A stainless steel pipe fitting inner wall cleanliness detection method based on intelligent sensor

Abstract

This invention relates to the field of pipe fitting inner wall inspection technology, and discloses a method for detecting the cleanliness of the inner wall of stainless steel pipe fittings based on intelligent sensors. The method includes: using a multi-sensor progressive fusion discrimination technology based on high-speed continuous imaging, polarized light, ultraviolet fluorescence, and mechanical micro-perturbation to achieve high-precision intelligent differentiation of bubbles and particles and a detection effect with an extremely low false judgment rate; using digital twin virtual-real response mapping and production process root cause tracing technology to achieve self-diagnosis of the detection system status, accurate location of abnormal process parameters, and pre-demonstration optimization of new specification strategies; and using thermal-fluid-solid-electric multi-physics field coupling evolution prediction and real-time feedback technology of upstream processes to achieve early warning of pollutant risks, preventive elimination of stubborn stains, and closed-loop self-evolutionary control of manufacturing quality.

Description

Technical Field

[0001] This invention relates to the field of pipe fitting inner wall inspection technology, specifically a method for detecting the cleanliness of the inner wall of stainless steel pipe fittings based on intelligent sensors. Background Technology

[0002] In modern industrial production, especially in industries with extremely high requirements for pipe cleanliness, such as pharmaceuticals, food and beverages, and semiconductors, the cleanliness of the inner wall of stainless steel pipes directly affects product quality, production efficiency, and even safety. Traditional cleanliness testing methods often suffer from drawbacks such as low efficiency, strong subjectivity, difficulty in real-time monitoring, and potential damage to pipes. The core of the stainless steel pipe inner wall cleanliness testing method based on intelligent sensors lies in using specific intelligent sensors to sense and quantify the microscopic state of the inner wall of the pipe, and intelligently analyzing this data to assess its cleanliness. This method can achieve quantitative and objective cleanliness assessment, reduce human error, avoid any damage to the pipe, ensure the integrity of the pipe, provide real-time feedback, and be easily integrated into existing production equipment or automated processes. At the same time, the test data can be recorded and stored, facilitating quality traceability and process improvement. Furthermore, appropriate sensors and analysis models can be selected or customized according to the characteristics of different industries and different contaminants. By integrating more sensor types and optimizing AI algorithms to improve the generalization ability and interpretability of the model, remote monitoring and predictive maintenance can be achieved. Existing methods for detecting the cleanliness of the inner walls of stainless steel pipe fittings based on smart sensors struggle to distinguish between bubbles and particles with high precision. They often lead to misjudgments due to similar optical characteristics. They cannot effectively integrate multi-source heterogeneous data such as high-speed vision, polarized light, fluorescence, and mechanical response to achieve precise spatiotemporal alignment. They lack the ability to predict detection strategies for new pipe fitting specifications, resulting in frequent blind spots and misjudgment hotspots. They are unable to trace the root causes of deviations in process parameters or equipment degradation in the production process, cannot predict the evolution trend and risk sequence of contaminants into stubborn stains, and fail to establish a real-time closed loop between detection feedback and upstream process optimization. They cannot achieve self-evolution and preventive intervention in quality control, thus their practicality has certain limitations. Summary of the Invention

[0003] This invention provides a method for detecting the cleanliness of the inner wall of stainless steel pipe fittings based on intelligent sensors, which helps to solve the problems mentioned in the background art.

[0004] This invention provides the following technical solution: a method for detecting the cleanliness of the inner wall of stainless steel pipe fittings based on intelligent sensors, comprising: Perform basic parameter acquisition and validity determination, construct testing entry thresholds, and screen out objects that are physically incompatible or have unreasonable processes; Perform multi-sensor data acquisition, build a multi-dimensional perception foundation, and form a high-fidelity multi-source dataset to provide sufficient information support for the intelligent differentiation of bubbles and particles; Perform time-series dynamic feature analysis and preliminary discrimination, establish preliminary discrimination rules for bubbles and particles, quickly divert high-confidence cases, and retain verification channels for fuzzy cases; Multispectral optical feature cross-validation is performed, and the problem of temporal ambiguity is solved by cross-validation of liquid film curvature optical response and liquid fluorescence edge effect, providing optical fingerprint identification for bubbles and particles; The final judgment of mechanical micro-perturbations breaks the deadlock of spectral verification, achieves the ultimate distinction between bubbles and particles, and retains a channel for manual intervention in abnormal cases; Perform virtual-real response mapping and difference comparison, build a virtual-real comparison bridge, map entity detection events to digital twins for ideal response prediction, verify the effectiveness of the model through multi-dimensional difference quantification, and identify system drift or process deviation. Perform root cause tracing and strategy pre-simulation optimization, locate process execution deviations or equipment performance degradation through twin playback, generate targeted corrective measures, and simulate the optimal detection strategy for new specifications. Perform multiphysics evolution analysis and final decision-making, predict the fate of pollutants through multiphysics coupling evolution, distinguish between emergency and routine treatment, generate targeted process optimization solutions and feed them back into the production system.

[0005] Preferably, the basic parameters are obtained and validity is determined, specifically as follows: Obtain the unique identification code of the pipe fitting, link it to the production order database, and automatically extract the file information; Measure the pipe fitting data and record the minimum inner diameter value; Randomly inspect the outer surface of pipe fittings to confirm whether the deviation between the nominal value and the measured value of surface roughness is within the allowable range, and mark the abnormal pipe fittings; Read the cleaning process records that the pipe fitting has undergone; Check each of the above collected parameters for any empty or missing values ​​to confirm that the key parameters have been fully acquired. Compare the minimum inner diameter of the pipe fitting with the minimum fit size of the sensor probe, and compare the length of the pipe fitting with the maximum detection stroke of the equipment to confirm physical compatibility; Determine whether the drying temperature is within the recommended range of the cleaning solution, whether the drying time reaches the standard calculated value for the pipe diameter, and mark any parameter combinations that deviate from the standard process. If the parameters are complete and within the range of equipment capabilities and process standards, output "Parameter valid" and execute multi-sensor data acquisition; If there are missing parameters, out-of-range parameters, or process deviations, output "Parameter Abnormality" and terminate the process, simultaneously generating an explanation of the cause of the abnormality and suggestions for handling it.

[0006] Preferably, multi-sensor data acquisition is performed, specifically as follows: Start the sensor assembly, perform zero-point calibration, light source intensity calibration, and gas path sealing check, and confirm the readiness status of each subsystem; The robotic arm clamps the pipe to establish a spatial coordinate system, adjusts the pipe axis to be coaxial with the sensor probe, sets the initial position of the probe at the pipe end face, and confirms that there is no physical interference. Set the acquisition parameters, trigger the continuous shooting mode, and the probe advances and rotates at a constant speed, simultaneously recording the time sequence image stream and position coordinates; During the advancement process, polarized light imaging is triggered at preset intervals; At critical locations, propulsion is paused, and a short-term pulsed airflow is applied through an integrated air nozzle, while simultaneously recording image sequences and mechanical sensor feedback before and after the disturbance. Establish a spatiotemporal correlation index based on a unified timestamp; Check image sharpness, illumination uniformity, and whether there is motion blur or overexposure; mark frames with abnormal quality and decide whether to resample locally or mark them as suspicious data. The verified multi-source data will be packaged and output in a standard format.

[0007] Preferably, the dynamic feature analysis and preliminary judgment of the execution time series are performed, specifically as follows: Denoising and illumination normalization are performed on continuous image sequences to identify and segment suspected target areas, eliminating interference from pipe wall background and fixed structure; Extract the edge contours of the target in each frame image, establish cross-frame contour correspondence, and record the change sequence of edge point coordinates over time; Analyze the high-frequency fluctuation components of the edge point coordinates, calculate the number and amplitude of edge fluctuations per unit time, and assess the dynamic activity level of the edge. Calculate the pixel area of ​​the target region in each frame, plot the area change curve over time, and identify the periodic pulsation pattern and its frequency characteristics; Using the target's center of gravity as a reference point, track its position change in the image coordinate system, fit the displacement trajectory curve, and compare it with the probe's motion parameters; Calculate the correlation coefficient between the target displacement and the probe's advancing and rotating motion to determine whether the target motion is synchronized with the probe motion or whether there are abnormal patterns such as lag, lead, or irrelevance. By combining three indicators—edge tremor frequency, area change period, and motion synchronization—and matching them with preset bubble feature patterns or particle feature patterns, a preliminary judgment result is output. Calculate the confidence level of the preliminary judgment; If all three indicators point to the same feature and the confidence level is higher than the threshold, then a confirmation message is output. If the indicators are contradictory or the confidence level is insufficient, they are marked as "temporally ambiguous" and multispectral optical feature cross-validation is performed for multispectral verification.

[0008] Preferably, multispectral optical feature cross-validation is performed, specifically as follows: Based on the analysis of the dynamic features of the execution time sequence and the preliminary judgment of the marked temporally ambiguous target location, the image set at the corresponding location is retrieved to ensure that the spatial coordinates of the images in the image set are consistent; By comparing the brightness distribution of parallel polarized and cross-polarized images, the birefringence halo pattern generated by the curved liquid film can be identified. If a typical halo pattern is detected and the polarization state difference is obvious, it is marked as "positive liquid film surface feature"; If the polarization image shows uniform reflection or metallic luster without obvious halo, it is marked as "positive solid surface feature"; If the feature falls between the two, it is marked as "polarization blur". By comparing ultraviolet fluorescence images with visible light images, we can identify whether there are fluorescent rings at the edge of the target and whether the internal region has no fluorescence or the fluorescence is significantly weaker than that at the edge. If a clear fluorescent ring is present at the edge but there is no fluorescence inside, it is marked as "positive for liquid-encapsulated gas characteristics"; If there is no fluorescence overall or uniform solid fluorescence, it is marked as "positive for solid particle characteristics"; If the fluorescence distribution is abnormal or interfered with by the background, it is marked as "fluorescence blurry"; Cross-validation was performed using the combined polarization determination results and the fluorescence determination results; If the polarization characteristics and fluorescence characteristics contradict each other, or if either is marked as ambiguous, it is determined as "spectral crossover unresolved", and mechanical perturbation final discrimination is performed to verify the mechanical perturbation. If the two spectra are consistent and clear, the final confirmation result will be output. Output one of three decisions: "bubble spectrum confirmation", "particle spectrum confirmation" or "final judgment of mechanical micro-perturbation", and simultaneously record the quantitative description of each spectral feature as a chain of evidence.

[0009] Preferably, the final determination of mechanical micro-perturbations is as follows: Based on the spatial coordinates of the unresolved target marked after cross-validation of multispectral optical features, the probe is driven to precisely retract to that position, visual tracking is re-established, and the target is ensured to be within the effective range of mechanical action. Assess the target size and adhesion status; A short-duration low-pressure pulse is applied, and the pressure gradient is gradually increased while the target's shape changes and displacement trajectory are recorded at high speed simultaneously. Frequency scanning is used to find the target resonance response point, an adjustable standing wave field is established, the aggregation or repulsion motion of the target between the sound pressure node and the antinode is observed, and its amplitude and phase response are recorded. Analyze high-speed image sequences to identify whether the target exhibits a response pattern; If a response pattern is detected, it is classified as a "fluid response pattern" and ultimately determined to be a bubble. "Bubble Confirmation" is output, and virtual-real response mapping and difference comparison are performed for digital twin traceability and process optimization. If no response mode occurs or only slight rolling occurs without deformation, it is classified as "solid response mode", and is ultimately determined to be particles. The output is "particle confirmation", and the process is transferred to the cleanliness level assessment process to count the size and quantity to determine the level. If the response is between the two or cannot be identified, it will be determined as a "fuzzy response", marked as "abnormal and unclear", the inspection status of the pipe fitting will be frozen, the manual re-inspection process will be triggered, and all original data will be retained for future reference. The complete discrimination chain, which includes performing time-series dynamic feature analysis and preliminary discrimination, performing multispectral optical feature cross-validation, and performing mechanical micro-perturbation final discrimination, is summarized and a stage report is generated.

