An infrared interactive detection method for defects of high-temperature hydraulic pipeline

Abstract

The application discloses a kind of high-temperature hydraulic pipeline defect infrared interactive detection methods, first based on interactive locking video stream real-time tracking segmentation, the high-temperature hydraulic pipeline specified by user is separated from environment background, generates continuous visual high-temperature hydraulic pipeline image, then through statistical abnormal area rough positioning, i.e. Suspected defect screening generates suspected defect screening result, again through non-supervised abnormal fine identification, i.e. False heat spot is eliminated, and defect detection result is obtained, finally through augmented reality interactive feedback, superimposed display and alarm prompt are carried out on the screen of interactive processing terminal.The application not only realizes high-precision online defect detection, but also greatly improves detection efficiency and user experience through visual interactive means.

Description

Technical Field

[0001] This invention belongs to the field of nondestructive testing technology, and more specifically, relates to an infrared interactive detection method for defects in high-temperature hydraulic pipelines. Background Technology

[0002] High-temperature hydraulic pipelines typically employ multi-layered composite high-pressure plastic hoses, which are widely used in high-pressure, high-temperature fluid transport systems in industrial equipment due to their corrosion resistance and good flexibility. For example... Figure 1 As shown, the structure of the high-pressure plastic hose from the inside out is as follows:

[0003] The innermost plastic layer that directly contacts the fluid (innermost layer): usually made of oil-resistant and corrosion-resistant modified plastic, directly contacting the high-temperature and high-pressure fluid flowing inside the pipe;

[0004] The intermediate steel mesh layer (intermediate layer) provides strength support: it is usually made of high-strength steel wire woven mesh and is used to withstand fluid pressure;

[0005] Outermost plastic protective layer (outermost layer): made of wear-resistant and aging-resistant plastic material.

[0006] In the actual operation of high-pressure, high-temperature fluid transport systems, cavitation can easily occur when impurities or air bubbles in the high-temperature, high-pressure liquid are injected into the pipe joints. The microjet and shock waves generated by cavitation can break through the innermost plastic layer, causing the high-temperature, high-pressure liquid to penetrate into the intermediate steel mesh layer. As the high-temperature fluid accumulates in the intermediate steel mesh layer, heat is rapidly conducted outwards, causing an abnormal temperature rise on the surface of the outermost plastic protective layer in this area. If such early defects are not detected in time, the damage will eventually lead to the outermost layer rupture, causing serious liquid or oil leaks, equipment downtime, and even safety hazards.

[0007] Currently, while there are various methods for detecting defects such as cavitation in high-temperature hydraulic pipelines, they still face many limitations in practical applications. Firstly, traditional contact-based detection methods, such as vibration monitoring or acoustic emission, suffer from significant attenuation of vibration and sound signals during transmission due to the inherent damping characteristics of plastic materials. This makes it difficult to sensitively capture the weak signals generated by the rupture of the innermost plastic layer of high-pressure plastic hoses. Secondly, while ultrasonic testing offers high precision, it typically requires the application of a coupling agent and point-by-point scanning, resulting in low detection efficiency and making it difficult to perform detection when high-temperature hydraulic pipelines are in operation and transporting liquids.

[0008] In fluid transport operations, thermal imagers are mostly handheld, resulting in continuous video jitter and field-of-view drift. This makes it difficult for operators to quickly identify the target—the high-temperature hydraulic pipeline—in complex industrial settings, and the algorithm cannot continuously lock onto the pipeline while the field of view shifts. Furthermore, the lack of intuitive visual feedback from the inspection results forces operators to rely on subjective experience to determine whether hotspots on the screen represent defects, significantly reducing defect detection efficiency and accuracy. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of existing infrared defect detection methods, such as difficulty in quickly and continuously identifying and locking high-temperature hydraulic pipelines in complex industrial settings and the lack of intuitive visual feedback of detection results, and to provide an infrared interactive detection method for defects in high-temperature hydraulic pipelines to improve the efficiency and accuracy of defect detection.

[0010] To achieve the above-mentioned objectives, the present invention provides an infrared interactive detection method for defects in high-temperature hydraulic pipelines, characterized by comprising the following steps:

[0011] (1) Real-time tracking and segmentation of video stream based on interactive locking: the high-temperature hydraulic pipeline specified by the user is segmented from the environmental background to generate a continuous and visualized high-temperature hydraulic pipeline image.

