Metal seal micro-leakage detection system and method based on infrared image recognition

Abstract

The scheme belongs to the field of image recognition seal leakage, and specifically relates to a metal seal micro-leakage detection system and method based on infrared image recognition. The metal seal micro-leakage detection method based on infrared image recognition comprises the following steps: S10: collect the detection environment temperature as the background temperature, determine the contrast temperature according to the background temperature and the resolution of the infrared camera; set the air pressure threshold, use the steam not lower than the contrast temperature to fill the closed space of the metal seal, and monitor the inflation air pressure of the inflation to ensure that the inflation air pressure is not higher than the air pressure threshold; S20: use infrared and ultrasonic multi-angle to obtain image data, process the collected infrared image data, and identify and label the leakage range. The scheme solves the problem of imbalance between reconstruction accuracy and operation speed of the prior art under the constraints of complex environmental noise and computing resources, improves the leakage detection sensitivity and positioning accuracy, and reduces the hardware dependence and computing amount.

Description

Technical Field

[0001] This solution belongs to the field of image recognition for leak detection of seals, specifically involving a system and method for detecting minute leaks in metal seals based on infrared image recognition. Background Technology

[0002] In fields such as new energy vehicles and high-end manufacturing, leak detection of metal seals is crucial. The increasing number of large-volume test components (such as battery packs and air conditioner evaporators) places higher demands on the automation, efficiency, and accuracy of testing. While direct pressure leak detection methods can quickly and quantitatively detect leaks, they require extended testing cycles for small leaks in large volumes and cannot pinpoint the leak location. Helium mass spectrometry can locate leaks, but its long testing time and high cost make it unsuitable for large-scale production.

[0003] Infrared imaging has become a development direction for detection technology due to its speed, non-destructive nature, and low cost. However, the cost of infrared equipment increases with resolution. Small leaks in the target area have few pixels and inconspicuous temperature difference characteristics, making them difficult to detect with low-resolution equipment. High-resolution cameras are expensive and cannot be used in large-scale production lines.

[0004] Chinese patent CN115994893A discloses a method for detecting small leaks in sealed components based on low-resolution infrared images. It uses an IMR-DAM super-resolution reconstruction network, combined with multi-scale residual modules and a dual-attention mechanism, to reconstruct the low-resolution infrared image and then detect leak points using bilateral filtering and frame difference variance method. This method attempts to achieve high-precision detection with a low-cost camera, but it suffers from the following drawbacks: First, interference from airflow and thermal radiation in industrial environments exacerbates noise in infrared images. Although bilateral filtering is used for noise reduction, it is difficult to completely eliminate the masking of subtle temperature differences by dynamic environmental noise, causing the frame difference variance method to easily misjudge noise as leak point coordinates when calculating grayscale changes. Second, the super-resolution network constructed by this method contains 16 multi-scale residual attention modules. Although the resolution is improved through the IMR-DAM structure, the complex network architecture leads to an exponential increase in computational cost. Furthermore, even with the introduction of a combined optimization strategy of MSE + perceptual loss + content loss, it is still impossible to avoid the loss of high-frequency details (such as leak point edges) during the reconstruction of the low-resolution image. The contradiction in this method reflects the imbalance between reconstruction accuracy and computational speed in existing technologies under complex environmental noise and computational resource constraints. Summary of the Invention

[0005] The purpose of this solution is to provide a system and method for detecting minute leaks in metal seals based on infrared image recognition, in order to solve the problem of imbalance between reconstruction accuracy and computing speed in existing technologies under complex environmental noise and computational resource constraints.

[0006] To achieve the above objectives, this solution provides a method for detecting minute leaks in metal seals based on infrared image recognition, comprising the following steps:

[0007] S10: Collect the ambient temperature as the background temperature, determine the comparison temperature based on the background temperature and the resolution of the infrared camera; set the air pressure threshold, use steam at a temperature not lower than the comparison temperature to fill the sealed space of the metal seal, and monitor the inflation pressure to ensure that the inflation pressure does not exceed the air pressure threshold.

[0008] S20: Use infrared and ultrasonic waves to acquire image data from multiple angles, process the acquired infrared image data, identify and mark the leakage range; obtain the acquisition time of the infrared image data based on the marking results, calculate the leakage rate based on the acquisition time and leakage range; identify and mark the leakage location in the ultrasonic image based on the marking content of the infrared image data as the initial leakage location;

[0009] S30: Establish a leakage prediction model based on leakage speed, leakage range, leakage location and inflation pressure. The leakage prediction model is a machine learning model. Analyze the correlation of data in each direction in the leakage prediction model. Use the leakage prediction model to predict the inflation pressure, leakage speed, leakage range and leakage location after a preset time based on the currently collected leakage speed, leakage range and inflation pressure.

[0010] S40: Collect the leakage rate, leakage range, and inflation pressure after a preset time. Calculate the prediction accuracy based on the collected and predicted results. Obtain the difference between the predicted leakage location and the initial leakage location as a correction difference. Modify the initial leakage location based on the prediction accuracy and the correction difference. Use the modified initial leakage location as the leakage detection result.

[0011] And, an infrared image recognition-based metal seal micro-leakage detection system.

[0012] The principle and technical effects of this solution are as follows: First, based on the principles of multi-sensor data fusion and machine learning prediction, this solution collects ambient temperature to determine a comparison temperature. Vapor at a temperature not lower than this comparison temperature is then injected into the metal seal while controlling the pressure. By regulating temperature and pressure, the temperature difference at the leak point is enhanced, making the leak more prominent in the infrared image. This eliminates the need for high-resolution equipment, solving the detection challenges of low-to-medium resolution infrared devices at their source. It also avoids the conflict between computational resources and accuracy in complex environmental noise conditions, thus improving detection sensitivity. Simultaneously, by fusing multi-modal data (infrared and ultrasonic), feature complementarity is achieved. This eliminates the need for highly complex super-resolution reconstruction of low-resolution infrared images, allowing direct acquisition of the thermal characteristics and physical location of the leak point. This avoids the trade-off between improving accuracy and sacrificing computational speed in single-modal methods, reducing reliance on high-performance computing hardware.

