillation and elastic weight consolidation. In this way, the model can learn from its own errors and continuously correct its discrimination logic. For example, when the model is "penalized" for misjudging a specific grain boundary reflection as a crack, its internal parameters will be adjusted so that it can make a more accurate judgment when encountering similar signal patterns in the future.

[0046] After incremental learning (S80), the updated and more powerful model will be redeployed to the AI ​​intelligent interpretation module for future AI prediction (S50). This forms a complete closed loop from "AI prediction" to "metallographic verification," then to "generating difference data," and finally updating the model through "incremental learning." Through this closed loop, the system's detection capability spirals upward with the continuous accumulation of detection cases and metallographic verification data, realizing the effective accumulation and intelligent reuse of expert experience and physical laws into the artificial intelligence system. By adopting a technical structure that correlates and compares AI interpretation results with the actual results of physicochemical metallographic analysis, and uses this to drive incremental learning of the model, the "data silo" problem formed by the separation of ultrasonic detection data and subsequent verification data is solved. This allows the detection system to continuously optimize and self-evolve using closed-loop feedback data, enabling the accumulation and reuse of knowledge and experience.

[0047] Furthermore, the core intelligent driving force of this system originates from the continuously running feedback loop and self-learning mechanism built into the model optimization and self-learning module. This mechanism begins with the data association unit, which acts as a bridge connecting non-destructive testing and destructive testing. By establishing a unique component identifier, it accurately maps and associates the ID of each ultrasonic C-scan image with the metallographic image ID and inspection report of its subsequent physicochemical testing. This crucial step successfully constructs a valuable closed-loop dataset of "AI prediction - metallographic verification." Specifically, when the intelligent interpretation module within the system determines that a C-scan image is "suspected crack," this determination, along with the image ID, is recorded by the system. After the component is cut open and subjected to metallographic (low or high magnification) analysis in subsequent processes, the laboratory technician enters the final metallographic analysis result (e.g., "confirmed crack" or "actually a pseudo-defect - grain boundary reaction") into the system. At this point, the data association unit completes the closed-loop link from the system's initial judgment to the physical truth.

[0048] Once the model is trained, we input a new, unlabeled C-scan image. The model extracts features from this image and compares these features with patterns of "cracks," "pores," and "inclusions" learned from massive training data. Finally, it outputs a probability, for example: Cracks: 92%, Pores: 5%, Inclusions: 3%. Head A1: Clutter level 16%, Bottom wave amplitude: 30%~60.5%; Tail B1: Clutter level 8%, Bottom wave amplitude: 55.5%~90.5%; Sloping 5° B2: Clutter level 18%; Sloping surface C1: Clutter level 24%. Based on this probability distribution, the system can automatically determine the most likely type of defect and provide a classification result and confidence level.

[0049] Building upon this foundation, the incremental learning unit is activated and begins to play a central role. This unit runs automatically periodically (e.g., weekly or monthly), systematically comparing all previous predictions from the intelligent interpretation module with the "true labels" provided by metallographic analysis. Through this comparison process, the system can automatically identify cases where there are biases or uncertainties in the system's judgments (including missed detections, false detections, and classification errors), thereby generating a "difference dataset" specifically for error correction. This dataset accurately records where the system "made mistakes" or "had doubts," serving as the most valuable nourishment driving model evolution.

[0050] Ultimately, the model optimization process utilizes this meticulously constructed "error correction dataset" to iteratively retrain the deep learning model in the system's intelligent interpretation module. This process is not simply parameter fine-tuning, but a "re-engineering" optimization combining incremental learning and transfer learning. It ensures that the model can learn new knowledge (learning from mistakes) while robustly retaining its ability to identify defects it has already mastered, thus effectively avoiding "catastrophic forgetting." Through this cyclical "prediction-verification-comparison-learning" closed loop, the system is endowed with the ability to continuously learn from its own successes and errors, driving the model to continuously optimize its performance parameters. In this way, when encountering similar microstructures or defect morphologies again, the accuracy and confidence of its judgment will be significantly improved, thereby achieving a spiral increase and self-evolution of the system's overall detection capabilities.

[0051] Example 2 This embodiment further illustrates the technical solution of this application using a specific application scenario and demonstrates some of its optional implementation methods.

[0052] Suppose this system is applied in an aero-engine manufacturing plant to inspect a low-pressure turbine shaft made of GH4169 nickel-based superalloy. In this scenario, the specific configuration and operational details of each module are as follows: In the data standardization and preprocessing module, for the acoustic characteristics of GH4169 alloy, the ultrasonic testing equipment is configured to use a 15 MHz high-frequency focusing probe to obtain higher resolution, with scanning parameters set to an axial step of 0.1 mm and a circumferential step of 0.15 degrees. In step S20, when extracting the A-scan signal amplitude, considering the possible minute metallurgical defects within the material, the system, in addition to extracting the maximum peak value, also calculates the area under the signal envelope as another feature. Accordingly, in step S40, the system generates two C-scan images: one based on the maximum peak value (more sensitive to sharp defects such as cracks), and the other based on energy integral (more sensitive to porosity defects). These two images are provided as multi-channel inputs to the AI ​​model.