[0012] 1.1) Video capture and real-time preview

[0013] The infrared thermal image video data of the high-temperature hydraulic pipeline in working condition is recorded by the thermal imaging acquisition unit, and the infrared thermal image video data is transmitted to the interactive processing terminal in real time for preview display.

[0014] 1.2) Interactive Selection and Initialization of High-Temperature Hydraulic Pipelines

[0015] On the interactive processing terminal, the user selects the high-temperature hydraulic pipeline to be detected in the preview display, generates initial target prompt information, namely the high-temperature hydraulic pipeline image instance, and sends it to the data processing unit.

[0016] 1.3) Dynamic tracking and segmentation based on interactive prompts

[0017] The data processing unit inputs the initial target prompt information into the pre-trained video segmentation model. The video segmentation model is then used to perform spatiotemporal correlation reasoning and frame-by-frame tracking on the video frames acquired by the subsequent thermal imaging acquisition unit. The model activates and locks the image instance of the high-temperature hydraulic pipeline selected by the user as the only tracking object, thereby segmenting the high-temperature hydraulic pipeline from the environmental background and generating a continuous and visualized high-temperature hydraulic pipeline image.

[0018] (2) Coarse localization of statistically abnormal regions, i.e. screening of suspected defects.

[0019] Statistical temperature distribution analysis is performed on the segmented and visualized high-temperature hydraulic pipeline images to quickly identify hot spots, i.e., suspected defects, and generate suspected defect screening results.

[0020] (3) Fine-grained identification of unsupervised anomalies, i.e., distinguishing true from false.

[0021] Unsupervised anomaly identification is performed on the suspected defect screening results to remove false hot spots that are not defects, thus obtaining the defect detection results;

[0022] (4) Augmented Reality Interactive Feedback

[0023] The visualized high-temperature hydraulic pipeline images and defect detection results are mapped back to the coordinates of the original video screen, and then overlaid and displayed on the screen of the interactive processing terminal with alarm prompts.

[0024] The objective of this invention is achieved as follows:

[0025] Addressing the challenges of existing technologies in detecting defects in high-temperature hydraulic pipelines under operating conditions, including difficulties with background interference and a lack of intuitive interactive feedback, this invention presents an infrared interactive defect detection method for high-temperature hydraulic pipelines. First, based on interactively locked video stream real-time tracking and segmentation, the user-specified high-temperature hydraulic pipeline is segmented from the environmental background, generating a continuous, visualized image of the high-temperature hydraulic pipeline. Then, a coarse localization of statistically abnormal regions (i.e., suspected defect screening) generates a suspected defect screening result. Next, unsupervised fine anomaly identification (i.e., false positives) eliminates non-defective false hot spots, yielding the defect detection result. Finally, augmented reality interactive feedback is used to overlay and display the results on the screen of the interactive processing terminal, along with alarm prompts. This invention not only achieves high-precision online defect detection but also significantly improves detection efficiency and user experience through visual interactive means.

[0026] Compared with existing technologies, the infrared interactive detection method for defects in high-temperature hydraulic pipelines of the present invention has the following advantages:

[0027] (1) Balancing interactivity and accuracy: This invention innovatively introduces a "human-computer interaction target selection" mechanism. Compared to fully automatic blind inspection, this method allows users to generate initial target prompt information by selecting the high-temperature hydraulic pipeline in the early stages of detection, and then lock and track the target based on this interactive prompt. This not only solves the problem of target loss caused by the movement of the field of view during handheld detection, but also effectively avoids false alarms caused by other irrelevant heat sources in the background (such as normal thermal pipelines next to it), achieving efficient detection of "point-and-test";

[0028] (2) Intuitive visual feedback: Real-time overlay display of video and graphics is achieved through the interactive processing terminal, transforming obscure temperature data into intuitive pipe outlines and defect marker boxes. Operators do not need professional thermal imaging analysis experience to quickly locate fault points through visual guidance on the screen, greatly improving the efficiency of on-site inspections and user experience.