[0010] Preferably, the virtual-to-real response mapping and difference comparison are performed, specifically as follows: Based on the basic parameters of the pipe fitting obtained through the execution of basic parameter acquisition and validity determination, the corresponding specification's geometric model, material attribute library, and standard process template are retrieved from the digital twin library to activate the virtual instance of the pipe fitting. The key attributes of the detected events confirmed after the final judgment of the mechanical micro-perturbation are transformed into twin input parameters, and the corresponding scene is reconstructed in virtual space; Drive the twin to execute a virtual detection process, simulate the ideal response under the same sensor configuration and motion parameters, and output virtual prediction data; Using spatiotemporal coordinates as an index, the actual response data of entity detection is aligned with the ideal response data predicted by twins at the frame level to establish a benchmark for point-by-point comparison. The system calculates four indicators—morphological difference, trajectory deviation, spectral similarity, and mechanical response deviation—to comprehensively evaluate the overall degree of consistency between the physical and virtual entities. Compare the overall difference with the preset tolerance range; If all four indicators are below their respective thresholds and the overall consistency meets the standard, it is judged as "virtual and real consistent", confirming the effectiveness of the twin model, and directly performing multiphysics evolution analysis and final decision-making for evolution prediction; If any indicator exceeds the standard or the overall consistency is insufficient, it is judged as "deviation between reality and virtuality". Anomaly root cause tracing and strategy pre-simulation optimization are performed to trace the root cause and further analyze the deviation characteristics. Output either "Virtual vs. Real" or "Virtual vs. Real" conclusions, along with a quantitative report of the difference and a deviation pattern label.

[0011] Preferably, the abnormal root cause tracing and strategy pre-simulation optimization are performed, specifically as follows: Load the complete manufacturing history of the pipe fitting into the twin, replay each process sequentially according to the timeline, and extract the process parameters of each process; The actual execution parameter curves are compared point by point with the standard process template to identify abnormal periods where the deviation exceeds the tolerance, and the timestamps and magnitudes of parameter mutations are marked. Analyze abnormal parameter combinations; Based on the identified anomaly type, the optimization knowledge base is invoked to generate an adjustment plan; If the process parameters are displayed as normal throughout, retrieve the sensor performance degradation model, compare the current light source output intensity with the initial calibration value, analyze the trend of lens transmittance change, and assess the degree of probe mechanical wear. If the light source intensity decreases by more than 20%, or the color temperature drifts significantly, or the lens contamination causes a decrease in contrast, or the probe bearing clearance exceeds the standard, then the system will identify "abnormal equipment status" and generate a maintenance work order and spare parts replacement suggestions. For pipe fittings with changes in specifications or materials that are being inspected for the first time, the detection effects of different sensor combinations, sampling densities, light source angles, and propulsion speeds are tested in batches within a twin to simulate the distribution of blind spots and heat maps of misjudged hotspots. If the simulation shows that the current strategy has an unacceptable blind spot or the false judgment rate exceeds the standard, then output "strategy optimization suggestion", adjust the parameters and return to execute multi-sensor data acquisition to re-execute the complete detection process; If the pre-run verifies that the current strategy is comprehensive and the misjudgment rate is controllable, then "Strategy Confirmed" is output, and multiphysics evolution analysis and final decision-making are performed.

[0012] Preferably, multiphysics evolution analysis and final decision-making are performed, specifically as follows: Key physical property parameters are extracted from cases of consistent virtual and real bubbles confirmed after performing virtual-real response mapping and difference comparison. Based on the extracted parameters, a heat-mass transfer coupling model was constructed to simulate the complete phase transition path of the liquid film from liquid to solid state under the current drying conditions, and to track the spatiotemporal evolution of solute precipitation and solvent volatilization. Based on the phase transition simulation results, the morphology of the final dried residue is predicted; Based on the initial thickness of the liquid film and the current drying rate, calculate the critical time required for complete drying in theory. Compare the theoretically required critical time for complete drying with the actual drying time obtained from the acquisition of basic parameters and effectiveness determination. If the actual drying time is lower than the critical value, an "insufficient drying" judgment will be output, quantifying the time gap. Multiphysics simulations were performed to model the asymmetric coalescence, rupture, or stable attachment modes of bubbles under complex boundary conditions. If the simulation shows that the local thermal boundary layer is too thick, causing abnormal bubble aggregation, the output will be "significant thermal boundary layer effect". It is recommended to optimize the hot air nozzle layout or increase turbulence to promote mixing. If the display shows that the electrostatic potential gradient drives the bubbles to migrate and attach to a specific wall surface, the output will be "significant electrostatic interference". It is recommended to increase the ion wind to neutralize or adjust the ambient humidity. For particle cases confirmed after performing virtual-real response mapping and difference comparison, calculate the adhesion strength and critical resuspension shear stress; If the current flow field shear stress is close to or exceeds the critical value, the risk of "false adhesion" will be output. It is recommended to reduce the movement speed of the detection probe or increase the pretreatment inert gas purging. Establish a time-varying evolution model for pollutants to simulate the time scale and path of the transformation from current bubble residue or particle attachment to "fresh residue - aged residue - stubborn stains"; If it is predicted that the pipe fittings will be converted into a stubborn state of chemical bonding or mechanical interlocking before the next process, an "emergency handling" suggestion will be output, triggering priority rework or isolation of the batch of pipe fittings. If it is predicted that the current clearable state can be maintained for more than the safety buffer period, then the "normal processing" suggestion will be output and the process will be handled according to standard logistics. The analysis results of the entire process, including execution time-series dynamic feature analysis and preliminary judgment, execution multispectral optical feature cross-validation, execution mechanical micro-perturbation final judgment, execution virtual-real response mapping and difference comparison, and execution anomaly root cause tracing and strategy pre-simulation optimization, are integrated to generate a comprehensive optimization scheme. This scheme is fed back to the upstream production process through the manufacturing execution system. The complete parameter set, evolution path, prediction results and actual outcome of this case are associated and archived in the digital sample library for continuous training and optimization of the twin model.

[0013] The present invention has the following beneficial effects:

[0014] 1. This method for detecting the cleanliness of the inner wall of stainless steel pipe fittings based on intelligent sensors first completes the comprehensive acquisition and validity verification of basic information of the pipe fittings in the basic sensing and data acquisition layer. This includes the integrity verification and physical compatibility confirmation of geometric dimensions, material properties, surface condition, and upstream process parameters, ensuring that the object being tested meets the equipment's capability range and process standards, and filtering out invalid objects that are out of specification or missing parameters. Subsequently, a multi-sensor collaborative acquisition mechanism is activated, coordinating subsystems such as high-speed continuous imaging, polarization light feature extraction, ultraviolet fluorescence induction, and mechanical micro-perturbation verification. Multimodal sensing data is acquired simultaneously and a strict spatiotemporal alignment relationship is established, forming a multi-source heterogeneous dataset covering time-series image streams, spectral feature sets, and mechanical response curves. This provides high-fidelity and high-completeness raw data support for subsequent in-depth analysis. Through pre-parameter verification and multimodal data acquisition, invalid objects being tested are filtered out from the source, ensuring the quality and completeness of the input data, laying a reliable foundation for accurate identification, and improving the robustness and efficiency of the overall detection process.

[0015] 2. This method for detecting the cleanliness of the inner wall of stainless steel pipe fittings based on intelligent sensors employs a progressive multimodal fusion strategy in its intelligent identification and judgment layers. First, it performs pixel-level temporal analysis on continuous image sequences to extract microscopic features such as edge dynamic tremor frequency, area periodic pulsation, and overall displacement trajectory, establishing preliminary discrimination hypotheses for bubbles and particles. For complex cases with ambiguous temporal features, a dual-spectral cross-validation mechanism of polarized light and ultraviolet fluorescence is introduced. Optical ambiguities are resolved using the birefringence halo effect of the liquid film surface and the fluorescence edge effect. Finally, mechanical micro-perturbation tests using controllable airflow pulses or ultrasonic standing wave fields are conducted, achieving final arbitration based on the essential difference between fluid deformation and rupture and solid rigidity maintenance. The three-layer judgment mechanism is interconnected, gradually eliminating uncertainty and constructing a complete evidence chain and credible decision-making path from suspected target screening to deterministic classification. Through progressive verification in the three dimensions of temporal, spectral, and mechanical dimensions, the method significantly improves the distinction accuracy between bubbles and particles, effectively overcoming the limitations of a single sensor, greatly reducing the risk of misjudgment and missed detection, and ensuring the reliability of the classification results.

[0016] 3. This method for detecting the cleanliness of the inner wall of stainless steel pipe fittings based on intelligent sensors constructs a virtual-real symbiotic mechanism through digital twin verification and optimization layers. It maps physical detection events to a virtual space in real time for ideal response simulation and comparison. Through multi-dimensional difference quantification, it identifies system drift, process deviation, or equipment degradation, deeply traces back the spatiotemporal data of the production process, accurately locates the timing of process parameter mutations and sensor performance degradation trajectories, distinguishes between process execution anomalies and equipment condition deterioration, and pre-simulates optimal detection strategies for new specifications to eliminate blind spots. Furthermore, based on thermal-fluid-solid-electric multi-physics field coupling simulation, it tracks the phase transition evolution, coalescence and fracture dynamics, and aging transformation paths of pollutants, predicts the temporal risk of their development into stubborn stains, generates targeted process optimization schemes, and feeds them back to the upstream production system in real time. Simultaneously, it archives the evolutionary laws into a knowledge base to achieve model self-evolution, completing a value loop from quality inspection to process optimization. Through virtual-real symbiosis and evolutionary prediction, it achieves a shift from passive detection to proactive prevention, continuously optimizing upstream process parameters, constructing a self-evolving quality assurance system, and improving the stability and controllability of the overall manufacturing process. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method for detecting the cleanliness of the inner wall of stainless steel pipe fittings based on intelligent sensors according to the present invention. Detailed Implementation

