A data analysis-based steel cylinder production quality traceability method and system

Abstract

This application discloses a data analysis-based method and system for tracing the production quality of steel cylinders. The method includes: applying an alternating magnetic field to the steel cylinder; rotating and scanning the steel cylinder to obtain magnetic response data of the weld area; determining the magnetic flux leakage signal of the steel cylinder and obtaining high-dimensional magnetic fingerprint features from the magnetic flux leakage signal; acquiring high-dimensional magnetic fingerprint features under different operating conditions; determining a structural stability feature subset and a state-sensitive feature subset based on the high-dimensional magnetic fingerprint features under different operating conditions; generating a magnetic fingerprint relationship template by combining the structural stability feature subset; constructing a corresponding healthy baseline manifold based on the magnetic fingerprint relationship template; determining the steel cylinder quality residual vector by combining the state-sensitive feature subset and the healthy baseline manifold; determining the steel cylinder quality degradation type and degree by using the steel cylinder quality residual vector; and generating a production quality traceability report by combining the steel cylinder quality degradation type and degree. This application can effectively improve the accuracy of steel cylinder production quality traceability.

Description

Technical Field

[0001] This application relates to the field of gas cylinder production quality traceability, and in particular to a gas cylinder production quality traceability method and system based on data analysis. Background Technology

[0002] With the rapid development of large-scale industrial production and networked logistics, steel cylinders, as the core carrier for the storage and transfer of high-pressure gases, are directly related to the production safety of key fields such as chemical, energy, and medical industries in terms of production quality and the reliability of full life-cycle traceability. They play an irreplaceable role in ensuring the stability of the industrial chain and supply chain and reducing the incidence of safety accidents.

[0003] Currently, the traceability of gas cylinder production quality mainly relies on external markings on the cylinder's surface. These markings, such as stamped serial numbers, QR codes, or RFID tags, link the cylinder's production batch, manufacturing parameters, and inspection information to the external marking. Subsequent processes involve reading these markings to complete information retrieval and traceability. However, this external marking-based traceability method is essentially a "separable" identity binding mechanism, its traceability capability entirely dependent on the integrity and readability of the marking itself. During the long-term service life of the cylinder, external markings are susceptible to damage or loss due to environmental factors such as wear, corrosion, high temperatures, and impacts. They may also be distorted due to transfer, replacement, or duplication. Once the external marking fails, even if the cylinder itself remains, its production quality information is difficult to accurately recover using existing technologies, thus disrupting the traceability chain and affecting the accuracy of gas cylinder production quality traceability.

[0004] There is currently no good solution to the above problems. Summary of the Invention

[0005] This application provides a data analysis-based method and system for tracing the quality of steel cylinder production, which improves the accuracy of tracing the quality of steel cylinder production.

[0006] To achieve the above objectives, the embodiments of this application adopt the following technical solutions: Firstly, a data analysis-based method for tracing the quality of steel cylinder production is provided, which includes: After the target cylinder is welded, a uniform alternating magnetic field is applied to the target cylinder. The target cylinder is rotated and scanned using a pre-set rotating device to obtain the magnetic response data of the weld area of ​​the target cylinder. The magnetic flux leakage signal of the target cylinder is determined based on the magnetic response data of the weld area, and the magnetic flux leakage signal is converted into a high-dimensional magnetic fingerprint feature of the target cylinder in a high-dimensional feature space. Acquire high-dimensional magnetic fingerprint features under different working conditions; Stability analysis was performed based on the high-dimensional magnetic fingerprint characteristics under different working conditions to determine the structural stability feature subset and state-sensitive feature subset of the target gas cylinder; A magnetic fingerprint relationship template for the weld seam region of the target gas cylinder is generated by combining a subset of structurally stable features in a high-dimensional feature space. The magnetic fingerprint relationship template is used to determine the intrinsic identity information of the target gas cylinder. Construct the corresponding health baseline manifold based on the magnetic fingerprint relationship template; Orthogonally project the state-sensitive feature subset and the healthy baseline manifold to determine the cylinder quality residual vector; The type and degree of quality degradation of the target cylinder are determined by the cylinder quality residual vector; A production quality traceability report for the target steel cylinder is generated by combining the type and degree of quality degradation of the cylinder.

[0007] In one possible implementation of the first aspect, determining the magnetic flux leakage signal of the target cylinder based on the magnetic response data of the weld region, and converting the magnetic flux leakage signal into a high-dimensional magnetic fingerprint feature of the target cylinder in a high-dimensional feature space, includes: The magnetic response data of the weld area is preprocessed to obtain the preprocessed magnetic response data of the weld area; The physical properties of the weld area of ​​the target steel cylinder are distinguished to determine the physical property zoning structure of the weld area of ​​the target steel cylinder; Based on the physical characteristics of the partition structure, the magnetic response data of the pre-processed weld area is partitioned and differentiated to obtain the magnetic flux leakage signal. Based on the physical characteristics of the partition structure, feature extraction is performed on the magnetic flux leakage signal to obtain a set of magnetic flux leakage feature vectors; The set of magnetic flux leakage feature vectors is mapped to a high-dimensional feature space using a preset nonlinear embedding mapping function, and the set of magnetic flux leakage feature vectors in the high-dimensional feature space is subjected to cluster analysis to determine the high-dimensional magnetic fingerprint features of the target cylinder.

[0008] In one possible implementation of the first aspect, the step of performing partitioned differential matching on the preprocessed weld region magnetic response data according to the physical characteristic partitioning structure to obtain the magnetic flux leakage signal includes: Based on the physical property partitioning structure of the target gas cylinder weld area, the preprocessed magnetic response data of the weld area is divided into multiple sub-signal segments, where each sub-signal segment corresponds to a physical property partitioning structure of the weld. Analyze the noise characteristics of each physical property partition structure; Based on the noise characteristics of the physical partition structure, differentiated noise reduction processing is applied to the sub-signal segments. The sub-signal segments after differential noise reduction are physically partitioned and sequentially spliced ​​to obtain the reconstructed magnetic response signal, which reflects the microscopic physical structure of the weld. Magnetic flux leakage signals are determined based on the microscopic physical structure of the weld.

[0009] In one possible implementation of the first aspect, the step of determining the structural stability feature subset and state-sensitive feature subset of the target gas cylinder based on the stability analysis of high-dimensional magnetic fingerprint features under different operating conditions includes: Each high-dimensional magnetic fingerprint feature is labeled with its operating condition to obtain the corresponding operating condition label for each high-dimensional magnetic fingerprint feature; Calculate the mutual information between each high-dimensional magnetic fingerprint feature and its corresponding operating condition label. The mutual information is used to characterize the degree of correlation between the high-dimensional magnetic fingerprint feature and the operating condition changes. Calculate the inter-class divergence of high-dimensional magnetic fingerprint features under the same working conditions. The inter-class divergence is used to characterize the distinguishing ability of high-dimensional magnetic fingerprint features between different working condition categories. The state sensitivity score is obtained by weighted fusion of mutual information and inter-class divergence. Based on the state sensitivity score, the high-dimensional magnetic fingerprint features are divided to determine the structural stability feature subset and the state sensitivity feature subset of the target gas cylinder.

[0010] In one possible implementation of the first aspect, generating a magnetic fingerprint relationship template for the target gas cylinder weld region by combining a subset of structurally stable features in a high-dimensional feature space includes: In a high-dimensional feature space, a high-dimensional reference point cloud is determined by a structurally stable subset of features. A topological evolution analysis is performed on a high-dimensional reference point cloud to determine the topological features of the high-dimensional feature point cloud, and a persistent image of the weld structure is generated based on the topological features. The persistent image of the weld structure is used as a magnetic fingerprint relationship template for the weld area of ​​the target gas cylinder.

[0011] In one possible implementation of the first aspect, the step of performing topological evolution analysis on the high-dimensional reference point cloud to determine the topological features of the high-dimensional feature point cloud, and generating a persistent weld structure map based on the topological features, includes: Obtain the distance relationships between high-dimensional reference point clouds and high-dimensional feature spaces; The geometric structure scaling function of the high-dimensional reference point cloud is determined based on the distance relationship between the high-dimensional reference point cloud and the high-dimensional feature space. The geometric structure scale function is discretized according to the preset scale growth rule to generate a scale parameter sequence; Sort the scale parameters in the scale parameter sequence in ascending order; For any scale parameter in a scale parameter sequence, construct the corresponding simple complex topology based on the high-dimensional reference point cloud; Based on the scale parameters arranged in ascending order, the topological evolution of simple complex topologies is analyzed to determine the generation and disappearance behaviors of simple complex topologies. The evolution of the topology of high-dimensional feature point clouds under different scale parameters is determined by using simple complex topology. Within a preset time window, stable topological features existing in multiple scale parameters are identified based on the topological evolution results. The generation and disappearance behaviors corresponding to each stable topological feature are mapped to generate a persistent graph of the weld structure.

[0012] In one possible implementation of the first aspect, determining the geometric structure scaling function of the high-dimensional reference point cloud based on the distance relationship between the high-dimensional reference point cloud and the high-dimensional feature space includes: Preprocess the high-dimensional reference point cloud; Construct an Euclidean distance matrix based on the distance relationships of the preprocessed high-dimensional reference point cloud; For any scale parameter, count the number of neighbor points in the high-dimensional reference point cloud with a preset distance radius, and calculate the local average number of neighbors to characterize the local connectivity of the point cloud. For any scale parameter, calculate the global discreteness of the high-dimensional reference point cloud; The geometric structure scale function is generated by combining the local average number of neighbors and the global dispersion for each scale parameter.

[0013] In one possible implementation of the first aspect, constructing the corresponding health baseline manifold based on the magnetic fingerprint relationship template includes: Anchored topological features are extracted from the magnetic fingerprint relationship template of the target gas cylinder to obtain a set of topological constraints; By establishing a mapping relationship between the topological constraint set and the state-sensitive feature subset, the health feature subset corresponding to the topological constraint set is obtained; The health feature subset is dimensionality reduced using a manifold learning algorithm, and the high-dimensional features are mapped to a low-dimensional space to obtain a low-dimensional feature point cloud using a set of topological constraints as a regularization term. A low-dimensional smooth manifold is obtained by fitting the low-dimensional feature point cloud; A low-dimensional smooth manifold is used as the healthy baseline manifold.

[0014] Secondly, this application provides a machine-readable storage medium storing instructions that cause a machine to execute the aforementioned data analysis-based gas cylinder production quality traceability method.

[0015] Thirdly, this application provides an electronic device, comprising: The memory is configured to store instructions; and The processor is configured to retrieve the instructions from the memory and, when executing the instructions, to implement the aforementioned data analysis-based method for tracing the quality of steel cylinder production.