[0013] Secondly, this solution utilizes machine learning models to perform correlation analysis and dynamic prediction of parameters such as leakage rate and range. It focuses computational resources on nonlinear fitting of key parameters, rather than pixel-level image reconstruction, significantly reducing computational load while maintaining detection accuracy. This solves the technical bottleneck of high-precision detection and computational resource constraints in complex industrial environments. Through iterative correction of actual detection data and prediction results, it dynamically optimizes leak point location without the need for fixed thresholds or complex post-processing. It adapts to leakage characteristics under different materials and air pressure conditions, avoiding missed detections caused by pixel aliasing in large-volume workpiece inspection, further improving the robustness and efficiency of detection in complex scenarios. Furthermore, by iteratively optimizing computational resource allocation through model iteration, this solution resolves the contradiction between high-precision reconstruction and computational cost in complex environments, improving both accuracy and efficiency.

[0014] Furthermore, this solution actively enhances the thermal signal intensity at the leak point during detection by coordinating the control of steam temperature and inflation pressure. When ultrasound is transmitted throughout the tested object, the discontinuity of the medium at the microcrack causes a nonlinear effect, triggering second harmonic vibrations. This leads to periodic compression and separation of the crack interface under high-frequency stress. This microscopic frictional heating manifests as characteristic patterns of abnormally high temperature in infrared imaging. Even without obvious leakage, the defect location can be accurately pinpointed through thermal signal differences. Further, through the coordinated excitation of inflation pressure and ultrasound, the crack exhibits a high-speed opening and closing state at a microscale under alternating stress, causing the sealing medium (such as gas) to periodically leak and be drawn into the crack gap. The heat changes carried by the leaking gas form dynamic heat flow characteristics in the infrared image, significantly enhancing the contrast between the defect area and the normal area. Compared to traditional single detection methods, this solution amplifies the thermal signal characteristics of microcracks through physical field coupling effects, improving the detection rate of microcracks and effectively solving the problem of insufficient sensitivity of conventional infrared detection to closed cracks. In the inspection of automotive braking system seals, this solution successfully identifies micro-fatigue cracks that traditional methods miss, improving the factory pass rate of seals and eliminating the risk of brake failure caused by minor defects from the source. It provides a breakthrough technical means for the quality control of seals in high-pressure and high-reliability scenarios.

[0015] Furthermore, the multi-sensor data fusion mechanism in this solution can form a redundant detection mode with cross-validation. When a single sensor is affected by factors such as material reflection or detection angle, another sensor can still provide effective detection data, significantly reducing the risk of missed detections due to equipment failure or environmental interference and improving the overall reliability of the system. The machine learning model has adaptive learning capabilities and can continuously optimize the prediction model by accumulating historical detection data under different working conditions. It can differentiate the leakage characteristics of workpieces of different specifications, such as new energy vehicle battery packs and aerospace sealing parts, without the need for frequent manual adjustment of detection parameters, significantly improving the detection efficiency and intelligence level of large-scale production lines. The dynamic correction mechanism based on the prediction results in this solution can predict the leakage development trend and quantify the risk level in advance, providing accurate maintenance timing suggestions for industrial maintenance. Compared with the post-detection mode, it can effectively reduce equipment downtime losses and safety hazards, realizing a technological upgrade from passive detection to proactive prevention.

[0016] Furthermore, this solution can also utilize the vibration of the seal caused by ultrasonic image data acquisition to raise the temperature of the seal, increase the temperature difference between the seal and the background, enhance the boundary difference between the seal and the background during the image recognition stage, and achieve a more accurate recognition effect.

[0017] In summary, this solution addresses the imbalance between reconstruction accuracy and computational speed in existing technologies under complex environmental noise and computational resource constraints. It also improves leakage detection sensitivity and location accuracy while reducing hardware dependence and computational load.

[0018] Furthermore, the vapor is atomized oil vapor or coating material vapor.

[0019] When this solution uses atomized oil vapor or coating material vapor for testing, the high-temperature environment allows for finer vapor particles. Combined with the uniform distribution effect produced by ultrasonic vibration, this results in a dense and uniform protective layer on the surface of metal seals. For example, in the sealing testing of new energy vehicle battery packs, high-temperature atomized oil vapor not only enhances the infrared characteristics of leak points through temperature differences, but its fine particles can also penetrate into the sealing gaps simultaneously under ultrasonic vibration, achieving lubrication coating while completing the testing. For precision seals in the aerospace field, coating material vapor, under the action of high temperature and vibration, can form a nanoscale protective film while testing airtightness. This shortens the time of the traditional step-by-step "testing-coating" process and improves the coating quality through dynamic and uniform coverage during the testing process. This integrated testing and maintenance approach shortens the seal production cycle, while the uniform material adhesion caused by high temperature and vibration enhances the bonding strength of the protective layer on the seal surface, achieving dual optimization of industrial testing and production quality.

[0020] Furthermore, in step S30, when analyzing the correlation of data in each direction in the leakage prediction model, historical detection data of leakage rate, leakage range, leakage location, and inflation pressure are first integrated. The data is then cleaned to remove outliers and standardized. Principal component analysis is then used to identify key parameters that significantly affect the leakage state, and the positive correlation between inflation pressure and leakage rate, as well as the spatial correlation between leakage range expansion and location change, are analyzed. Next, the feature importance evaluation mechanism of the machine learning model is used to quantify the contribution of each parameter to the leakage prediction results, and a parameter correlation network is constructed to visualize the mutual influence of data in each direction. Finally, cross-validation is used to verify the reliability of the correlation analysis results, ensuring that the model can accurately capture the dynamic correlation of each parameter during prediction.

[0021] This solution integrates historical data and removes outliers to eliminate noise caused by environmental interference or equipment fluctuations, allowing model training to focus more on real leak characteristics. Principal component analysis identifies key driving factors from multi-dimensional parameters such as inflation pressure and leakage rate; for example, it shows that increased pressure significantly accelerates leak expansion, providing accurate input for the prediction model. Feature importance assessment quantifies the contribution of each parameter, clarifying that the impact of pressure changes on leak location is heavier than other factors, helping engineers prioritize core variables. Visualized correlation networks intuitively present the interactions between parameters, such as how the expansion of the leak area triggers local pressure changes, forming a dynamic feedback mechanism. Cross-validation ensures the model remains stable and reliable across different batches of products, avoiding misjudgments due to material differences or process fluctuations. In summary, this solution improves leak prediction accuracy, shortens detection time, and the generated parameter correlation knowledge can directly guide seal design optimization, such as reducing material stress concentration in pressure-sensitive areas to reduce leak risk at the source.