[0018] Example 1: A method for detecting the cleanliness of the inner wall of stainless steel pipe fittings based on smart sensors, see reference. Figure 1 ,include: The system acquires basic parameters and determines their validity, establishes testing entry thresholds, and filters out objects that are physically incompatible or have unreasonable processes through identity information association, geometric measurement, process traceability, and multi-dimensional verification. This eliminates invalid testing from the source and ensures the reliability of data and the safety of equipment in subsequent steps. Perform multi-sensor data acquisition to build a multi-dimensional perception foundation. Through high-speed synchronous acquisition and spatiotemporal alignment of time series, spectral features, and mechanical response, a high-fidelity multi-source dataset is formed, providing sufficient information support for the intelligent differentiation of bubbles and particles. Perform time-series dynamic feature analysis and preliminary discrimination. Extract three core features—edge dynamics, area period, and motion synchronization—through pixel-level time-series analysis. Establish preliminary discrimination rules for bubbles and particles. Quickly divert high-confidence cases and retain verification channels for ambiguous cases to achieve a balance between efficiency and accuracy. By performing cross-validation of multispectral optical features and utilizing the complementary dual-spectral properties of polarized light and fluorescence, the problem of temporal ambiguity is solved through cross-validation of liquid film curvature optical response and liquid fluorescence edge effect, providing molecular-level optical fingerprint identification for bubbles and particles. The final judgment is made by performing mechanical micro-perturbation. By controlling the physical micro-perturbation of airflow or ultrasonic field, the inherent mechanical response difference between fluid and solid is utilized to break the deadlock of spectral verification, achieve the ultimate judgment of bubbles and particles, and retain the channel for manual intervention in abnormal cases. Perform virtual-real response mapping and difference comparison, build a virtual-real comparison bridge, map entity detection events to digital twins for ideal response prediction, verify the effectiveness of the model through multi-dimensional difference quantification, identify system drift or process deviation, and lay the foundation for accurate traceability or reliable prediction. Perform root cause tracing and strategy pre-simulation optimization for anomalies. By playing back the twin to locate process execution deviations or equipment performance degradation, generate targeted corrective measures, and pre-simulate the optimal detection strategy for new specifications, realize closed-loop management from anomaly diagnosis to strategy evolution, and ensure the continuous and reliable operation of the detection system. Perform multiphysics evolution analysis and final decision-making, predict the fate of pollutants through multiphysics coupling evolution, distinguish between emergency and routine treatment, generate targeted process optimization solutions and feed them back into the production system, and realize the ultimate transformation of detection value into a closed loop of manufacturing quality.

[0019] Specifically, the acquisition of basic parameters and the determination of validity are performed as follows: The unique identification code of the pipe fitting is obtained by scanning barcode or reading RFID, and linked to the production order database to automatically extract file information such as specifications, production batch, and production date. Pipe data is measured using laser rangefinders or mechanical calipers, and the minimum inner diameter value is recorded to determine the compatibility of the sensor probe. The pipe data includes the actual outer diameter, inner diameter, wall thickness, and total length. By visually inspecting or using a roughness tester to randomly inspect the outer surface of the pipe fittings, confirm whether the deviation between the nominal value and the measured value of the surface roughness is within the allowable range. Mark the abnormal pipe fittings, that is, the pipe fittings whose measured value of surface roughness deviates from the nominal value by more than the allowable range (such as ±20%). This may be due to processing defects or handling damage and they need to be isolated. The system reads the cleaning process records that the pipe fitting has undergone from the manufacturing execution system, including the type and concentration of cleaning fluid, the number of rinsings, the drying temperature profile, the drying duration, and the cooling method. The manufacturing execution system is a production information management system that connects the enterprise planning layer and the workshop control layer, and records manufacturing data such as process parameters, equipment status, and material flow in real time. Check each of the above collected parameters for any missing or empty values, and confirm that the key parameters (inner diameter, drying time) have been completely obtained. Compare the minimum inner diameter of the pipe fitting with the minimum fit size of the sensor probe, and compare the length of the pipe fitting with the maximum detection stroke of the equipment to confirm physical compatibility: First, measure the minimum inner diameter of the pipe fitting; second, check the minimum fit size in the technical specifications of the sensor probe; third, compare the two to ensure that the probe can be safely inserted and a gap is maintained; finally, compare the length of the pipe fitting with the maximum detection stroke of the equipment to ensure full coverage without overtravel. If all four comparisons are met, compatibility is confirmed. Determine whether the drying temperature is within the recommended range of the cleaning solution and whether the drying time reaches the standard calculated value for the pipe diameter. Mark the parameter combinations that deviate from the standard process. The standard calculated value refers to the theoretical shortest drying time calculated by the drying kinetic model based on heat and mass transfer theory, cleaning solution physical properties (viscosity, surface tension, boiling point), pipe geometric parameters (inner diameter, wall thickness, roughness), and environmental conditions (temperature, humidity, air pressure). If the parameters are complete and within the range of equipment capabilities and process standards, output "Parameter valid" and execute multi-sensor data acquisition; If there are missing parameters, out-of-range parameters, or process deviations, output "parameter abnormality" and terminate the process, and simultaneously generate an explanation of the cause of the abnormality and suggestions for handling it. The criteria for "within the range of equipment capabilities and process standards" and "out of range or process deviation" are determined through a six-dimensional assessment: First, the completeness of parameters is checked (no missing items); second, geometric compatibility is checked (inner diameter greater than the minimum probe size, length less than the stroke); third, temperature compliance is checked (within the recommended range of cleaning fluid); fourth, time adequacy is checked (meeting the standard calculated value); fifth, roughness consistency is checked (actual deviation from nominal deviation is within tolerance); and finally, logical rationality is checked (no contradictions between parameters). Passing all six dimensions indicates the equipment is within the range; failure in any one dimension indicates the equipment is out of range or deviating from the standard. For example, when a batch of stainless steel pipe fittings enters the inspection station, the system automatically associates the following information after scanning the barcode: material 304 stainless steel, nominal inner diameter 15mm, length 800mm, roughness Ra0.8μm, cleaning solution is alkaline water-based cleaning agent, drying temperature 80℃, drying time 8 minutes. The minimum inner diameter measured on-site is 14.5mm, and it is confirmed that the minimum fit size of the sensor probe is 12mm, which is physically compatible. However, it is found that the standard drying temperature of the cleaning agent should be 60-70℃. The current 80℃ is too high and may cause the residual liquid to form a film quickly, which is not conducive to volatilization. In addition, the standard drying time should be 12 minutes, and the current 8 minutes is insufficient. The system judges "process parameters deviate" and outputs "parameter abnormality". It is recommended to adjust the drying process to 65℃ and 12 minutes and then resubmit for inspection, rather than directly entering the inspection process and causing the risk of misjudgment.

[0020] Example 2 is an improvement on Example 1. This method for detecting the cleanliness of the inner wall of stainless steel pipe fittings based on intelligent sensors performs multi-sensor data acquisition, specifically as follows: Start the high-speed camera, polarization light source, ultraviolet light source, airflow nozzle, ultrasonic transducer and other sensor components, and perform zero-point calibration, light source intensity calibration and gas path sealing check to confirm the readiness of each subsystem. The steps for zero-point calibration, light source intensity calibration and gas path sealing check are as follows: First, move the probe to the mechanical origin, confirm the position through the limit switch and clear the coordinates to complete the zero-point calibration; second, use a standard white board to collect the reflected light intensity, compare it with the factory reference value and adjust the light source drive current to the standard output to complete the intensity calibration; finally, close the gas path valve and monitor the pressure sensor reading change rate. If the pressure drop is less than the threshold, the sealing is deemed qualified. If all three items pass, the system is confirmed to be ready. The robotic arm clamps the pipe to establish a spatial coordinate system, adjusts the pipe axis to be coaxial with the sensor probe, sets the initial position of the probe at the pipe end face, and confirms that there is no physical interference. The steps to confirm no physical interference are as follows: First, establish a three-dimensional envelope model of the pipe and the probe. Second, perform motion simulation in virtual space to calculate the minimum distance between each component during the propulsion and rotation process. Third, set a safety clearance threshold (e.g., 5 mm) and check if there is any instantaneous distance less than the threshold during the entire stroke. If so, adjust the clamping angle of the pipe or the initial position of the probe, and re-simulate until the minimum distance of the entire stroke is greater than the threshold. Finally, perform a low-speed test run in actual space to confirm no contact. Set the acquisition parameters, trigger the continuous shooting mode, and the probe advances and rotates at a constant speed, synchronously recording the time-series image stream and position coordinates. The acquisition parameters include camera frame rate, exposure time, and gain parameters. The time-series image stream refers to the image sequence acquired continuously in chronological order. Each frame of the image is accompanied by a precise timestamp, forming a continuous time-image data stream, which is used to record the dynamic change process of the target. During the advancement process, polarized light imaging is triggered at a preset interval: two images, one parallel polarized and one cross-polarized, are acquired. The ultraviolet light source is then switched to acquire a fluorescence excitation image, and the spatial position of each image is marked. The preset interval refers to the spatial sampling interval set according to the detection accuracy requirements (e.g., triggering once every 10 mm of advancement). Its function is to control the amount of data while ensuring detection coverage, avoiding redundancy caused by oversampling or missed detection caused by insufficient sampling. At key locations, advance is paused, and short-term pulsed airflow is applied through integrated air nozzles, or an ultrasonic transducer is activated to establish a standing wave field. Simultaneously, image sequences before and after the disturbance and feedback from mechanical sensors are recorded. Key locations refer to the spatial coordinates marked as "temporally ambiguous" or "suspected target" in the execution of temporal dynamic feature analysis and preliminary judgment. Areas with abnormal features but which cannot be clearly determined are identified through image analysis and used as candidate points for mechanical verification. Using a unified timestamp as a benchmark, high-speed images, polarization images, fluorescence images, mechanical response data, and probe pose data are synchronized at the frame level to establish a spatiotemporal correlation index. Frame-level synchronization refers to unifying the time benchmark of each data source with microsecond-level precision to ensure that high-speed images (obtained through temporal dynamic feature analysis and preliminary discrimination), polarization images (obtained through multispectral optical feature cross-validation), fluorescence images (obtained through multispectral optical feature cross-validation), mechanical responses (obtained through final discrimination of mechanical micro-perturbations), and probe poses (obtained through multi-sensor data acquisition) are strictly aligned in time, establishing the correspondence between multi-source data at the same moment. The process involves checking image sharpness, illumination uniformity, and the presence of motion blur or overexposure. Frames with abnormal quality are marked to determine whether to resample locally or mark them as suspicious data. First, the image gradient magnitude is calculated to assess sharpness; if the average gradient value is below a threshold, it is considered blurry. Second, the standard deviation of the grayscale histogram is analyzed to assess uniformity; if the variance is too large, it is considered uneven illumination. Third, the displacement of edge pixels is detected to assess motion blur; if the displacement exceeds 3 pixels, it is considered dynamic blur. Finally, the grayscale values ​​of highlighted areas are checked to assess overexposure; if they are close to the maximum value, they are considered overexposed. After marking abnormal frames, if the problem is localized, resampling is instructed; if the problem is global, it is marked as suspicious. The verified multi-source data is packaged and output in a standard format, including the original image, metadata tags, spatiotemporal index, and quality rating; For example, to test a stainless steel straight tube with an inner diameter of 25 mm and a length of 600 mm, the probe advances at a speed of 50 mm per second and rotates twice per second. A high-speed camera continuously captures images at 100 frames per second, forming a time-series image stream of 6000 frames. The probe pauses every 100 mm and switches between polarized light source and ultraviolet light source, acquiring a pair of polarized images and a fluorescence image from each source, resulting in a total of 6 sets of multispectral data. A 50-millisecond pulsed airflow is applied at the mid-section of the tube at 300 mm, and the bubble deformation response is observed. The air pressure curve and the images before and after it are recorded simultaneously. The final result is a multi-source dataset containing 6000 high-speed sequences, 12 polarized images, 6 fluorescence images, and 1 set of mechanical responses. Each data point is accompanied by a position label accurate to 0.1 mm to ensure accurate retrospective analysis.