[0016] The above technical solution achieves several improvements. First, by applying a uniform alternating magnetic field to the welded target cylinder, the weld area is placed in a repeatable magnetic excitation environment. This excites differences in magnetic response corresponding to microstructure and material inhomogeneities within the weld, making signals from minute defects easier to detect and improving detection sensitivity. A rotating scanning method achieves full circumferential coverage of the magnetic response in the weld area, improving the integrity of the magnetic response data. Converting the original magnetic response signal into a magnetic flux leakage signal highly sensitive to the weld's microstructure allows for better differentiation between different types of defects and normal states, improving detection accuracy and reliability. Based on high-dimensional magnetic fingerprint features under different operating conditions, the internal structural changes of the cylinder under various life-cycle operating conditions can be clearly identified, facilitating more accurate identification and differentiation between normal changes and quality defects under different conditions. After determining the structural stability feature subset and the state-sensitive feature subset, the internal identity and quality status of the cylinder can be clearly defined, aiding in the accurate identification of the cylinder's intrinsic identity and quality status. By combining a subset of structurally stable features to generate a magnetic fingerprint relationship template, the intrinsic identity information of the gas cylinder can be determined, enabling unique identification and improving traceability efficiency. Constructing a health baseline manifold allows for a direct assessment of the cylinder's health status, enabling timely detection of potential quality issues. Orthogonal projection of the state-sensitive feature subset and the health baseline manifold allows for a quantitative comparison between the state-sensitive features and the health baseline, generating a quality residual vector that provides a basis for quantitative analysis of quality degradation. Finally, analyzing the quality residual vector allows for timely detection and assessment of quality degradation, facilitating preventative maintenance measures and preventing safety accidents during cylinder use. Furthermore, generating a production quality traceability report helps quickly trace the root cause of quality problems, facilitating targeted improvement measures and enhancing production quality and management levels.

[0017] Other features and advantages of the embodiments of this application will be described in detail in the following detailed description section. Attached Figure Description

[0018] Figure 1 A flowchart illustrating a data analysis-based method for tracing the quality of steel cylinder production, provided as an embodiment of this application; Figure 2 This is a flowchart illustrating a method for determining the high-dimensional magnetic fingerprint features of a target gas cylinder, as provided in an embodiment of this application. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only for illustration and explanation of the embodiments of this application and are not intended to limit the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0020] It should be noted that if the embodiments of this application involve directional indicators (such as up, down, left, right, front, back, etc.), the directional indicators are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicators will also change accordingly.

[0021] Furthermore, if the embodiments of this application involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed in this application.

[0022] Figure 1 The illustration shows a flowchart of a data analysis-based method for tracing the quality of steel cylinder production according to an embodiment of this application. Figure 1 As shown in the figure, this application provides a method for tracing the quality of steel cylinder production based on data analysis. The method may include the following steps.

[0023] S101. After the target cylinder is welded, apply a uniform alternating magnetic field to the target cylinder. S102. Use a pre-set rotating device to perform a rotational scan on the target cylinder to obtain the magnetic response data of the weld area of ​​the target cylinder. S103. Determine the magnetic flux leakage signal of the target cylinder based on the magnetic response data of the weld area, and convert the magnetic flux leakage signal into a high-dimensional magnetic fingerprint feature of the target cylinder in a high-dimensional feature space. S104. Obtain high-dimensional magnetic fingerprint features under different working conditions; S105. Based on the high-dimensional magnetic fingerprint characteristics under different working conditions, perform stability analysis to determine the structural stability feature subset and state-sensitive feature subset of the target gas cylinder; S106. Generate a magnetic fingerprint relationship template for the weld seam region of the target steel cylinder by combining the structurally stable feature subset in the high-dimensional feature space. The magnetic fingerprint relationship template is used to determine the intrinsic identity information of the target steel cylinder. S107. Construct the corresponding health baseline manifold based on the magnetic fingerprint relationship template; S108. Perform orthogonal projection on the state-sensitive feature subset and the healthy baseline manifold to determine the cylinder quality residual vector; S109. Determine the type and degree of quality degradation of the target cylinder by using the cylinder quality residual vector; S110. Generate a production quality traceability report for the target cylinder by combining the type and degree of cylinder quality degradation.

[0024] In this embodiment, after the target cylinder completes the welding process, a uniform alternating magnetic field is applied to the target cylinder. A pre-set excitation device can be used to magnetically excite the target cylinder, so that the target cylinder as a whole is in an alternating magnetization state. The pre-set excitation device can be set on the outside of the target cylinder and arranged along the circumference of the target cylinder. By applying the uniform alternating magnetic field, the weld area and its adjacent heat-affected zone will produce magnetic response changes under alternating magnetization conditions, thereby stimulating the discontinuity of magnetic flux distribution in the weld area caused by welding process and other factors, and providing stable magnetic excitation conditions for obtaining magnetic response characteristics.

[0025] Subsequently, the target cylinder is rotated using a pre-set rotating device located at its bottom. Under a uniform alternating magnetic field, the rotating device drives the cylinder to rotate around its own axis. Simultaneously, a magnetic detection unit positioned near the weld area synchronously acquires the magnetic response signals generated during the rotation, thus obtaining continuously distributed magnetic response data along the circumference of the weld. This magnetic response data characterizes the changes in magnetic properties of the weld area of ​​the target cylinder under alternating magnetic excitation conditions. The magnetic response data includes, but is not limited to, instantaneous values ​​and peak values ​​of magnetic induction intensity, and changes in magnetic flux.

[0026] From the magnetic response data of the weld area, abnormal or abrupt magnetic field signals, i.e., magnetic flux leakage signals, are identified. First, the magnetic response data undergoes preprocessing such as filtering, normalization, and noise suppression. Then, within a preset window period, the amplitude standard deviation of the magnetic response data is calculated. Signal segments with amplitude standard deviations exceeding a preset amplitude threshold are identified as amplitude anomalies. Next, gradient changes are calculated from the magnetic response data, which can be achieved using first-order differencing. Gradient abrupt changes exceeding a preset gradient threshold are identified as potential anomaly signals. Finally, signal segments exhibiting both amplitude anomalies and gradient abrupt changes are marked as magnetic flux leakage signals, forming a continuous magnetic flux leakage signal sequence along the weld circumference. The preset amplitude threshold is determined by collecting magnetic response data of the weld area under the same detection conditions and performing statistical analysis on the amplitude of the magnetic response data to obtain the standard deviation statistical parameter of the amplitude distribution. Based on the standard deviation statistical parameter, the normal fluctuation range of the magnetic response amplitude is determined. The upper and lower limits of the normal fluctuation range are used as the upper and lower limits of the preset amplitude threshold. The preset gradient threshold is calculated by statistically analyzing the normal gradient values, sorting all normal gradient values, and selecting a specific percentile as the threshold, such as 95%. The specific percentile can be set according to the company's accuracy requirements. After obtaining the magnetic flux leakage signal sequence continuously distributed along the weld circumference, the magnetic flux leakage signal is divided according to the equal-angle segmentation rule of the weld circumference position. That is, the weld circumference is divided into several equal-angle intervals, and the magnetic flux leakage signal in each interval is a signal segment. A corresponding basic feature representation is constructed for each magnetic flux leakage signal segment. The basic feature representation reflects at least the amplitude change characteristics and gradient change characteristics of the magnetic flux leakage signal segment. The segmented magnetic flux leakage signal is transformed into a high-dimensional space using kernel functions. Existing radial basis functions can be used to transform the magnetic flux leakage signal into a high-dimensional feature space. This mapped high-dimensional feature space consists of multiple kernel-mapped response feature components, and its feature dimension is determined by the number of segments in the magnetic flux leakage signal and the number of kernel-mapped response feature components corresponding to each segment, rather than being a fixed dimension. The magnetic flux leakage signal mapped to the high-dimensional feature space is used as the high-dimensional magnetic fingerprint feature of the target gas cylinder. This high-dimensional magnetic fingerprint feature is used to characterize the target gas cylinder, perform quality traceability, and assess the consistency of weld structure.

[0027] Collecting high-dimensional magnetic fingerprint features under different operating conditions refers to acquiring magnetic response data of the weld area during the production process of the target gas cylinder at different stages or production environments to obtain corresponding high-dimensional magnetic fingerprint features. Different operating conditions include variations in environmental conditions and states of the gas cylinder during different production stages such as welding, cooling, inspection, storage, and transportation, including factors such as temperature, stress distribution, vibration, and magnetic field influences. In each process environment, a pre-set magnetic excitation device and a rotating scanning device are used to detect the weld area, obtaining magnetic flux leakage signals, thereby forming high-dimensional magnetic fingerprint features corresponding to each process or environment.

[0028] Each acquired high-dimensional magnetic fingerprint feature is labeled with a working condition, thus obtaining a corresponding working condition label for each feature. Then, the mutual information between each high-dimensional magnetic fingerprint feature and its corresponding working condition label is calculated. This mutual information characterizes the correlation between the high-dimensional magnetic fingerprint feature's response to changes in working conditions, i.e., it measures the sensitivity of the feature under different working conditions. Further, the inter-class divergence of the high-dimensional magnetic fingerprint features under the same working condition is calculated. This inter-class divergence characterizes the distinguishing ability of the high-dimensional magnetic fingerprint feature among different working condition categories, i.e., it measures the feature's ability to maintain consistency under different working conditions. Then, the mutual information and inter-class divergence are weighted and fused according to preset weights to obtain a state sensitivity score for each high-dimensional magnetic fingerprint feature. Finally, the high-dimensional magnetic fingerprint features are divided according to their state sensitivity scores. Features with low state sensitivity and stable performance under different working conditions are identified as a structurally stable feature subset, while features with high state sensitivity and significant response to changes in working conditions are identified as a state-sensitive feature subset.

[0029] Based on the structurally stable feature subset selected in the high-dimensional feature space, a high-dimensional reference point cloud of the target gas cylinder weld region is first constructed, where each high-dimensional reference point corresponds to a one-dimensional feature vector in the structurally stable feature subset. Subsequently, topological evolution analysis is performed on the high-dimensional reference point cloud to obtain its topological features and generate a persistent weld structure map. Specifically, the distance relationships of the high-dimensional reference point cloud in the high-dimensional feature space are obtained, and the geometric scale function of the high-dimensional reference point cloud is determined based on these distance relationships. The geometric scale function is discretized according to a preset scale growth rule to generate a scale parameter sequence, which is then arranged in ascending order. For any scale parameter, a corresponding simple complex topology is constructed based on the high-dimensional reference point cloud. Based on the ascending scale parameters, topological evolution analysis is performed on the simple complex topology to determine its generation and disappearance behaviors, and stable topological features existing under multiple scale parameters are identified within a preset time window. Furthermore, the generation and disappearance behaviors corresponding to each stable topological feature are mapped to generate a persistent weld structure graph, which is then used as a magnetic fingerprint relationship template for the weld region of the target gas cylinder. This magnetic fingerprint relationship template characterizes the intrinsic physical structural properties of the weld region of the target gas cylinder and is used to determine the intrinsic identity information of the target gas cylinder, thereby achieving unique identification and reliable traceability of the gas cylinder.

[0030] Next, anchored topological features are extracted from the magnetic fingerprint relationship template of the target gas cylinder to form a topological constraint set, which is used to constrain the key stable topological features of the weld structure. These anchored topological features are extracted from the previously constructed persistent graph of the weld structure and include, but are not limited to, highly persistent topological features, such as features that persist across multiple scale parameters in the persistent graph of the weld structure; and generation and disappearance behavior features, such as the time windows or behavior patterns of specific topological points during scale parameter changes. Subsequently, a mapping relationship is established between the topological constraint set and the state-sensitive feature subset to obtain the corresponding healthy feature subset, which is used to characterize the key feature distribution of the gas cylinder weld area under normal conditions. Then, the healthy feature subset is subjected to dimensionality reduction using a manifold learning algorithm, which can be an isometric mapping algorithm. In the healthy feature subset, each healthy feature sample point is used as a node, and its K nearest neighbor sample points are selected to construct a nearest neighbor graph structure. The neighborhood parameter K is set according to the size of the healthy feature samples, and those skilled in the art can adaptively adjust it according to the sample size. Finally, the Euclidean distance between adjacent sample points is used as the edge weight to form a weighted adjacency graph corresponding to the healthy feature subset. Furthermore, in the weighted adjacency graph, the connection path with the minimum total weight between any feature sample points is determined by progressively accumulating edge weights, thereby obtaining the approximate geodesic distance relationship between the feature sample points. Based on the approximate geodesic distance relationship, the health feature subset is subjected to low-dimensional embedding processing. This means using the approximate geodesic distance between feature sample points as a constraint condition to redetermine the position of each feature sample point in the low-dimensional space, so that the distance relationship between any feature sample points in the low-dimensional space can reflect its approximate geodesic distance relationship in the high-dimensional space as much as possible, thereby maintaining the overall structural distribution characteristics of the health feature subset during the dimensionality reduction process. The target dimension for dimensionality reduction can be set according to the inherent structural complexity of the weld region health features, preferably 2-dimensional or 3-dimensional (for example only), to reduce the computational complexity of low-dimensional feature modeling and subsequent health assessment processes while maintaining the overall geometric relationship of the weld region health features. Next, using the topological constraint set as a regularization term, the high-dimensional features are mapped to a low-dimensional space to obtain a low-dimensional feature point cloud. Subsequently, the low-dimensional feature point cloud is fitted to generate a low-dimensional smooth manifold, which can characterize the health baseline state of the weld area of ​​the target cylinder. Finally, the low-dimensional smooth manifold is used as the health baseline manifold of the target cylinder for subsequent cylinder condition monitoring, health assessment, and abnormal feature comparison.

