Pressure vessel health assessment system based on multi-modal data fusion

Abstract

The invention discloses a pressure vessel health assessment system based on multi-modal data fusion, and relates to the technical field of pressure vessel safety detection, and the system comprises a synchronous fusion module, a topology learning module, an event detection module and a causal assessment module. The synchronous fusion module obtains acoustic emission time domain signals and weld joint digital images, and a feature fusion set is generated through space-time registration; the topology learning module inputs the feature fusion set into a graph neural network of specific nodes and edge weights, and generates a graph embedding vector through graph convolution; the event detection module accumulates a graph embedding vector time sequence, calculates a cosine similarity identification mutation point, and extracts a damage event feature packet; the causal evaluation module constructs a multi-modal feature causal graph, analyzes direct and indirect influence factors, and outputs damage type labels and confidence. According to the system, multi-modal feature topological correlation learning and causal-driven damage evaluation are realized, and the accuracy of health detection of the pressure vessel is improved.

Description

Technical Field

[0001] This invention belongs to the field of pressure vessel safety inspection technology, specifically a pressure vessel health assessment system based on multimodal data fusion. Background Technology

[0002] In the current field of pressure vessel health assessment, most methods adopt single-modal detection. Acoustic emission detection focuses on feature extraction and anomaly identification of time-domain signals, while digital image detection focuses on geometric shape recognition of defects in weld areas. Some multimodal detection schemes only perform simple feature stitching processing on acoustic signals and image data, and the synchronous fusion stage only completes basic data integration operations.

[0003] Single-modal detection methods cannot simultaneously characterize the dynamic acoustic response and static geometric morphology of pressure vessel damage. Multimodal schemes using simple feature stitching fail to establish spatial relationships between detection units and defect regions, and cannot uncover structural-level correlations among multimodal features. Existing damage event detection methods often employ fixed threshold discrimination, without calculating cumulative similarity for temporal feature vectors, making it difficult to accurately capture the triggering moments of potential damage. Current damage assessment models rely on feature correlation analysis for classification, failing to distinguish the interactions between different modal features, and resulting in damage assessment results lacking an inherent logical basis for differentiation.

[0004] A graph neural network with specific node and edge weight definitions needs to be constructed to achieve multimodal feature topological association learning. The role factors of acoustic and geometric features need to be analyzed through causal graphs to complete damage assessment. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art; Therefore, this invention proposes a pressure vessel health assessment system based on multimodal data fusion, comprising: The synchronous fusion module is used to acquire acoustic emission time-domain signals and digital images of the weld area, construct a multimodal raw dataset, perform time synchronization and spatial registration on the multimodal raw dataset, and generate a feature fusion set. The topology learning module is used to input the feature fusion set into a pre-constructed graph neural network. The nodes of the graph neural network are composed of the location of the acoustic emission sensor and the center of the defect region in the digital image. The edge weights are determined by the arrival time difference of the acoustic emission signal and the spatial distance of the image. The topological association between multimodal features is learned through graph convolution operations to generate a graph embedding vector containing the structural health status. The event detection module is used to perform temporal accumulation on the graph embedding vectors, calculate the cosine similarity between graph embedding vectors at different time steps, identify abrupt change points in the graph embedding vectors, take the time corresponding to the abrupt change point as the trigger time of the potential damage event, and extract multimodal features before and after the trigger time to form a damage event feature package. The causal assessment module is used to drive a causal reasoning-based damage assessment model by taking the damage event feature package as input. The causal reasoning-based damage assessment model constructs a causal graph between multimodal features, analyzes the direct and indirect influencing factors of acoustic feature changes on geometric feature evolution, and outputs assessment results including damage type labels and confidence levels.