[0021] This embodiment also provides execution timing dynamic feature analysis and preliminary discrimination, specifically as follows: The process involves denoising and normalizing the illumination of continuous image sequences to identify and segment suspected target areas, eliminating interference from pipe wall background and fixed structures. The continuous image sequence refers to a time-ordered set of images captured continuously by a high-speed camera at a high frame rate (e.g., 100 frames per second) during temporal dynamic feature analysis and preliminary discrimination. Synchronous recording is achieved through trigger signals. The steps for identifying and segmenting suspected target areas are as follows: First, Gaussian filtering is applied to remove image noise. Second, adaptive threshold segmentation or edge detection algorithms are used to extract areas with significant contrast to the background. Third, morphological opening and closing operations are used to remove small noise points and connect broken edges. Finally, connected regions that match pollutant characteristics (rather than pipe wall texture) are selected as suspected targets based on geometric features such as area, roundness, and aspect ratio. Extract the edge contours of the target in each frame image, establish cross-frame contour correspondence, and record the change sequence of edge point coordinates over time. The edge contour of the target refers to the set of pixels that form the boundary between the suspected target area and the background. The steps to establish cross-frame contour correspondence are as follows: First, extract the set of edge points of the target in the current frame and calculate the centroid coordinates. Second, set a search window with the centroid as the center in the next frame. Third, calculate the shape similarity (such as Hu moment distance) and proximity of each candidate target in the window. Finally, select the target with the highest similarity and the closest distance to establish a correspondence. If the match fails, mark it as a newly appeared or disappeared target. To analyze the high-frequency fluctuation components of edge point coordinates, calculate the number and amplitude of edge undulations per unit time, and assess the dynamic activity level of the edges: First, perform a fast Fourier transform on the coordinate sequence of a single edge point to extract frequency domain information. Second, set a high-frequency cutoff frequency (e.g., above 10 Hz) to filter out low-frequency rigid body displacements and retain the flutter components. Third, count the number of fluctuation cycles with amplitudes exceeding the threshold per unit time to obtain the number of undulations. Calculate the average peak-to-valley difference of these cycles to obtain the flutter amplitude. Finally, take the average of all edge points to obtain the overall edge dynamic activity index. Calculate the pixel area of ​​the target region in each frame, plot the area change curve over time, and identify the periodic pulsation pattern and its frequency characteristics: First, count the pixels of the target region in each frame image to obtain the area value. Second, plot the area-time curve according to the time series. Third, apply the peak detection algorithm to identify local maxima and minima, calculate the time interval between adjacent peaks to obtain the pulsation period, take its reciprocal to obtain the frequency, and finally statistically analyze the stability of the frequency and the regularity of the amplitude. If the period is relatively constant, it is determined to be a periodic pulsation. Using the target's centroid as a reference point, its position change in the image coordinate system is tracked, a displacement trajectory curve is fitted, and compared with the probe's motion parameters: First, the centroid coordinates of the target area in each frame are calculated as the centroid. Second, the centroids of all frames are connected to form a discrete trajectory. Third, the least squares method or spline interpolation is applied to fit a smooth curve, the tangent direction of the curve is calculated, and finally, the direction is vector-decomposed with the probe's advancing direction (axial) and rotating direction (circumferential) to obtain the axial displacement component and the circumferential displacement component. Calculate the correlation coefficient between the target displacement and the probe's advance and rotation motion to determine whether the target motion is synchronized with the probe motion or whether there are abnormal modes such as lag, lead, or irrelevance: First, extract the axial and circumferential components of the target displacement curve; second, extract the probe's advance speed curve and rotation angle curve; third, calculate the Pearson correlation coefficient between the target's axial displacement and the probe's advance speed, and the Pearson correlation coefficient between the target's circumferential displacement and the probe's rotation angle, respectively; finally, set a synchronization threshold (e.g., 0.8). If both coefficients are greater than the threshold, synchronization is determined; if the axial coefficient is low and the circumferential coefficient is high, sliding is determined; if both are low, independent motion is determined. By combining three indicators—edge jitter frequency, area change period, and motion synchronization—and matching them with preset bubble or particle feature patterns, a preliminary judgment result is output. First, feature templates for bubbles (high jitter, periodic pulsation, asynchronous) and feature templates for particles (low jitter, constant area, synchronous) are established. Second, the similarity between the measured three indicators and the two templates is calculated (e.g., Euclidean distance or cosine similarity). Third, the difference in similarity with the two templates is compared. If the similarity with the bubble template is significantly higher than that with the particle template, it is judged as a bubble; otherwise, it is a particle. Finally, the category with the highest matching score is output as the preliminary judgment result. Calculate the confidence level of the preliminary judgment: The confidence level is equal to the weight of the edge tremor index multiplied by the score of the index, plus the weight of the area change index multiplied by the score of the index, plus the weight of the synchronicity index multiplied by the score of the index, where the sum of each weight is one, and the scores of each index are normalized to the range of zero to one according to the degree of matching with the template. If all three indicators point to the same feature and the confidence level is higher than the threshold, then a confirmation message is output, which is either "preliminary confirmation of bubbles" or "preliminary confirmation of particles". If the indicators are contradictory (i.e., the three indicators are consistent but do not point to the same type of feature) or the confidence level is insufficient, it is marked as "temporal ambiguity" and multispectral optical feature cross-validation is performed for multispectral verification. The steps to determine whether the three indicators point to the same type of feature are as follows: First, check whether the edge flutter indicator indicates bubbles (high value) or particles (low value); second, check whether the area change indicator indicates bubbles (periodic) or particles (stable); and finally, check whether the synchronicity indicator indicates bubbles (asynchronous) or particles (synchronous). If all three indicators point to the same category, they are considered to be consistent. For example, analyzing a suspicious circular target in 100 consecutive frames of images of the inner wall of a pipe fitting, edge tracking shows that the contour points fluctuate 12 times per second with an amplitude of about 3 pixels, exhibiting high-frequency jitter. The area curve shows periodic contraction and expansion 4 times per second with an amplitude of about 15%. The displacement trajectory shows that the target slowly drifts towards the pipe opening, opposite to the probe's direction of advancement, with a correlation coefficient of -0.3, indicating significant asynchrony. All three indicators conform to the bubble characteristic pattern, and the system outputs "Preliminary confirmation of bubble" with a confidence level of 92%. Digital twin tracing is then performed directly using virtual-real response mapping and difference comparison. In another case, the target edge is stable without jitter, the area is constant, and the displacement trajectory is completely synchronized with the probe rotation, with a correlation coefficient of 0.95, resulting in the output "Preliminary confirmation of particle". In the third case, the edge has slight jitter but the area is stable, and the motion synchronization is moderate. The three indicators contradict each other, with a confidence level of only 45%, and it is marked as "temporal ambiguity". Multispectral optical feature cross-validation is performed to seek multispectral evidence support.

[0022] This embodiment also provides a method for performing cross-validation of multispectral optical features, specifically: Based on the analysis of the dynamic characteristics of the execution time sequence and the preliminary judgment of the marked temporally ambiguous target position, the image set of the corresponding position is retrieved to ensure that the spatial coordinates of the images in the image set are consistent. The image set includes parallel polarization images, cross polarization images, ultraviolet fluorescence images and visible light reference images. By comparing the brightness distribution of parallel polarized and cross-polarized images, the birefringence halo pattern generated by the curved liquid film is identified: First, the two images are spatially registered to ensure pixel alignment. Second, the brightness difference of the corresponding pixels is calculated to generate a difference image. Third, the difference image is thresholded to extract the bright ring region. The geometry (circular or elliptical) and width of the ring are analyzed. If the ring is concentric and the intensity difference of the ring is significant under the two polarization states, it is determined to be a birefringence halo feature of the curved liquid film. The surface of the bubble forms a concentric or elliptical halo due to the curvature change, and the intensity difference of the halo is significant under the two polarization states. If a typical halo pattern is detected and the polarization state difference is obvious, it is marked as "positive liquid film surface feature"; If the polarization image shows uniform reflection or metallic luster without obvious halo, it is marked as "positive solid surface feature"; If the feature falls between the two, it is marked as "polarization blur". By comparing ultraviolet fluorescence images with visible light images, we can identify whether there are fluorescent rings at the edge of the target and whether the internal region has no fluorescence or significantly weaker fluorescence than the edge: First, register the fluorescence image with the visible light image. Second, determine the target edge position in the visible light image. Third, extract the gray value profile line at the corresponding edge position in the fluorescence image and check whether the profile line shows a distribution pattern of bright edges and dark interior. If the edge gray value is significantly higher than the internal background and shows a continuous ring, it is determined that there is a fluorescent ring. If a clear fluorescent ring is present at the edge but there is no fluorescence inside, it is marked as "positive for liquid-encapsulated gas characteristics"; If there is no fluorescence overall or uniform solid fluorescence, it is marked as "positive for solid particle characteristics"; If the fluorescence distribution is abnormal or interfered with by the background, it is marked as "fluorescence blurry"; Cross-validation is performed by combining polarization determination results and fluorescence determination results: First, the output labels of the polarization analysis module (positive for liquid film surface features, positive for solid surface features, or ambiguous) are read; second, the output labels of the fluorescence analysis module (positive for liquid-encapsulated gas features, positive for solid particle features, or ambiguous) are read; then, a decision matrix is ​​established: if both modules are positive and of the same category, the confirmation of that category is strengthened; if both are negative, the confirmation of the opposite category is strengthened; if one is positive and the other is negative, or if either is ambiguous, it is marked as conflict; finally, the cross-validation conclusion is output. If the polarization characteristics and fluorescence characteristics contradict each other, or if either is marked as ambiguous, it is determined as "spectral crossover unresolved", and mechanical perturbation final discrimination is performed to verify the mechanical perturbation. If the dual spectra are consistent and clear, the final confirmation result is output: First, check whether the cross-validation conclusion is a consistent judgment without conflict. Second, select the output category (bubble or particle) according to the consistent judgment result. Third, package the judgment result with the original data of performing time-series dynamic feature analysis and preliminary discrimination and performing multispectral optical feature cross-validation to form an evidence chain. Finally, output the confirmation result to perform virtual-real response mapping and difference comparison (bubble) or cleanliness assessment process (particle). If there is a conflict, output the unresolved signal to perform mechanical micro-perturbation final discrimination. Output one of three decisions: "bubble spectrum confirmation", "particle spectrum confirmation" or "final judgment of mechanical micro-perturbation", and simultaneously record the quantitative description of each spectral feature as a chain of evidence. For example, a time-series ambiguous target is located in a depression near the weld of a pipe fitting. Polarization image analysis shows that under parallel polarization, the target center is bright and the edges are dark, while under cross polarization, bright concentric rings appear at the edges, with a difference of 78%, marked as "positive liquid film surface feature". Fluorescence image analysis shows that under 365nm ultraviolet excitation, the target edge shows a bright blue-green ring, while the central area is completely dark, with a ring width of about 0.3mm, marked as "positive liquid-encapsulated gas feature". The cross-matching of the two spectra points to the bubble, and the system outputs "bubble spectrum confirmation". Skipping the final judgment of mechanical micro-perturbation, it directly performs virtual-real response mapping and difference comparison. In another case, the polarization image shows a uniform metallic luster without curvature halo, and the fluorescence image shows no overall response. Both spectra are negative, and the output is "particle spectrum confirmation". In a third case, the polarization shows a weak halo, but the fluorescence is interfered with by the weld reflection and cannot be identified, marked as "fluorescence fuzzy". The final judgment of mechanical micro-perturbation is performed to seek mechanical evidence.