[0031] The state-sensitive feature subset reflects the set of features that may undergo minute changes during the production or use of the gas cylinder weld. In this embodiment, the healthy baseline manifold is a point set composed of a low-dimensional feature point cloud and its smoothed fitting results, representing the feature distribution of the gas cylinder weld in a healthy and normal state. The healthy baseline manifold is not limited to an analytical function form, but exists as a point set composed of discrete feature points and their local continuous structures. Orthogonal projection refers to mapping the state-sensitive feature vector along a direction perpendicular to the local tangent space of the healthy baseline manifold to the healthy baseline manifold, minimizing the distance between the state-sensitive feature vector and its projection point on the healthy baseline manifold. By using orthogonal projection, the influence of feature changes along the direction of the healthy baseline manifold on the deviation calculation can be avoided, thereby accurately quantifying the degree of abnormal deviation of the target gas cylinder weld area relative to the healthy state. The formula for calculating the orthogonal projection point is as follows: Where x is the state-sensitive eigenvector; p is the projection reference point of the healthy baseline manifold at a local location; U is a set of orthogonal basis vectors of the local tangent space of the healthy baseline manifold at the reference point p, and each column of it is an orthogonal basis vector. The orthogonal basis vectors are the reference vectors extracted from the healthy sample data using principal component analysis. T X is the transpose of U; X is the orthogonal projection point of the final calculated state-sensitive feature vector onto the healthy baseline manifold. The subset of state-sensitive features is projected onto the space of the healthy baseline manifold. The projected point represents the position of the feature under ideal health conditions. The deviation between the projection and the actual feature is the residual vector, which equals the actual state vector minus the point projected onto the healthy baseline manifold. This residual vector is used to quantify the deviation of the cylinder weld from the healthy state. In other words, each state-sensitive feature vector is projected onto the nearest point found in the healthy baseline manifold. The nearest point can be found using minimum Euclidean distance or other distance metrics, and the projection direction is perpendicular to the tangent space of the manifold to ensure orthogonal projection. Next, the difference vector between each state-sensitive feature vector and its projection point on the healthy baseline manifold is calculated. This difference vector is the cylinder's quality residual vector. By analyzing the magnitude and direction of the quality residual vector, the deviation of the target cylinder weld area from the healthy state can be quantified.

[0032] Subsequently, the residual vector is mapped to a predefined degradation type category to determine the quality degradation type of the target cylinder; that is, the residual vector is compared with the predefined degradation type category to determine the quality degradation type. The predefined degradation type category is constructed during experimental calibration by collecting the corresponding magnetic fingerprint features and residual vectors for different types of typical degradation states of the target cylinder's weld, forming typical residual vector samples for each degradation state, thus building the degradation type category. Quality degradation types include, but are not limited to, weld microcrack degradation, material aging degradation, weld pit or defect degradation, and localized hardening degradation.

[0033] Subsequently, the type and degree of quality degradation are correlated with the production information of the target cylinder, including the production batch, welding process parameters, material information, and production time. A production quality traceability report is generated based on this correlation information. This report records the quality status and degradation characteristics of the target cylinder during production and use. The quality status and degradation characteristics are then correlated with the database in the system, and compared with the unique identifier information in the database. This allows for traceability to specific production batches, production lines, process parameters, and other production source information. The database in the system contains cylinder production information, and at least stores the unique identifier information of the cylinder along with corresponding production batch information, production line identifier information, and key process parameter range information. This method enables systematic management and traceable monitoring of the quality during the production process of the target cylinder, thereby ensuring the accuracy of cylinder production quality.

[0034] By applying a uniform alternating magnetic field to the target cylinder after welding, the weld area is placed in a repeatable magnetic excitation environment. This excites differences in magnetic response corresponding to microstructure and material inhomogeneities within the weld, making signals from minute defects easier to detect and improving detection sensitivity. A rotating scanning method achieves full circumferential coverage of the magnetic response in the weld area, improving the integrity of the magnetic response data. Converting the raw magnetic response signal into a magnetic flux leakage signal highly sensitive to the weld's microstructure allows for better differentiation between different types of defects and normal states, improving detection accuracy and reliability. Based on high-dimensional magnetic fingerprint features under different operating conditions, the internal structural changes of the cylinder under various life-cycle operating conditions can be clearly identified, facilitating more accurate identification and differentiation between normal changes and quality defects under different operating conditions. After determining the structural stability feature subset and the state-sensitive feature subset, the internal identity and quality status of the cylinder can be clearly defined, aiding in the accurate identification of the cylinder's intrinsic identity and quality status. By combining a subset of structurally stable features to generate a magnetic fingerprint relationship template, the intrinsic identity information of the gas cylinder can be determined, enabling unique identification and improving traceability efficiency. Constructing a health baseline manifold allows for a direct assessment of the cylinder's health status, enabling timely detection of potential quality issues. Orthogonal projection of the state-sensitive feature subset and the health baseline manifold allows for a quantitative comparison between the state-sensitive features and the health baseline, generating a quality residual vector that provides a basis for quantitative analysis of quality degradation. Finally, analyzing the quality residual vector allows for timely detection and assessment of quality degradation, facilitating preventative maintenance measures and preventing safety accidents during cylinder use. Furthermore, generating a production quality traceability report helps quickly trace the root cause of quality problems, facilitating targeted improvement measures and enhancing production quality and management levels.

[0035] Figure 2This schematically illustrates a flowchart of a method for determining the high-dimensional magnetic fingerprint features of a target gas cylinder according to an embodiment of this application. Figure 2 As shown, in one embodiment of this example, the magnetic flux leakage signal of the target cylinder is determined based on the magnetic response data of the weld area, and the magnetic flux leakage signal is converted into a high-dimensional magnetic fingerprint feature of the target cylinder in a high-dimensional feature space, including: S210. Preprocess the magnetic response data of the weld area to obtain the preprocessed magnetic response data of the weld area. S220. Determine the physical property zoning structure of the weld area of ​​the target cylinder by differentiating its physical properties. S230. Based on the physical characteristics of the partition structure, the magnetic response data of the pre-processed weld area is partitioned and differentiated to obtain the magnetic flux leakage signal. S240. Based on the physical characteristic partitioning structure, feature extraction is performed on the magnetic flux leakage signal to obtain a set of magnetic flux leakage feature vectors. S250. The set of magnetic flux leakage feature vectors is mapped to a high-dimensional feature space using a preset nonlinear embedding mapping function, and the set of magnetic flux leakage feature vectors in the high-dimensional feature space is subjected to cluster analysis to determine the high-dimensional magnetic fingerprint features of the target cylinder.

[0036] The magnetic response data acquired from the weld area of ​​the target gas cylinder undergoes preprocessing. This preprocessing includes denoising, baseline correction, and amplitude normalization of the raw magnetic response data to reduce interference from environmental magnetic fields and sensor noise. Simultaneously, based on the location and distribution characteristics of the weld on the cylinder surface, the magnetic response data is truncated and spatially aligned to remove magnetic response information from non-weld areas. Through these preprocessing steps, the obtained magnetic response data from the weld area can stably reflect the inherent magnetic properties of the weld, improving the accuracy of prediction.

[0037] In this embodiment, the physical property zoning structure refers to dividing the gas cylinder into multiple detection areas based on the differences in physical properties of different parts of the cylinder. Each area corresponds to a unique set of physical property parameters. During the welding process, the high-temperature heating and cooling of the gas cylinder leads to significant differences in the physical properties of the weld metal zone, heat-affected zone, and base metal zone. The weld metal zone refers to the weld area formed after the filler metal is melted during welding, i.e., the actual volume of weld metal deposited. The heat-affected zone refers to the area where the base metal is locally heated but not melted due to the heat conducted by the weld pool during welding. The base metal zone refers to the original gas cylinder metal area outside the weld and heat-affected zone, unaffected by welding heat. The surface of the target gas cylinder is divided into a grid to form a two-dimensional sampling grid, with each grid corresponding to a unique spatial coordinate. A magnetically strongly correlated sensor, such as a Hall sensor, is placed around the target gas cylinder to detect the relative permeability of each two-dimensional sampling grid. A pre-set image acquisition device, such as a microhardness tester, is used to collect the Vickers hardness value and the grain diameter of the weld metal zone for each two-dimensional sampling grid. After normalizing the above parameters, a weighted fusion is performed to obtain the corresponding comprehensive evaluation value. The region with the highest comprehensive evaluation value for Vickers hardness, weld metal grain diameter, and relative permeability is identified as the weld metal region. When the comprehensive evaluation value of a certain physical property region is the lowest, and its physical property parameters exhibit a continuous and stable spatial distribution, that physical property region is identified as the base metal region. When the comprehensive evaluation value of a certain physical property region lies between the comprehensive evaluation values ​​corresponding to the weld metal region and the base metal region, that physical property region is identified as the heat-affected zone.

[0038] Subsequently, the preprocessed magnetic response data of the weld area is matched to physical characteristic partitions based on spatial location. This means mapping the preprocessed magnetic response data of the weld area to corresponding physical characteristic partitions according to their spatial location, forming multiple subsets of magnetic response data for each partition. Each subset of magnetic response data has a specific magnetic response matching benchmark, used to characterize the magnetic response characteristics of that partition under normal conditions. The baseline calculation method for the weld metal zone is to take the average magnetic induction intensity of all defect-free data points within the zone as the magnetic response matching benchmark. The mean of the magnetic response data for each partition of the heat-affected zone is directly calculated and used as the magnetic response baseline for the heat-affected zone. The magnetic response baseline of the base metal zone is used to characterize the magnetic response characteristics of the base metal zone under normal conditions, and the data is normalized. Then, the average value of the normalized magnetic response data is directly calculated, and a simple smoothing process can be performed on the average value to form the magnetic response baseline for the base metal zone. Next, within the same physical characteristic partition, the current magnetic response data is compared with the matching benchmark corresponding to that partition. If a value exceeds the magnetic response matching benchmark, that portion of the magnetic response data is marked as an abnormal magnetic response. The magnetic response data marked as abnormal within the partition are filtered and enhanced after aggregation to obtain the magnetic flux leakage signal characterizing the local magnetic anomaly of the weld.