[0053] In the AI-powered intelligent interpretation module, the defect detection submodule's model has been specifically trained for common defect types in the GH4169 alloy. Its category labels include, but are not limited to, "microcracks," "lack of fusion," "porosity," and "high-density hard inclusions." Furthermore, the report generator in the graphic and text recording submodule uses a template customized based on the manufacturer's internal quality control document QMS-0078B. The generated reports can be directly uploaded to the factory's manufacturing execution system, achieving seamless integration with production management processes.

[0054] In the model optimization and self-learning module, this embodiment employs a more proactive learning strategy. In addition to utilizing destructive metallographic analysis results, the system also incorporates non-destructive re-inspection results. For cases marked as "requiring manual verification" by the intelligent judgment and application interface module, after a senior engineer performs verification using a higher-precision detection method (such as phased array ultrasound), their final confirmed conclusion (e.g., "confirmed as a false defect; the signal originates from reflection by coarse grains") is also recorded in the system. These non-destructive inspection results confirmed by experts are also added to the differential dataset as high-quality labeled data for incremental model learning. This approach allows for more frequent and lower-cost acquisition of high-quality labeled data without damaging components, thereby accelerating the iterative evolution of the model.

[0055] In the intelligent judgment and application interface module, this embodiment employs a dynamic confidence threshold. The system can adjust the decision threshold based on the type and size of the defect. For example, for crack-like defects with a length greater than 1 mm, the confidence threshold is set to an extremely high 99.5%, and any judgment below this value will force manual review; while for small, single pores, the threshold can be appropriately relaxed to 90%. This dynamic threshold strategy allows the system to further improve overall automated processing efficiency while ensuring that critical defects are not missed.

[0056] As an alternative implementation, the defect detection submodule within the AI ​​intelligent interpretation module can also employ an instance segmentation-based model, such as Mask R-CNN, to replace the aforementioned fusion architecture. Such models, in addition to outputting the defect's bounding box and category, can also output a pixel-level precise contour mask. Based on this mask, the quantization feature calculation engine of the image and text recording submodule can calculate more accurate defect morphological parameters, such as the defect's true area, perimeter, and maximum Feret diameter, providing more detailed data for subsequent defect hazard assessment.

[0057] As an alternative implementation, the incremental learning (S80) process of the model optimization and self-learning module can be conducted online. That is, whenever a "difficult case" is manually reviewed or metallographically verified and obtains a valid label, the system does not store it in a waiting queue for batch processing, but immediately triggers a small-scale model fine-tuning. This online learning method enables the model to respond and absorb new knowledge more quickly, and is particularly suitable for scenarios where defect patterns change rapidly during the initial stages of new production line commissioning or process changes.

[0058] In summary, the system and method provided in this application organically integrate data processing, intelligent interpretation, automatic recording, and closed-loop self-learning to construct a fully automated, self-evolving, and highly reliable intelligent inspection system. This system can significantly improve the efficiency, accuracy, and standardization of non-destructive testing of key industrial components.

[0059] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0060] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A fully automated intelligent system for ultrasonic immersion testing of low-pressure turbine shafts, characterized in that, include: The system includes a data standardization and preprocessing module, an AI intelligent interpretation module, a model optimization and self-learning module, and an intelligent judgment and application interface module. The system also enables data interaction with an external physicochemical metallographic analysis module to construct a closed-loop learning and optimization path. The data standardization and preprocessing module is used to convert the raw waveform signals acquired by the ultrasound equipment into standardized visual images. The AI ​​intelligent interpretation module is used to analyze the visual image to identify defects, extract and quantify information, and automatically generate structured detection information. The model optimization and self-learning module is used to associate the interpretation results of the AI ​​intelligent interpretation module with the verification results of the physicochemical metallographic analysis module, generate a difference dataset, and drive the deep learning model to incrementally learn and perform iterative optimization of the model. The intelligent judgment and application interface module is used to schedule the entire detection process and execute intelligent decisions based on the confidence level of the judgment results.

2. The fully automated intelligent system for ultrasonic immersion testing of low-pressure turbine shafts according to claim 1, characterized in that, The data standardization and preprocessing module converts the original waveform signal into a standardized C-scan grayscale image, including the following steps: Analyze the probe scanning path and step spacing in the raw data to create a two-dimensional mesh corresponding to the physical space; For the A-scan signal at each scanning position, extract the characteristic value of the echo signal within the preset gate; The feature values ​​extracted from all scan points are constructed into a two-dimensional matrix and normalized to generate a standardized C-scan grayscale image with traceable physical coordinate information.