[0022] Furthermore, in step S30, when analyzing the correlation of data in each direction in the leakage prediction model, spatiotemporal coupled principal component analysis is used to consider the spatial characteristics and time series of the leakage location, and a spatiotemporal correlation matrix is ​​constructed. , As shown in formula (2) below:

[0023] (2),

[0024] in, The distance is the Euclidean distance in three-dimensional space. For time intervals; With covariance matrix Fusion: ,right Eigenvalue decomposition is performed to obtain principal components that take into account both spatiotemporal information, thereby more accurately capturing the dynamic correlation between leaked parameters.

[0025] When evaluating feature importance and constructing the correlation network, the causal dependency between the parameters is proposed by using the Causal Information Flow Network (CIFN), as shown in the following formula (3):

[0026] (3),

[0027] in Information entropy, calculated The causal effect of inflation pressure on leakage rate is quantified, and a causal information flow network is constructed based on this, with nodes as parameters and edge weights determined by the following formula (4):

[0028] (4),

[0029] Visualize the causal transmission path between parameters.

[0030] In the scenario of turbine blade seal inspection for aero-engines, this solution demonstrates significant effectiveness. Spatiotemporal coupled principal component analysis can deeply uncover the leakage evolution patterns of different parts of the blade over flight time. For example, it can accurately locate the progressive leakage caused by high-temperature fatigue at the blade tenon joint, avoiding the omission of early potential problems due to neglecting the complex spatial structure and temporal changes of the blade. The causal information flow network can clearly identify the causal relationships between various parameters, intuitively presenting the inherent logic of how pressure fluctuations in the turbine cavity cause wear of the sealing ring, leading to increased leakage. During inspection, this solution quickly identifies the causal chain of blade tip sealing coating peeling due to sudden pressure changes, helping engineers adjust assembly processes in a timely manner. The synergy of these two approaches not only improves the accuracy of leakage parameter prediction but also provides maintenance personnel with a "roadmap" for fault tracing through visualized causal transmission paths, significantly shortening the maintenance cycle.

[0031] Furthermore, in step S40, when calculating the prediction accuracy, the collected results and prediction results are first preprocessed to unify the data format and units to ensure comparability. Then, a comparative analysis is performed from multiple dimensions, including the changing trend of leakage speed, the expansion pattern of leakage range, and the spatial distribution of leakage location. Next, corresponding evaluation indicators are set for each dimension to evaluate the degree of agreement between the collected results and the prediction results. Finally, weights are assigned according to the importance of different dimensions, and the prediction accuracy is calculated comprehensively.

[0032] This solution utilizes dynamic adaptive standardization to process heterogeneous data from different batches of seals, effectively eliminating abnormal interference caused by environmental temperature, humidity, and process fluctuations. This ensures that parameters such as leakage rate and range remain comparable even under complex operating conditions. The spatiotemporal attention contrast network accurately captures the dynamic leakage characteristics of seals under pressure changes, such as detecting differences in leakage expansion patterns at battery pack corners due to stress concentration, avoiding the missed detection of local anomalies by traditional methods. The dynamic weighted decision tree automatically adjusts the weights of evaluation indicators based on real-time operating conditions, increasing the weight of air pressure parameters during the high-temperature phase of fast charging, significantly enhancing the predictive ability for leaks caused by thermal expansion and contraction. The synergy of these three components elevates detection accuracy to a high level, helping companies reduce battery pack sealing failure rates to a low level, significantly reducing after-sales maintenance costs. Simultaneously, the generated multi-dimensional evaluation report provides data support for seal design optimization, such as guiding improvements to the sealing structure at battery pack corners, further enhancing product reliability.

[0033] Furthermore, it also includes assessing the confidence level of the predicted accuracy, with the following specific steps:

[0034] S41a: Collect historical data similar to the current detection conditions to construct a reference dataset;

[0035] S41b: Compare the current accurate prediction value with the results in the reference dataset to analyze the position of the current accurate prediction value in the distribution of historical results;

[0036] S41c: Evaluate the quality of the current detection data, including data integrity, noise level, and outlier handling.

[0037] S41d: Evaluate the adaptability of the leakage prediction model to the current operating conditions by combining the training performance of the leakage prediction model;

[0038] S41e: Consider the potential impact of environmental factors on various prediction data, including ambient temperature, humidity, and vibration;

[0039] S41f: Combining steps S41b to S41e, the confidence level of the predicted accuracy is graded and evaluated to determine the reliability level of the predicted accuracy.

[0040] This solution constructs a reference dataset by collecting historical data from similar operating conditions. This provides a reliable reference for current predictions based on past experience, avoiding biases caused by isolated assessments. By comparing the position of the current predicted value within the historical distribution, it visually demonstrates the degree of deviation from the norm, aiding in the assessment of prediction stability. A comprehensive evaluation of the quality of the current detection data effectively eliminates the risk of misjudgment caused by missing data, noise interference, or improper handling of outliers. Evaluating the operational adaptability of the leakage prediction model in conjunction with its training performance allows for the timely identification of potential model defects under special operating conditions. Simultaneously, fully considering the impact of environmental factors on the prediction data accurately captures the interference of external conditions such as temperature, humidity, and vibration on the performance of the seals. Finally, by comprehensively evaluating the results from various dimensions, a tiered judgment is made to clarify the reliability level of the predicted accuracy. This helps technicians quickly identify high-risk prediction results and prioritize handling anomalies with low confidence levels. This avoids resource waste caused by over-reliance on unreliable predictions and effectively prevents the risk of battery pack seal failure due to misjudgments.

[0041] Furthermore, in step S40, when modifying the initial leak location based on the predicted accuracy and correction difference, the following steps are included:

[0042] S42a: Set the correction adjustment coefficient based on the reliability level of the predicted accuracy value; the higher the level, the smaller the coefficient.

[0043] S42b: Multiply the correction difference by the adjustment factor to obtain the adjusted correction difference;

[0044] S42c: Correct the initial leakage location based on the adjustment results; positive differences move towards the predicted location, while negative differences move in the opposite direction.

