Vacuum box helium leakage positioning method and system based on machine learning
Machine learning technology was used to purify and reduce the dimensionality of helium concentration data in the vacuum chamber, and a concentration gradient vector field was constructed. The gradient descent algorithm was then used to track the helium leak source, which solved the problem of accurately locating the helium leak source in the vacuum chamber and improved the safety and reliability of the system.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHENZHEN WALSHENG INTELLIGENT CONTROL TECH CO LTD
- Filing Date
- 2025-12-15
- Publication Date
- 2026-05-19
AI Technical Summary
Existing technologies cannot accurately identify the location of helium leaks in vacuum chambers, mainly because the complexity and multidimensionality of helium concentration data make feature extraction difficult, and it is impossible to accurately locate the direction and rate of helium diffusion.
A machine learning-based approach is used to pre-process the helium concentration data collected by multiple sensor nodes in the vacuum chamber, perform dimensionality reduction and feature vector extraction, generate an initial gradient distribution matrix of sampling points, construct a three-dimensional concentration gradient vector field, and use the gradient descent algorithm to trace back and iterate multiple times to determine the location of the leak source.
It enables precise location of helium leak sources in vacuum chambers, improving the safety and reliability of vacuum systems.
Smart Images
Figure CN122062848A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of electronic digital data processing technology, and in particular to a method and system for locating helium leaks in a vacuum chamber based on machine learning. Background Technology
[0002] Vacuum chamber helium leak detection is a core technology for industrial safety monitoring. Helium is an ideal tracer gas, and the accuracy of its leak detection is directly related to equipment safety and the reliability of product quality control.
[0003] The fundamental challenge in locating helium leaks within a vacuum chamber lies in the complexity and multidimensionality of helium concentration data. Multi-point helium concentration data often contains a large amount of redundant information and noise interference. The leak source location information contained within this complex helium concentration data exhibits highly nonlinear and multi-coupling characteristics. Due to the lack of effective feature extraction methods, the system cannot accurately identify key diffusion mode features from massive concentration data. This difficulty in feature identification further leads to inaccuracies in the construction of the concentration gradient vector field. Traditional helium leak detection methods, which employ fixed threshold judgments and simple signal processing techniques, cannot accurately determine the main direction and rate of helium diffusion, thus causing deviations in leak source location.
[0004] It is evident that there is an urgent need for those skilled in the art to provide a solution that can accurately locate the nitrogen leak in a vacuum chamber. Summary of the Invention
[0005] The purpose of this invention is to provide a machine learning-based method and system for locating helium leaks in a vacuum chamber, which can solve the problem in the prior art of being unable to accurately locate the location of nitrogen leaks in a vacuum chamber.
[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a machine learning-based method for locating helium leaks in a vacuum chamber, wherein the method includes: The nitrogen concentration dataset collected by multiple sensor nodes inside the vacuum chamber is subjected to a pre-set purification process to obtain the target nitrogen concentration dataset. After dimensionality reduction processing of the target nitrogen concentration dataset, feature vector extraction is performed to obtain the target concentration feature vector. An initial gradient distribution matrix for sampling points is generated based on the target concentration feature vector; The target diffusion mode type and diffusion parameters are determined based on the initial gradient distribution matrix of the sampling points; A three-dimensional concentration gradient vector field is constructed based on the target diffusion mode type and diffusion parameters, and the gradient information between sampling points is supplemented by the difference algorithm to obtain the target gradient vector field; The gradient descent algorithm is used to trace backward along the direction of maximum concentration gradient, and the convergence point of the target gradient vector field is determined through multiple iterations. The location of the convergence point is verified according to the preset algorithm to obtain the location of the nitrogen leak source.
[0007] Optionally, the step of performing a pre-defined purification process on the nitrogen concentration dataset collected by multiple sensor nodes within the vacuum chamber to obtain the target nitrogen concentration dataset includes: The nitrogen concentration dataset collected by multiple sensor nodes in the vacuum chamber is filtered and smoothed to obtain a smoothed nitrogen concentration dataset. Identify abnormal fluctuation signals in the smoothed nitrogen concentration dataset; The median filtering algorithm was used to replace outliers in the nitrogen concentration data corresponding to the abnormal fluctuation signals to obtain the purification dataset. The purification dataset was analyzed to identify abnormal sensors; The purification data corresponding to the abnormal sensors in the purification dataset are removed to obtain the target nitrogen concentration dataset.
[0008] Optionally, the step of extracting feature vectors from the target nitrogen concentration dataset after dimensionality reduction to obtain the target concentration feature vector includes: The target nitrogen concentration dataset is preprocessed to obtain a concentration data matrix; Determine the covariance matrix based on the concentration data matrix; The covariance matrix is subjected to eigenvalue decomposition to extract eigenvalues and corresponding eigenvectors; The concentration data matrix is linearly transformed using the feature vector to generate a dimension-reduced target concentration feature vector.
[0009] Optionally, after the step of performing a linear transformation on the concentration data matrix using the feature vector to generate a dimension-reduced target concentration feature vector, the method further includes: Determine whether the dimension of the target concentration feature vector is higher than a preset threshold; If so, sort the feature values, select the feature vectors corresponding to the first preset number of feature values in the sort, and construct the dimensionality-reduced concentration feature vector; The concentration change direction is analyzed on the dimensionality-reduced concentration feature vector to extract the target concentration features and generate the target concentration feature vector.