[0006] Further, the acquisition of acoustic emission time-domain signals and digital images of the weld area, the construction of a multimodal raw dataset, the performance of time synchronization and spatial registration on the multimodal raw dataset, and the generation of a feature fusion set include: Acoustic emission time-domain signals are acquired by an array of acoustic emission sensors arranged on the surface of the pressure vessel, and digital images of the weld area are acquired by a digital X-ray imaging system of the pressure vessel, forming a multimodal raw dataset containing acoustic and visual features. The multimodal raw dataset is time-synchronized and spatially registered to establish an association index between acoustic emission events and corresponding spatial locations in digital images. The acoustic emission time-domain signal is decomposed by wavelet packet to extract energy entropy features, and the digital image is morphologically segmented to extract geometric moment features of defect edges, forming a feature fusion set that includes time-frequency domain acoustic features and spatial domain geometric features. The process involves acquiring acoustic emission time-domain signals through an array of acoustic emission sensors arranged on the surface of the pressure vessel, and simultaneously acquiring digital images of the weld area through a digital X-ray imaging system of the pressure vessel, forming a multimodal raw dataset containing both acoustic and visual features, including: The acoustic emission sensor array is activated in acquisition mode, and the sampling frequency is set to a preset high-frequency value. The elastic wave signal generated during the operation of the pressure vessel is continuously received, and the elastic wave signal is converted into a discrete voltage time-domain signal as the acoustic emission time-domain signal. The scanning mode of the digital X-ray imaging system is activated simultaneously to perform multi-angle exposure imaging of the weld area of ​​the pressure vessel and obtain a grayscale image sequence containing the internal structure of the weld. The grayscale image sequence is stitched and corrected to eliminate geometric distortion caused by imaging angle deviation and generate a digital image covering the entire weld area. The acoustic emission time-domain signal and digital image are aligned according to timestamps. Data frames with mismatched timestamps are removed, and the acoustic emission time-domain signal and digital image with consistent timestamps are retained to form a multimodal original dataset containing acoustic and visual features.

[0007] Furthermore, the step of performing time synchronization and spatial registration on the multimodal raw dataset to establish an association index between acoustic emission events and corresponding spatial locations in digital images includes: A unified clock reference is selected in the multimodal raw dataset, and the clock signals of all acoustic emission sensors and the clock signal of the digital X-ray imaging system are calibrated to the clock reference. A Cartesian coordinate system is established in the digital image, with the geometric center of the pressure vessel as the origin and the weld direction as the coordinate axis direction, to calibrate the conversion relationship between the pixel coordinates of the digital image and the actual physical coordinates. In the acoustic emission time domain signal, the arrival time and sensor number of the acoustic emission event are identified by the peak detection algorithm. Combined with the physical installation coordinates of the acoustic emission sensor, the spatial coordinates of the acoustic emission source of the acoustic emission event are calculated. By using the nearest neighbor matching algorithm, the spatial coordinates of the acoustic emission source are mapped to the Cartesian coordinate system of the digital image. The nearest pixel is found as the associated pixel, and the correspondence between the acoustic emission event and the associated pixel is established to generate an association index.

[0008] Furthermore, the step of extracting energy entropy features from the acoustic emission time-domain signal through wavelet packet decomposition and extracting geometric moment features of defect edges from the digital image through morphological segmentation, forming a feature fusion set containing both time-frequency domain acoustic features and spatial domain geometric features, includes: By selecting preset wavelet basis functions and decomposition levels, wavelet packet decomposition is performed on the acoustic emission time-domain signal to obtain the energy spectrum of different frequency bands; Calculate the ratio of energy in each frequency band to the total energy, construct an energy distribution vector, and calculate the Shannon entropy of the energy distribution vector as an energy entropy feature; Digital images are subjected to grayscale and binarization processing, and image noise is removed by a combination of opening and closing morphological filters. Defect regions in an image are identified using a connected component labeling algorithm. The contour edges of the defect regions are extracted, and the second-order geometric moments of the contour edges are calculated as geometric moment features. The energy entropy features and geometric moment features are arranged in the order of their associated indices and combined to form a feature fusion set that includes time-frequency domain acoustic features and spatial domain geometric features.

[0009] Further, the feature fusion set is input into a pre-constructed graph neural network. The nodes of the graph neural network are formed by the location of the acoustic emission sensor and the center of the defect region in the digital image. The edge weights are determined by the arrival time difference of the acoustic emission signal and the spatial distance of the image. The topological relationships between multimodal features are learned through graph convolution operations to generate a graph embedding vector containing the structural health status, including: Initialize the adjacency matrix of the graph neural network, and use the coordinates of the acoustic emission sensor position and the center of the defect region as the node set of the graph; For each node pair, the arrival time difference of the acoustic emission signal from one sensor to another is calculated as the time weight, and the pixel distance between the two nodes in the digital image is calculated as the spatial weight. The time weight and spatial weight are normalized and then added together to obtain the edge weight. The energy entropy features and geometric moment features in the feature fusion set are mapped to the corresponding nodes, which serve as the initial feature vectors of the nodes. Perform multi-layer graph convolution operations. Each layer of graph convolution operations performs a weighted summation and non-linear activation on the initial feature vector of a node and the feature vectors of its neighboring nodes to update the node's feature vector. After all graph convolution operations are completed, the feature vectors of all nodes are globally pooled to obtain a fixed-dimensional graph embedding vector, which contains structural health information.