[0029] (3) Robustness under dynamic working conditions: This invention directly utilizes the natural temperature rise of the high-temperature hydraulic pipeline during operation for detection, without the need to stop the machine. The introduced video segmentation model has spatiotemporal memory capability, which can accurately separate the high-temperature hydraulic pipeline in dynamic working environment (such as shaking, deformation) from the complex background, providing clean data for subsequent analysis;

[0030] (4) High reliability of two-level discrimination: The mechanism of “statistical suspected defect screening” combined with “unsupervised anomaly fine identification” can deeply analyze the heat conduction characteristics caused by defects (such as hot spots of specific shapes), effectively distinguish real defects from false anomalies such as surface reflection and oil stains, and significantly reduce the false alarm rate;

[0031] (5) Integrated portable terminal design: The terminal solution proposed in this invention integrates data acquisition, calculation and interaction functions, making it easy to carry to various corners of the industrial site (such as narrow spaces or high places), and supports sensitive and fast mobile detection and data archiving of dispersed pipelines. Attached Figure Description

[0032] Figure 1 This is a cross-section of a multi-layered composite high-pressure plastic hose and a schematic diagram of the heat conduction mechanism during cavitation.

[0033] Figure 2 This is a flowchart of a specific embodiment of the infrared interactive detection method for defects in high-temperature hydraulic pipelines according to the present invention;

[0034] Figure 3 This is a schematic diagram illustrating the working principle and application system of the infrared interactive detection method for defects in high-temperature hydraulic pipelines according to the present invention.

[0035] Figure 4 This is an overall structural block diagram of the infrared interactive detection method for defects in high-temperature hydraulic pipelines according to the present invention;

[0036] Figure 5 This is an example diagram of interactive selection for high-temperature hydraulic pipelines;

[0037] Figure 6 It generates a continuous, visualized example image of a high-temperature hydraulic pipeline.

[0038] Figure 7 This is an example of the defect detection results for a high-temperature hydraulic pipeline. Detailed Implementation

[0039] The specific embodiments of the present invention will now be described with reference to the accompanying drawings to enable those skilled in the art to better understand the invention. It should be particularly noted that in the following description, detailed descriptions of known functions and designs that might obscure the main content of the invention will be omitted here.

[0040] In this embodiment, as Figure 1 As shown, a high-pressure plastic hose with a multi-layer composite structure is used as a high-temperature hydraulic pipeline, and cavitation is used as a defect for detailed explanation.

[0041] The physical mechanism of cavitation detection: This invention does not require an external thermal excitation source, but directly utilizes the natural temperature rise generated when a high-pressure plastic hose is transporting high-temperature, high-pressure fluids (such as hydraulic oil in a hydraulic system). When air bubbles or impurities are mixed into the fluid inside the hose, cavitation occurs under high-speed flow. The microjets and shock waves generated by cavitation first damage the inner lining, causing it to rupture. Once the inner lining fails, the high-temperature, high-pressure fluid will rapidly penetrate and fill the reinforcing layer (wire mesh). Since the thermal conductivity of metal is far superior to that of plastic, the high-temperature fluid accumulated in the reinforcing layer will rapidly conduct heat outward, resulting in localized hot spots on the surface of the protective layer (outermost layer) corresponding to the damage point, with temperatures significantly higher than the surrounding area. This invention captures this abnormal temperature rise characteristic conducted from the inside out through a thermal imaging acquisition unit, thereby achieving early warning of cavitation defects before the outermost layer physically ruptures and leaks oil.

[0042] Figure 2 , 3 4 is a flowchart, working principle and application diagram, and overall structural block diagram of a specific implementation method of the high-temperature hydraulic pipeline defect infrared interactive detection method of the present invention.

[0043] To address the issues of complex industrial environments and persistent shaking and field-of-view drift in handheld video footage, this embodiment employs a video segmentation network based on an improved SAM2 (Segment Anything Model 2) architecture, combined with user interactive input, to process the infrared video stream in real time.

[0044] In this embodiment, as Figure 2 As shown, the infrared interactive detection method for defects in high-temperature hydraulic pipelines of the present invention includes the following steps:

[0045] Step S1: Based on interactive locking, real-time tracking and segmentation of the video stream is used to segment the user-specified high-temperature hydraulic pipeline from the environmental background, generating a continuous and visualized image of the high-temperature hydraulic pipeline.