[0045] S42d: After correction, assess whether the new location conforms to physical laws and the structure of the seal. At the same time, conduct a secondary verification based on the actual characteristics of the infrared image data. Modify the initial leak location again based on the verification results.

[0046] This solution dynamically sets correction adjustment coefficients based on the reliability level of the predicted accuracy, avoiding overcorrection for high-reliability predictions and undercorrection for low-reliability predictions. For example, when the detection system provides a high-reliability prediction, the adjustment coefficient automatically approaches 0, ensuring the correction range remains within the minimum necessary range. The adaptive weighted fusion model fully considers the spatiotemporal characteristics of the leak location, accurately capturing the location offset patterns caused by pipeline vibration or thermal expansion and contraction. The physical constraint optimization model forces the correction results to conform to the structural characteristics of the sealing component, eliminating unreasonable location offsets. For example, at pipeline flange connections, it effectively prevents the corrected position from exceeding the actual sealing area. Multimodal consistency verification combined with infrared image features performs secondary verification, promptly detecting misjudgments caused by detection errors or environmental interference, reducing leak location positioning errors to a low level. This approach improves the accuracy of leak point location, significantly reducing the troubleshooting time and costs for maintenance personnel.

[0047] Furthermore, when modifying the initial leak location, the weights of each dimension are first clarified, and the comparison results of each dimension are weighted. By comparing with the preset screening threshold, the correction differences that should be ignored or adopted are determined. For the adopted correction differences, the modification coefficient is determined and adjusted according to the reliability level of the predicted accuracy value. After adjustment, the rationality of the correction difference is evaluated. If it is not reasonable, it is readjusted until it is reasonable. Finally, the initial leak location is modified according to the reasonable correction difference, and the screened correction differences are verified in combination with the structural characteristics and physical laws of the sealing component.

[0048] In the detection of minute leaks in metal seals, this solution significantly improves the accuracy and engineering applicability through a collaborative mechanism of multi-dimensional weight fusion, adaptive screening, reliability grading optimization, and two-way verification. The solution first clarifies and weights the multi-dimensional characteristics of the leak, accurately filtering out invalid correction differences caused by environmental interference or noise. Then, it dynamically adjusts the correction magnitude based on the reliability of the predicted accuracy value; high-reliability predictions reduce corrections to avoid over-adjustment, while low-reliability predictions enhance corrections to improve accuracy. Finally, through dual verification of rationality assessment and the structural characteristics of the seal, it ensures that the corrected leak location conforms to physical laws.

[0049] Furthermore, when filtering correction discrepancies, the temperature, pressure, and vibration frequency in the detection environment are first collected in real time to construct an environmental feature vector. The environmental feature vector is then input into a pre-trained Gaussian mixture model, and the current working condition category is identified by calculating the posterior probability of the samples. Based on the identified working condition category, a dynamic threshold function is called to adjust the filtering threshold. Then, the comparison results of each dimension are weighted and fused with their corresponding weights to obtain a comprehensive score. Subsequently, the weighted result is compared element-wise with the adjusted threshold vector to generate a filtering mask, which is used to filter correction discrepancy vectors. Finally, the current environmental parameters and filtering results are stored in a historical database, and the parameters of the Gaussian mixture model are updated through online incremental learning to continuously optimize the accuracy of working condition identification.

[0050] This solution collects environmental parameters such as temperature, pressure, and vibration frequency in real time. Utilizing an unsupervised learning model, it accurately identifies operating conditions and dynamically adjusts screening thresholds for leakage rate and infrared feature differences based on these conditions. This effectively filters out false correction discrepancies caused by environmental changes or equipment interference. Simultaneously, it weights the comparison results across various dimensions and compares the weighted results with dynamic thresholds to ensure that the collected correction discrepancies accurately reflect the leakage situation. Furthermore, it adjusts the correction magnitude based on the reliability level of the predicted accuracy value and verifies the results based on the structural characteristics of the thin-walled curved surface of the seal. Ultimately, it controls the leakage location error within a low range, improving the accuracy of good product inspection and significantly reducing the risk of instrument short circuits and infection due to seal failure, thus avoiding product recall losses. Attached Figure Description

[0051] Figure 1 This is a flowchart of a method for detecting minute leaks in metal seals based on infrared image recognition, as described in an embodiment of the present invention.

[0052] Figure 2 This is a flowchart illustrating the evaluation of the confidence level of the predicted accuracy in an embodiment of the present invention.

[0053] Figure 3 This is a flowchart illustrating the modification of the initial leak location based on the predicted accuracy and correction difference in an embodiment of the present invention. Detailed Implementation

[0054] The following will describe the concept and technical effects of the present invention clearly and completely with reference to embodiments, so as to fully understand the purpose, features and effects of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are all within the scope of protection of the present invention.

[0055] like Figure 1 As shown, a method for detecting minute leaks in metal seals based on infrared image recognition includes the following steps:

[0056] S10: Collect the ambient temperature as the background temperature, determine the comparison temperature based on the background temperature and the resolution of the infrared camera; set the air pressure threshold, use steam at a temperature not lower than the comparison temperature to fill the sealed space of the metal seal, and monitor the inflation pressure to ensure that the inflation pressure does not exceed the air pressure threshold.

[0057] S20: Use infrared and ultrasonic waves to acquire image data from multiple angles, process the acquired infrared image data, identify and mark the leakage range; obtain the acquisition time of the infrared image data based on the marking results, calculate the leakage rate based on the acquisition time and leakage range; identify and mark the leakage location in the ultrasonic image based on the marking content of the infrared image data as the initial leakage location;

[0058] S30: Establish a leakage prediction model based on leakage speed, leakage range, leakage location and inflation pressure. The leakage prediction model is a machine learning model. Analyze the correlation of data in each direction in the leakage prediction model. Use the leakage prediction model to predict the inflation pressure, leakage speed, leakage range and leakage location after a preset time based on the currently collected leakage speed, leakage range and inflation pressure.

[0059] S40: Collect the leakage rate, leakage range, and inflation pressure after a preset time. Calculate the prediction accuracy based on the collected and predicted results. Obtain the difference between the predicted leakage location and the initial leakage location as a correction difference. Modify the initial leakage location based on the prediction accuracy and the correction difference. Use the modified initial leakage location as the leakage detection result.