[0023] This embodiment also provides a final determination of the mechanical micro-perturbation, specifically as follows: Based on the spatial coordinates of the unresolved target marked after performing multispectral optical feature cross-validation, the probe is driven to precisely retract to that position, and visual tracking is re-established to ensure that the target is within the effective range of mechanical action. The spatial coordinates of the unresolved target are the three-dimensional position coordinates (axial distance, circumferential angle, and radial depth) of the target marked as "spectral cross-validation unresolved" inside the pipe during the multispectral optical feature cross-validation. They are calculated by combining the image pixel coordinates with the probe pose data through coordinate transformation. Assess target size and attachment status: For isolated targets larger than 0.5 mm, airflow pulse is preferred, while for small targets or densely distributed areas, ultrasonic standing wave field is used to avoid the airflow blowing away the target and causing loss of position. Small targets or densely distributed areas are obtained through image analysis: First, calculate the equivalent diameter of the suspected target in step three. If it is less than 0.5 mm, it is determined to be a small target. Second, count the number of targets per unit area. If it exceeds a preset threshold (e.g., 5 targets per square centimeter), it is determined to be a densely distributed area. Short-duration low-pressure pulses are applied via an integrated micro-nozzle, with the pressure gradient increasing progressively. Simultaneously, the target's morphological changes and displacement trajectory are recorded at high speed. The application time is 50 to 200 milliseconds. The progressively increasing pressure gradient means starting from an initial low pressure (e.g., 0.02 MPa), increasing by a fixed step size (e.g., 0.02 MPa) at each level, and gradually applying pressure up to the highest pressure (e.g., 0.1 MPa). The pressure is maintained at each level to observe the response before increasing the pressure again. This avoids sudden high pressure causing the target to disappear instantly and making the process impossible to observe. The target morphological changes and displacement trajectory refer to the changes in the target's contour shape under mechanical action (e.g., becoming longer, flatter, or breaking) and the path of the target's center of mass moving in space. Start the ultrasonic transducer, scan the frequency to find the target resonance response point, establish an adjustable standing wave field, observe the target's aggregation or repulsion motion between the sound pressure node and the antinode, and record its amplitude and phase response: First, scan the ultrasonic frequency from low frequency to high frequency with a certain step size (e.g., 100 Hz), and observe the target vibration amplitude at the same time. Record the frequency at which the amplitude is the maximum as the resonance point. Second, fix this frequency and adjust the position of the reflector to establish a standing wave field, determine the position of the sound pressure node (the point of minimum amplitude) and the antinode (the point of maximum amplitude), and observe the target's motion behavior at the node and antinode again, recording whether it aggregates towards the node or is repelled by the antinode. Finally, measure the amplitude of the target vibration and its phase relationship with the sound pressure wave (in phase or out of phase). Analyze high-speed image sequences to identify whether the target exhibits the following response patterns: overall deformation or local depression, slippage or detachment along the pipe wall, surface cracking or bursting and disappearance, no visible changes or only slight shaking; If any of the response modes of deformation, movement, rupture, or detachment is detected, it is classified as a "fluid response mode", and finally determined to be a bubble. "Bubble confirmation" is output, and virtual and real response mapping and difference comparison are performed for digital twin traceability and process optimization. If no response mode occurs (i.e., no displacement) or only slight rolling without deformation, it is classified as "solid response mode," ultimately determined to be particles, outputting "particle confirmed," and proceeding to the cleanliness level assessment process. The size and quantity are counted to determine the level. Slight rolling without deformation is determined as follows: First, observe whether the target contour remains rigid (aspect ratio change less than 5%); second, measure the target displacement (less than one diameter); third, confirm no local contour depressions or expansions. Meeting these three points determines slight rolling without deformation. The cleanliness level assessment process proceeds as follows: First, measure the equivalent diameter of all confirmed particles; second, count the number of particles for each level according to standard classifications (e.g., greater than 100 micrometers, 50 to 100 micrometers, less than 50 micrometers); third, compare with the cleanliness standard table (e.g., allowable number of particles per square centimeter); finally, determine the pipe fitting cleanliness level (e.g., Grade A, Grade B, Grade C). If the response is between the two or cannot be identified, it will be determined as a "fuzzy response", marked as "abnormal and unclear", the inspection status of the pipe fitting will be frozen, the manual re-inspection process will be triggered, and all original data will be retained for future reference. The complete discrimination chain, which includes performing time-series dynamic feature analysis and preliminary discrimination, performing multispectral optical feature cross-validation, and performing mechanical micro-perturbation final discrimination, is summarized and a stage report is generated. This report includes a three-level decision path of preliminary judgment, spectral verification, and mechanical confirmation, providing structured input for subsequent analysis. For example, a target at a weld seam depression, after cross-verification using multispectral optical features, was determined to be "fluorescently blurred." The final determination was then performed using mechanical micro-perturbation. The system employed a pulsed airflow approach, applying a low-pressure airflow for 100 milliseconds. High-speed imaging showed the target instantly elongating and deforming into an ellipsoidal shape, then sliding approximately 2 millimeters along the airflow direction before finally detaching from the pipe wall, floating upwards, and disappearing. The entire process took 300 milliseconds, and the response mode was clearly categorized as "fluid response," outputting "bubble confirmation." Virtual-real response mapping and difference comparison were then performed to trace the cause of the bubble. In another case, the target under the same pulsed airflow... The system was completely still. After switching to ultrasonic standing wave, it only produced a slight vibration with an amplitude of less than 0.1 mm. There was no displacement or deformation, which was classified as "solid response". The output was "particle confirmation". The system then proceeded to cleanliness assessment to determine if the particle size exceeded the standard. In the third case, the target edge trembled slightly under the action of airflow but remained stable overall. Under the action of ultrasound, it exhibited irregular vibrations but no regular displacement, which could not be clearly classified. It was marked as "abnormal and unclear". The system paused the automatic process and notified the quality inspector to bring an endoscope for on-site verification. Finally, it was manually confirmed as a loosely attached fiber cluster, which was neither a typical bubble nor a dense particle, and was treated as a special contaminant.

[0024] Example 3 is an improvement on Example 2. In this example, virtual-real response mapping and difference comparison are performed, specifically as follows: Based on the basic parameters of the pipe fitting obtained through the execution of basic parameter acquisition and validity determination, the geometric model, material attribute library, and standard process template of the corresponding specification (i.e., the specification parameters (nominal diameter, wall thickness, material grade) of the stainless steel pipe fitting to be inspected) are retrieved from the digital twin library. The virtual instance of the pipe fitting is then activated. Here, the pipe fitting refers to the physical pipe fitting currently entering the inspection process, and the virtual instance refers to the virtual mapping object created in the digital twin system based on the actual parameters of the pipe fitting, which has independent state and behavior capabilities. The activation steps are as follows: first, retrieve the basic template that matches the specification from the digital twin library; second, inject the actual parameters (measured dimensions, actual roughness) obtained through the execution of basic parameter acquisition and validity determination into the template to update the geometry and attributes; third, initialize the virtual sensor model and virtual process environment; and finally, start the instance clock to enable it to have time-series evolution capabilities. The key attributes of the detected events confirmed after the final judgment of the mechanical perturbation are transformed into twin input parameters, including target type, size, location coordinates, detection time, temporal feature description, spectral feature label, and mechanical response mode. The corresponding scene is then reconstructed in virtual space: First, the structured deconstruction data output from the final judgment of the mechanical perturbation is parsed (target type, size value, location coordinates, timestamp, temporal feature label, spectral feature code, and mechanical response mode). Second, unit unification and coordinate system transformation are performed (e.g., converting image pixel coordinates to world coordinate system millimeter values). Third, discrete labels are converted into continuous parameters (e.g., converting "high jitter" into a specific frequency range). Finally, the data is encapsulated into an input vector according to the twin interface protocol and injected into the virtual scene. The virtual twin is driven to execute a virtual detection process, simulating the ideal response under the same sensor configuration and motion parameters, and outputting virtual prediction data: First, the same sensor model (camera parameters, light source type, mechanical probe specifications) as the physical entity is configured in the virtual instance. Second, the same motion trajectory (propulsion speed, rotation speed, and stopping position) as the physical entity is set. Then, the virtual simulation engine is started to calculate physical processes such as light propagation, flow field distribution, and mechanical action. Finally, the prediction data such as images, spectra, and mechanical feedback generated by the virtual sensor are recorded. The virtual prediction data includes the predicted target shape, expected displacement trajectory, theoretical spectral response curve, and standard mechanical feedback amplitude. Using spatiotemporal coordinates as an index, the actual response data of entity detection is aligned with the ideal response data predicted by twins at the frame level to establish a benchmark for point-by-point comparison. Four indicators are calculated: morphological difference, trajectory deviation, spectral similarity, and mechanical response deviation, to comprehensively evaluate the overall fit between the entity and the virtual model. Morphological difference is the relative deviation between the size of the detected entity and the virtual predicted size. Trajectory deviation is the Hausdorff distance or root mean square error between the displacement path of the entity and the virtual predicted path. Spectral similarity is the correlation coefficient or cosine similarity between the spectral curve of the entity and the spectral curve of the virtual model. Mechanical response deviation is the absolute difference and relative error between the mechanical feedback amplitude of the entity and the amplitude of the virtual prediction. Compare the overall difference with the preset tolerance range; If all four indicators are below their respective thresholds and the overall consistency meets the standard, it is judged as "virtual and real consistent", confirming the effectiveness of the twin model, and directly performing multiphysics evolution analysis and final decision-making for evolution prediction; If any indicator exceeds the standard or the overall consistency is insufficient, it is judged as "virtual-real deviation". Anomaly root cause tracing and strategy pre-run optimization are performed to trace the root cause and further analyze the deviation characteristics: First, the values ​​of the four difference indicators output by the virtual-real response mapping and difference comparison are read. Second, classification rules are established: If the morphological difference degree and trajectory deviation degree both exceed the standard, but the spectral similarity degree and mechanical response deviation are normal, it is classified as geometric positioning deviation. If the spectral similarity degree and mechanical response deviation exceed the standard, but the morphological trajectory is normal, it is classified as sensor performance deviation. If multiple indicators fluctuate irregularly or all exceed the standard, it is classified as systemic anomaly. Finally, the deviation mode label is output. The system outputs either a "virtual-real consistency" or "virtual-real deviation" conclusion, along with a difference quantification report and deviation pattern label. This provides a basis for decision-making, including root cause tracing, strategy pre-simulation optimization, multiphysics evolution analysis, and final decision-making. The difference quantification report includes the specific values ​​of four indicators, the identifier of the out-of-range item, and the percentage of deviation, calculated by performing virtual-real response mapping and difference comparison. The deviation pattern label refers to the category identifier assigned according to the above classification rules (geometric positioning, sensor performance, systemic anomaly). For example, a confirmed bubble event is mapped to a twin: the physical detection record shows a bubble diameter of 1.2 mm, located 150 mm from the tube opening, with a drift trajectory towards the tube opening, a polarized halo intensity of 78%, and a deformation and rupture time of 120 ms under an airflow pulse. The twin predicts that a bubble of the same size and location, under standard drying processes, is expected to have a diameter of 0.8 to 1.0 mm, a drift trajectory towards the tube tail, a halo intensity of 65% to 75%, and a rupture time of 80 to 100 ms. The comparison shows that the physical bubble is 20% oversized, drifts in the opposite direction, has a slightly higher halo intensity, and a delayed rupture time. The overall difference was 32%, exceeding the 25% tolerance threshold, and was judged as "deviation between virtual and real". The deviation mode was "geometric positioning deviation" combined with "sensor performance deviation". Anomaly root cause tracing and strategy pre-simulation optimization were performed. It was found that the actual drying temperature of this batch was 10°C higher than the standard, which caused the viscosity of the residual liquid to decrease and the bubbles to expand easily. At the same time, the probe advance speed was too fast, which caused the airflow entrainment direction to change. In another case, the response of the solid particle was highly consistent with the twin prediction, with a difference of 8%, and was judged as "consistent between virtual and real". Multiphysics evolution analysis and final decision prediction were directly performed to predict the evolution risk of the particle in subsequent processes.