[0046] Step S1.1: Video capture and real-time preview

[0047] like Figure 3 The diagram shows an infrared interactive detection system for defects in high-temperature hydraulic pipelines constructed based on this invention. It includes a thermal imaging acquisition unit, an interactive processing terminal, and a data processing unit, which are sequentially connected via a high-speed data interface (such as Gigabit Ethernet or USB 3.0) to form a closed-loop interactive detection system. The thermal imaging acquisition unit and the interactive processing terminal constitute a handheld integrated instrument. The thermal imaging acquisition unit records infrared thermal video data of the high-temperature hydraulic pipeline in its operating state, and transmits the infrared thermal video data to the interactive processing terminal in real time for preview display.

[0048] In this embodiment, the thermal imaging acquisition unit serves as the front-end sensing device and employs an uncooled focal plane array infrared thermal imager. To meet the requirement of capturing the minute temperature differences in the early stages of cavitation in high-pressure plastic hoses, the thermal sensitivity (NETD) of the thermal imaging acquisition unit is preferably ≤40mK, with a resolution of 640×480 pixels or higher and a frame rate of 25Hz. This unit is used to record infrared thermal video data of the high-pressure plastic hose as it operates while performing fluid transport tasks, showing the process of its temperature rising.

[0049] Interactive processing terminal: Serving as the system's human-machine interface, it employs an industrial flat panel or display screen with capacitive touch functionality. Its functions include:

[0050] Real-time interactive preview: Displays the raw infrared video stream and provides a touch interface for users to click or select on the video screen to specify the target hose to be detected (i.e., "interactive target initialization").

[0051] Visual overlay feedback: Receive the analysis results from the data processing unit and overlay a semi-transparent segmentation mask and defect highlighting boxes on the original video in real time to achieve intuitive pipe outlines and defect marking boxes.

[0052] Step S1.2: Interactive selection and initialization of high-temperature hydraulic pipelines

[0053] like Figure 5 As shown, on the interactive processing terminal, the user selects the high-temperature hydraulic pipeline to be detected in the preview display, generates initial target prompt information, namely the high-temperature hydraulic pipeline image instance, and sends it to the data processing unit.

[0054] Step S1.3: Dynamic tracking and segmentation based on interactive prompts

[0055] The data processing unit inputs the initial target prompt information into a pre-trained video segmentation model. This model then performs spatiotemporal correlation reasoning and frame-by-frame tracking on subsequent video frames acquired by the thermal imaging acquisition unit. It activates and locks the user-selected high-temperature hydraulic pipeline image instance as the sole tracking object, thereby segmenting the high-temperature hydraulic pipeline from the environmental background and generating continuous, visualized images of the high-temperature hydraulic pipeline, such as... Figure 6 As shown.

[0056] In this embodiment, this step has the following three specific characteristics:

[0057] (1) Interactive target initialization and locking: Unlike fully automated blind detection, this step introduces a human-computer interaction process:

[0058] User selection (Prompt input): At the start of the detection phase (first frame of the video or any other time), the user selects (clicks) the high-temperature hydraulic pipeline to be detected in the preview display via the touchscreen of the interactive processing terminal. The interactive processing terminal captures the coordinates of this operation, generates an initial target prompt message (Prompt), and sends it to the data processing unit.

[0059] Target Locking: The data processing unit receives the initial target prompt information (Prompt), inputs it into the prompt encoder of the SAM 2 model, thereby activating and locking the user-specified high-temperature hydraulic pipeline instance as the sole tracking object.

[0060] (2) Model Construction and Improvement Innovation: Using the SAM2 basic large model as the video segmentation model, the following targeted improvements were made for handheld infrared detection scenarios to ensure tracking stability based on initial target cue information (Prompt):

[0061] 2.1) Thermal Infrared Domain Adaptation Fine-tuning: Since the native SAM 2 basic model is trained on a visible light dataset, its edge sensitivity to single-channel infrared images is insufficient. This embodiment embeds a lightweight thermal imaging feature adaptation layer into the image encoder of the SAM 2 basic model. Low-rank adaptation (LoRA) fine-tuning is performed on the adaptation layer using a labeled "high-temperature hydraulic pipe infrared dataset," enabling the SAM 2 basic model to accurately capture the edge gradients of high-temperature hydraulic pipes under low contrast, ensuring the identification of the high-temperature hydraulic pipe outline even in handheld shooting with unstable focus or motion blur.