[0010] Furthermore, the graph embedding vectors are accumulated over time, and the cosine similarity between graph embedding vectors at different time steps is calculated to identify abrupt changes in the graph embedding vectors. The time corresponding to the abrupt change is taken as the trigger time of the potential damage event. Multimodal features before and after the trigger time are extracted to form a damage event feature package, including: Arrange the graph embedding vectors in chronological order to form a graph embedding vector sequence; For each element in the graph embedding vector sequence, calculate its cosine similarity with the previous element to obtain the cosine similarity sequence; Apply sliding window statistics to the cosine similarity sequence, calculate the mean and standard deviation of the cosine similarity within the window, and mark the corresponding time point as a change point when the cosine similarity at a certain time is less than the mean minus a certain number of times the standard deviation. Centered on the mutation point, time periods of preset length are selected both forward and backward. Acoustic emission time-domain signals and digital images within the corresponding time periods are extracted from the multimodal raw dataset. At the same time, energy entropy features and geometric moment features within the corresponding time periods are extracted from the feature fusion set and combined to form a damage event feature package.

[0011] Furthermore, using the damage event feature package as input, a damage assessment model based on causal reasoning is driven. This model constructs a causal graph among multimodal features, analyzes the direct and indirect influencing factors of acoustic feature changes on geometric feature evolution, and outputs assessment results including damage type labels and confidence levels, including: The energy entropy change trend of acoustic emission signals and the geometric moment change trend of digital images are extracted from the feature package of damage events and used as input variables for causal inference. Based on domain knowledge and historical data statistics, we construct causal relationship hypotheses among multimodal features and form an initial causal graph, where nodes represent feature variables and directed edges represent causal relationships. A Bayesian network learning method is used to learn and correct the edge weights in the initial causal graph using data from the damage event feature package, thereby determining the final causal graph structure. In the final causal diagram, the direct and indirect causal effects of acoustic feature changes on geometric feature evolution are calculated using intervention analysis algorithms, and the main influencing factors are identified. Based on the value range of the main influencing factors and the intensity of the causal effect, and in accordance with the preset damage type determination rules, the damage type label and the corresponding confidence level value are output to form the evaluation result.

[0012] Furthermore, the system also includes: The model coupling module is used to map the evaluation results to the three-dimensional finite element model of the pressure vessel. According to the failure mode corresponding to the damage type label, the corresponding load condition and boundary condition are activated in the three-dimensional finite element model to calculate the stress concentration factor and plastic strain increment of the damaged area. The iterative correction module is used to feed back the stress concentration factor and the plastic strain increment to the damage assessment model based on causal reasoning, correct the weight coefficients of the influencing factors in the causal graph, and re-output the adjusted assessment results. The trajectory analysis module is used to perform time series analysis on the adjusted assessment results, extract the upward trend slope and fluctuation amplitude of the damage severity index in the assessment results, and construct the damage evolution trajectory curve. The life assessment module is used to compare the damage evolution trajectory curve with the pressure vessel design life curve, identify the degree of deviation between the damage evolution trajectory curve and the pressure vessel design life curve, classify the health level according to the degree of deviation, and generate an assessment report containing the health level and health index. The evaluation results are mapped to a three-dimensional finite element model of the pressure vessel. Based on the failure mode corresponding to the damage type label, the corresponding load conditions and boundary conditions are activated in the three-dimensional finite element model to calculate the stress concentration factor and plastic strain increment of the damaged region, including: Read the three-dimensional geometric model and material property parameters of the pressure vessel and import them into the three-dimensional finite element model; Based on the damage type label in the assessment results, a preset failure mode library is called to match the corresponding load conditions and boundary conditions. The load conditions include internal pressure load, temperature load and mechanical vibration load. In the three-dimensional finite element model, the location of the damaged area is located, and the mesh is refined at the location to improve the calculation accuracy; Submit the solution calculation to obtain the stress distribution cloud map and strain distribution cloud map of the damaged area; Extract the maximum stress value and the nominal stress value of the damaged area from the stress distribution cloud map, and calculate the ratio of the maximum stress value to the nominal stress value as the stress concentration factor; extract the equivalent plastic strain value of the damaged area from the strain distribution cloud map as the plastic strain increment.