3. The fully automated intelligent system for ultrasonic immersion testing of low-pressure turbine shafts according to claim 1, characterized in that, The AI ​​intelligent interpretation module includes a defect detection submodule and a graphic and text recording submodule; The defect detection submodule adopts a fusion architecture of convolutional neural network and visual Transformer. The convolutional neural network is used to extract local subtle features of defects, while the visual Transformer is used to analyze the global context of the image through a self-attention mechanism, thereby achieving accurate identification of defects and outputting the bounding box, category and confidence score of the defects. The image and text recording submodule integrates an OCR engine, a quantitative feature calculation engine, and a report generator. These are used to identify the scan parameter text embedded in the image and parse it into structured key-value pairs, to perform quantitative feature calculation on the defect areas identified by artificial intelligence, and the report generator to automatically integrate all the extracted and calculated information to generate a structured inspection report that meets industry standards.

4. The fully automated intelligent system for ultrasonic immersion testing of low-pressure turbine shafts according to claim 3, characterized in that, The quantification features include the geometric features, intensity features, and location features of the defect; The geometric features include defect area, aspect ratio, perimeter, and roundness; The positional features include principal axis direction angle and centroid coordinates that can be converted into physical coordinates; The intensity features include the average gray value, maximum gray value, and gray standard deviation of the pixels within the defect bounding box.

5. The fully automated intelligent system for ultrasonic immersion testing of low-pressure turbine shafts according to claim 1, characterized in that, The model optimization and self-learning module includes a data association unit, a difference analysis unit, and an incremental learning unit. The data association unit uses the unique identifier of the detected component to accurately map and associate the AI ​​prediction record with the real label of the physicochemical metallographic analysis; The difference analysis unit filters out inconsistent cases of missed detections, false detections, and classification errors, and generates a difference dataset. The incremental learning unit employs knowledge distillation or elastic weight consolidation techniques to incrementally learn the deep learning model using differential datasets.

6. The fully automated intelligent system for ultrasonic immersion testing of low-pressure turbine shafts according to claim 1, characterized in that, The intelligent decision-making process includes verifying the confidence score for each defect in the report. When the confidence level of all defects is higher than the system's preset confidence level threshold, the automatic judgment result is determined to be reliable, and the report is directly archived or automatically distributed to relevant quality management personnel. If the confidence level of any defect in the report is lower than the preset threshold, the case will be automatically marked as requiring manual review and pushed to the engineer for final decision. At the same time, the case will also be automatically sent to the waiting queue of the model optimization and self-learning module as data to drive model evolution.

7. The fully automated intelligent system for ultrasonic immersion testing of low-pressure turbine shafts according to claim 5, characterized in that, The true label of the physicochemical metallographic analysis module includes relevant data on the nature, size, and location of defects confirmed by physical sectioning and metallographic analysis; The data association unit uses a unique component identifier to associate AI prediction records with real labels in the database.

8. The fully automated intelligent system for ultrasonic immersion testing of low-pressure turbine shafts according to claim 2, characterized in that, For the cylindrical component of the low-pressure turbine shaft, the analyzed scanning path information is converted into an equivalent plane unfolding or polar coordinate representation of the axial position and circumferential angle. The actual scanning position of the probe is placed at the corresponding position of the two-dimensional virtual space grid by interpolation or direct mapping.

9. A fully automated ultrasonic water immersion method for detecting low-pressure turbine shafts according to claim 1, characterized in that, Includes the following steps: Step S1: The intelligent judgment and application interface module receives the raw test data of the low-pressure turbine shaft collected by the ultrasonic testing equipment, starts the automated processing flow, and transmits the data to the data standardization and preprocessing module; Step S2: The data standardization and preprocessing module analyzes, extracts features, constructs matrices and normalizes the raw detection data to generate standardized C-scan grayscale images and coordinate mapping information, which are then transmitted to the AI ​​intelligent interpretation module. Step S3: The AI ​​intelligent interpretation module performs defect detection, text recognition, and quantitative feature calculation on the standardized C-scan grayscale image, integrates the information to generate a structured inspection report, and feeds it back to the intelligent judgment and application interface module; Step S4: The intelligent judgment and application interface module check the defect confidence in the report. If all of them are higher than the preset threshold, the report is archived and distributed; if there are defects below the threshold, they are marked as difficult cases and pushed to the manual review and model optimization and self-learning module. Step S5: The model optimization and self-learning module associates the AI ​​interpretation results of difficult cases with the verification results of the physicochemical metallographic analysis module, generates a difference dataset, and performs incremental learning on the deep learning model. The updated model is then redeployed to the AI ​​intelligent interpretation module.

10. The fully automated intelligent system for ultrasonic immersion testing of low-pressure turbine shafts according to claim 1, characterized in that, The raw detection data mentioned in step S1 is a binary format file, which contains A-scan signals recorded in chronological order, context information and metadata; The data standardization and preprocessing module converts the original waveform signal into a standardized C-scan grayscale image, including the following steps: Analyze the probe scanning path and step spacing in the raw data to create a two-dimensional mesh corresponding to the physical space; For the A-scan signal at each scanning position, extract the characteristic value of the echo signal within the preset gate; The feature values ​​extracted from all scan points are constructed into a two-dimensional matrix and normalized to generate a standardized C-scan grayscale image with traceable physical coordinate information.

Images