[0062] 2.2) Spatiotemporal Memory Mechanism Against Field-of-View Drift: To address global motion in the frame caused by handheld devices, the SAM2 basic model utilizes its unique memory bank and memory attention module. The SAM2 basic model stores the historical features of the high-temperature hydraulic pipeline in its memory bank. Even if the high-temperature hydraulic pipeline shifts significantly in the frame due to operator hand tremors, or if the shooting angle changes due to movement, the SAM2 basic model can continuously "re-identify" and lock onto the high-temperature hydraulic pipeline in consecutive frames through spatiotemporal context association, achieving a stable tracking effect similar to a "visual gimbal."

[0063] 2.3) Dynamic Occlusion Perception and Recovery: During handheld detection, high-temperature hydraulic pipes may be temporarily obscured by the operator's limbs or other equipment. The improved SAM 2 basic large model enhances occlusion perception capabilities. When a target reappears after being temporarily obscured (or the field of view shifts back), it can quickly recover the target using long-term memory, ensuring the continuity of the segmentation mask.

[0064] (3) Automated processing:

[0065] 3.1) Robust Inference Across All Scenes: The SAM 2 basic model receives a continuous stream of jittery infrared video and outputs a binarized mask for each frame. Regardless of how the background changes with camera movement, the SAM 2 basic model always outputs only the mask for the high-temperature hydraulic pipe area.

[0066] Background dynamic culling: The generated dynamic mask is bitwise ANDed with the original thermal video frame. At this point, even if the background is constantly moving and changing, the calculation result only retains the temperature data of the high-temperature hydraulic pipeline area.

[0067] Technical effect: This step utilizes the powerful spatiotemporal tracking capabilities of the SAM2 basic large model and the guidance of user interaction commands to realize a "virtual visual gimbal". It overcomes the problem of unstable field of view caused by handheld device acquisition, ensures that the system only detects the target that the user wants to measure, eliminates interference from other irrelevant heat sources in the background, and generates a hose temperature data sequence with position alignment and a clean background.

[0068] Step S2: Coarse localization of statistically abnormal regions, i.e., screening of suspected defects.

[0069] Statistical temperature distribution analysis is performed on the segmented and visualized high-temperature hydraulic pipeline images to quickly identify hot spots, i.e., suspected defects, and generate suspected defect screening results.

[0070] The clean high-temperature hydraulic pipeline area (visualized high-temperature hydraulic pipeline image) segmented and output in step S1 is subjected to statistical temperature distribution analysis to quickly identify potential hot spots prone to cavitation, while minimizing sensor noise and ambient stray light interference. Considering the non-uniform lighting and natural temperature gradient along the hose during handheld inspection, this embodiment employs an adaptive local statistical and spatiotemporal joint screening strategy, specifically including the following sub-steps:

[0071] Step S2.1: Temperature data normalization

[0072] Temperature data within visualized high-temperature hydraulic pipeline images are mapped to a normalized grayscale matrix, and the temperature histogram distribution of the grayscale matrix is ​​calculated. Under normal operating conditions, the surface temperature of a high-temperature hydraulic pipeline typically follows a normal distribution or exhibits a smooth gradient distribution due to fluid temperature drop. Hot spots caused by cavitation appear as outliers at the high-temperature tail end of the histogram.

[0073] Step S2.2: Adaptive Local Threshold Segmentation

[0074] To overcome the natural temperature difference that may exist along the length of the high-temperature hydraulic pipeline (e.g., the temperature at the inlet is higher than that at the outlet) and the uneven lighting caused by the handheld shooting angle, this embodiment abandons the traditional global fixed threshold and adopts an adaptive local threshold algorithm based on sliding windows for image segmentation.

[0075] Algorithm logic: Define a size of A sliding window is used to traverse the grayscale matrix.

[0076] Discrimination formula: For the center pixel of the sliding window Calculate the average temperature of local neighboring pixels within the window. and standard deviation If a pixel satisfies the discrimination condition:

[0077]

[0078] in, For pixel temperature, This is the sensitivity coefficient (usually taken as 3~5, used to define the significance of the anomaly). If the value is a bias constant (used to filter out small temperature fluctuations), then the pixel is retained and recorded as 1; otherwise, the pixel is not retained and recorded as 0. This yields the binarized image of the high-temperature hydraulic pipeline segmentation.

[0079] Technical effect: This method can adaptively handle temperature gradient changes on the surface of high-temperature hydraulic pipelines. Even If one end is generally hot, it will not be falsely reported as long as there are no relatively local hot spots in that area; conversely, even if there is a relatively local abnormal temperature rise at the colder end, it can be sensitively detected.