[0013] Further, the step of feeding back the stress concentration factor and the plastic strain increment to the damage assessment model based on causal reasoning, correcting the weight coefficients of the influencing factors in the causal graph, and re-outputting the adjusted assessment results includes: The stress concentration factor and the plastic strain increment are added as new observed variables to the causal graph of the damage assessment model based on causal reasoning. Calculate the mutual information values ​​between the stress concentration factor and the plastic strain increment and the original characteristic variables to measure the correlation strength between them; Based on the mutual information value, adjust the weight coefficient of the corresponding edge in the causal graph, increasing the weight of edges with strong correlation and decreasing the weight of edges with weak correlation. Using the causal graph with corrected weighting coefficients, the data in the damage event feature package is recalculated to update the damage type label and confidence value, and the adjusted evaluation results are output.

[0014] Furthermore, the adjusted assessment results are subjected to time series analysis to extract the upward trend slope and fluctuation amplitude of the damage severity index in the assessment results, and to construct a damage evolution trajectory curve, including: The damage severity index is extracted from the adjusted assessment results. The damage severity index is calculated by multiplying the confidence level value and the stress concentration factor. Arrange the severity of injury indicators in chronological order to form a time series of injury severity. A linear regression algorithm is applied to the time series of damage severity to fit a straight line, and the slope of the straight line is the slope of the upward trend. The sliding window variance calculation method is applied to the time series of damage severity to calculate the variance value within each window. The set of all variance values ​​is the fluctuation amplitude. Using time as the horizontal axis and the severity of damage as the vertical axis, the upward trend slope and fluctuation amplitude are plotted on the coordinate system, and the data points are connected to form a damage evolution trajectory curve.

[0015] Compared with the prior art, the beneficial effects of the present invention are: The nodes of the graph neural network are composed of the locations of acoustic emission sensors and the centers of defect regions in the digital image. The edge weights are determined by the arrival time difference of the acoustic emission signals and the spatial distance of the image. Graph convolution operations are used to learn the topological relationships between multimodal features, generating graph embedding vectors that contain structural health information. This node configuration integrates the spatial attributes of the acoustic detection hardware location and the geometric center of the visual defect. The fusion calculation of edge weights characterizes the physical correlation characteristics of multimodal data, graph convolution operations extract the implicit correlations of different modal features in the spatial structure, and the graph embedding vectors can carry the overall state information of the pressure vessel structure, achieving a structured representation of multimodal features.

[0016] Using damage event feature packages as input, a causal graph is constructed among multimodal features to analyze the direct and indirect influencing factors of acoustic feature changes on geometric feature evolution. The output includes damage type labels and confidence levels. The causal graph eliminates irrelevant interference factors among multimodal features, and the distinction between direct and indirect influencing factors clarifies the interaction paths between acoustic and geometric features. The evaluation model outputs corresponding labels and confidence levels based on causal relationships, and the evaluation results reflect the intrinsic interaction logic between features, differing from the superficial correlation discrimination method of conventional correlation analysis. Attached Figure Description

[0017] Figure 1 This is a timing diagram of a pressure vessel health assessment system based on multimodal data fusion as described in this invention; Figure 2 A flowchart for time synchronization and spatial registration methods; Figure 3 A flowchart of graph neural network topology learning and graph embedding vector generation methods; Figure 4 A heatmap of the intensity of causal effects characteristic of multimodal features; Figure 5 This is the time-series evolution curve of stress concentration factor and plastic strain increment in the damaged area of ​​the pressure vessel. Detailed Implementation

[0018] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] See Figure 1The synchronization fusion module is responsible for acquiring acoustic emission time-domain signals and digital images of the weld area during the operation of the pressure vessel, constructing a multimodal raw dataset, and performing time synchronization and spatial registration on this dataset. Then, energy entropy features and geometric moment features are extracted to form a feature fusion set. The topology learning module receives this feature fusion set and inputs it into a pre-constructed graph neural network. The nodes of this network are composed of the locations of the acoustic emission sensors and the centers of the defect areas identified in the digital images. The edge weights between nodes are determined by the arrival time difference of the acoustic emission signals and the spatial distance between the images. Through multi-layer graph convolution operations, this module learns the topological relationships between multimodal features, ultimately generating a graph embedding vector representing the structural health status. The event detection module dynamically monitors the temporally arranged graph embedding vectors. By calculating the cosine similarity between adjacent graph embedding vectors and analyzing their statistical characteristics, it identifies abrupt change points representing potential damage events, and then extracts multimodal features within the time period before and after the abrupt change point to form a damage event feature package. The causal assessment module is driven by the damage event feature package and calls a damage assessment model based on causal reasoning. This model constructs and learns the causal graph between multimodal features, quantifies the direct and indirect causal effects of acoustic feature changes on geometric feature evolution, and finally outputs assessment results containing specific damage type labels and their confidence levels.