[0010] Optionally, the step of generating an initial extraction distribution matrix for sampling points based on the target concentration feature vector includes: The spatiotemporal variation pattern of nitrogen concentration is determined by analyzing the target concentration feature vector using time series analysis methods. When the spatiotemporal variation of nitrogen concentration exhibits a directional trend, the concentration gradient value between each sampling point is calculated; An initial gradient distribution matrix is generated based on the concentration gradient values between each sampling point.
[0011] Optionally, the step of determining the target diffusion mode type and diffusion parameters based on the initial gradient distribution matrix of the sampling points includes: The initial gradient distribution matrix of the sampling points is preprocessed to obtain the processed gradient distribution matrix; The processed gradient distribution matrix is subjected to pattern recognition using a support vector machine algorithm to obtain the first diffusion pattern category. If the first diffusion mode category matches the point source diffusion characteristics, the diffusion parameters are extracted; The diffusion parameters were analyzed to obtain their distribution patterns. The gradient characteristics of point source diffusion are verified based on the parameter distribution law. If the verification is successful, the first diffusion mode category is determined to be the target diffusion mode type.
[0012] Optionally, the step of verifying the convergence point location according to a preset algorithm to obtain the location of the nitrogen leak source includes: For the convergence point location, a verification method is used to calculate the confidence level of each tracking path, and the weight value of each path is obtained. Select a tracking path with a confidence level higher than the preset confidence level to generate the boundary coordinates of the convergence region; The weighted average algorithm is used to calculate the average position of each path weight to obtain the preliminary coordinates of the leakage source. The preliminary coordinates are then determined as the updated convergence point position, and the next round of coordinate position convergence is performed. The coordinate position was converged through multiple iterations to obtain the location of the nitrogen leak source.
[0013] This invention also provides a machine learning-based system for locating helium leaks in a vacuum chamber, wherein the system includes: The purification module is used to perform preset purification processing on the nitrogen concentration dataset collected by multiple sensor nodes in the vacuum chamber to obtain the target nitrogen concentration dataset. The feature extraction module is used to perform dimensionality reduction processing on the target nitrogen concentration dataset and then extract the feature vector to obtain the target concentration feature vector. The generation module is used to generate an initial gradient distribution matrix of sampling points based on the target concentration feature vector; The type determination module is used to determine the target diffusion mode type and diffusion parameters based on the initial gradient distribution matrix of the sampling points; The vector field determination module is used to construct a three-dimensional concentration gradient vector field based on the target diffusion mode type and diffusion parameters, and to supplement the gradient information between sampling points using an interpolation algorithm to obtain the target gradient vector field. The location determination module is used to perform reverse tracking along the direction of maximum concentration gradient using a gradient descent algorithm, and determine the convergence point location of the target gradient vector field through multiple iterations. The location verification module is used to verify the location of the convergence point according to a preset algorithm to obtain the location of the nitrogen leak source.
[0014] Optionally, the purification module includes: The first submodule is used to filter and smooth the nitrogen concentration dataset collected by multiple sensor nodes in the vacuum chamber to obtain a smoothed nitrogen concentration dataset. The second submodule is used to identify abnormal fluctuation signals in the smoothed nitrogen concentration dataset; The third submodule is used to replace outliers in the nitrogen concentration data corresponding to the abnormal fluctuation signal using a median filtering algorithm to obtain a purification dataset. The fourth submodule is used to analyze the purification dataset to identify abnormal sensors; The fifth submodule is used to remove the purification data corresponding to the abnormal sensors from the purification dataset to obtain the target nitrogen concentration dataset.
[0015] Optionally, the feature extraction module includes: The sixth submodule is used to preprocess the target nitrogen concentration dataset to obtain a concentration data matrix; The seventh submodule is used to determine the covariance matrix based on the concentration data matrix; The eighth submodule is used to perform eigenvalue decomposition on the covariance matrix and extract eigenvalues and corresponding eigenvectors; The ninth submodule is used to perform a linear transformation on the concentration data matrix using the feature vector to generate a dimension-reduced target concentration feature vector.
[0016] Optionally, the feature extraction module further includes: The judgment submodule is used to determine whether the dimension of the target concentration feature vector is higher than a preset threshold. If so, sort the feature values, select the feature vectors corresponding to the first preset number of feature values in the sort, and construct the dimensionality-reduced concentration feature vector; The feature vector generation submodule is used to analyze the concentration change direction of the dimensionality-reduced concentration feature vector and extract the target concentration features to generate the target concentration feature vector.
[0017] Optionally, the generation module is specifically used for: The spatiotemporal variation pattern of nitrogen concentration is determined by analyzing the target concentration feature vector using time series analysis methods. When the spatiotemporal variation of nitrogen concentration exhibits a directional trend, the concentration gradient value between each sampling point is calculated; An initial gradient distribution matrix is generated based on the concentration gradient values between each sampling point.
[0018] Optionally, the type determination module is specifically used for: The initial gradient distribution matrix of the sampling points is preprocessed to obtain the processed gradient distribution matrix; The processed gradient distribution matrix is subjected to pattern recognition using a support vector machine algorithm to obtain the first diffusion pattern category. If the first diffusion mode category matches the point source diffusion characteristics, the diffusion parameters are extracted; The diffusion parameters were analyzed to obtain their distribution patterns. The gradient characteristics of point source diffusion are verified based on the parameter distribution law. If the verification is successful, the first diffusion mode category is determined to be the target diffusion mode type.