[0080] Step S2.3: Optimize image morphological operations

[0081] The binarized image after thresholding and segmentation, following morphological filtering and geometric feature constraints, may contain isolated noise points or thin, elongated bright lines (non-cavitation features) caused by surface scratches. In this implementation, image morphological operations are used for optimization:

[0082] Opening operation: Erosion followed by dilation to eliminate isolated thermal noise points smaller than a preset area threshold (e.g., 3×3 pixels).

[0083] Closure operation: First expand and then corrode, filling the tiny cavitation cavitation hot spots and connecting the suspected fractured areas to form a complete connected domain.

[0084] Morphological filtering and geometric feature constraints: Calculate the circularity and aspect ratio of each connected region. Since heat conduction caused by cavitation usually presents as an approximately circular or elliptical diffusion pattern, regions with excessively large aspect ratios (appearing as thin strips, suspected of being reflective or scratched) are eliminated, and only connected regions that conform to the thermal diffusion morphology characteristics of defects such as cavitation are retained as suspected anomalous regions.

[0085] Step S2.4: Time-domain stability check

[0086] To address the potential for momentary glare and flicker during handheld shooting, a temporal domain verification step is added, leveraging the temporal continuity provided by the SAM 2 basic large model segmentation.

[0087] Verification logic: It not only analyzes the current frame, but also backtracks to the previous frame. Historical data for frames (e.g., the first 5 frames).

[0088] Judgment rule: Only when a suspected abnormal area is in a continuous Only when all defects are detected in the frame, and the displacement of the centroid position in the high-temperature hydraulic pipeline coordinate system is less than the preset threshold, are they marked as suspected defects and sent to step S3, fine identification of mechanical unsupervised anomalies.

[0089] Effect: Effectively filters out momentary interference caused by rapid camera movement or angle changes, ensuring that the heat source sent to the next stage is a continuous physical heat source.

[0090] Step S3: Unsupervised fine-grained anomaly identification, i.e., removing false positives and retaining true positives.

[0091] Unsupervised anomaly identification is performed on the suspected defect screening results to eliminate false hot spots that are not defects, thus obtaining the defect detection results.

[0092] In this embodiment, unsupervised fine-grained anomaly identification based on the Dinomaly2 unified framework is employed.

[0093] Although the suspected defects identified in step S2 have eliminated background interference, they may still contain non-cavitation pseudo-hot spots. To adapt to the current situation of mixed inspection of multiple types of high-temperature hydraulic pipelines in industrial sites and the extreme scarcity of cavitation defect samples, this embodiment adopts a full-spectrum system-unsupervised anomaly detection framework (i.e., the improved Dinomaly2 architecture) for fine discrimination.

[0094] This step abandons the cumbersome process of training a dedicated model for a single category, and adopts a design of "general feature extraction + context reconstruction", which specifically includes the following sub-processes:

[0095] Step S3.1: The general visual feature extraction system normalizes the suspected defects extracted in step S2 and inputs them into the pre-trained basic visual large model (such as ViT-L / 14).

[0096] General Representation: Unlike traditional methods that train small CNNs from scratch, this embodiment utilizes general visual representations learned by the base model on large-scale datasets. This gives the model an inherently strong understanding of texture, edges, and thermal gradients in infrared images, enabling it to extract robust high-dimensional feature vectors even with few samples (e.g., only 8 normal samples).

[0097] Parameter freezing: During inference, the parameters of the feature extraction network, i.e. the basic visual model, are kept frozen and used only to output multi-scale feature maps, ensuring the stability and generalization ability of the features.

[0098] Step S3.2: Feature Reconstruction

[0099] The extracted high-dimensional features are input into a feature reconstruction network, which is based on the "less is more" design philosophy and consists of the following key components to prevent the model from "memorizing" the input image (i.e., preventing identity mapping):

[0100] Context recentering: Before reconstruction, the feature reconstruction network first performs context recentering on the input features. By calculating the center of the normal sample feature distribution and aligning the input features to this center, the distribution offset between different batches and models of high-temperature hydraulic pipelines is eliminated. This step enables the system to use the same model to simultaneously detect multiple high-temperature hydraulic pipelines of different specifications.