[0039] Based on the physical characteristics and partitioning structure of the weld area, features of the magnetic flux leakage signal in each partition are extracted to obtain a set of magnetic flux leakage feature vectors. The magnetic flux leakage signal is then partitioned into weld metal region, heat-affected zone, and base material region, forming multiple partitioned magnetic flux leakage signal subsets. Subsequently, multi-dimensional features such as amplitude, gradient, frequency domain, and spatial distribution are extracted from each partitioned signal subset, and these features are combined to form a partitioned magnetic flux leakage feature vector. Finally, the feature vectors of all partitions are integrated to obtain a set of magnetic flux leakage feature vectors, which are used for subsequent high-dimensional magnetic fingerprint feature construction and target cylinder status analysis.

[0040] A nonlinear embedding mapping function can be used to map feature vectors to a high-dimensional space, amplifying previously indistinguishable minute differences. The preset nonlinear embedding mapping function can be a Gaussian kernel function, as is used in existing technologies. The set of magnetic flux leakage feature vectors is transformed into a high-dimensional space using the Gaussian kernel function. Subsequently, cluster analysis is performed on the high-dimensional feature vector set, and K-means can be used to determine the cluster structure and distribution centers of the high-dimensional feature vectors. Finally, the statistical characteristics of the cluster centers or clusters are used as the high-dimensional magnetic fingerprint features of the target gas cylinder to characterize the overall magnetic response mode of the gas cylinder's weld area. The high-dimensional magnetic fingerprint features are used to characterize the microstructural differences in the weld area of ​​the target gas cylinder and serve as the inherent identity information of the gas cylinder.

[0041] By generating high-dimensional magnetic fingerprint features that can uniquely characterize the differences in microstructure and overall magnetic response mode of the weld area of ​​the target gas cylinder, it is possible not only to accurately reflect the differences in magnetic properties of the weld area, but also to effectively improve the accuracy of gas cylinder detection and contribute to the accuracy of gas cylinder quality traceability.

[0042] In one embodiment of this invention, the magnetic response data of the preprocessed weld region is partitioned and differentiated according to a physical characteristic partitioning structure to obtain a magnetic flux leakage signal, including: S310. Based on the physical characteristic partitioning structure of the target gas cylinder weld area, the preprocessed magnetic response data of the weld area is divided into multiple sub-signal segments, where each sub-signal segment corresponds to a physical characteristic partitioning structure of the weld. S320. Analyze the noise characteristics of each physical characteristic partition structure; S330. Based on the noise characteristics of the physical partition structure, implement partitioned differential noise reduction processing for sub-signal segments; S340. The sub-signal segments after differential noise reduction are physically partitioned and sequentially spliced ​​to obtain the reconstructed magnetic response signal, so as to reflect the micro-physical structure of the weld. S350, determining magnetic flux leakage signals based on the microscopic physical structure of the weld.

[0043] The spatial location and extent information of the weld metal zone, heat-affected zone, and base metal zone within the weld area are obtained. Subsequently, the preprocessed magnetic response data sequence is segmented according to the location and extent of each zone to form sub-signal segments corresponding to each zone. Each sub-signal segment records the zone type and the index information of the data in the original magnetic response sequence, so as to facilitate subsequent differential analysis and magnetic flux leakage feature extraction of each zone, thereby achieving independent characterization and accurate analysis of the microstructural features of each zone in the weld area.

[0044] This analysis examines the magnetic response signal fluctuations caused by non-defect factors in the weld metal zone, heat-affected zone, and base metal zone. These noises are not due to equipment malfunctions, but rather normal signal fluctuations caused by the inherent physical characteristics of each zone or the testing conditions. The noise characteristics of different zones differ significantly, requiring zone-specific analysis for accurate response. Specifically, the noise in the weld metal zone originates from the microstructural inhomogeneity during the solidification process of the weld pool. The cast grains formed during the cooling process exhibit orientation differences and localized component segregation, leading to slight fluctuations in magnetic permeability. The noise is uniformly distributed throughout the zone without spatial aggregation. In the heat-affected zone, where melting has not occurred, the noise primarily stems from incomplete microstructural transformation caused by local temperature gradients, the alternating distribution of coarsened and refined grains, and the spatially uneven release of residual welding stress. The noise characteristics of the heat-affected zone exhibit moderate amplitude fluctuations, spatially continuous segmented undulations, a banded distribution along the weld direction, and significantly higher noise levels in some local areas compared to adjacent areas, but overall remain within the normal range. Noise in the base material area mainly originates from texture orientation formed during the rolling or forging process of raw materials, environmental magnetic field disturbances, and slight deviations in the detection posture. The noise in the base material area is characterized by a generally small noise amplitude, slow signal change, no obvious peaks, and spatially manifests as smooth, long-wave undulations without localized aggregation or abrupt changes. Because different physical characteristics of the zoning structures result in different noise characteristics, targeted noise reduction processing is necessary to improve the accuracy of quality traceability.

[0045] Next, differentiated noise reduction processing is implemented for the sub-signal segments corresponding to the magnetic response signal. Specifically, the noise characteristics in the weld metal sub-signal segment are mainly manifested as local abrupt changes and isolated anomalies. Therefore, in the weld metal sub-signal segment, the local average value is calculated, and it is determined whether a single sampling point deviates significantly from its adjacent points. For isolated points that deviate significantly, the median value of two adjacent points is used for replacement, thereby achieving noise reduction in the weld metal area. The noise in the heat-affected zone (HAZ) is characterized by continuous but non-real structural fluctuations. To reduce noise in the HAZ, the sub-signal segments of the HAZ can be smoothed, with the sliding window length set to be greater than that of the weld area and less than that of the base material area, preserving the overall trend of the signal change to eliminate irregular fluctuations, thus achieving noise reduction in the HAZ. The noise in the base material area is characterized by slow changes and low-frequency drift. To reduce noise in the base material area, the sub-signal segments of the base material area can be averaged with a large window to extract the background change amount, and the background amount is subtracted from the original signal to obtain an approximately stable base material response.

[0046] After completing the partitioned differential noise reduction processing, the sub-signal segments are spliced ​​together according to the actual physical order of the weld on the gas cylinder. The splicing order is determined based on the physical partitions of the weld and the scanning path, such as base material area, heat-affected zone, weld metal area, heat-affected zone, and base material area. Slight smoothing is applied at the boundaries of each partition to eliminate signal abrupt changes that may be caused by partitioning, thus obtaining a reconstructed signal that is both low-noise and retains the microscopic physical structure characteristics of the weld. Finally, the magnetic flux leakage signal is determined based on the microscopic physical structure of the weld. That is, the reconstructed signal obtained after preprocessing and partitioned differential noise reduction is the magnetic flux leakage signal. This signal reflects the microscopic structural characteristics of each partition of the weld. The magnetic flux leakage signal refers to the local magnetic flux leakage generated on the surface due to the microscopic discontinuities or material structure differences within the weld when a magnetic field is applied to the weld area. It can be used for weld defect detection, microstructure characterization, high-dimensional magnetic fingerprint feature extraction, and weld quality traceability and identification.

[0047] By using a physical property partitioning structure to perform refined noise reduction on the magnetic response data of the weld area, the detection accuracy of magnetic flux leakage signals can be effectively improved, providing a more accurate basis for weld quality inspection, helping to detect potential weld defects in a timely manner, and ensuring the safety and reliability of welded structures such as gas cylinders.

[0048] In one embodiment of this invention, stability analysis is performed based on the high-dimensional magnetic fingerprint characteristics under different operating conditions to determine the structural stability feature subset and state-sensitive feature subset of the target gas cylinder, including: S410. Mark each high-dimensional magnetic fingerprint feature with a working condition label to obtain the working condition label corresponding to each high-dimensional magnetic fingerprint feature; S420. Calculate the mutual information between each high-dimensional magnetic fingerprint feature and the corresponding operating condition label. The mutual information is used to characterize the degree of correlation between the high-dimensional magnetic fingerprint feature and the operating condition change. S430. Calculate the inter-class divergence of high-dimensional magnetic fingerprint features under the same working conditions. The inter-class divergence is used to characterize the ability of high-dimensional magnetic fingerprint features to distinguish between different working condition categories. S440. The state sensitivity score is obtained by weighted fusion of mutual information and inter-class divergence. S450. Based on the state sensitivity score, the high-dimensional magnetic fingerprint features are divided to determine the structural stability feature subset and the state sensitivity feature subset of the target gas cylinder.

[0049] Each high-dimensional magnetic fingerprint feature is labeled with a corresponding working condition tag. In this embodiment, the working condition tag includes, but is not limited to, parameters such as welding current, welding speed, heating temperature, residual stress state, welding position, and welding sequence. During the magnetic response signal acquisition process, various parameters of the weld under different working conditions are recorded and associated with the extracted high-dimensional magnetic fingerprint feature vector to obtain the working condition tag corresponding to each high-dimensional magnetic fingerprint feature. The working condition tag can be used for the classification and comparison of high-dimensional magnetic fingerprint features under different working conditions, abnormal weld identification, weld position, or welding sequence, etc.

[0050] Next, mutual information is calculated for each high-dimensional magnetic fingerprint feature and its corresponding working condition label to characterize the degree of correlation between the high-dimensional magnetic fingerprint feature and changes in weld working conditions. Specifically, the data of each high-dimensional feature and its corresponding working condition label are discretized. This can be achieved using existing techniques such as equal-width binning and equal-frequency binning to map continuous values ​​to discrete categories. Then, the joint occurrence frequency of each feature category and working condition category is counted. The joint probability distribution is equal to the ratio of the feature category to the total number of samples. The marginal probability distributions of the feature category and the working condition category are then calculated using existing marginal probability distribution formulas. Finally, the mutual information value between each feature and the working condition label is calculated according to the mutual information formula, thus quantitatively reflecting the sensitivity of the high-dimensional magnetic fingerprint feature to changes in working conditions.

[0051] Inter-class divergence is calculated for high-dimensional magnetic fingerprint features under the same operating conditions to characterize the discriminative ability of each high-dimensional magnetic fingerprint feature among different operating condition categories. Specifically, each high-dimensional feature is grouped according to operating condition categories, and the mean of the feature in each operating condition category and the overall mean of all samples are calculated. Then, the weighted squared difference between the mean of each operating condition category and the overall mean is calculated, and the sum of the squared weighted differences of all operating condition categories is used as the inter-class divergence value of the feature. By calculating the inter-class divergence, the degree of difference of each high-dimensional magnetic fingerprint feature among different operating condition categories can be quantitatively reflected. The larger the inter-class divergence value, the stronger the discriminative ability of the feature among different operating condition categories. For example, taking welding current as an example, the samples corresponding to the high-dimensional feature f1 are divided into low current group, medium current group, and high current group. The low current group contains feature values ​​of 0.12, 0.15, and 0.10; the medium current group contains feature values ​​of 0.35, 0.32, and 0.36; and the high current group contains feature values ​​of 0.60, 0.62, and 0.58. Subsequently, the feature mean of each working condition group and the overall mean of all samples are calculated. Then, the weighted squared difference between the mean of each working condition group and the overall mean is calculated, and the weighted squared differences of all working condition groups are summed to obtain the inter-class divergence value of the feature.