[0019] Optionally, the location verification module is specifically used for: For the convergence point location, a verification method is used to calculate the confidence level of each tracking path, and the weight value of each path is obtained. Select a tracking path with a confidence level higher than the preset confidence level to generate the boundary coordinates of the convergence region; The weighted average algorithm is used to calculate the average position of each path weight to obtain the preliminary coordinates of the leakage source. The preliminary coordinates are then determined as the updated convergence point position, and the next round of coordinate position convergence is performed. The coordinate position was converged through multiple iterations to obtain the location of the nitrogen leak source.
[0020] This invention provides an electronic device, which includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor. When the program or instructions are executed by the processor, they implement the steps of any of the above-described machine learning-based vacuum chamber helium leak localization methods.
[0021] This invention provides a readable storage medium storing a program or instructions, which, when executed by a processor, implement the steps of any of the above-described guide methods for the blind.
[0022] The machine learning-based helium leak localization scheme for vacuum chambers provided in this invention involves: pre-purifying a nitrogen concentration dataset collected by multiple sensor nodes within the vacuum chamber to obtain a target nitrogen concentration dataset; extracting feature vectors from the target nitrogen concentration dataset after dimensionality reduction to obtain a target concentration feature vector; generating an initial gradient distribution matrix of sampling points based on the target concentration feature vector; determining the target diffusion mode type and diffusion parameters based on the initial gradient distribution matrix; constructing a three-dimensional concentration gradient vector field based on the target diffusion mode type and diffusion parameters, and supplementing the gradient information between sampling points using an interpolation algorithm to obtain the target gradient vector field; using a gradient descent algorithm to perform reverse tracing along the direction of maximum concentration gradient, iterating multiple times to determine the convergence point of the target gradient vector field; and verifying the convergence point position according to a pre-defined algorithm to obtain the location of the nitrogen leak source. This machine learning-based helium leak localization scheme for vacuum chambers effectively analyzes the features of multi-point helium concentration data using machine learning technology, accurately extracts the main directional and rate features in the diffusion mode, and constructs a reliable concentration gradient vector field, resulting in a more accurate and reliable determination of the nitrogen leak source location. Attached Figure Description
[0023] Figure 1 This is a flowchart illustrating the steps of a machine learning-based method for locating helium leaks in a vacuum chamber, according to an embodiment of this application. Figure 2 This is a structural block diagram illustrating a machine learning-based vacuum chamber helium leak location system according to an embodiment of this application; Figure 3 This is a structural block diagram illustrating an embodiment of an electronic device according to this application. Detailed Implementation
[0024] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.
[0025] The machine learning-based helium leak localization scheme for vacuum chambers provided in this application collects helium concentration data through multiple sensor nodes, performs filtering preprocessing and dimensionality reduction analysis on the data to extract the spatiotemporal variation patterns of concentration, calculates the concentration gradient distribution matrix, uses support vector machines to identify diffusion patterns, constructs a three-dimensional concentration gradient vector field, and uses a gradient descent algorithm for backtracking to ultimately determine the precise location of the leak source. This scheme can effectively handle helium concentration data in complex environments, and through multidimensional data analysis and machine learning algorithms, achieves rapid and accurate location of the leak source, thereby improving the safety and reliability of the vacuum system.
[0026] The following description, in conjunction with the accompanying drawings, details the machine learning-based helium leak location scheme for vacuum chambers provided in this application through specific embodiments and application scenarios.
[0027] As attached Figure 1 As shown, the machine learning-based method for locating helium leaks in a vacuum chamber according to an embodiment of this application includes the following steps: Step 101: Perform a preset purification process on the nitrogen concentration dataset collected by multiple sensor nodes in the vacuum chamber to obtain the target nitrogen concentration dataset.
[0028] The machine learning-based method for locating helium leaks in a vacuum chamber provided in this application can be applied to electronic devices. The electronic devices store corresponding computer programs, and when the computer programs are run, they execute the machine learning-based method for locating helium leaks in a vacuum chamber.
[0029] In one optional embodiment, the method of performing a preset purification process on the nitrogen concentration dataset collected by multiple sensor nodes inside the vacuum chamber to obtain the target nitrogen concentration dataset may include the following sub-steps: Sub-step 1: Filter and smooth the nitrogen concentration dataset collected by multiple sensor nodes in the vacuum chamber to obtain a smoothed nitrogen concentration dataset; Sub-step 2: Identify anomalous fluctuation signals in the smoothed nitrogen concentration dataset; In actual implementation, the difference between adjacent data points in the smoothed nitrogen concentration dataset can be calculated. If the difference exceeds a preset threshold, it is judged as an abnormal fluctuation signal, and the identified abnormal fluctuation signal is marked.
[0030] Sub-step 3: Use the median filtering algorithm to replace outliers in the nitrogen concentration data corresponding to the abnormal fluctuation signals to obtain the purification dataset; Sub-step 4: Analyze the cleanup dataset to identify abnormal sensors; In practical implementation, the mean and variance of helium concentration for each sensor node can be calculated based on the purification dataset to obtain the concentration distribution characteristics. These characteristics are then used to determine the data deviation of the sensor nodes. For each sensor node (also referred to as a sensor), if its corresponding deviation exceeds a preset deviation threshold, it is marked as an abnormal sensor.
[0031] Sub-step 5: Remove the purification data corresponding to abnormal sensors from the purification dataset to obtain the target nitrogen concentration dataset.