[0025] This embodiment also provides the following: performing anomaly root cause tracing and strategy pre-simulation optimization, specifically: Load the complete manufacturing history of the pipe fitting (i.e., the specific stainless steel pipe fitting currently undergoing the anomaly root cause tracing and strategy pre-simulation optimization tracing process, i.e. the physical pipe fitting that is determined to be "virtual-real deviation" in the virtual-real response mapping and difference comparison) into the twin, and replay each process of cleaning fluid injection, ultrasonic cleaning, rinsing, drying and cooling in sequence according to the time axis, and extract the process parameters of each process, including real-time curves of temperature, pressure, flow rate, time and vacuum degree; The actual execution parameter curve is compared point by point with the standard process template to identify abnormal periods where the deviation exceeds the tolerance. The timestamp and magnitude of parameter mutations are marked: First, the standard process template curve (standard values ​​of temperature, pressure, etc. that change over time) is loaded. Second, the point-by-point difference between the actual curve and the standard curve is calculated. Third, a tolerance threshold is set (e.g., temperature ±5 degrees Celsius, pressure ±10 kPa). Periods where the difference exceeds the threshold are marked as abnormal periods. Finally, the start and end timestamps of the period and the maximum deviation magnitude between the actual value and the standard value are recorded. Analyze abnormal parameter combinations: If the drying section temperature is lower than the standard lower limit or the heating rate is insufficient, or the drying time does not reach the theoretical calculation value, or the vacuum degree fluctuates periodically, it will be identified as "process execution deviation" and the conclusion "process parameter abnormal" will be output. For each anomaly identified, the optimization knowledge base is invoked to generate adjustment solutions: if the temperature is insufficient, it is recommended to increase the setpoint or check the heater power; if the time is insufficient, it is recommended to extend the drying section or optimize the airflow distribution; if the vacuum fluctuates, it is recommended to repair the vacuum pump or sealing system. The optimization knowledge base is a structured database that stores expert experience and process rules. It contains the mapping relationship between various process anomalies (such as insufficient temperature, insufficient time, and vacuum fluctuations) and corresponding optimization strategies (such as increasing the setpoint, extending the time, and repairing the equipment), as well as adjustment parameters from historical successful cases. If the process parameters playback shows that the entire process is normal, then retrieve the sensor performance degradation model, compare the current light source output intensity with the initial calibration value, analyze the trend of lens transmittance change, and assess the degree of probe mechanical wear: First, retrieve the lens calibration data (standard white board reflectance test value) from each maintenance, then fit the time series to obtain the transmittance degradation slope, predict the current transmittance, then retrieve the historical data of probe positioning accuracy (repeated positioning error), analyze the error growth trend, and finally combine the two trends to determine the performance level. Among them, the sensor performance degradation model is a mathematical model that describes the degradation law of key sensor performance indicators (light source intensity, lens transmittance, mechanical accuracy) with the duration of use or environmental factors, and is established based on historical calibration data. If the light source intensity decreases by more than 20%, or the color temperature drifts significantly, or lens contamination causes a decrease in contrast, or the probe bearing clearance exceeds the standard, the system will be flagged as "abnormal equipment status," and a maintenance work order and spare parts replacement recommendations will be generated. The following methods are used to determine if the light source intensity decreases by more than 20%, or the color temperature drifts significantly, or the lens contamination causes a decrease in contrast, or the probe bearing clearance exceeds the standard: First, measure the ratio of the current light source output to the initial calibration value. If it is less than 80%, the decrease is considered to exceed 20%. Second, measure the light source color temperature value. If the deviation from the calibration value exceeds 500K, the color temperature is considered to drift. Third, measure the contrast ratio (MTF value) of the standard target. If it is lower than 70% of the initial value, contamination is considered. Finally, measure the probe radial runout. If it exceeds 150% of the design clearance, the bearing clearance is considered to exceed the standard. For pipe fittings with changes in specifications or materials that are being inspected for the first time, the detection effects of different sensor combinations, sampling densities, light source angles, and propulsion speeds are tested in batches within a twin to simulate blind zone distribution and misjudgment hotspot heat maps: First, an orthogonal experimental table is designed to arrange and combine various parameter levels (such as combinations of high-speed camera + polarized light, low-speed camera + fluorescence, etc., as well as different sampling intervals, light source incident angles, and propulsion speeds). Second, virtual detection is performed in batches within the twin, recording the coverage area (the proportion of detected areas) and misjudgment rate (the proportion of bubbles misjudged as particles) under each configuration. Third, blind zone locations (spatial areas that have never been detected) and high-incidence misjudgment locations (areas where incorrect judgments are concentrated) are statistically analyzed. Finally, a spatial heat map is generated using color coding to visualize the risk distribution. If the simulation shows that the current strategy has an unacceptable blind spot or the false judgment rate exceeds the standard, then output "strategy optimization suggestion", adjust the parameters and return to execute multi-sensor data acquisition to re-execute the complete detection process; If the pre-run verifies that the current strategy is comprehensive and the misjudgment rate is controllable, then output "Strategy Confirmed" and perform multiphysics evolution analysis and final decision-making evolution analysis. The criteria for "the pre-test shows that the current strategy has unacceptable blind spots or the misjudgment rate exceeds the standard" and "the pre-test verifies that the current strategy is comprehensive and the misjudgment rate is controllable" are determined as follows: First, set acceptable standards (such as blind spot area less than 1% and misjudgment rate less than 5%). Second, read the blind spot distribution data and misjudgment statistics generated by the execution anomaly root cause tracing and strategy pre-test optimization. Third, calculate the actual blind spot ratio and weighted average misjudgment rate. Finally, compare with the standard. If any indicator is worse than the standard, it is determined that there is an unacceptable risk. If all indicators are better than the standard, it is determined that the verification is passed. For example, after performing virtual-real response mapping and difference comparison on a certain pipe fitting, it was determined that there was a "virtual-real deviation". After performing abnormal root cause tracing and strategy pre-simulation optimization, the playback showed that the standard parameters of the drying process were 80℃ constant temperature for 15 minutes. The actual curve showed that the temperature rose normally to 80℃ for the first 8 minutes. In the 9th minute, the temperature dropped sharply to 45℃ for 4 minutes due to equipment failure. Then, the temperature was urgently raised to 90℃ to catch up. The total effective drying time was only 11 minutes. The "abnormal process parameters" were identified as temperature interruption and insufficient effective time, and optimization suggestions were generated: overhaul the temperature control system of the drying oven and add a supplementary drying process for the same batch of pipe fittings. In another case, the process parameters were normal throughout, but the diagnosis found that the intensity of the camera light source had decreased by 35% after more than 8,000 hours of use, and there was cleaning fluid splash contamination at the front of the lens. The output "abnormal equipment status" triggered the replacement of the light source module and lens cleaning. The third case was a pre-test of a new specification thin-walled pipe fitting. It was found that the existing strategy had a 15 mm blind zone on the inside of the elbow and the bubble misjudgment rate reached 12%. The output "strategy optimization suggestion" was generated. After adjusting the probe rotation speed and the light source angle, the multi-sensor data acquisition was re-executed. The blind zone was eliminated and the misjudgment rate was reduced to 3%. Finally, the output "strategy confirmed" was generated to perform multi-physics evolution analysis and final decision.