[0052] After obtaining the mutual information and inter-class divergence, a mutual information value is calculated for each high-dimensional feature to characterize its response strength to changes in working conditions; an inter-class divergence value is calculated to characterize its discriminative ability among different working condition categories. Subsequently, the mutual information and inter-class divergence values ​​are normalized to the same dimension and weighted and summed according to corresponding preset weight coefficients to obtain the state sensitivity score for each high-dimensional magnetic fingerprint feature. The state sensitivity score is a comprehensive index obtained after quantitatively evaluating each high-dimensional magnetic fingerprint feature, used to characterize the feature's response sensitivity and discriminative ability under changes in weld working conditions. Response sensitivity refers to the strength of the feature's response to changes in working conditions, i.e., the degree to which the feature value changes with the working condition category, reflecting the feature's sensitivity to changes in weld state; discriminative ability refers to the degree of difference between features in different working condition categories, i.e., whether the feature can effectively distinguish different working condition categories, reflecting the feature's discriminative ability. The higher the state sensitivity score, the stronger the sensitivity of the feature to changes in working conditions. The preset mutual information weight coefficient and preset inter-class divergence weight coefficient can be set based on the relative importance of mutual information and inter-class divergence in feature evaluation. For example, if mutual information is more critical in reflecting the sensitivity of a feature to changes in operating conditions during actual weld inspection, the preset mutual information weight coefficient can be set to a larger value to enhance its influence, while the preset inter-class divergence weight coefficient can be set to a smaller value to reduce its influence. The relative importance of mutual information and inter-class divergence in feature evaluation can be determined based on the Pierre correlation coefficient, and the specific values ​​can be set according to the company's production requirements.

[0053] Based on the state sensitivity score, the high-dimensional magnetic fingerprint features are divided into a structural stability feature subset and a state sensitivity feature subset to determine the target gas cylinder. The structural stability feature subset reflects the inherent microstructural characteristics of the weld, while the state sensitivity feature subset reflects the service status or environmental changes of the weld. Specifically, after calculating the state sensitivity score for each high-dimensional magnetic fingerprint feature, a preset state sensitivity score threshold is established. High-dimensional features with state sensitivity scores below this threshold are classified into the structural stability feature subset to characterize the stable physical structural characteristics of the target gas cylinder weld. High-dimensional features with state sensitivity scores above this threshold are classified into the state sensitivity feature subset to characterize the response characteristics of the target gas cylinder weld under different operating conditions. In this embodiment, the preset state sensitivity score threshold can be set using the mean method, calculating the overall mean of the state sensitivity scores of all high-dimensional magnetic fingerprint features and using this mean as the preset threshold. By determining the state sensitivity feature subset and the structural stability feature subset, it can be used for weld condition monitoring, anomaly identification, and quality traceability.

[0054] By identifying the structural stability feature subset and the state-sensitive feature subset of the target gas cylinder, it is possible to effectively identify the feature subset that has little impact on the structural stability of the gas cylinder and the feature subset that is highly sensitive to changes in operating conditions. This provides more targeted and accurate feature basis for the condition monitoring and fault diagnosis of the gas cylinder, helps to provide early warning of potential safety hazards, and ensures the safe operation of the gas cylinder.

[0055] In one embodiment of this invention, a magnetic fingerprint relationship template for the target gas cylinder weld region is generated by combining a subset of structurally stable features in a high-dimensional feature space, including: S510. Determine the high-dimensional reference point cloud in the high-dimensional feature space by using structurally stable feature subsets; S520. Perform topological evolution analysis on the high-dimensional reference point cloud, determine the topological features of the high-dimensional feature point cloud, and generate a persistent image of the weld structure based on the topological features. S530. Use the persistent weld structure diagram as a magnetic fingerprint relationship template for the weld area of ​​the target gas cylinder.

[0056] In a high-dimensional feature space, a high-dimensional reference point cloud is determined by using a subset of structurally stable features. Specifically, for each sample point corresponding to the previously mapped subset of structurally stable features in the high-dimensional feature space, only the structurally stable feature dimension is retained. The mapping results of each sample point in this dimensional space are then aggregated to determine the high-dimensional reference point cloud. The high-dimensional reference point cloud can accurately reflect the feature distribution of the target object in a stable state and provide a reliable basis for subsequent structural state identification, anomaly detection, and source tracing.

[0057] A topological evolution analysis is performed on a high-dimensional reference point cloud to determine its topological features, and a persistent weld structure map is generated based on these features. First, high-dimensional magnetic fingerprint feature vectors from all healthy steel cylinders are collected to form a reference point cloud dataset. This dataset undergoes standardization preprocessing, such as Z-score normalization. Second, the distance metric between any two points in the preprocessed reference point cloud dataset is calculated. Next, an incremental set of scale parameters is introduced, representing the neighborhood coverage between sample points. At each scale parameter, the distance between any two sample points in the reference point cloud is assessed. If the distance between two points is less than or equal to the current scale parameter, the two sample points are considered to be adjacent at that scale parameter, and a point cloud neighborhood structure is constructed based on this adjacency relationship. As the scale parameter gradually increases, the adjacency relationship between sample points in the point cloud gradually expands from local connectivity to global connectivity, thus forming a multi-scale neighborhood structure sequence reflecting the structural evolution characteristics of the high-dimensional reference point cloud. In a multi-scale neighborhood structure sequence, the evolution of topological elements is tracked and analyzed, including at least the generation and merging of connected components in the point cloud, and the generation and filling of ring structures in the point cloud. The generation scale is recorded at the scale parameter where a topological element first appears; the disappearance scale is recorded at the scale parameter where the topological element disappears or is covered by a higher-order structure. That is, during multi-scale analysis of the point cloud, different structures (such as a connected region or a ring structure) will first appear at a certain scale and eventually disappear or be merged as the scale continues to increase. The scale value corresponding to the initial appearance of the structure is taken as the generation scale, and the scale value corresponding to the disappearance of the structure is taken as the disappearance scale. Recording these two scale values ​​together yields a scale pair describing the interval in which the structure exists. The set formed by summing all the scale pairs corresponding to all structures is the persistent graph of the weld structure. Structures with a large scale range indicate that they exist consistently over a wide scale range and possess strong stability, typically corresponding to consistent and reliable structural characteristics within the weld. Conversely, structures that appear only briefly within a very small scale range indicate poor stability and are often caused by local disturbances, detection noise, or minute random variations. The above steps can be implemented in MATLAB software.

[0058] After preprocessing, feature extraction, and high-dimensional mapping, the magnetic response signal of the target gas cylinder weld area can generate a high-dimensional feature point cloud in a high-dimensional feature space. The topological evolution of the high-dimensional feature point cloud is described by a persistent weld structure map to reflect the stable topological features of the target weld. Subsequently, when performing magnetic response detection on a new gas cylinder weld, its corresponding high-dimensional feature point cloud can be compared with the persistent weld structure map template. By analyzing the consistency or deviation in the topological structure of the two, the structural state of the weld can be identified as normal, thereby realizing the detection of weld anomalies, quality evaluation, and source tracing.

[0059] By determining the magnetic fingerprint relationship template of the target gas cylinder weld area, the inherent structural relationship of the weld area in the high-dimensional feature space can be effectively captured. This provides a stable and reliable reference template for long-term monitoring and condition assessment of the weld, which helps to improve the accuracy and efficiency of weld detection, reduce the false judgment rate, and ensure the safety and reliability of the gas cylinder weld.

[0060] In one embodiment of this invention, a topological evolution analysis is performed on a high-dimensional reference point cloud to determine the topological features of the high-dimensional feature point cloud, and a persistent weld structure map is generated based on the topological features, including: S610. Obtain the distance relationship between the high-dimensional reference point cloud and the high-dimensional feature space; S620. Construct the geometric structure scaling function of the high-dimensional reference point cloud based on the distance relationship between the high-dimensional reference point cloud and the high-dimensional feature space. S630. Discretize the geometric structure scale function according to the preset scale growth rule to generate a scale parameter sequence; S640. Sort the scale parameters in the scale parameter sequence in ascending order; S650. For any scale parameter in a scale parameter sequence, construct the corresponding simple complex topology based on a high-dimensional reference point cloud. S660. Based on the scale parameters after ascending order, perform topological evolution analysis on the simple complex topology to determine the generation and disappearance behavior of the simple complex topology. S670. Determine the topological evolution of high-dimensional feature point clouds under different scale parameters by using simple complex topological structures. S680. Within a preset time window, identify stable topological features existing in multiple scale parameters based on the topological evolution results. S690. Map the generation behavior and disappearance behavior corresponding to each stable topological feature to generate a persistent graph of the weld structure.

[0061] The distance relationships of a high-dimensional reference point cloud in a high-dimensional feature space are obtained. The high-dimensional reference point cloud consists of sample points of the target object in the high-dimensional feature space, with each sample point corresponding to a high-dimensional vector of a structurally stable feature subset, where each dimension represents a structurally stable feature of the target object. Subsequently, based on the Euclidean distance method, the distance between any two sample points in the high-dimensional reference point cloud is calculated to obtain the spatial relationships between points in the point cloud. By calculating pairwise distances between all sample points in the reference point cloud, a distance matrix of the high-dimensional reference point cloud is formed, which comprehensively reflects the relative positional relationships of sample points in the high-dimensional feature space and the similarity between point clouds. These distance relationships can be used to subsequently construct geometric structure scaling functions, generate simplex topological structures and persistent maps of weld structures, improving the accuracy of high-dimensional topological feature recognition of the weld region of the target object.

[0062] A geometric scaling function for the high-dimensional reference point cloud is constructed based on the distance relationships between the high-dimensional reference point cloud and the high-dimensional feature space. Specifically, the high-dimensional reference point cloud undergoes preprocessing, including feature normalization. Then, based on the distance relationships between sample points in the preprocessed high-dimensional reference point cloud, an Euclidean distance matrix is ​​constructed to characterize the spatial distance between any two sample points in the high-dimensional feature space. Further, for any scale parameter, based on the Euclidean distance matrix and using the scale parameter as a preset distance radius, the number of neighbor points contained within this distance radius for each sample point in the high-dimensional reference point cloud is counted. The neighbor counts of all sample points are then summarized to obtain the local average neighbor count for the corresponding scale parameter, which characterizes the local connectivity of the high-dimensional reference point cloud at that scale. Simultaneously, for the same scale parameter, the global dispersion of the high-dimensional reference point cloud is calculated based on the Euclidean distance matrix to reflect the overall looseness of the point cloud's distribution in the high-dimensional feature space. Finally, the local average neighbor count obtained under each scale parameter is fused with the global dispersion to generate a geometric structure scaling function that varies with the scale parameter. The geometric structure scaling function can comprehensively reflect the local connectivity characteristics and global structural features of the high-dimensional reference point cloud at different scales.

[0063] After obtaining the geometric structure scaling function of the high-dimensional reference point cloud, the geometric structure scaling function is discretized according to a preset scaling rule to generate a sequence of scaling parameters. Initial values ​​of the scaling parameters are determined based on the distance distribution characteristics of the high-dimensional reference point cloud, and a pre-set scaling rule is established. This pre-set scaling rule can be a constant-step linear scaling rule or an exponential scaling rule, etc. A constant-step linear scaling rule starts with the preset initial scaling parameters and increases the scaling parameters linearly at a fixed scaling step, generating multiple discrete scaling parameters. An exponential scaling rule uses the preset initial scaling parameters as a baseline and increases the scaling parameters exponentially according to a preset scaling factor, maintaining high resolution at small scales and rapidly covering the overall structure of the high-dimensional reference point cloud at large scales. Subsequently, multiple discrete scaling parameters are generated sequentially within the scaling parameter range according to the scaling rule, and each discrete scaling parameter is substituted into the geometric structure scaling function for sampling, thereby converting the continuously changing geometric structure scaling function into structural response results under multiple discrete scaling parameters. Finally, the discrete scale parameters are sorted according to their numerical values ​​to form a scale parameter sequence. This scale parameter sequence is used to guide the subsequent scale-by-scale construction of simple complex topologies and the topology evolution analysis, providing a scale basis for the generation of persistent weld structure diagrams.