[0032] Assume there are 5 sensor nodes deployed at different locations inside the vacuum chamber. Each sensor node collects data every 5 minutes, resulting in a set of raw data. For example, the concentration values collected at a certain time point might be 12.5 ppm, 13.2 ppm, 18.7 ppm, 12.8 ppm, and 13.0 ppm. For this data, a preprocessing module first performs filtering. The system automatically performs preliminary analysis, calculating the mean and standard deviation. The mean is 14.04 ppm, and the standard deviation is 2.52. To determine if there are any abnormal fluctuations, a threshold is set at the mean plus or minus twice the standard deviation, i.e., 9.0 ppm to 19.08 ppm. The value of 18.7 ppm is found to be close to the upper limit, suggesting a possible abnormal fluctuation. The system automatically triggers a median filtering algorithm to remove noise. With a window size of 5, the data from the current data point and the two time points before and after it (assuming the first two points are 12.6 ppm and 12.9 ppm, and the last two points are 13.1 ppm and 12.7 ppm) are sorted. The median value of 12.9 ppm is taken as the purified value of the current point, replacing the original value of 18.7 ppm. The final purified dataset is: 12.5 ppm, 13.2 ppm, 12.9 ppm, 12.8 ppm, and 13.0 ppm.
[0033] By processing the raw data collected by the sensor nodes to mitigate abnormal fluctuations, the impact of these fluctuations can be effectively reduced, ensuring data smoothness. Further analysis of the purified dataset revealed a mean of 12.88 ppm and a standard deviation reduced to 0.26, indicating significantly improved data stability and providing a reliable foundation for subsequent helium concentration distribution analysis.
[0034] In a preferred embodiment, the processed data can be correlated with the airflow distribution model inside the vacuum chamber to deduce that the concentration anomaly may be related to local airflow disturbance. The system can automatically generate a report to prompt optimization of sensor layout or adjustment of environmental parameters inside the vacuum chamber.
[0035] Step 102: After dimensionality reduction of the target nitrogen concentration dataset, feature vector extraction is performed to obtain the target concentration feature vector.
[0036] In an optional embodiment, the method of extracting feature vectors after dimensionality reduction of the target nitrogen concentration dataset to obtain the target concentration feature vector may include the following sub-steps: Sub-step 1: Preprocess the target nitrogen concentration dataset to obtain a concentration data matrix; Sub-step 2: Determine the covariance matrix based on the concentration data matrix; Sub-step 3: Perform eigenvalue decomposition on the covariance matrix to extract eigenvalues and corresponding eigenvectors; Sub-step 4: Perform a linear transformation on the concentration data matrix using the feature vector to generate a dimension-reduced target concentration feature vector.
[0037] After performing a linear transformation on the concentration data matrix using the feature vectors to generate a dimension-reduced target concentration feature vector, the following process can be executed: Determine whether the dimension of the target concentration feature vector is higher than a preset threshold; if so, sort the feature values, select the feature vectors corresponding to the preset number of feature values at the top of the sort, and construct the dimensionality-reduced concentration feature vector; analyze the concentration change direction of the dimensionality-reduced concentration feature vector, extract the target concentration features, and generate the target concentration feature vector.
[0038] The preset threshold can be set by those skilled in the art according to actual needs, and no specific restrictions are imposed on it in this embodiment. The dimensionality reduction process of the target concentration feature vector is repeated until the dimension of the obtained target concentration feature vector is lower than or equal to the preset threshold.
[0039] In practical implementation, principal component analysis (PCA) can be used to reduce the dimensionality of the purified helium concentration dataset (i.e., the target nitrogen concentration dataset). For example, suppose there is a helium concentration dataset containing 5 samples, each with 3 features, and the data matrix is X=[[1.2, 2.1, 3.0], [1.5, 2.3, 3.2], [1.8, 2.5, 3.4], [2.0, 2.7, 3.6], [2.3, 2.9, 3.8]]. First, the data is standardized using information technology, and the mean and standard deviation of each dimension are calculated. For example, the mean of the first dimension is 1.76, and the standard deviation is 0.42, resulting in the standardized matrix Z. Next, the covariance matrix C is calculated using a linear algebra library (such as NumPy), yielding C = [[1.0, 0.98, 0.97], [0.98, 1.0, 0.99], [0.97, 0.99, 1.0]], reflecting the correlation between dimensions. Eigenvalues and eigenvectors of C are calculated using an eigenvalue decomposition algorithm. The eigenvalues are [2.94, 0.05, 0.01], and the corresponding eigenvectors are V1 = [0.58, 0.58, 0.58], V2 = [-0.71, 0.0, 0.71], and V3 = [-0.41, 0.82, -0.41]. Based on the magnitude of the eigenvalues, the first two principal components (with a cumulative contribution of approximately 98%) are selected, while the third principal component is ignored, thus achieving dimensionality reduction. Finally, the standardized data Z is projected onto the selected feature vectors V1 and V2 to obtain the dimensionality-reduced concentration feature vector, such as [2.1, 0.3] for the first sample. Through the above operations, the main direction of change in helium concentration data can be extracted, reducing dimensionality while retaining key information.
[0040] Step 103: Generate the initial gradient distribution matrix of the sampling points based on the target concentration feature vector.
[0041] An optional method for generating an initial extraction distribution matrix for sampling points based on the target concentration feature vector may include the following sub-steps: Sub-step 1: Analyze the target concentration feature vector using time series analysis to determine the spatiotemporal variation pattern of nitrogen concentration; Sub-step 2: When the spatiotemporal variation of nitrogen concentration shows a directional trend, calculate the concentration gradient value between each sampling point; In practical implementation, the concentration gradient values between each sampling point can be used to form a concentration gradient distribution dataset. Using the concentration gradient distribution dataset, an initial gradient distribution matrix can be constructed. The gradient distribution pattern between each sampling point can be determined through the initial gradient distribution matrix.