[0060] Among them, vapor is atomized oil vapor or coating material vapor.

[0061] Specifically, the atomizing oil includes polyalphaolefin (PAO) synthetic oil, silicone oil, and other materials that are resistant to high and low temperatures, oxidation, and chemical stability; the coating materials include nano-silica, molybdenum disulfide, and other materials. In this embodiment of the solution, PAO atomized oil vapor is used for airtightness testing of the gearbox housing, and is injected at a high temperature of 120°C. This not only enhances the indication of leak points through temperature difference but also forms a lubricating film with a thickness of approximately 5 μm on the sealing surface, reducing friction during subsequent assembly.

[0062] In step S30, when analyzing the correlation of data in each direction in the leakage prediction model, historical detection data of leakage speed, leakage range, leakage location, and inflation pressure are first integrated. The data is then cleaned to remove outliers and standardized. Principal component analysis is then used to identify key parameters that significantly affect the leakage status, and the positive correlation between inflation pressure and leakage speed, as well as the spatial correlation between leakage range expansion and location change, are analyzed. Next, the feature importance evaluation mechanism of the machine learning model is used to quantify the contribution of each parameter to the leakage prediction results, and a parameter correlation network is constructed to visualize the mutual influence of data in each direction. Finally, cross-validation is used to verify the reliability of the correlation analysis results, ensuring that the model can accurately capture the dynamic correlation of each parameter during prediction.

[0063] Specifically, during data integration and cleaning, a dynamic weighted filtering method is introduced to handle outliers, assigning dynamic weights to each sample based on the time-series characteristics of the data. , The calculation formula is shown in formula (1) below:

[0064] (1),

[0065] in For sample collection time, This is the time decay coefficient, used to weight the data for a weighted average, automatically mitigating the impact of outliers. Standardization employs an adaptive power transform. ; where q is dynamically determined by the kurtosis and skewness of the data, improving the standardization of the data distribution.

[0066] During the correlation analysis, spatiotemporal coupled principal component analysis (ST-PCA) was used to consider the spatial characteristics and time series of the leakage location, and a spatiotemporal correlation matrix was constructed. , As shown in formula (2) below:

[0067] (2),

[0068] in, The first in the spatiotemporal correlation matrix The sample and the first The degree of association between individual samples.

[0069] To leak location and The three-dimensional Euclidean distance, Sample collection time and The time interval is used to measure the differences in time series data. This represents the total number of samples. With covariance matrix Fusion: ,right By performing eigenvalue decomposition, we can obtain principal components that take into account both spatiotemporal information and more accurately capture the dynamic correlation between leaked parameters.

[0070] When evaluating feature importance and constructing the correlation network, the causal dependency between the parameters is proposed by using the Causal Information Flow Network (CIFN), as shown in the following formula (3):

[0071] (3),

[0072] in For information entropy, In order to meet the conditions Below, variables For variables Conditional causal entropy, a measure right The degree of causal influence; In order to meet the conditions Down Information entropy, representation Uncertainty; In order to meet the conditions and Down Information entropy, representing known and hour Uncertainty; through calculation The causal effect of inflation pressure on leakage rate is quantified, and a causal information flow network is constructed based on this, with nodes as parameters and edge weights determined by the following formula (4):

[0073] (4),

[0074] For nodes in a causal information flow network arrive The edge weight, representing right The weight of causal influence; In order to exclude and After that, under other variable conditions right The conditional causal entropy; m is the total number of parameters (such as leakage rate, air pressure, range, etc.). It is used in step S30 to construct the parameter correlation network and visualize the mutual influence of data in each direction.

[0075] During cross-validation, dynamic Bayesian validation (DBV) is used in combination with Bayesian inference and a sliding window mechanism. The reliability of the model is updated by the posterior probability, as shown in the following formula (5):

[0076] (5),

[0077] Given the observed data Data, this represents the reliable posterior probability of the model. This is the likelihood function of the observed data Data when the model is reliable. The prior probability for model reliability (set based on historical performance). This is the likelihood function of the observed data Data when the model is unreliable. The prior probability of model unreliability ( ). Among them, the prior probability Based on the historical model performance specification, the likelihood function The prediction error is dynamically calculated within the sliding window. The reliability of the model is determined at that time.

[0078] In this embodiment of the solution, during the detection of minor leaks in the annular seal between the turbine blade and the casing on the aero-engine turbine blade production line, 1000 sets of historical data were collected, including inflation pressure (0.5-1.5MPa), leakage velocity (0.1-5mL / min), leakage location (3D coordinates), and leakage range (pixel area). A dynamic weighted filtering method was applied to automatically identify and remove outliers caused by sensor fluctuations, such as samples where the pressure suddenly increased during a test but no corresponding leakage change occurred. Spatiotemporal coupled principal component analysis revealed that pressure changes in the leading edge region of the blade had a significantly greater impact on the leakage velocity than those in the trailing edge region, revealing the influence of geometric structure on leakage. A causal information flow network was constructed to visualize the causal chain where increased pressure first causes seal deformation (manifested as positional shift), leading to an expansion of the leakage range. A random forest model was used to quantify parameter importance, finding that pressure contributes 42% to the leakage velocity, making it a priority monitoring parameter. Five-fold cross-validation was employed to ensure that the model's prediction error rate was below 5% across different batches of blade inspections. In actual testing, 200℃ atomized silicone oil vapor is injected into the sealed cavity, and leak detection and lubrication coating are performed simultaneously. Real-time data is collected and input into the prediction model. When the predicted leak expansion exceeds the threshold, the air pressure parameters are automatically adjusted and high-risk areas are marked.

[0079] In step S40, when calculating the prediction accuracy, the collected results and prediction results are first preprocessed to unify the data format and units, ensuring comparability. Then, a comparative analysis is performed from multiple dimensions, including the changing trend of leakage rate, the expansion pattern of leakage range, and the spatial distribution of leakage location. Next, corresponding evaluation indicators are set for each dimension to evaluate the degree of agreement between the collected results and the prediction results. Finally, weights are assigned according to the importance of different dimensions, and the prediction accuracy is calculated comprehensively.