[0064] After obtaining the scale parameter sequence, the scale parameters in the sequence are arranged in ascending order of their numerical values. Specifically, multiple scale parameters generated by discretization are compared and sorted, increasing from the smallest scale to the largest scale, thus forming an ordered sequence of scale parameters with a clear sequential order. By arranging the scale parameters in ascending order, the simple complex topologies subsequently constructed based on different scale parameters can be gradually generated and evolved in order of increasing scale, accurately reflecting the structural change process of the high-dimensional reference point cloud from a locally connected state to a globally connected state.

[0065] For any scale parameter in the scale parameter sequence, a simplex topology corresponding to that scale parameter is constructed based on the distance relationship between the high-dimensional reference point cloud and the high-dimensional feature space. The simplex topology is used to describe the spatial connectivity and structural morphology of the high-dimensional reference point cloud under different scale parameters. The simplex topology is constructed sequentially from zero-dimensional simplex, one-dimensional simplex, two-dimensional simplex, and higher-dimensional simplex. The zero-dimensional simplex corresponds to a single reference point in the high-dimensional reference point cloud and is used to characterize the local stability characteristics of the weld structure. The one-dimensional simplex consists of two zero-dimensional simplexes with a distance smaller than the current scale parameter and is used to characterize the connectivity between points. The two-dimensional and higher-dimensional simplexes consist of a set of multiple zero-dimensional simplexes that satisfy the distance condition pairwise and are used to characterize the aggregation morphology of the local and global high-dimensional structures. In other words, under the current scale parameters, each reference point in the high-dimensional reference point cloud is treated as a zero-dimensional simplex. The zero-dimensional simplex is used to characterize the local magnetic response of the weld structure in a steady state. Secondly, when the distance between two zero-dimensional simplexes is less than or equal to the current scale parameter, a one-dimensional simplex is constructed between the corresponding zero-dimensional simplexes to describe the connection relationship between points. Subsequently, when the distance between any two pairs of the three zero-dimensional simplexes is less than or equal to the scale parameter, a two-dimensional simplex is constructed on the set of three points to characterize the local three-point convergence surface. Further, when multiple zero-dimensional simplexes simultaneously satisfy the distance condition, a higher-dimensional simplex is constructed on the corresponding set of reference points to characterize the coupling and aggregation features of the high-dimensional structure. By sequentially constructing simplexes from zero-dimensional to higher-dimensional, a simplex complex topology is formed, thereby determining the local and global spatial relationships of the high-dimensional reference point cloud under the current scale parameters.

[0066] Based on the scale parameters arranged in ascending order, a topological evolution analysis is performed on the simplex complex topology to determine its generation and disappearance behaviors. Specifically, the scale parameter sequence is obtained sequentially in ascending order, and a corresponding simplex complex topology is constructed under each scale parameter condition. Then, the simplex complex structure at the current scale is compared with the structure at the previous scale. For simplexes appearing for the first time at the current scale, their generation behavior and generation scale are recorded. For simplexes that no longer exist at the current scale or have merged into higher-dimensional simplexes, their disappearance behavior and disappearance scale are recorded. Generation and disappearance behaviors are used to characterize the topological evolution process of the high-dimensional reference point cloud at different spatial scales, thereby identifying the stable topological features of the high-dimensional feature point cloud and further using them to construct a persistent map of the weld structure. By sequentially analyzing the changes in the simplex complex structure under all scale parameters, a topological evolution record of the simplex complex topology is obtained. This topological evolution record reflects the generation and disappearance of the structure of the high-dimensional reference point cloud at different spatial scales, providing a foundation for the construction and extraction of the persistent map of the weld structure.

[0067] The topological evolution of high-dimensional feature point clouds at different scale parameters is determined by using simplex complex topology. This involves comparing the simplex structure at each scale with the structure at the previous scale. For simplexes appearing for the first time at the current scale, their generation behavior and generation scale are recorded. For simplexes that disappear or merge into higher-dimensional simplexes at the current scale, their disappearance behavior and disappearance scale are recorded. By statistically summarizing the generation and disappearance behaviors of simplexes at all scale parameters, the topological evolution of high-dimensional feature point clouds at different scale parameters is obtained. The topological evolution results can reflect the evolutionary process of high-dimensional feature point clouds from local connection to global connectivity.

[0068] Next, within a preset time window, stable topological features existing across multiple scale parameters are identified based on the topological evolution results. A preset time window is set to determine the scale range for judging topological stability. The preset time window can be based on a proportional scaling method, i.e., setting a fixed proportion of the window period according to the total range of the scale parameter sequence. For example, 5% to 10% of the total range of the scale parameter sequence (only an example) can be used as the window period to ensure that topological structures that persist within the selected window period can be identified as stable topological features. The specific setting can be determined according to the enterprise's accuracy requirements. Subsequently, each simplex and its combination in the topological evolution results is traversed to determine whether each simplex persists within the scale window, i.e., it does not disappear from its generation scale to its disappearance scale. For simplexes or combinations of simplexes that persist within the preset time window, they are marked as stable topological features, and their corresponding scale range and persistence scale length are recorded. Through the above steps, topological features that remain stable under different scale parameters in the high-dimensional feature point cloud can be obtained. Stable topological features can accurately reflect the steady-state connection mode of weld structures under multi-scale conditions, providing basic data for the generation of persistent weld structure diagrams and the stability analysis of weld structures.

[0069] Based on the identified stable topological features and their generation and disappearance scales, nodes are established for the persistent graph of the weld structure. Each node corresponds to a stable topological feature, and its generation, disappearance, and persistence scale lengths are recorded. Subsequently, connections are established between nodes based on topological structure or spatial relationships, forming topological associations between nodes. Finally, all nodes and their connections are constructed into a complete persistent graph of the weld structure, where the scale length of each node reflects the stability of the corresponding topological feature under different scale parameters. The persistent graph of the weld structure can accurately reflect the stable topological features of high-dimensional feature point clouds under different scale conditions.

[0070] By generating persistent weld structure maps, the stable topological characteristics of the weld region can be accurately reflected, providing a solid foundation for long-term monitoring of weld conditions. This helps improve the accuracy and reliability of weld inspection and provides strong protection for the safe operation of gas cylinders.

[0071] In one embodiment of this example, a geometrical scaling function for the high-dimensional reference point cloud is constructed based on the distance relationships between the high-dimensional reference point cloud and the high-dimensional feature space, including: S710. Preprocess the high-dimensional reference point cloud; S720. Construct an Euclidean distance matrix based on the distance relationships of the preprocessed high-dimensional reference point cloud; S730. For any scale parameter, count the number of neighbor points in the high-dimensional reference point cloud with a preset distance radius, and calculate the local average number of neighbors to characterize the local connectivity of the point cloud. S740. For any scale parameter, calculate the global discreteness of the high-dimensional reference point cloud. S750: Combine the local average number of neighbors and the global dispersion under each scale parameter to generate the geometric structure scale function.

[0072] First, the high-dimensional reference point cloud is preprocessed, including but not limited to filtering out anomalies and noise points to eliminate interference caused by measurement errors or non-defect factors; the high-dimensional feature vectors are normalized or standardized to unify the scale of each feature component and avoid certain features from dominating the distance calculation; and the point cloud is locally smoothed or clustered as needed to enhance the stability of the local topology.

[0073] Secondly, an Euclidean distance matrix is ​​constructed based on the distance relationships of the preprocessed high-dimensional reference point cloud. Each preprocessed high-dimensional reference point cloud is a high-dimensional eigenvector, representing the local magnetic response characteristics of the weld region. Subsequently, for any two high-dimensional reference points in the point cloud, their Euclidean distance is calculated as a measure of their similarity or spatial aggregation. Furthermore, the pairwise Euclidean distances between all points are organized into an N×N Euclidean distance matrix, where N is the number of points in the high-dimensional reference point cloud, the diagonal elements are zero, and the matrix is ​​symmetric. The Euclidean distance matrix can completely characterize the distance relationships between points in the high-dimensional reference point cloud and is used for subsequent construction of geometric scaling functions, simplex topology, and topological evolution analysis.

[0074] For any given scale parameter, the number of neighboring points within a preset distance radius of each point in the high-dimensional reference point cloud is counted, and the local average neighbor count is calculated to characterize the local connectivity of the point cloud. A preprocessed high-dimensional reference point cloud is obtained, and the scale parameter is used as the preset distance radius. Subsequently, for each reference point in the point cloud, the number of neighboring points within the preset distance radius is counted, and the average of the neighbor counts for all reference points is calculated to obtain the local average neighbor count. The local average neighbor count reflects the degree of local connectivity of the high-dimensional reference point cloud at the current scale.

[0075] Next, the preprocessed high-dimensional reference point cloud and current scale parameters are obtained. The overall average value of the point cloud is then calculated by either calculating the Euclidean distance between any two reference points in the point cloud or counting the number of neighbors for each reference point within the stated scale radius. Subsequently, the global dispersion is calculated based on this average value. The global dispersion reflects the overall sparse or clustered state of the high-dimensional reference point cloud at the current scale. The global dispersion can be combined with the local average neighbor count to construct a geometric scaling function.

[0076] The local average neighbor count reflects the local connectivity of the high-dimensional reference point cloud at the current scale; the global dispersion reflects the overall sparse or clustered distribution of the high-dimensional reference point cloud at the current scale. The local average neighbor count and global dispersion corresponding to each scale parameter can be used to construct a geometric structure scaling function using normalized ratios. The specific expression is shown below:

[0077] Where S is the function value of the geometric structure scaling function at scale ri; n i The local average number of neighbors is denoted by Gi; the global dispersion is Gi; n max This represents the maximum number of neighbors that a high-dimensional reference point cloud can have at the selected scale parameters.

[0078] The local average number of neighbors and its maximum value are normalized for each scale parameter to obtain the local connectivity normalization index. Then, the global dispersion is normalized for its maximum number of neighbors, and the complement value is taken to reflect the overall clustering degree. Next, the normalized local connectivity index and the global clustering index are multiplied to obtain the geometric structure scale function value for that scale parameter. The above steps are repeated for all scale parameters to form a complete geometric structure scale function sequence. The geometric structure scale function sequence can comprehensively reflect the local connectivity and overall distribution characteristics of the high-dimensional reference point cloud at different scales.

[0079] By constructing a geometric structure scaling function for a high-dimensional reference point cloud, the geometric structure characteristics of the high-dimensional reference point cloud at different scales can be clearly defined. This provides an accurate quantitative description for subsequent topology analysis and the generation of persistent weld structure maps, which helps to improve the accuracy and reliability of weld inspection and provides strong support for the health monitoring of gas cylinder welds.

[0080] In one embodiment of this example, constructing a corresponding health baseline manifold based on the magnetic fingerprint relationship template includes: S810. Extract anchoring topological features from the magnetic fingerprint relationship template of the target gas cylinder to obtain the topological constraint set; S820. Establish a mapping relationship between the topological constraint set and the state-sensitive feature subset to obtain the health feature subset corresponding to the topological constraint set; S830. The health feature subset is subjected to dimensionality reduction using the manifold learning algorithm, and the high-dimensional features are mapped to the low-dimensional space to obtain the low-dimensional feature point cloud using the topological constraint set as the regularization term. S840. Fit the low-dimensional feature point cloud to obtain a low-dimensional smooth manifold; S850, using a low-dimensional smooth manifold as the healthy baseline manifold.