[0026] This embodiment also provides the following: performing multiphysics evolution analysis and final decision-making, specifically: From the cases of consistent virtual and real bubbles confirmed after performing virtual-real response mapping and difference comparison, key physical property parameters such as the initial size distribution of the bubble, spatial coordinates, estimated value of the surrounding liquid film thickness, local temperature field distribution, flow field shear stress intensity, wall electrostatic potential distribution gradient, and surfactant concentration and ionic strength of the cleaning fluid are extracted. Based on the extracted parameters, a heat-mass transfer coupling model was constructed to simulate the complete phase transition path of the liquid film from liquid to solid state under the current drying conditions, and to track the spatiotemporal evolution of solute precipitation and solvent evaporation. First, a three-dimensional mesh model of the inner wall of the pipe was established, and the initial liquid film thickness distribution was set. Then, the temperature field boundary conditions (hot air temperature, flow rate) and mass transfer boundary conditions (humidity, air pressure) were input, and the energy conservation equation was solved to obtain the temperature distribution. At the same time, the mass conservation equation was solved to obtain the solvent evaporation rate and solute concentration evolution. Then, the precipitation time was determined according to the concentration exceeding the solubility, and the crystal nucleation and growth process was tracked. Finally, the sequence of changes in liquid film thickness, concentration, and phase state over time was recorded. Based on the phase change simulation results, the morphology of the final dried residue is predicted: First, the material distribution at the end of the phase change path is read. Second, the spatial distribution uniformity of the residue is analyzed (concentration variance is calculated). If the variance is less than the threshold and the thickness is uniform, it is judged as a risk of uniform film. If the variance is large and it is distributed in a discrete point shape, it is judged as a risk of spot residue. If the whole is rapidly solidified to form a three-dimensional network crystal structure, it is judged as a risk of crystal cluster contamination. If it is completely volatilized and there is no residue, it is judged as ideal drying. Based on the initial thickness of the liquid film and the current drying rate, calculate the critical time required for complete drying in theory: First, determine or estimate the initial average thickness of the liquid film; second, calculate the mass transfer coefficient according to the current drying conditions (temperature, wind speed, humidity); apply the drying rate equation (drying rate is proportional to the remaining moisture content); and calculate the time required to reduce the initial moisture content to the target moisture content through integration to obtain the critical drying time. Compare the theoretically required critical time for complete drying with the actual drying time obtained from the acquisition of basic parameters and effectiveness determination. If the actual drying time is lower than the critical value, an "insufficient drying" judgment will be output, quantifying the time gap. Multiphysics simulations are performed by coupling temperature field, flow field, and electric field to simulate asymmetric coalescence, rupture, or stable attachment modes of bubbles under complex boundary conditions. First, an initial field including temperature gradient, flow velocity distribution, and electrostatic potential distribution is established. Second, the bubble interface is set as a deformable boundary. The Navier-Stokes equations are solved to obtain the drag force of the flow field on the bubble. The heat conduction equation is solved to obtain the Marangoni force caused by the temperature gradient. The electric field equation is solved to obtain the electrostatic force. Then, the forces and motion of the bubble are calculated by combining the forces. The coalescence process when multiple bubbles meet or the rupture process near the wall are simulated. Finally, the evolution history of the bubble position, shape, and number is recorded. If the simulation shows that the local thermal boundary layer is too thick, causing abnormal bubble aggregation, the output will be "significant thermal boundary layer effect". It is recommended to optimize the hot air nozzle layout or increase turbulence to promote mixing. If the display shows that the electrostatic potential gradient drives the bubbles to migrate and attach to a specific wall surface, the output will be "significant electrostatic interference". It is recommended to increase the ion wind to neutralize or adjust the ambient humidity. Among them, "simulation shows that excessive local thermal boundary layer leads to abnormal bubble aggregation" is determined by analyzing the temperature gradient in the near-wall region being less than the thickness in the mainstream region. "Showing that electrostatic potential gradient drives bubbles to migrate and attach to a specific wall" is determined by analyzing the consistency between the bubble trajectory and the normal of the equipotential line. By analyzing the temperature field data, the temperature gradient changes in the near-wall region and the mainstream region are compared. If it is found that the temperature gradient in the near-wall region is significantly less than that in the mainstream region and the thermal boundary layer thickness exceeds a set threshold, it is determined that the excessive thermal boundary layer leads to bubble aggregation. At the same time, the bubble trajectory is tracked and the angle between its direction of movement and the normal direction of the electrostatic equipotential line is calculated. If the trajectory continues to point towards a specific wall along the normal and the angle is less than the tolerance, it is determined that the electrostatic potential gradient drives the bubble migration and attachment. For particle cases confirmed after performing virtual-real response mapping and difference comparison, the adhesion strength and resuspension critical shear stress are calculated based on the combined effects of van der Waals forces, electrostatic forces, and liquid bridging forces between the particle and the wall, using molecular dynamics simulation: the adhesion strength equals the contribution of van der Waals forces plus the contribution of electrostatic forces plus the contribution of liquid bridging forces. Among them, the van der Waals forces are inversely proportional to the fourth power of the distance between the particle and the wall, the electrostatic forces are directly proportional to the product of surface charge density, and the liquid bridging forces are directly proportional to the surface tension of the liquid and the cosine of the contact angle; the resuspension critical shear stress equals the adhesion strength divided by the product of the particle's windward area and the moment coefficient. When the shear stress applied by the fluid exceeds this critical value, the particle will be resuspended. If the current flow field shear stress is close to or exceeds the critical value, the risk of "false adhesion" will be output. It is recommended to reduce the movement speed of the detection probe or increase the pretreatment inert gas purging. A time-varying evolution model of pollutants was established to simulate the time scale and path of the transformation from the current state of bubble residue or particle adhesion to "fresh residue - aged residue - stubborn stains". First, the current state parameters of pollutants (chemical composition, adhesion mode, and degree of aging) were determined. Second, a chemical reaction kinetic equation was established to describe the oxidation, polymerization, and hydrolysis process of the residual liquid. A physical adsorption model was established to describe the transformation of intermolecular forces from weak adsorption to chemical bonding. Third, the time step was set and integral solutions were performed to obtain the time-varying curves of the chemical composition and adhesion strength of pollutants. Finally, the three stages of fresh, aged, and stubborn stains were divided according to the change of the curve slope. If it is predicted that the pipe fittings will be converted into a stubborn state of chemical bonding or mechanical interlocking before the next process, an "emergency handling" suggestion will be output, triggering priority rework or isolation of the batch of pipe fittings. If it is predicted that the current clearable state can be maintained for more than the safety buffer period, then the "normal processing" suggestion will be output and the process will be handled according to standard logistics. Among them, "predicting that it will transform into a chemically bonded or mechanically interlocked stubborn state before the next process" is determined by comparing the transformation time constant with the time interval of the next process. If the time constant is less than the interval, it is determined that the transformation will occur. "predicting that the current removable state can be maintained for more than the safety buffer period" is determined by calculating whether the transformation time from the current state to the stubborn state is greater than the safety buffer period (usually 1.5 times the time of the next process). A time constant model of pollutant state evolution is established to calculate the time constant required for the transformation from the current state to the chemically bonded or mechanically interlocked stubborn state. This time constant is then numerically compared with the time interval of the next process. If the time constant is less than the process interval, it is determined that the pollutant will transform into a stubborn state before the next process. At the same time, the total transformation time from the current state to the stubborn state is calculated and compared with the set safety buffer period (usually 1.5 times the time of the next process). If the transformation time is greater than the buffer period, it is determined that the current removable state can be maintained. The analysis results of the entire process, including dynamic feature analysis and preliminary judgment of execution time sequence, cross-validation of execution multispectral optical features, final judgment of execution mechanical micro-perturbation, execution virtual and real response mapping and difference comparison, and execution anomaly root cause tracing and strategy pre-simulation optimization, are integrated to generate a comprehensive optimization scheme that includes drying process optimization, hot air system improvement, electrostatic control, and detection parameter adjustment. This scheme is fed back to the upstream production process through the manufacturing execution system. The complete parameter set, evolution path, prediction results and actual outcome of this case are linked and archived in the digital sample library to continuously train and optimize the twin model. For example, in a case involving a bubble with consistent virtual and real characteristics, multiphysics evolution analysis and final decision-making were performed. Extracted parameters showed a bubble diameter of 0.5 to 1.5 mm, located in a straight section 200 mm from the nozzle, with a liquid film thickness of approximately 10 micrometers, a local temperature of 75°C (lower than the set value of 80°C), a flow field shear of 0.2 Pa, an electrostatic potential gradient of 150 V / m, and a high concentration of silicate in the cleaning solution. Phase transition simulation showed rapid silicate nucleation, predicting the formation of cluster-like residues rather than ideal volatilization. The critical drying time was 18 minutes, but in reality, it was only 12 minutes, resulting in the outputs "insufficient drying" and "risk of cluster contamination." Multiphysics coupling revealed an 8 mm thermal boundary layer thickness, causing bubble retention. Electrostatic potential drives bubbles to migrate towards the wall, outputting dual optimization suggestions: extend drying to 20 minutes, optimize hot air to a rotating jet to destroy the boundary layer, and increase ion wind to neutralize static electricity. Particle case simulation shows that the adhesion strength of a certain metal shavings is only 0.05 N, and the current probe movement generates 0.08 N of shear, outputting "false adhesion" risk, suggesting that the probe speed be reduced by 30% or pre-purged with nitrogen. Aging prediction shows that the crystal cluster residue will siliconize into stubborn stains within 4 hours, outputting "emergency treatment", and the batch should be reworked and re-acid-washed immediately. The final comprehensive solution is fed back to the drying oven control system, and the case data is archived in the enrich twin crystal cluster evolution model.

[0027] In this embodiment, a multi-sensor progressive fusion discrimination technology combining high-speed continuous imaging, polarized light, ultraviolet fluorescence, and mechanical micro-perturbation is used to achieve high-precision intelligent differentiation of bubbles and particles and an extremely low false positive rate. Through digital twin virtual-real response mapping and production process root cause tracing technology, the detection system achieves self-diagnosis of its status, accurate location of abnormal process parameters, and pre-simulation optimization of new specification strategies. Through thermal-fluid-solid-electric multi-physics field coupling evolution prediction and real-time feedback technology of upstream processes, the system achieves early warning of pollutant risks, preventive removal of stubborn stains, and closed-loop self-evolutionary control of manufacturing quality.

Claims

1. A method for detecting the cleanliness of the inner wall of stainless steel pipe fittings based on intelligent sensors, characterized in that: include: Perform basic parameter acquisition and validity determination, construct testing entry thresholds, and screen out objects that are physically incompatible or have unreasonable processes; Perform multi-sensor data acquisition, build a multi-dimensional perception foundation, and form a high-fidelity multi-source dataset to provide sufficient information support for the intelligent differentiation of bubbles and particles; Perform time-series dynamic feature analysis and preliminary discrimination, establish preliminary discrimination rules for bubbles and particles, quickly divert high-confidence cases, and retain verification channels for fuzzy cases; Multispectral optical feature cross-validation is performed, and the problem of temporal ambiguity is solved by cross-validation of liquid film curvature optical response and liquid fluorescence edge effect, providing optical fingerprint identification for bubbles and particles; The final judgment of mechanical micro-perturbations breaks the deadlock of spectral verification, achieves the ultimate distinction between bubbles and particles, and retains a channel for manual intervention in abnormal cases; Perform virtual-real response mapping and difference comparison, build a virtual-real comparison bridge, map entity detection events to digital twins for ideal response prediction, verify the effectiveness of the model through multi-dimensional difference quantification, and identify system drift or process deviation. Perform root cause tracing and strategy pre-simulation optimization, locate process execution deviations or equipment performance degradation through twin playback, generate targeted corrective measures, and simulate the optimal detection strategy for new specifications. Perform multiphysics evolution analysis and final decision-making, predict the fate of pollutants through multiphysics coupling evolution, distinguish between emergency and routine treatment, generate targeted process optimization solutions and feed them back into the production system.

2. The method for detecting the cleanliness of the inner wall of stainless steel pipe fittings based on intelligent sensors according to claim 1, characterized in that: The basic parameters are retrieved and their validity is determined, specifically as follows: Obtain the unique identification code of the pipe fitting, link it to the production order database, and automatically extract the file information; Measure the pipe fitting data and record the minimum inner diameter value; Randomly inspect the outer surface of pipe fittings to confirm whether the deviation between the nominal value and the measured value of surface roughness is within the allowable range, and mark the abnormal pipe fittings; Read the cleaning process records that the pipe fitting has undergone; Check each of the above collected parameters for any empty or missing values ​​to confirm that the key parameters have been fully acquired. Compare the minimum inner diameter of the pipe fitting with the minimum fit size of the sensor probe, and compare the length of the pipe fitting with the maximum detection stroke of the equipment to confirm physical compatibility; Determine whether the drying temperature is within the recommended range of the cleaning solution, whether the drying time reaches the standard calculated value for the pipe diameter, and mark any parameter combinations that deviate from the standard process. If the parameters are complete and within the range of equipment capabilities and process standards, output "Parameter valid" and execute multi-sensor data acquisition; If there are missing parameters, out-of-range parameters, or process deviations, output "Parameter Abnormality" and terminate the process, simultaneously generating an explanation of the cause of the abnormality and suggestions for handling it.

3. The method for detecting the cleanliness of the inner wall of stainless steel pipe fittings based on intelligent sensors according to claim 1, characterized in that: Perform multi-sensor data acquisition, specifically: Start the sensor assembly, perform zero-point calibration, light source intensity calibration, and gas path sealing check, and confirm the readiness status of each subsystem; The robotic arm clamps the pipe to establish a spatial coordinate system, adjusts the pipe axis to be coaxial with the sensor probe, sets the initial position of the probe at the pipe end face, and confirms that there is no physical interference. Set the acquisition parameters, trigger the continuous shooting mode, and the probe advances and rotates at a constant speed, simultaneously recording the time sequence image stream and position coordinates; During the advancement process, polarized light imaging is triggered at preset intervals; At critical locations, propulsion is paused, and a short-term pulsed airflow is applied through an integrated air nozzle, while simultaneously recording image sequences and mechanical sensor feedback before and after the disturbance. Establish a spatiotemporal correlation index based on a unified timestamp; Check image sharpness, illumination uniformity, and whether there is motion blur or overexposure; mark frames with abnormal quality and decide whether to resample locally or mark them as suspicious data. The verified multi-source data will be packaged and output in a standard format.