[0020] In one embodiment of the present invention, a synchronous fusion module acquires acoustic emission time-domain signals and digital images of the weld area to construct a multimodal raw dataset. Acoustic emission time-domain signals are acquired by an acoustic emission sensor array deployed on the surface of the pressure vessel. The acquisition mode of the acoustic emission sensor array is activated, and the sampling frequency is set to a preset high-frequency value. This preset high-frequency value ensures that the high-frequency elastic wave components generated during the operation of the pressure vessel can be captured. The acoustic emission sensor array continuously receives the elastic wave signals and converts them into discrete voltage time-domain signals, which serve as the acoustic emission time-domain signals. Simultaneously, digital images of the weld area are acquired by a digital X-ray imaging system of the pressure vessel. The scanning mode of the digital X-ray imaging system is synchronously activated, performing multi-angle exposure imaging on the weld area of ​​the pressure vessel to obtain a grayscale image sequence containing information about the internal structure of the weld. The grayscale image sequence is then stitched and corrected. The correction process eliminates geometric distortion caused by differences in imaging angles, generating a digital image covering the entire weld area with unified spatial coordinates. The acoustic emission time-domain signal and digital image are aligned according to their respective timestamps. The timestamp alignment operation will discard data frames whose timestamps do not match, and retain acoustic emission time-domain signal and digital image frames with completely consistent timestamps. These retained data are combined to form a multimodal raw dataset containing acoustic and visual features.

[0021] In some embodiments, temporal synchronization and spatial registration are performed on the multimodal raw dataset to establish an association index between acoustic emission events and their corresponding spatial locations in digital images. See also Figure 2 A high-precision external clock source was selected as a unified clock reference from the multimodal raw dataset. The internal clock signals of all acoustic emission sensors and the internal clock signal of the digital X-ray imaging system were synchronized and calibrated to this unified clock reference. A Cartesian coordinate system was established in the digital image, with the geometric center of the pressure vessel cylinder as the origin and the extension direction of the main weld around the cylinder as the X-axis. The linear transformation relationship between the pixel coordinates of the digital image and the actual physical space coordinates was pre-calibrated using a calibration plate. In the acoustic emission time-domain signal, the arrival time of the waveform of each acoustic emission event and the sensor number of the identified waveform were identified using a peak detection algorithm. Combined with the known physical installation coordinates of the acoustic emission sensors on the pressure vessel surface, a time-difference positioning algorithm was used to calculate the spatial coordinates of the acoustic emission source in the pressure vessel shell. The nearest neighbor matching algorithm maps the calculated spatial coordinates of the acoustic emission source to the digital image in the Cartesian coordinate system. The pixel point in the pixel matrix of the digital image that is closest to the Euclidean distance of the acoustic emission source's spatial coordinates is found. This pixel point is defined as the associated pixel point, thus establishing a one-to-one correspondence between the acoustic emission event and the associated pixel point in the digital image. This correspondence is recorded as the association index.

[0022] In practice, the acoustic emission time-domain signal is decomposed using wavelet packets to extract energy entropy features. The 'db4' wavelet from the Daubechies wavelet family is selected as the preset wavelet basis function, and the decomposition level is set to 5 levels. A complete wavelet packet decomposition is performed on each segment of the acoustic emission time-domain signal to obtain the energy spectrum covering different frequency sub-bands. The ratio of the energy value of the frequency band corresponding to each wavelet packet decomposition node to the total signal energy is calculated, and the set of all ratios constitutes an energy distribution vector. The Shannon entropy of this energy distribution vector is calculated and used as the energy entropy feature characterizing the complexity of the time-frequency energy distribution of the acoustic emission signal. The calculation formula is: in: Representing the The energy percentage of each wavelet packet decomposition frequency band. This represents a base-2 logarithmic operation. Geometric moment features of defect edges are extracted after morphological segmentation of the digital image. The digital image is then converted to grayscale, transforming color information into grayscale intensity information, followed by binarization using the Otsu adaptive thresholding algorithm. A combined morphological filter, employing opening followed by closing operations, is used to process the binary image. The opening operation uses a circular structuring element with a radius of 3 pixels to remove small white noise points, while the closing operation uses the same structuring element to fill tiny black holes. All defect regions in the image are identified and marked using a region growing-based connected component labeling algorithm, and the pixel coordinate sequence of the contour edge of each marked defect region is extracted. The second-order central moment of the contour edge pixel coordinate sequence is calculated as a geometric moment feature, which describes the shape and extension characteristics of the defect edge. Finally, the extracted energy entropy features and geometric moment features are arranged and combined according to the order specified by the previously generated association index to form the final feature fusion set. The feature fusion set is a structured dataset where each data sample simultaneously contains time-frequency domain acoustic features and spatial domain geometric features.