[0042] Sub-step 3: Generate an initial gradient distribution matrix based on the concentration gradient values between each sampling point.
[0043] This study analyzes the correlation between spatiotemporal variations and directional trends in helium distribution patterns, identifying key change patterns within the distribution. Using these key change patterns and time-series data, regression analysis is applied to predict future trends in helium concentration, yielding predicted change characteristics. Based on these predicted characteristics, the dynamic relationship between concentration characteristics and gradient distribution is analyzed to determine the existence of anomalous distribution areas. If anomalous areas are significant, relevant sampling points are marked to define the anomalous distribution range. Data on the anomalous distribution range is obtained, and combined with spatiotemporal variation patterns, a concentration characteristic correction model for the anomalous areas is constructed, resulting in a corrected concentration distribution dataset.
[0044] In practical implementation, to extract the spatiotemporal variation pattern of helium concentration feature vectors, the concentration data is first processed using time series analysis. Assuming the input data consists of daily helium concentration values (unit: ppm) at 20 sampling points within a certain region over 30 days, for example, sampling point A has a concentration of 5.2 on day 1, 5.3 on day 2, and so on, forming a 20×30 concentration matrix. An autoregressive moving average model is used for time series analysis, with the model parameters set to ARIMA(1,1,1). Non-stationarity is eliminated through differencing, and the time series stationarity of the concentration at each sampling point is calculated, yielding the trend and fluctuation terms for each point. For example, the trend analysis of sampling point A shows a daily increase of 0.02 ppm, indicating an upward trend. Next, the spatial distribution pattern is analyzed. Kriging interpolation is used to spatially interpolate the concentration values at the 20 sampling points, generating a 100×100 grid concentration map. The interpolation parameter is a spherical variogram with a range of 50 meters, resulting in a continuous concentration distribution map that reveals the spatial pattern of decreasing concentration from the center to the edge of the region. If the concentration change exhibits a directional trend, such as increasing along a northwest-southeast direction, the concentration gradient is calculated. Adjacent sampling point pairs (e.g., A and B, 10 meters apart, concentration difference 0.5 ppm) are selected, and the gradient value is calculated using the gradient formula (concentration difference / distance), yielding a gradient of 0.05 ppm / m from A to B. An initial gradient distribution matrix (10×10) is constructed by iterating through all adjacent point pairs, with matrix elements representing the gradient values between adjacent points and non-adjacent points set to 0. For example, matrix elements M(1,2)=0.05 and M(2,1)=-0.05, reflecting the directionality. Singular value decomposition (SVD) is performed on the matrix to extract the dominant gradient direction. The decomposition results show that the direction vector corresponding to the first singular value is [0.7, -0.7], confirming that northwest-southeast is the dominant direction. Finally, the spatiotemporal variation of helium concentration is quantified using the gradient matrix and spatial interpolation plot, providing data support for subsequent analysis.
[0045] Step 104: Determine the target diffusion mode type and diffusion parameters based on the initial gradient distribution matrix of the sampling points.
[0046] An optional method for determining the target diffusion mode type and diffusion parameters based on the initial gradient distribution matrix of the sampling points is as follows: The initial gradient distribution matrix of the sampling points is preprocessed to obtain a processed gradient distribution matrix; a support vector machine algorithm is used to perform pattern recognition on the processed gradient distribution matrix to obtain a first diffusion mode category; if the first diffusion mode category conforms to the point source diffusion characteristics, diffusion parameters are extracted; the diffusion parameters are analyzed to obtain the parameter distribution pattern; the gradient characteristics of point source diffusion are verified based on the parameter distribution pattern, and if the verification is successful, the first diffusion mode category is determined as the target diffusion mode type.
[0047] Step 105: Construct a three-dimensional concentration gradient vector field based on the target diffusion mode type and diffusion parameters, and use the difference algorithm to supplement the gradient information between sampling points to obtain the target gradient vector field.
[0048] In practical implementation, the finite difference method can be used to calculate the initial three-dimensional concentration distribution from the target diffusion mode and diffusion parameters to obtain the initial concentration field. Using the initial concentration field, the concentration gradient at each sampling point is calculated based on the finite difference method to obtain the gradient vector field. If there are gaps between sampling points in the gradient vector field, the Kriging interpolation algorithm is used to supplement the missing gradient information, generating a continuous concentration gradient vector field. The gradient direction and amplitude at each point are calculated based on the continuous concentration gradient vector field to determine the vector field distribution characteristics. A stereo algorithm is used to process the vector field distribution characteristics to generate a three-dimensional concentration gradient visualization distribution, obtaining a complete distribution image. If there are local anomalies in the complete distribution image, the anomaly region is smoothed using a mean filtering algorithm to obtain an optimized distribution image. Based on the optimized distribution image, the concentration gradient change trend is extracted to generate the target gradient vector field.
[0049] Step 106: Use the gradient descent algorithm to trace backward along the direction of maximum concentration gradient, and iterate multiple times to determine the convergence point of the target gradient vector field.
[0050] In the specific implementation process, concentration gradient data can be obtained for the target gradient vector field. By analyzing the vector distribution characteristics of the concentration gradient data, the first region with significant concentration gradient changes can be identified. Based on this region, a gradient descent algorithm is used to trace backward along the direction of maximum gradient, calculating the trend of gradient vector change. Iterative calculations are performed based on the gradient vector change trend to gradually approach the convergence point, obtaining preliminary convergence point location data. For the preliminary convergence point location data, local features of the surrounding vector distribution are obtained. If the local features are consistent with the direction of maximum concentration gradient, the location tracking range is further refined. Based on the refined location tracking range and the iterative calculation results, a set of candidate locations for the leakage source is determined. The matching degree between each location and the gradient vector field is analyzed using the candidate location set to determine the optimal leakage source coordinates. Based on the optimal leakage source coordinates and the global characteristics of the concentration gradient vector distribution, the coordinate determination result is verified, and the convergence point location of the target gradient vector field is obtained.