[0080] Specifically, a dynamic scaling factor is introduced to standardize the collected and predicted results, avoiding the sensitivity of traditional Z-scores to outliers. For each data dimension d, the dynamic scaling factor... Dynamic calculation based on local data distribution , The calculation formula is shown in formula (6) below:

[0081] (6),

[0082] in, is the dynamic scaling factor for the d-th data dimension, used to standardize the collected and predicted results. The interquartile range of the data in the d-th dimension measures the degree of data dispersion. For the data collection results in the d-th dimension, This is the predicted result. This is a median function, reducing the impact of outliers. The standardized data are as follows: .

[0083] By using the above methods, local data characteristics can be enhanced, thereby improving the comparability of data under different operating conditions.

[0084] In one embodiment of this solution, during the testing of a batch of battery pack seals, a slight adjustment in the manufacturing process caused a change in the elastic coefficient of some seals, resulting in intermittent abnormal data from the leakage velocity sensor. The Dynamic Adaptive Standardization (DAS) method was employed to first calculate the interquartile range of the leakage velocity data for this batch, and then generate a scaling factor by combining the median difference between the collected and predicted data. After processing with this factor, the interference of abnormal data on the overall trend was significantly reduced, and heterogeneous data collected from different testing devices were accurately aligned, improving the accuracy of subsequent analysis.

[0085] In multi-dimensional analysis, a dynamic time warping (DTW)-attention mechanism is used to align velocity sequences and calculate similarity: Attention weights are dynamically allocated based on the intensity of local fluctuations in the sequence, highlighting changes at key time points; topological encoding transforms the two-dimensional leakage range image into a feature vector. Morphological differences are assessed using the Structural Similarity Index (SSIM): The concept of spatial density field is introduced to transform discrete location points into probability density distributions. and The Wasserstein distance metric is used to measure the distribution difference. .

[0086] In this embodiment of the solution, the detection scenario for the liquid cargo tank seals of a large LNG carrier involves cylindrical and massive liquid cargo tanks with seals distributed at the tank seams. During detection, it is necessary to simultaneously monitor the changes in leakage at different heights and circumferential positions within the tank, as well as the expansion trend over time. Traditional methods struggle to accurately capture the dynamic characteristics of leakage caused by ship swaying and changes in cryogenic medium pressure. A Spatiotemporal Attention Comparison Network (SACN) is used to detect the liquid cargo tank seals. In the detection of a specific tank section, dynamic time warping and an attention mechanism are used to align the time axis deviation of leakage velocity data caused by ship swaying, accurately identifying the trend error of the predicted results in the bottom region of the tank. Analysis of infrared thermal images using topological structure encoding reveals a significant difference between the predicted leakage range at the corners of the tank seams and the actual range. The spatial density field method is used to compare the probability distribution of the leakage location, pinpointing the offset between the predicted and actual locations along the tank circumference. Ultimately, SACN, through multi-dimensional comparison, improves the leakage detection accuracy to a high level, effectively avoiding the risk of LNG leakage due to seal failure.

[0087] When evaluating indicators and assigning weights, a dynamic weight decision tree is constructed, and the dimensional weights are automatically adjusted according to operating parameters (such as inflation pressure and sealing material). The weight update formula for each decision node is shown in formula (7) below:

[0088] (7),

[0089] in, The weight of the d-th dimension is used to comprehensively calculate the prediction accuracy. D is the total number of dimensions (such as leakage rate, range, location, etc.). d represents the information gain of dimension d, which measures the contribution of that dimension to the prediction. This represents the historical prediction confidence level of the d-th dimension under the current operating conditions. The final prediction accuracy value. for: .

[0090] In this embodiment of the solution, during vacuum chamber sealing inspection in a semiconductor chip manufacturing workshop, an extremely high vacuum level must be maintained inside the chamber, placing stringent requirements on the airtightness of the sealing components. The chamber pressure and temperature conditions vary significantly across different process stages, and traditional fixed-weight evaluation methods cannot adapt to the dynamic changes in parameter importance. In a pre-lithography inspection, chamber preheating caused thermal expansion of the sealing components. At this time, pressure fluctuations significantly increased the impact of leakage, but the original model still used leakage rate as the core indicator, resulting in inaccurate predictions. By introducing a Dynamic Weighted Decision Tree (DWDT), the weights are automatically adjusted based on real-time operating parameters (e.g., chamber temperature 150°C, pressure 5Pa). By calculating the information gain and historical confidence level of each dimension, the leakage rate weight is reduced from 0.5 to 0.2, while the inflation pressure weight is increased from 0.1 to 0.6. The final calculated prediction accuracy reaches 0.91, a significant improvement over traditional methods. Based on this, technicians can promptly identify the problem of partial sealing failure caused by thermal expansion, avoiding a decrease in chip production yield due to substandard vacuum levels.

[0091] More specifically, this also includes assessing the confidence level of the predicted accuracy, such as... Figure 2 As shown, the specific steps are as follows:

[0092] S41a: Collect historical data similar to the current detection conditions to construct a reference dataset;

[0093] S41b: Compare the current accurate prediction value with the results in the reference dataset to analyze the position of the current accurate prediction value in the distribution of historical results;

[0094] S41c: Evaluate the quality of the current detection data, including data integrity, noise level, and outlier handling.

[0095] S41d: Evaluate the adaptability of the leakage prediction model to the current operating conditions by combining the training performance of the leakage prediction model;

[0096] S41e: Consider the potential impact of environmental factors on various prediction data, including ambient temperature, humidity, and vibration;

[0097] S41f: Combining steps S41b to S41e, the confidence level of the predicted accuracy is graded and evaluated to determine the reliability level of the predicted accuracy.

[0098] In step S40, when modifying the initial leak location based on the predicted accuracy and correction difference, such as... Figure 3 As shown, it includes the following steps:

[0099] S42a: Set the correction adjustment coefficient based on the reliability level of the predicted accuracy value; the higher the level, the smaller the coefficient.

[0100] S42b: Multiply the correction difference by the adjustment factor to obtain the adjusted correction difference;

[0101] S42c: Correct the initial leakage location based on the adjustment results; positive differences move towards the predicted location, while negative differences move in the opposite direction.

[0102] S42d: After correction, assess whether the new location conforms to physical laws and the structure of the seal. At the same time, conduct a secondary verification based on the actual characteristics of the infrared image data. Modify the initial leak location again based on the verification results.