[0101] Basic Dropout and Simplified Attention: To force the network to learn the inherent patterns of normal high-temperature hydraulic pipelines rather than simply copying pixels, the reconstruction network introduces a random dropout strategy, coupled with a simplified self-attention mechanism. This design disrupts the feature continuity of anomalous regions (cavitation hotspots), preventing the reconstruction network from using surrounding information to "reconstruct" the anomalous textures caused by cavitation.

[0102] Relaxed optimization objective: During the training phase (using only normal samples), a relaxed feature distance loss function is employed. This allows for minor differences in the reconstruction results in non-critical regions, thereby improving the model's tolerance to normal industrial environmental noise (such as slight manufacturing marks on the surface of high-temperature hydraulic pipes) and focusing on capturing genuine structural anomalies.

[0103] Step S3.3: Anomaly Scoring and Multimodal Adaptation

[0104] Feature difference calculation: Calculate input features With reconstructed network output features The cosine or Euclidean distance between them is used to generate a feature-level anomaly map.

[0105] Multimodal adaptation: Thanks to the natural adaptability of the Dinomaly2 architecture to multimodal data such as RGB-3D and RGB-IR, this embodiment directly applies the architecture to process infrared thermal imaging data without the need for special modifications to the network structure.

[0106] Final determination: The heat conduction caused by cavitation-induced inner layer rupture manifests as an "unnatural" heat diffusion pattern on the thermal image. This pattern does not exist in the feature space of normal samples, and therefore will be severely distorted by the reconstruction network, resulting in a significant increase in the feature distance. When the feature distance exceeds the adaptive threshold, it is determined to be a true defect.

[0107] Step S4: Augmented Reality Interactive Feedback

[0108] The visualized high-temperature hydraulic pipeline images and defect detection results are mapped back to the coordinates of the original video screen, and then overlaid and displayed on the screen of the interactive processing terminal with alarm prompts.

[0109] This invention can achieve scene adaptation and multi-point inspection operation, and supports multi-stage traversal inspection of multiple spatially discrete high-temperature hydraulic pipelines.

[0110] Supports multi-point interaction: The user holds the terminal and first aligns it with the high-temperature hydraulic pipeline at position A, then clicks the screen to lock the target for detection. Then, the user clicks the "Reset / Next Target" button on the terminal interface to move to position B and aligns it with the second high-temperature hydraulic pipeline to re-select and lock the target.

[0111] Unified model inference: Thanks to the context recentering mechanism in step S3, there is no need to recalibrate for the environmental differences between location A and location B (such as different background temperatures), and the same general model can be used to complete the detection continuously.

[0112] This invention features interactive result feedback and alarm functions:

[0113] Visual overlay feedback: On the real-time video, a semi-transparent green layer is used to outline the contour of the tracked high-temperature hydraulic pipeline; when defects such as cavitation are detected, a red highlighted mark box or a local thermal rendering layer is overlaid at the corresponding coordinate position.

[0114] Dynamic interactive alarm: An anomaly scoring trend chart is displayed on the side of the interface. Once a defect is confirmed, an audible and visual alarm is automatically triggered. At this time, the "Data Archive" button on the interface is activated, and the user can click it to save the currently marked video clip and analysis results into the database, completing the closed-loop interaction.

[0115] Initialization configuration and "cold start" training of this invention:

[0116] To ensure the accuracy of the unsupervised algorithm, the following configuration is performed before formal testing: Uncooled focal plane array infrared thermal imager parameter calibration: The data processing unit sets the emissivity parameter of the uncooled focal plane array infrared thermal imager to 0.92~0.96 to adapt to the plastic material; and enables the "scene-based automatic NUC" mode, which automatically closes the shutter for correction every preset time (e.g., 5 minutes) to ensure image purity.

[0117] "Cold start" on-site training: In the initial detection phase, operators record a video (approximately 1-2 minutes) of a normal hose. The data processing unit automatically extracts normal samples using the SAM 2 segmentation model, constructs a dataset, and trains the feature reconstruction network of the unsupervised anomaly detection model (Dinomaly2). This process is completed locally, without needing to upload to the cloud, enabling rapid system deployment.

[0118] Although the illustrative specific embodiments of the present invention have been described above to enable those skilled in the art to understand the invention, it should be understood that the invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the concept of the present invention are protected.