[0051] Step 107: Verify the convergence point location according to the preset algorithm to obtain the location of the nitrogen leak source.
[0052] In one optional embodiment, the method of verifying the convergence point location according to a preset algorithm to obtain the location of the nitrogen leak source may include the following sub-steps: Sub-step 1: For the convergence point location, use the verification method to calculate the confidence of each tracking path and obtain the weight value of each path; In practical implementation, a multi-path tracing method can be used to generate convergence point locations for multiple target gradient vector fields. For the convergence point locations of the generated target gradient vector fields, a verification method is used to calculate the confidence level of each tracing path, thereby obtaining the weight value of each path.
[0053] Sub-step 2: Select a tracking path with a confidence level higher than the preset confidence level, and generate the boundary coordinates of the convergence region; Sub-step 3: Calculate the average position of each path weight using the weighted average algorithm to obtain the preliminary coordinates of the leakage source. Determine the preliminary coordinates as the updated convergence point position and start the next round of coordinate position convergence. Sub-step 4: Iterate multiple times to converge the coordinate position and obtain the location of the nitrogen leak source.
[0054] Based on the tracking paths within the convergence region, a weighted average algorithm is used to calculate the average position of each path's weight, yielding the preliminary coordinates of the leak source. An iterative optimization algorithm refines these preliminary coordinates, adjusting them to minimize concentration distribution errors and determining the precise location coordinates. Helium concentration data for the precise location coordinates is acquired, and an anomaly detection algorithm is used to determine if any leak anomalies exist, resulting in the final leak source confirmation. Based on the final leak source confirmation, the location of the nitrogen leak source is determined.
[0055] The machine learning-based method for locating helium leaks in a vacuum chamber provided in this invention involves: pre-purifying a nitrogen concentration dataset collected by multiple sensor nodes within the vacuum chamber to obtain a target nitrogen concentration dataset; extracting feature vectors from the target nitrogen concentration dataset after dimensionality reduction to obtain a target concentration feature vector; generating an initial gradient distribution matrix of sampling points based on the target concentration feature vector; determining the target diffusion mode type and diffusion parameters based on the initial gradient distribution matrix; constructing a three-dimensional concentration gradient vector field based on the target diffusion mode type and diffusion parameters, and supplementing the gradient information between sampling points using an interpolation algorithm to obtain the target gradient vector field; using a gradient descent algorithm to perform reverse tracking along the direction of maximum concentration gradient, iterating multiple times to determine the convergence point of the target gradient vector field; and verifying the convergence point position according to a pre-defined algorithm to obtain the location of the nitrogen leak source. This machine learning-based method for locating helium leaks in a vacuum chamber effectively analyzes the features of multi-point helium concentration data using machine learning technology, accurately extracts the main directional and rate features in the diffusion mode, and constructs a reliable concentration gradient vector field, resulting in a more accurate and reliable determination of the nitrogen leak source location.
[0056] Figure 2 The structural block diagram of a machine learning-based vacuum chamber helium leak location system according to an embodiment of this application is shown.
[0057] The machine learning-based vacuum chamber helium leak location system provided in this application includes the following functional modules: Purification module 201 is used to perform preset purification processing on the nitrogen concentration dataset collected by multiple sensor nodes in the vacuum chamber to obtain the target nitrogen concentration dataset. Feature extraction module 202 is used to perform dimensionality reduction processing on the target nitrogen concentration dataset and then extract feature vectors to obtain target concentration feature vectors; The generation module 203 is used to generate an initial gradient distribution matrix of sampling points based on the target concentration feature vector; The type determination module 204 is used to determine the target diffusion mode type and diffusion parameters based on the initial gradient distribution matrix of the sampling points; The vector field determination module 205 is used to construct a three-dimensional concentration gradient vector field according to the target diffusion mode type and diffusion parameters, and to supplement the gradient information between sampling points using a difference algorithm to obtain the target gradient vector field. The position determination module 206 is used to perform reverse tracking along the direction of maximum concentration gradient using a gradient descent algorithm, and determine the convergence point position of the target gradient vector field through multiple iterations. The location verification module 207 is used to verify the location of the convergence point according to a preset algorithm to obtain the location of the nitrogen leak source.
[0058] Optionally, the purification module includes: The first submodule is used to filter and smooth the nitrogen concentration dataset collected by multiple sensor nodes in the vacuum chamber to obtain a smoothed nitrogen concentration dataset. The second submodule is used to identify abnormal fluctuation signals in the smoothed nitrogen concentration dataset; The third submodule is used to replace outliers in the nitrogen concentration data corresponding to the abnormal fluctuation signal using a median filtering algorithm to obtain a purification dataset. The fourth submodule is used to analyze the purification dataset to identify abnormal sensors; The fifth submodule is used to remove the purification data corresponding to the abnormal sensors from the purification dataset to obtain the target nitrogen concentration dataset.
[0059] Optionally, the feature extraction module includes: The sixth submodule is used to preprocess the target nitrogen concentration dataset to obtain a concentration data matrix; The seventh submodule is used to determine the covariance matrix based on the concentration data matrix; The eighth submodule is used to perform eigenvalue decomposition on the covariance matrix and extract eigenvalues and corresponding eigenvectors; The ninth submodule is used to perform a linear transformation on the concentration data matrix using the feature vector to generate a dimension-reduced target concentration feature vector.