[0103] Specifically, when modifying the initial leak location, the weight of each dimension is first clarified, and the comparison results of each dimension are weighted. By comparing with the preset screening threshold, the correction difference that should be ignored or adopted is determined. For the adopted correction difference, the modification coefficient is determined and adjusted according to the reliability level of the predicted accuracy value. After adjustment, the rationality of the correction difference is evaluated. If it is not reasonable, it is readjusted until it is reasonable. Finally, the initial leak location is modified according to the reasonable correction difference, and the screened correction difference is verified in combination with the structural characteristics and physical laws of the sealing component.

[0104] In this embodiment of the solution, the quality inspection of the pacemaker housing seal is crucial because seal failure could lead to short circuits or infection risks, necessitating extremely high accuracy in detecting even minor leaks. During inspection, this solution first determines the weight of each dimension based on the leakage rate, the infrared thermographic characteristics of the leakage area, and the location of micro-cracks on the seal surface. The comparison results are then weighted to successfully filter out spurious correction discrepancies caused by electrostatic interference from production equipment. For the correction discrepancies used, adjustments are made based on the reliability level of the predicted accuracy value. If the prediction reliability is high, only a minor correction is made to the initial leak location; if the reliability is low, the correction magnitude is increased by incorporating historical production data.

[0105] After adjustments, the rationality of the correction discrepancies was assessed, and the thin-walled, curved surface structure of the pacemaker housing seal was used to verify the selected correction discrepancies. Ultimately, the location error of the leak was controlled within a low range, improving the accuracy of good product testing for the seal, effectively ensuring the safety and reliability of the medical device, and avoiding the risk of product recalls due to sealing issues.

[0106] More specifically, when filtering for correction discrepancies, the system first collects real-time data on temperature, pressure, and vibration frequency in the detection environment to construct an environmental feature vector. This feature vector is then input into a pre-trained Gaussian mixture model, which calculates the posterior probability of the samples to identify the current operating condition category. Based on the identified operating condition category, a dynamic threshold function is called to adjust the filtering threshold. Specifically, the leakage rate threshold tightens with increasing pressure, while the infrared feature difference threshold increases with increasing vibration frequency. Simultaneously, weighted coefficient thresholds are set to filter low-confidence dimensions. The comparison results of each dimension are then weighted and fused with their corresponding weights to obtain a comprehensive score. Subsequently, the weighted result is compared element-wise with the adjusted threshold vector to generate a filtering mask, which is used to filter correction discrepancy vectors. Finally, the current environmental parameters and filtering results are stored in a historical database, and the Gaussian mixture model parameters are updated through online incremental learning to continuously optimize the operating condition identification accuracy.

[0107] In this embodiment of the solution, environmental parameters at the detection station are collected in real time on the production line, including workshop temperature, cleanroom air pressure, and vibration frequency generated by equipment operation. After inputting these environmental parameters into a pre-trained unsupervised learning model, this solution can accurately identify the current production conditions, such as distinguishing between strong vibration conditions during high-speed equipment operation and weak vibration conditions during low-speed debugging. The threshold for filtering and correcting differences is dynamically adjusted based on the condition identification results. When the cleanroom air pressure fluctuates, the leakage rate threshold is automatically tightened to prevent minor leaks from being masked by air pressure changes; if the vibration frequency of the detection equipment increases, the screening threshold for infrared image feature differences is immediately increased to prevent misjudgments caused by artifacts generated by equipment vibration. In the weighted processing stage, weights are assigned according to the importance of dimensions such as leakage rate, infrared thermal image features, and the location of micro-cracks in the seal, and the comparison results of each dimension are weighted and fused to obtain a comprehensive evaluation value.

[0108] Subsequently, the weighted results were compared with the adjusted thresholds to identify and ignore spurious correction differences caused by factors such as electrostatic interference, using only the true and valid correction information. The used correction differences were adjusted based on the reliability level of the predicted accuracy: if the prediction reliability was high, only minor corrections were made to the initial leak location; if the reliability was low, the correction magnitude was increased by combining historical test data from the same batch of products. Finally, considering the special structural characteristics of the thin-walled, curved surface of the pacemaker's housing seal, the screened correction differences were further verified to ensure accurate leak location.

[0109] The above descriptions are merely embodiments of the present invention, and common knowledge regarding specific structures and characteristics is not elaborated upon here. It should be noted that those skilled in the art can make various modifications and improvements without departing from the structure of the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A method for detecting a slight leakage of a metal seal based on infrared image recognition, characterized in that, The method comprises the following steps: S10: Collect the ambient temperature as the background temperature, and determine the contrast temperature according to the background temperature and the resolution of the infrared camera; Set a pressure threshold, fill the sealed space of the metal sealing member with steam not lower than the contrast temperature, and monitor the inflation pressure of the inflation to ensure that the inflation pressure is not higher than the pressure threshold; S20: Obtain image data by using infrared rays and ultrasonic waves from multiple angles, process the collected infrared image data, identify and label the leakage range therein, obtain the collection time of the infrared image data according to the labeling result, calculate the leakage speed according to the collection time and the leakage range, and identify and label the leakage position in the ultrasonic wave image as the initial leakage position according to the labeling content of the infrared image data; S30: Establish a leakage prediction model based on the leakage speed, the leakage range, the leakage position and the inflation pressure, the leakage prediction model is a machine learning model, analyze the correlation of the data in the leakage prediction model, and use the leakage prediction model to predict the inflation pressure, the leakage speed, the leakage range and the leakage position after a preset time according to the current collected leakage speed, the leakage range and the inflation pressure; S40: Collect the leakage speed, the leakage range and the inflation pressure after the preset time, calculate the prediction accuracy value according to the collection result and the prediction result, obtain the difference between the predicted leakage position and the initial leakage position as the correction difference, modify the initial leakage position according to the prediction accuracy value and the correction difference, and take the modified initial leakage position as the leakage detection result.

2. The method for detecting the micro-leakage of metal seal based on infrared image recognition according to claim 1, characterized in that: The steam is atomized oil steam or film-coated material steam.