4. The method for detecting the cleanliness of the inner wall of stainless steel pipe fittings based on intelligent sensors according to claim 1, characterized in that: Execution time-series dynamic feature analysis and preliminary judgment are as follows: Denoising and illumination normalization are performed on continuous image sequences to identify and segment suspected target areas, eliminating interference from pipe wall background and fixed structure; Extract the edge contours of the target in each frame image, establish cross-frame contour correspondence, and record the change sequence of edge point coordinates over time; Analyze the high-frequency fluctuation components of the edge point coordinates, calculate the number and amplitude of edge fluctuations per unit time, and assess the dynamic activity level of the edge. Calculate the pixel area of ​​the target region in each frame, plot the area change curve over time, and identify the periodic pulsation pattern and its frequency characteristics; Using the target's center of gravity as a reference point, track its position change in the image coordinate system, fit the displacement trajectory curve, and compare it with the probe's motion parameters; Calculate the correlation coefficient between the target displacement and the probe's advancing and rotating motion to determine whether the target motion is synchronized with the probe motion or whether there are abnormal patterns such as lag, lead, or irrelevance. By combining three indicators—edge tremor frequency, area change period, and motion synchronization—and matching them with preset bubble feature patterns or particle feature patterns, a preliminary judgment result is output. Calculate the confidence level of the preliminary judgment; If all three indicators point to the same feature and the confidence level is higher than the threshold, then a confirmation message is output. If the indicators are contradictory or the confidence level is insufficient, they are marked as "temporally ambiguous" and multispectral optical feature cross-validation is performed for multispectral verification.

5. The method for detecting the cleanliness of the inner wall of stainless steel pipe fittings based on intelligent sensors according to claim 1, characterized in that: Perform multispectral optical feature cross-validation, specifically as follows: Based on the analysis of the dynamic features of the execution time sequence and the preliminary judgment of the marked temporally ambiguous target location, the image set at the corresponding location is retrieved to ensure that the spatial coordinates of the images in the image set are consistent; By comparing the brightness distribution of parallel polarized and cross-polarized images, the birefringence halo pattern generated by the curved liquid film can be identified. If a typical halo pattern is detected and the polarization state difference is obvious, it is marked as "positive liquid film surface feature"; If the polarization image shows uniform reflection or metallic luster without obvious halo, it is marked as "positive solid surface feature"; If the feature falls between the two, it is marked as "polarization blur". By comparing ultraviolet fluorescence images with visible light images, we can identify whether there are fluorescent rings at the edge of the target and whether the internal region has no fluorescence or the fluorescence is significantly weaker than that at the edge. If a clear fluorescent ring is present at the edge but there is no fluorescence inside, it is marked as "positive for liquid-encapsulated gas characteristics"; If there is no fluorescence overall or uniform solid fluorescence, it is marked as "positive for solid particle characteristics"; If the fluorescence distribution is abnormal or interfered with by the background, it is marked as "fluorescence blurry"; Cross-validation was performed using the combined polarization determination results and the fluorescence determination results; If the polarization characteristics and fluorescence characteristics contradict each other, or if either is marked as ambiguous, it is determined as "spectral crossover unresolved", and mechanical perturbation final discrimination is performed to verify the mechanical perturbation. If the two spectra are consistent and clear, the final confirmation result will be output. Output one of three decisions: "bubble spectrum confirmation", "particle spectrum confirmation" or "final judgment of mechanical micro-perturbation", and simultaneously record the quantitative description of each spectral feature as a chain of evidence.

6. The method for detecting the cleanliness of the inner wall of stainless steel pipe fittings based on intelligent sensors according to claim 1, characterized in that: The final determination of the mechanical perturbation is as follows: Based on the spatial coordinates of the unresolved target marked after cross-validation of multispectral optical features, the probe is driven to precisely retract to that position, visual tracking is re-established, and the target is ensured to be within the effective range of mechanical action. Assess the target size and adhesion status; A short-duration low-pressure pulse is applied, and the pressure gradient is gradually increased while the target's shape changes and displacement trajectory are recorded at high speed simultaneously. Frequency scanning is used to find the target resonance response point, an adjustable standing wave field is established, the aggregation or repulsion motion of the target between the sound pressure node and the antinode is observed, and its amplitude and phase response are recorded. Analyze high-speed image sequences to identify whether the target exhibits a response pattern; If a response pattern is detected, it is classified as a "fluid response pattern" and ultimately determined to be a bubble. "Bubble Confirmation" is output, and virtual-real response mapping and difference comparison are performed for digital twin traceability and process optimization. If no response mode occurs or only slight rolling occurs without deformation, it is classified as "solid response mode", and is ultimately determined to be particles. The output is "particle confirmation", and the process is transferred to the cleanliness level assessment process to count the size and quantity to determine the level. If the response is between the two or cannot be identified, it will be determined as a "fuzzy response", marked as "abnormal and unclear", the inspection status of the pipe fitting will be frozen, the manual re-inspection process will be triggered, and all original data will be retained for future reference. The complete discrimination chain, which includes performing time-series dynamic feature analysis and preliminary discrimination, performing multispectral optical feature cross-validation, and performing mechanical micro-perturbation final discrimination, is summarized and a stage report is generated.

7. The method for detecting the cleanliness of the inner wall of stainless steel pipe fittings based on intelligent sensors according to claim 1, characterized in that: Perform virtual-to-real response mapping and difference comparison, specifically as follows: Based on the basic parameters of the pipe fitting obtained through the execution of basic parameter acquisition and validity determination, the corresponding specification's geometric model, material attribute library, and standard process template are retrieved from the digital twin library to activate the virtual instance of the pipe fitting. The key attributes of the detected events confirmed after the final judgment of the mechanical micro-perturbation are transformed into twin input parameters, and the corresponding scene is reconstructed in virtual space; Drive the twin to execute a virtual detection process, simulate the ideal response under the same sensor configuration and motion parameters, and output virtual prediction data; Using spatiotemporal coordinates as an index, the actual response data of entity detection is aligned with the ideal response data predicted by twins at the frame level to establish a benchmark for point-by-point comparison. The system calculates four indicators—morphological difference, trajectory deviation, spectral similarity, and mechanical response deviation—to comprehensively evaluate the overall degree of consistency between the physical and virtual entities. Compare the overall difference with the preset tolerance range; If all four indicators are below their respective thresholds and the overall consistency meets the standard, it is judged as "virtual and real consistent", confirming the effectiveness of the twin model, and directly performing multiphysics evolution analysis and final decision to predict evolution; If any indicator exceeds the standard or the overall consistency is insufficient, it is judged as "deviation between reality and virtuality". Anomaly root cause tracing and strategy pre-simulation optimization are performed to trace the root cause and further analyze the deviation characteristics. Output either "Virtual vs. Real" or "Virtual vs. Real" conclusions, along with a quantitative report of the difference and a deviation pattern label.

8. The method for detecting the cleanliness of the inner wall of stainless steel pipe fittings based on intelligent sensors according to claim 1, characterized in that: Performing root cause analysis and strategy pre-simulation optimization, specifically: Load the complete manufacturing history of the pipe fitting into the twin, replay each process sequentially according to the timeline, and extract the process parameters of each process; The actual execution parameter curves are compared point by point with the standard process template to identify abnormal periods where the deviation exceeds the tolerance, and the timestamps and magnitudes of parameter mutations are marked. Analyze abnormal parameter combinations; Based on the identified anomaly type, the optimization knowledge base is invoked to generate an adjustment plan; If the process parameters are displayed as normal throughout, retrieve the sensor performance degradation model, compare the current light source output intensity with the initial calibration value, analyze the trend of lens transmittance change, and assess the degree of probe mechanical wear. If the light source intensity decreases by more than 20%, or the color temperature drifts significantly, or the lens contamination causes a decrease in contrast, or the probe bearing clearance exceeds the standard, then the system will identify "abnormal equipment status" and generate a maintenance work order and spare parts replacement suggestions. For pipe fittings with changes in specifications or materials that are being inspected for the first time, the detection effects of different sensor combinations, sampling densities, light source angles, and propulsion speeds are tested in batches within a twin to simulate the distribution of blind spots and heat maps of misjudged hotspots. If the simulation shows that the current strategy has an unacceptable blind spot or the false judgment rate exceeds the standard, then output "strategy optimization suggestion", adjust the parameters and return to execute multi-sensor data acquisition to re-execute the complete detection process; If the pre-run verifies that the current strategy is comprehensive and the misjudgment rate is controllable, then "Strategy Confirmed" is output, and multiphysics evolution analysis and final decision-making are performed.

9. The method for detecting the cleanliness of the inner wall of stainless steel pipe fittings based on intelligent sensors according to claim 1, characterized in that: Perform multiphysics evolution analysis and final decision-making, specifically as follows: Key physical property parameters are extracted from cases of consistent virtual and real bubbles confirmed after performing virtual-real response mapping and difference comparison. Based on the extracted parameters, a heat-mass transfer coupling model was constructed to simulate the complete phase transition path of the liquid film from liquid to solid state under the current drying conditions, and to track the spatiotemporal evolution of solute precipitation and solvent volatilization. Based on the phase transition simulation results, the morphology of the final dried residue is predicted; Based on the initial thickness of the liquid film and the current drying rate, calculate the critical time required for complete drying in theory. Compare the theoretically required critical time for complete drying with the actual drying time obtained from the acquisition of basic parameters and effectiveness determination. If the actual drying time is lower than the critical value, an "insufficient drying" judgment will be output, quantifying the time gap; Multiphysics simulations were performed to model the asymmetric coalescence, rupture, or stable attachment modes of bubbles under complex boundary conditions. If the simulation shows that the local thermal boundary layer is too thick, causing abnormal bubble aggregation, the output will be "significant thermal boundary layer effect". It is recommended to optimize the hot air nozzle layout or increase turbulence to promote mixing. If the display shows that the electrostatic potential gradient drives the bubbles to migrate and attach to a specific wall surface, the output will be "significant electrostatic interference". It is recommended to increase the ion wind to neutralize or adjust the ambient humidity. For particle cases confirmed after performing virtual-real response mapping and difference comparison, calculate the adhesion strength and critical resuspension shear stress; If the current flow field shear stress is close to or exceeds the critical value, the risk of "false adhesion" will be output. It is recommended to reduce the movement speed of the detection probe or increase the pretreatment inert gas purging. Establish a time-varying evolution model for pollutants to simulate the time scale and path of the transformation from current bubble residue or particle attachment to "fresh residue - aged residue - stubborn stain". If it is predicted that the pipe fittings will be converted into a stubborn state of chemical bonding or mechanical interlocking before the next process, an "emergency handling" suggestion will be output, triggering priority rework or isolation of the batch of pipe fittings. If it is predicted that the current clearable state can be maintained for more than the safety buffer period, then the "normal processing" suggestion will be output and the process will be handled according to standard logistics. The analysis results of the entire process, including execution time-series dynamic feature analysis and preliminary judgment, execution multispectral optical feature cross-validation, execution mechanical micro-perturbation final judgment, execution virtual-real response mapping and difference comparison, and execution anomaly root cause tracing and strategy pre-simulation optimization, are integrated to generate a comprehensive optimization scheme. This scheme is fed back to the upstream production process through the manufacturing execution system. The complete parameter set, evolution path, prediction results and actual outcome of this case are associated and archived in the digital sample library for continuous training and optimization of the twin model.

Images