Claims

1. A method for infrared interactive detection of defects in high-temperature hydraulic pipelines, characterized in that, Includes the following steps: (1) Real-time tracking and segmentation of video stream based on interactive locking, segmenting the user-specified high-temperature hydraulic pipeline from the environmental background, and generating continuous visualized high-temperature hydraulic pipeline images; 1.1) Video capture and real-time preview; The infrared thermal image video data of the high-temperature hydraulic pipeline in working condition is recorded by the thermal imaging acquisition unit, and the infrared thermal image video data is transmitted to the interactive processing terminal in real time for preview display. 1.2) Interactive selection and initialization of high-temperature hydraulic pipelines; On the interactive processing terminal, the user selects the high-temperature hydraulic pipeline to be detected in the preview display, generates initial target prompt information, namely the high-temperature hydraulic pipeline image instance, and sends it to the data processing unit. 1.3) Dynamic tracking and segmentation based on interactive prompts; The data processing unit inputs the initial target prompt information into the pre-trained video segmentation model. The video segmentation model is then used to perform spatiotemporal correlation reasoning and frame-by-frame tracking on the video frames acquired by the subsequent thermal imaging acquisition unit. The model activates and locks the image instance of the high-temperature hydraulic pipeline selected by the user as the only tracking object, thereby segmenting the high-temperature hydraulic pipeline from the environmental background and generating a continuous and visualized high-temperature hydraulic pipeline image. (2) Coarse localization of statistically abnormal regions, i.e., screening of suspected defects; Statistical temperature distribution analysis is performed on the segmented and visualized high-temperature hydraulic pipeline images to quickly identify hot spots, i.e., suspected defects, and generate suspected defect screening results. (3) Fine-grained identification of unsupervised anomalies, i.e., distinguishing the true from the false; Unsupervised anomaly identification is performed on the suspected defect screening results to remove false hot spots that are not defects, thus obtaining the defect detection results; (4) Augmented reality interactive feedback; The visualized high-temperature hydraulic pipeline images and defect detection results are mapped back to the coordinates of the original video screen, and then overlaid and displayed on the screen of the interactive processing terminal with alarm prompts.

2. The infrared interactive detection method for defects in high-temperature hydraulic pipelines according to claim 1, characterized in that, Step (2) involves performing a statistical temperature distribution analysis on the segmented and visualized high-temperature hydraulic pipeline image to quickly identify hot spots, i.e., suspected defects, and generating the suspected defect screening results as follows: 2.1) Temperature data normalization; The temperature data within the visualized high-temperature hydraulic pipeline image is mapped into a normalized grayscale matrix; 2.2) Adaptive local threshold segmentation; Define a size of A sliding window is used to traverse the grayscale matrix, and for the center pixel of the sliding window... Calculate the average temperature of local neighboring pixels within the window. and standard deviation If the pixel meets the discrimination condition: ； in, For pixel temperature, This is the sensitivity coefficient. To filter out the bias constant of tiny temperature fluctuations, the pixel is retained and recorded as 1; otherwise, the pixel is not retained and recorded as 0. This yields the binarized image of the high-temperature hydraulic pipeline after segmentation. 2.3) Optimize image morphological operations; Opening operation: Erosion followed by expansion to eliminate isolated thermal noise points smaller than a preset area threshold; Closure operation: First expand and then corrode to fill the tiny cavitation hot spots inside, connect the suspected broken areas, and make them form a complete connected domain; Morphological filtering and geometric feature constraints: Calculate the roundness and aspect ratio of each connected domain, remove regions with excessive aspect ratios, and retain only connected domains that conform to the morphological characteristics of defect thermal diffusion as suspected abnormal regions; 2.4) Time-domain stability verification; Pre-retrogression Historical data of frames, when a suspected anomaly area is in continuous A suspected defect is only identified when all defects are detected in the frame and the displacement of its centroid position in the high-temperature hydraulic pipeline coordinate system is less than a preset threshold.

3. The infrared interactive detection method for defects in high-temperature hydraulic pipelines according to claim 1, characterized in that, The video segmentation model described in step 1.3) adopts the SAM 2 basic large model.

4. The infrared interactive detection method for defects in high-temperature hydraulic pipelines according to claim 1, characterized in that, Step (3) involves performing unsupervised fine-grained anomaly identification on the suspected defect screening results, eliminating false hotspots that are not defects, and obtaining the defect detection results as follows: Unsupervised anomaly identification based on the Dinomaly2 unified framework was used to identify and eliminate false hotspots that were not defects, thus obtaining defect detection results.

Images