3. The method for detecting the micro-leakage of metal seal based on infrared image recognition according to claim 2, characterized in that: In the step S30, when analyzing the correlation of the data in the leakage prediction model, first, integrate the historical detection data of the leakage speed, the leakage range, the leakage position and the inflation pressure, clean the data to eliminate abnormal values and standardize the processing; then identify the key parameters that significantly affect the leakage state through principal component analysis, analyze the positive correlation between the inflation pressure and the leakage speed, and the spatial correlation between the expansion of the leakage range and the change of the position; then use the feature importance evaluation mechanism of the machine learning model to quantify the contribution of each parameter to the leakage prediction result, construct a parameter correlation network to visualize the mutual influence of the data in each direction; finally, verify the reliability of the correlation analysis result through cross-validation to ensure that the model prediction can accurately capture the dynamic correlation of each parameter.

4. The method for detecting micro-leakage of metal seal based on infrared image recognition according to claim 3, characterized in that: When analyzing the relevance of each direction data in the leakage prediction model in the step S30, the spatial characteristics and time series of the leakage position are considered by using the space-time coupling principal component analysis to construct a space-time correlation matrix , As shown in the following formula (2): （2）， wherein, is the correlation degree between the i-th sample and the j-th sample in the spatio-temporal correlation matrix; is the correlation degree between the i-th sample and the j-th sample in the spatio-temporal correlation matrix; is the correlation degree between the i-th sample and the j-th sample in the spatio-temporal correlation matrix is the three-dimensional spatial Euclidean distance of the leak location and , is the time interval for measuring the time series difference and , is the total number of samples; and is fused with the covariance matrix : Eigenvalue decomposition is performed on to obtain principal components that take into account the spatial and temporal information and more accurately capture the dynamic correlation between the leakage parameters. When evaluating the feature importance and constructing the correlation network, the conditional causal entropy is used to measure the causal dependence between parameters through a causal information flow network, as shown in the following formula (3): （3）， in For information entropy, In order to meet the conditions Below, variables For variables Conditional causal entropy, a measure right The degree of causal influence; In order to meet the conditions Down Information entropy, representation Uncertainty; In order to meet the conditions and Down Information entropy, representing known and hour Uncertainty; through calculation The causal effect of inflation pressure on leakage rate is quantified, and a causal information flow network is constructed based on this, with nodes as parameters and edge weights determined by the following formula (4): （4）， For nodes in a causal information flow network arrive The edge weight, representing right The weight of causal influence; In order to exclude and After that, under other variable conditions right The conditional causal entropy; m is the total number of parameters; The causal conduction path between parameters is visualized.

5. The method for detecting the micro-leakage of metal seal based on infrared image recognition according to claim 4, characterized in that: In the step S40, when calculating the prediction accuracy value, first, preprocess the collection result and the prediction result to unify the data format and dimension, and ensure their comparability; Then, compare and analyze from multiple dimensions, including the change trend of the leakage speed, the expansion form of the leakage range and the spatial distribution of the leakage position; then set corresponding evaluation indexes for each dimension to evaluate the degree of coincidence between the collection result and the prediction result; and then allocate weights according to the importance of different dimensions to comprehensively calculate the prediction accuracy value.

6. The method for detecting the micro-leakage of metal seal based on infrared image recognition according to claim 5, characterized in that: The confidence of the prediction accuracy value is also evaluated, and the specific steps are as follows: S41a: Collect historical data similar to the current detection condition, and construct a reference data set; S41b: Compare the current prediction accuracy value with the results in the reference data set, and analyze the position of the current prediction accuracy value in the historical result distribution; S41c: Evaluate the quality of the current detection data, including data integrity, noise level, and abnormal value processing of the current data; S41d: Evaluate the adaptability of the leakage prediction model to the current condition in combination with the training performance of the leakage prediction model; S41e: Consider the potential influence of environmental factors on each prediction data, including environmental temperature, humidity, and vibration; S41f: Based on steps S41b to S41e, perform a hierarchical evaluation of the confidence of the prediction accuracy value, and determine the reliability level of the prediction accuracy value.

7. The method for detecting the micro-leakage of metal seal based on infrared image recognition according to claim 6, characterized in that: In step S40, when modifying the initial leakage position according to the prediction accuracy value and the correction difference, the following steps are included: S42a: Set a correction adjustment coefficient according to the reliability level of the prediction accuracy value, and the higher the level, the smaller the coefficient; S42b: Multiply the correction difference by the adjustment coefficient to obtain the adjusted correction difference; S42c: Modify the initial leakage position according to the adjustment result, with positive difference moving towards the predicted position and negative difference moving in the opposite direction; S42d: After modification, evaluate whether the new position meets the physical law and the structure of the sealing element, and perform secondary verification in combination with the actual features of the infrared image data, and modify the initial leakage position again according to the verification result.

8. The method for detecting the micro-leakage of metal seal based on infrared image recognition according to claim 7, characterized in that: In step S40, when modifying the initial leakage position, first determine the weight of each dimension, weight the comparison results of each dimension, and compare them with the preset screening threshold to determine whether to ignore or use the correction difference; for the used correction difference, determine the modification coefficient according to the reliability level of the prediction accuracy value and adjust it; after adjustment, evaluate the reasonableness of the correction difference, and if it is not reasonable, adjust it again until it is reasonable; finally, modify the initial leakage position according to the reasonable correction difference, and verify the screened correction difference in combination with the structure characteristics of the sealing element and the physical law.

9. The method for detecting micro-leakage of metal seal based on infrared image recognition according to claim 8, characterized in that: When screening the correction difference, first collect the temperature, pressure, and vibration frequency in the detection environment to construct an environmental feature vector; input the environmental feature vector into the pre-trained Gaussian mixture model to identify the current condition category by calculating the sample posterior probability; Based on the identified condition category, call the dynamic threshold function to adjust the screening threshold; then weight and fuse the comparison results of each dimension with the corresponding weight to obtain a comprehensive score; then compare the weighted results with the adjusted threshold vector element by element to generate a screening mask, and filter the correction difference vector according to the mask; finally, store the current environmental parameters and screening results in the historical database, and update the Gaussian mixture model parameters through online incremental learning to continuously optimize the condition recognition accuracy.

10. A system for detecting a micro-leakage of a metal seal based on infrared image recognition, characterized in that, The metal sealing element micro-leakage detection method based on infrared image recognition of any one of claims 1-9 is used.

Images