[0060] Optionally, the feature extraction module further includes: The judgment submodule is used to determine whether the dimension of the target concentration feature vector is higher than a preset threshold. If so, sort the feature values, select the feature vectors corresponding to the first preset number of feature values in the sort, and construct the dimensionality-reduced concentration feature vector; The feature vector generation submodule is used to analyze the concentration change direction of the dimensionality-reduced concentration feature vector and extract the target concentration features to generate the target concentration feature vector.
[0061] Optionally, the generation module is specifically used for: The spatiotemporal variation pattern of nitrogen concentration is determined by analyzing the target concentration feature vector using time series analysis methods. When the spatiotemporal variation of nitrogen concentration exhibits a directional trend, the concentration gradient value between each sampling point is calculated; An initial gradient distribution matrix is generated based on the concentration gradient values between each sampling point.
[0062] Optionally, the type determination module is specifically used for: The initial gradient distribution matrix of the sampling points is preprocessed to obtain the processed gradient distribution matrix; The processed gradient distribution matrix is subjected to pattern recognition using a support vector machine algorithm to obtain the first diffusion pattern category. If the first diffusion mode category matches the point source diffusion characteristics, the diffusion parameters are extracted; The diffusion parameters were analyzed to obtain their distribution patterns. The gradient characteristics of point source diffusion are verified based on the parameter distribution law. If the verification is successful, the first diffusion mode category is determined to be the target diffusion mode type.
[0063] Optionally, the location verification module is specifically used for: For the convergence point location, a verification method is used to calculate the confidence level of each tracking path, and the weight value of each path is obtained. Select a tracking path with a confidence level higher than the preset confidence level to generate the boundary coordinates of the convergence region; The weighted average algorithm is used to calculate the average position of each path weight to obtain the preliminary coordinates of the leakage source. The preliminary coordinates are then determined as the updated convergence point position, and the next round of coordinate position convergence is performed. The coordinate position was converged through multiple iterations to obtain the location of the nitrogen leak source.
[0064] The machine learning-based vacuum chamber helium leak location system provided in this application performs a pre-processed purification of nitrogen concentration datasets collected by multiple sensor nodes within the vacuum chamber to obtain a target nitrogen concentration dataset. After dimensionality reduction, feature vectors are extracted from the target nitrogen concentration dataset to obtain a target concentration feature vector. An initial gradient distribution matrix of sampling points is generated based on the target concentration feature vector. The target diffusion mode type and diffusion parameters are determined based on the initial gradient distribution matrix. A three-dimensional concentration gradient vector field is constructed based on the target diffusion mode type and diffusion parameters, and a difference algorithm is used to supplement the gradient information between sampling points to obtain the target gradient vector field. A gradient descent algorithm is used to trace backward along the direction of maximum concentration gradient, iterating multiple times to determine the convergence point of the target gradient vector field. The convergence point is verified according to a pre-set algorithm to obtain the location of the nitrogen leak source. This machine learning-based vacuum chamber helium leak location system effectively analyzes the features of multi-point helium concentration data using machine learning technology, accurately extracts the main directional and rate features in the diffusion mode, and constructs a reliable concentration gradient vector field, resulting in a more accurate and reliable determination of the nitrogen leak source location.
[0065] In the embodiments of this application Figure 2 The machine learning-based vacuum chamber helium leak location system shown can be installed in a user-end electronic device or in a server. The electronic device or server equipped with this system can be a device with an operating system. This operating system can be Android, iOS, or other possible operating systems; this application does not specifically limit its implementation.
[0066] The embodiments provided in this application Figure 2 The machine learning-based vacuum chamber helium leak location system shown can achieve... Figure 1 The various processes implemented in the method implementation examples will not be described again here to avoid repetition.
[0067] Optionally, refer to Figure 3 The present application also provides an electronic device 300, including a processor 301, a memory 302, and a program or instructions stored in the memory and executable on the processor. When the program or instructions are executed by the processor, they implement the processes performed by the above-mentioned machine learning-based vacuum chamber helium leak location system and achieve the same technical effect. To avoid repetition, they will not be described again here.
[0068] It should be noted that the electronic device in this application embodiment includes the server described above.
[0069] The processor is the processor in the electronic device described in the above embodiments. The readable storage medium includes computer-readable storage media, such as computer read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.
[0070] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.
[0071] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A machine learning-based method for locating helium leaks in a vacuum chamber, characterized in that, include: The nitrogen concentration dataset collected by multiple sensor nodes inside the vacuum chamber is subjected to a pre-set purification process to obtain the target nitrogen concentration dataset. After dimensionality reduction processing of the target nitrogen concentration dataset, feature vector extraction is performed to obtain the target concentration feature vector. An initial gradient distribution matrix for sampling points is generated based on the target concentration feature vector; The target diffusion mode type and diffusion parameters are determined based on the initial gradient distribution matrix of the sampling points; A three-dimensional concentration gradient vector field is constructed based on the target diffusion mode type and diffusion parameters, and the gradient information between sampling points is supplemented by the difference algorithm to obtain the target gradient vector field; The gradient descent algorithm is used to trace backward along the direction of maximum concentration gradient, and the convergence point of the target gradient vector field is determined through multiple iterations. The location of the convergence point is verified according to the preset algorithm to obtain the location of the nitrogen leak source.