[0081] Anchored topological features are extracted from the magnetic fingerprint relationship template of the target gas cylinder to obtain a topological constraint set. The magnetic fingerprint relationship template of the target gas cylinder is a previously generated persistent weld structure map or high-dimensional magnetic fingerprint template, used to characterize the internal structure and state of the weld. Anchored topological features are the most stable, typical, or critical topological features selected from the template. These features maintain consistency under multi-scale analysis and are not easily affected by noise or local fluctuations; they are essentially topological fixed points in the point cloud. The topological constraint set is a collection of these anchored topological features, used to compare and match the high-dimensional features or newly acquired magnetic fingerprints of the target gas cylinder. According to preset screening rules, topological features that persist across multiple scales, whose generation and disappearance behaviors meet preset thresholds, and exhibit local connectivity are selected from the template and used as anchored topological features. Furthermore, the anchored topological features are aggregated to form a topological constraint set, where each topological feature contains a corresponding high-dimensional point set, generation and disappearance scale parameters, and topological characteristic indicators. The topological constraint set can be used to guide the matching of newly acquired magnetic fingerprints of the target gas cylinder with the template, weld structure stability analysis, and defect identification. The preset screening rules can be to select only topological features that persist across multiple scale parameters. For example, a feature is considered stable if it exists at more than a certain proportion (e.g., 70% is just an example) of a scale parameter. This embodiment innovatively couples the magnetic fingerprint relationship template with the health baseline manifold, with the core objective of achieving individualized health benchmark calibration. Specifically, the topological constraint set is extracted from the magnetic fingerprint relationship template, representing the inherent and stable physical structural features of the target cylinder (such as weld geometry, material distribution texture, etc.). When constructing the health baseline manifold, this topological constraint set is introduced as a regularization term, which constrains the manifold's geometry, ensuring consistency with the inherent identity features of the target cylinder in low-dimensional space. This avoids the false alarm problem caused by the inability of traditional general health baselines to distinguish between individual manufacturing differences and actual damage degradation. By locking the identity topology, the health baseline manifold can focus on representing abnormal changes that deviate from the inherent structure, thereby significantly improving the detection sensitivity for early, minor defects.

[0082] A mapping relationship is established between the topological constraint set and the state-sensitive feature subset to obtain the health feature subset corresponding to the topological constraint set. Specifically, the topological constraint set and the corresponding state-sensitive feature subset of the target gas cylinder are obtained. The topological constraint set includes anchored topological features of the weld region at multiple scales, and the state-sensitive feature subset includes high-dimensional features that can reflect changes in the structural state of the gas cylinder. Subsequently, for each anchored topological feature in the topological constraint set, its similarity with each feature in the state-sensitive feature subset is calculated, and the state-sensitive feature with the strongest correlation is selected as the mapping result. Further, the mapping results are aggregated to form a health feature subset, where each health feature corresponds to at least one topological constraint feature, which can reflect the structural health state of the weld region. The health feature subset can be used for structural health evaluation of gas cylinder welds, matching of newly acquired magnetic fingerprints with templates, defect identification, and structural integrity analysis.

[0083] A subset of health features is dimensionality-reduced using a manifold learning algorithm, with a set of topological constraints used as a regularization term to map high-dimensional features to a low-dimensional space, resulting in a low-dimensional feature point cloud. The health feature subset includes state-sensitive features associated with the weld health status, and the topological constraint set includes anchored topological features. Subsequently, a suitable manifold learning algorithm, such as locally linear embedding, is selected, and the health feature subset is used as input data. During the optimization process of the manifold learning algorithm, the set of topological constraints is introduced as a regularization term. By setting the weighting coefficients of the loss function and the constraint term, the low-dimensional mapping maintains the local and global structural relationships of the health feature subset while preserving the proximity and connectivity of the anchored topological features in the low-dimensional space. Further, dimensionality reduction is performed to map the high-dimensional health features to a low-dimensional feature point cloud. The weighting coefficients can be fixed constants used to balance the structure preservation objective and the topological constraint objective of the manifold learning algorithm. The loss function represents the optimization objective of the manifold learning algorithm itself, used to maintain the local or global structure of the high-dimensional data in the low-dimensional space; it can be a locally linear embedding function as used in existing technologies, with the expression:

[0084] Where Loss represents the loss function; yi is the low-dimensional mapping point; w ij The high-dimensional neighbor reconstruction weights are used to represent the linear reconstruction contribution of each healthy feature point in the high-dimensional space to its neighbors. They can be obtained by solving a local linear reconstruction optimization problem, where the reconstruction error of each high-dimensional feature point and its neighbors is minimized while satisfying the normalization constraint of the reconstruction weights. Alternatively, the weights can be directly calculated using exponential weighting or normalization methods based on the distance or similarity between the high-dimensional feature point and its neighbors. The neighbor reconstruction weights can also be determined by constructing an adjacency graph and using the edge weights of the graph structure.

[0085] A low-dimensional smooth manifold is obtained by fitting a low-dimensional feature point cloud. This low-dimensional feature point cloud can be obtained through a manifold learning algorithm, where each point corresponds to the mapping position of a subset of the target cylinder's health features in the low-dimensional space. A suitable fitting method is selected, such as spline curves or Gaussian process regression. Local fitting is performed on the low-dimensional point cloud to form a continuous local smooth surface. Furthermore, the local fitting results are combined into a global low-dimensional manifold, and a smoothing regularization term is added during the fitting process to control the surface curvature and gradient continuity, thereby preventing overfitting and local jumps. Finally, a continuous low-dimensional smooth manifold is obtained, which can smoothly reflect the structural relationships of the low-dimensional feature point cloud both locally and globally. This low-dimensional smooth manifold can be used to characterize the potential low-dimensional structure of weld health features.

[0086] After obtaining the low-dimensional smooth manifold, it is used as the health baseline manifold. The low-dimensional smooth manifold reflects the low-dimensional feature structure of the weld area of ​​the target gas cylinder in a healthy state, and can continuously and smoothly characterize the potential distribution characteristics of the healthy feature subset in the low-dimensional space. The health baseline manifold serves as a standard reference template and can be used to compare with newly acquired low-dimensional feature points to evaluate the structural health status of the gas cylinder weld area and identify potential defects or abnormal features.

[0087] By constructing a corresponding health baseline manifold using a magnetic fingerprint relationship template, complex high-dimensional features can be effectively mapped to a low-dimensional space through topological constraints and dimensionality reduction processing, while retaining key health status information. This provides a stable and reliable low-dimensional reference model for the health monitoring of gas cylinder welds, which helps to achieve efficient monitoring and accurate assessment of weld conditions.

[0088] A brief explanation will be given using the example of a periodic inspection station for liquefied petroleum gas cylinders.

[0089] After the target cylinder is welded, a uniform alternating magnetic field is applied to it. When the cylinder enters the initial position of the inspection station, it is placed in a surrounding magnetization device composed of Helmholtz coils. This device is supplied with a sinusoidal alternating current with a frequency of 1-5Hz and controllable intensity, which generates a uniform and stable low-frequency alternating magnetic field environment around the cylinder to magnetize the weld area.

[0090] A pre-set rotating device is used to scan the target cylinder, obtaining magnetic response data of the weld area. The cylinder is conveyed by rollers and driven to rotate uniformly around its axis (e.g., 10 rpm). A high-sensitivity magnetic induction intensity sensor array is fixedly installed directly above the weld. During one rotation of the cylinder, the sensor array continuously collects the magnetic induction intensity change data at every point on the circumference of the weld at a high-frequency sampling rate (e.g., 1 kHz), obtaining a complete waveform sequence of "magnetic response data of the weld area".

[0091] The magnetic flux leakage signal of the target gas cylinder is determined based on the magnetic response data of the weld area, and the magnetic flux leakage signal is converted into a high-dimensional magnetic fingerprint feature of the target gas cylinder in a high-dimensional feature space. The magnetic flux leakage signal is determined by bandpass filtering (to remove power frequency interference and high-frequency noise) and baseline correction of the collected raw magnetic response data. The local gradient and amplitude statistics of the signal are calculated using a sliding window. Signal segments with significant gradient abrupt changes and amplitudes exceeding the statistical threshold of a healthy weld are identified as "magnetic flux leakage signals," typically corresponding to microscopic discontinuities in the weld or heat-affected zone (such as micropores, slag inclusions, or stress concentration areas).

[0092] The process involves converting the identified magnetic flux leakage signal segments into high-dimensional magnetic fingerprint features. These segments are arranged sequentially according to their position on the weld circumference. A set of time-domain and frequency-domain features is extracted for each segment, including peak value, root mean square value, impulse factor, waveform factor, spectral centroid, and spectral variance. These features from all signal segments along the entire circumference are then concatenated sequentially to form a feature vector of several hundred dimensions. This high-dimensional vector represents the "high-dimensional magnetic fingerprint feature" of the gas cylinder in its current state, comprehensively encoding the spatial distribution information of the weld microstructure.

[0093] The system acquires high-dimensional magnetic fingerprint features under different operating conditions; in this embodiment, "different operating conditions" refers to different stages of the gas cylinder's lifecycle. The system acquires magnetic fingerprints at the following two key points: Operating Condition A (Factory Standard Operating Condition): After the gas cylinder is manufactured and passes the hydrostatic test, the first magnetic fingerprint is collected and archived before leaving the factory. Operating Condition B (Periodic Inspection Operating Condition): When the gas cylinder returns to the inspection station in the Nth year of its service (e.g., a 3-year or 5-year periodic inspection cycle), magnetic fingerprints are collected again under the same testing conditions.

[0094] Stability analysis was conducted based on the high-dimensional magnetic fingerprint characteristics under different operating conditions to determine the structural stability feature subset and state-sensitive feature subset of the target gas cylinder. Two high-dimensional magnetic fingerprint feature vectors obtained from the same gas cylinder under operating conditions A and B were compared and analyzed. Stability analysis: The difference or correlation coefficient between the two vectors in each feature dimension was calculated. Feature dimensions with highly consistent values ​​and minimal changes (e.g., change rate <5%) under both operating conditions were considered structurally stable features less affected by aging and external environmental factors. These features reflect the inherent microstructural "identity" information of the weld, formed during the manufacturing process, such as the inherent magnetic differences between the base material and the welding material, and the main welding patterns. Sensitivity analysis: Feature dimensions with significant changes in values ​​under both operating conditions (e.g., change rate >20% or low correlation coefficient) were considered state-sensitive features. These features are more sensitive to fatigue accumulation, microcrack initiation, corrosion, and other damage during use. Subset division: Based on a preset threshold (e.g., a change rate of 10%), all high-dimensional features are divided into a structurally stable feature subset (for identity recognition) and a state-sensitive feature subset (for health status assessment).

[0095] A magnetic fingerprint relationship template for the weld area of ​​the target gas cylinder is generated by combining a subset of structurally stable features in a high-dimensional feature space. This template is used to determine the intrinsic identity information of the target gas cylinder. A subset of structurally stable features is extracted from the feature vector obtained under operating condition A (factory standard). This subset constitutes a compact numerical template representing the unique and inherent identity of the gas cylinder, called the "magnetic fingerprint relationship template." This template is bound to the gas cylinder's factory serial number and stored in a central database. When the gas cylinder returns for inspection, the system only needs to quickly match the subset of structurally stable features extracted from the current inspection data with the template in the database (e.g., calculating Euclidean distance or cosine similarity). If the matching degree exceeds a threshold (e.g., 95%), its intrinsic identity information can be accurately determined, achieving traceability without physical identification.