2. The method according to claim 1, characterized in that, The steps for performing a pre-defined purification process on the nitrogen concentration dataset collected by multiple sensor nodes within the vacuum chamber to obtain the target nitrogen concentration dataset include: The nitrogen concentration dataset collected by multiple sensor nodes in the vacuum chamber is filtered and smoothed to obtain a smoothed nitrogen concentration dataset. Identify abnormal fluctuation signals in the smoothed nitrogen concentration dataset; The median filtering algorithm was used to replace outliers in the nitrogen concentration data corresponding to the abnormal fluctuation signals to obtain the purification dataset. The purification dataset was analyzed to identify abnormal sensors; The purification data corresponding to the abnormal sensors in the purification dataset are removed to obtain the target nitrogen concentration dataset.
3. The method according to claim 1, characterized in that, The steps of performing dimensionality reduction processing on the target nitrogen concentration dataset and then extracting feature vectors to obtain the target concentration feature vectors include: The target nitrogen concentration dataset is preprocessed to obtain a concentration data matrix; Determine the covariance matrix based on the concentration data matrix; The covariance matrix is subjected to eigenvalue decomposition to extract eigenvalues and corresponding eigenvectors; The concentration data matrix is linearly transformed using the feature vector to generate a dimension-reduced target concentration feature vector.
4. The method according to claim 3, characterized in that, After the step of performing a linear transformation on the concentration data matrix using the feature vector to generate a dimensionality-reduced target concentration feature vector, the method further includes: Determine whether the dimension of the target concentration feature vector is higher than a preset threshold; If so, sort the feature values, select the feature vectors corresponding to the first preset number of feature values in the sort, and construct the dimensionality-reduced concentration feature vector; The concentration change direction is analyzed on the dimensionality-reduced concentration feature vector to extract the target concentration features and generate the target concentration feature vector.
5. The method according to claim 1, characterized in that, The step of generating an initial extraction distribution matrix for sampling points based on the target concentration feature vector includes: The spatiotemporal variation pattern of nitrogen concentration is determined by analyzing the target concentration feature vector using time series analysis methods. When the spatiotemporal variation of nitrogen concentration exhibits a directional trend, the concentration gradient value between each sampling point is calculated; An initial gradient distribution matrix is generated based on the concentration gradient values between each sampling point.
6. The method according to claim 1, characterized in that, The steps for determining the target diffusion mode type and diffusion parameters based on the initial gradient distribution matrix of the sampling points include: The initial gradient distribution matrix of the sampling points is preprocessed to obtain the processed gradient distribution matrix; The processed gradient distribution matrix is subjected to pattern recognition using a support vector machine algorithm to obtain the first diffusion pattern category. If the first diffusion mode category matches the point source diffusion characteristics, the diffusion parameters are extracted; The diffusion parameters were analyzed to obtain their distribution patterns. The gradient characteristics of point source diffusion are verified based on the parameter distribution law. If the verification is successful, the first diffusion mode category is determined to be the target diffusion mode type.
7. The method according to claim 1, characterized in that, The step of verifying the convergence point location according to a preset algorithm to obtain the location of the nitrogen leak source includes: For the convergence point location, a verification method is used to calculate the confidence level of each tracking path, and the weight value of each path is obtained. Select a tracking path with a confidence level higher than the preset confidence level to generate the boundary coordinates of the convergence region; The weighted average algorithm is used to calculate the average position of each path weight to obtain the preliminary coordinates of the leakage source. The preliminary coordinates are then determined as the updated convergence point position, and the next round of coordinate position convergence is performed. The coordinate position was converged through multiple iterations to obtain the location of the nitrogen leak source.
8. A machine learning-based system for locating helium leaks in a vacuum chamber, characterized in that, include: The purification module is used to perform preset purification processing on the nitrogen concentration dataset collected by multiple sensor nodes in the vacuum chamber to obtain the target nitrogen concentration dataset. The feature extraction module is used to perform dimensionality reduction processing on the target nitrogen concentration dataset and then extract the feature vector to obtain the target concentration feature vector. The generation module is used to generate an initial gradient distribution matrix of sampling points based on the target concentration feature vector; The type determination module is used to determine the target diffusion mode type and diffusion parameters based on the initial gradient distribution matrix of the sampling points; The vector field determination module is used to construct a three-dimensional concentration gradient vector field based on the target diffusion mode type and diffusion parameters, and to supplement the gradient information between sampling points using an interpolation algorithm to obtain the target gradient vector field. The location determination module is used to perform reverse tracking along the direction of maximum concentration gradient using a gradient descent algorithm, and determine the convergence point location of the target gradient vector field through multiple iterations. The location verification module is used to verify the location of the convergence point according to a preset algorithm to obtain the location of the nitrogen leak source.
9. The system according to claim 8, characterized in that, The purification module includes: The first submodule is used to filter and smooth the nitrogen concentration dataset collected by multiple sensor nodes in the vacuum chamber to obtain a smoothed nitrogen concentration dataset. The second submodule is used to identify abnormal fluctuation signals in the smoothed nitrogen concentration dataset; The third submodule is used to replace outliers in the nitrogen concentration data corresponding to the abnormal fluctuation signal using a median filtering algorithm to obtain a purification dataset. The fourth submodule is used to analyze the purification dataset to identify abnormal sensors; The fifth submodule is used to remove the purification data corresponding to the abnormal sensors from the purification dataset to obtain the target nitrogen concentration dataset.
10. An electronic device, characterized in that, The electronic device includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions are executed by the processor to perform the steps of any one of the machine learning-based vacuum chamber helium leak location methods according to claims 1-7.