[0096] Constructing a corresponding health baseline manifold based on the magnetic fingerprint relationship template; this step aims to establish a reference benchmark for the "health status" of the gas cylinder. Collect a batch of gas cylinder data of the same model, acquired under operating condition A (factory health status), whose structural stability characteristics are similar to the gas cylinder template (belonging to the same manufacturing batch or process). Extract a subset of their state-sensitive features from this data. Use a manifold learning algorithm (such as Isomap or t-SNE) to reduce the dimensionality of these high-dimensional "health status-sensitive features" to a low-dimensional (such as 2D or 3D) space. In this low-dimensional space, the state feature points of all healthy gas cylinders will cluster to form a continuous, low-dimensional surface or region, which is the "health baseline manifold". This manifold specifically refers to the standard health state space composed of healthy samples with similar inherent identity (template) to the target gas cylinder.

[0097] Orthogonally project the state-sensitive feature subset and the health baseline manifold to determine the cylinder quality residual vector. For the current cylinder to be inspected (whose identity has been confirmed through template matching), extract its state-sensitive feature subset and map it into the same low-dimensional space as the "health baseline manifold" to obtain a feature point P_current. Calculate the vector difference between P_current and its nearest point P_healthy on the "health baseline manifold": residual vector R = P_current - P_healthy. The direction and magnitude of this residual vector quantify the degree of deviation of the current cylinder state from its corresponding health benchmark.

[0098] The quality degradation type and degree of the target gas cylinder are determined by analyzing the residual vector of the cylinder's quality. Determining the degradation type involves first establishing a residual vector pattern library based on historical data and failure analysis. For example, residuals primarily along a positive axis in a low-dimensional space might correspond to "uniform fatigue," residuals in a specific direction might correspond to "localized cracks," and another pattern might correspond to "corrosion." The current gas cylinder's residual vector R is matched against the pattern library, and the most likely quality degradation type (e.g., "micro-crack tendency in the weld heat-affected zone") is determined using nearest neighbor classification or a neural network model. Determining the degree of degradation involves calculating the modulus |R| of the residual vector R. This modulus, after normalization, can be mapped to a quality degradation degree index between 0 and 1. For example, |R| less than 0.3 is considered "slight," 0.3-0.6 is considered "moderate concern," and greater than 0.6 is considered "significant degradation requiring close inspection."

[0099] The system generates a production quality traceability report for the target gas cylinder based on the type and degree of quality degradation. The system automatically generates a structured report, which includes at least the following: 1. Cylinder Identification Information: The original production batch / number matched with the magnetic fingerprint template.

[0100] 2. Conclusion of this test: Identified quality degradation type (e.g., micro-fatigue accumulation) and quantified degradation degree (e.g., index 0.45, moderate).

[0101] 3. Source analysis: Based on the historical process data of this production batch, report the manufacturing factors that may be related (e.g., "the cooling rate after welding in this batch is relatively fast, which can easily lead to higher residual stress, and is related to the fatigue tendency currently detected").

[0102] 4. Maintenance Recommendations: Based on the type and extent of degradation, recommendations such as "continue to use", "shorten the next inspection cycle", "recommend local non-destructive testing (such as ultrasonic testing)" or "immediate scrapping" are given.

[0103] The report is also updated to the cylinder's lifetime electronic file, completing the traceability and archiving of quality information for this inspection.

[0104] This application also provides a machine-readable storage medium storing instructions that cause a machine to execute the above-described data analysis-based method and system for tracing the quality of steel cylinder production.

[0105] This application also provides an electronic device, including: The memory is configured to store instructions; and The processor is configured to retrieve instructions from memory and, when executing instructions, to implement the aforementioned data analysis-based method and system for tracing the quality of steel cylinder production.

[0106] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0107] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0108] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0109] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0110] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0111] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, like read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0112] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0113] It should also be noted that 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 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.

[0114] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for tracing the quality of steel cylinder production based on data analysis, characterized in that, include: After the target cylinder is welded, a uniform alternating magnetic field is applied to the target cylinder. The target cylinder is rotated and scanned using a pre-set rotating device to obtain the magnetic response data of the weld area of ​​the target cylinder. The magnetic flux leakage signal of the target cylinder is determined based on the magnetic response data of the weld area, and the magnetic flux leakage signal is converted into a high-dimensional magnetic fingerprint feature of the target cylinder in a high-dimensional feature space. Acquire high-dimensional magnetic fingerprint features under different working conditions; Stability analysis was performed based on the high-dimensional magnetic fingerprint characteristics under different working conditions to determine the structural stability feature subset and state-sensitive feature subset of the target gas cylinder; A magnetic fingerprint relationship template for the weld seam region of the target gas cylinder is generated by combining a subset of structurally stable features in a high-dimensional feature space. The magnetic fingerprint relationship template is used to determine the intrinsic identity information of the target gas cylinder. Construct the corresponding health baseline manifold based on the magnetic fingerprint relationship template; Orthogonally project the state-sensitive feature subset and the healthy baseline manifold to determine the cylinder quality residual vector; The type and degree of quality degradation of the target cylinder are determined by the cylinder quality residual vector; A production quality traceability report for the target steel cylinder is generated by combining the type and degree of quality degradation of the cylinder.

2. The method according to claim 1, characterized in that, The step of determining the magnetic flux leakage signal of the target gas cylinder based on the magnetic response data of the weld area, and converting the magnetic flux leakage signal into a high-dimensional magnetic fingerprint feature of the target gas cylinder in a high-dimensional feature space, includes: The magnetic response data of the weld area is preprocessed to obtain the preprocessed magnetic response data of the weld area; The physical properties of the weld area of ​​the target steel cylinder are distinguished to determine the physical property zoning structure of the weld area of ​​the target steel cylinder; Based on the physical characteristics of the partition structure, the magnetic response data of the pre-processed weld area is partitioned and differentiated to obtain the magnetic flux leakage signal. Based on the physical characteristics of the partition structure, feature extraction is performed on the magnetic flux leakage signal to obtain a set of magnetic flux leakage feature vectors; The set of magnetic flux leakage feature vectors is mapped to a high-dimensional feature space using a preset nonlinear embedding mapping function, and the set of magnetic flux leakage feature vectors in the high-dimensional feature space is subjected to cluster analysis to determine the high-dimensional magnetic fingerprint features of the target cylinder.

3. The method according to claim 2, characterized in that, The step of performing partitioned differential matching on the preprocessed weld region magnetic response data based on the physical characteristic partitioning structure to obtain the magnetic flux leakage signal includes: Based on the physical property partitioning structure of the target gas cylinder weld area, the preprocessed magnetic response data of the weld area is divided into multiple sub-signal segments, where each sub-signal segment corresponds to a physical property partitioning structure of the weld. Analyze the noise characteristics of each physical property partition structure; Based on the noise characteristics of the physical partition structure, differentiated noise reduction processing is applied to the sub-signal segments. The sub-signal segments after differential noise reduction are physically partitioned and sequentially spliced ​​to obtain the reconstructed magnetic response signal, which reflects the microscopic physical structure of the weld. Magnetic flux leakage signals are determined based on the microscopic physical structure of the weld.

4. The method according to claim 1, characterized in that, The determination of the structural stability feature subset and state-sensitive feature subset of the target gas cylinder based on the stability analysis of high-dimensional magnetic fingerprint features under different operating conditions includes: Each high-dimensional magnetic fingerprint feature is labeled with its operating condition to obtain the corresponding operating condition label for each high-dimensional magnetic fingerprint feature; Calculate the mutual information between each high-dimensional magnetic fingerprint feature and its corresponding operating condition label. The mutual information is used to characterize the degree of correlation between the high-dimensional magnetic fingerprint feature and the operating condition changes. Calculate the inter-class divergence of high-dimensional magnetic fingerprint features under the same working conditions. The inter-class divergence is used to characterize the distinguishing ability of high-dimensional magnetic fingerprint features between different working condition categories. The state sensitivity score is obtained by weighted fusion of mutual information and inter-class divergence. Based on the state sensitivity score, the high-dimensional magnetic fingerprint features are divided to determine the structural stability feature subset and the state sensitivity feature subset of the target gas cylinder.

5. The method according to claim 1, characterized in that, The generation of the magnetic fingerprint relationship template for the target gas cylinder weld region by combining a subset of structurally stable features in a high-dimensional feature space includes: In a high-dimensional feature space, a high-dimensional reference point cloud is determined by a structurally stable subset of features. A topological evolution analysis is performed on a high-dimensional reference point cloud to determine the topological features of the high-dimensional feature point cloud, and a persistent image of the weld structure is generated based on the topological features. The persistent image of the weld structure is used as a magnetic fingerprint relationship template for the weld area of ​​the target gas cylinder.

6. The method according to claim 5, characterized in that, The step of performing topological evolution analysis on a high-dimensional reference point cloud to determine the topological features of the high-dimensional feature point cloud, and generating a persistent weld structure map based on the topological features, includes: Obtain the distance relationships between high-dimensional reference point clouds and high-dimensional feature spaces; The geometric structure scaling function of the high-dimensional reference point cloud is determined based on the distance relationship between the high-dimensional reference point cloud and the high-dimensional feature space. The geometric structure scale function is discretized according to the preset scale growth rule to generate a scale parameter sequence; Sort the scale parameters in the scale parameter sequence in ascending order; For any scale parameter in a scale parameter sequence, construct the corresponding simple complex topology based on the high-dimensional reference point cloud; Based on the scale parameters arranged in ascending order, the topological evolution of simple complex topologies is analyzed to determine the generation and disappearance behaviors of simple complex topologies. The evolution of the topology of high-dimensional feature point clouds under different scale parameters is determined by using simple complex topology. Within a preset time window, stable topological features existing in multiple scale parameters are identified based on the topological evolution results. The generation and disappearance behaviors corresponding to each stable topological feature are mapped to generate a persistent graph of the weld structure.

7. The method according to claim 6, characterized in that, The step of determining the geometric structure scaling function of the high-dimensional reference point cloud based on the distance relationship between the high-dimensional reference point cloud and the high-dimensional feature space includes: Preprocess the high-dimensional reference point cloud; Construct an Euclidean distance matrix based on the distance relationships of the preprocessed high-dimensional reference point cloud; For any scale parameter, count the number of neighbor points in the high-dimensional reference point cloud with a preset distance radius, and calculate the local average number of neighbors to characterize the local connectivity of the point cloud. For any scale parameter, calculate the global discreteness of the high-dimensional reference point cloud; The geometric structure scale function is generated by combining the local average number of neighbors and the global dispersion for each scale parameter.

8. The method according to claim 1, characterized in that, The construction of the corresponding health baseline manifold based on the magnetic fingerprint relationship template includes: Anchored topological features are extracted from the magnetic fingerprint relationship template of the target gas cylinder to obtain a set of topological constraints; By establishing a mapping relationship between the topological constraint set and the state-sensitive feature subset, the health feature subset corresponding to the topological constraint set is obtained; The health feature subset is dimensionality reduced using a manifold learning algorithm, and the high-dimensional features are mapped to a low-dimensional space to obtain a low-dimensional feature point cloud using a set of topological constraints as a regularization term. A low-dimensional smooth manifold is obtained by fitting the low-dimensional feature point cloud; A low-dimensional smooth manifold is used as the healthy baseline manifold.

9. A machine-readable storage medium, characterized in that, The machine-readable storage medium stores instructions for causing the machine to execute a data analysis-based method for tracing the quality of steel cylinder production according to any one of claims 1 to 8.

10. An electronic device, characterized in that, include: The memory is configured to store instructions; as well as The processor is configured to retrieve the instructions from the memory and, when executing the instructions, to implement a data analysis-based method for tracing the quality of steel cylinder production according to any one of claims 1 to 8.

Images