[0023] It is understandable that time synchronization ensures a strict temporal correspondence between acoustic emission events and digital images, while spatial registration establishes a spatial mapping between acoustic signal sources and visual image regions. Association indexing is the key link connecting multimodal data. Wavelet packet decomposition can extract stable frequency band energy distribution features from non-stationary acoustic emission time-domain signals, and morphological segmentation can stably separate the geometric morphological features of defects from complex digital image backgrounds. Energy entropy features quantify the uniformity of the frequency band distribution of signal energy, and geometric moment features quantify the shape inertia characteristics of defect edges. The generation process of the feature fusion set unifies sensor information with different physical meanings and dimensions into a spatiotemporally consistent feature expression framework, providing well-structured and information-rich input data for subsequent topological association learning.

[0024] In one embodiment of the invention, the topology learning module inputs a feature fusion set into a pre-constructed graph neural network. See also... Figure 3The adjacency matrix of the graph neural network is initialized, storing the connection relationships and weights between nodes in the graph. The set of nodes in the graph consists of two types of coordinates: the first type is the physical installation coordinates of the acoustic emission sensors fixed on the surface of the pressure vessel; the second type is the physical coordinates obtained by transforming the pixel coordinates of the centers of various defect regions identified from the digital image. All coordinates are represented in a unified container reference coordinate system. For any pair of nodes in the graph, the edge weight between the pair is calculated. The edge weight is determined by both temporal and spatial weights. The temporal weight is derived from the arrival time difference of the acoustic emission signal from one sensor node to another, calculated by analyzing the waveform of the same event in the acoustic emission time domain signal at the times it arrives at different sensors. The spatial weight is derived from the pixel distance between the two nodes in the digital image coordinate system, calculated by euclidean distance between the pixel coordinates of the two nodes. The calculated arrival time difference and pixel distance are normalized respectively, scaling the original values ​​to the range of 0 to 1. Then, the normalized time weight and the normalized spatial weight are weighted and added together to obtain the final edge weight value representing the strength of the association between node pairs. in: Represents the edge weight between node pairs. Represents the normalized arrival time difference. Represents the normalized pixel distance. and It is an adjustable coefficient used to adjust the relative importance of time and space factors, and satisfies The energy entropy features and geometric moment features in the feature fusion set are mapped to the corresponding graph nodes, which serve as the initial feature vectors for the corresponding nodes. One acoustic emission sensor node corresponds to one energy entropy feature scalar, and one defect region center node corresponds to one feature vector composed of geometric moment features.

[0025] In some embodiments, multi-layer graph convolution operations are performed to learn the topological relationships between multimodal features. Each layer of graph convolution operations performs the same computational steps, aggregating the current feature vector of each node in the graph with the feature vectors of all its neighboring nodes. The aggregation operation is specifically a weighted summation, where the feature vectors of each neighboring node are multiplied by their corresponding edge weights and then summed. The result of the weighted summation is then linearly combined with the node's own feature vector. The resulting feature vector is then passed through a non-linear activation function, which introduces non-linear expressive power into the model. The feature vector after the non-linear activation function becomes the new feature vector for that node in the next layer of graph convolution. By stacking multiple layers of such graph convolution operations, the feature vector of each node in the graph can continuously aggregate information from its multi-hop neighboring nodes, thereby learning complex, graph-topology-based relationships between node features. After all preset layers of graph convolution operations are completed, a global pooling operation is performed on the final feature vectors of all nodes. The global pooling operation integrates the feature vectors of all nodes in the graph into a fixed-dimensional vector. The integration method can be taking the average, maximum, or sum of the feature vectors of all nodes. This fixed-dimensional vector generated by the global pooling operation is called the graph embedding vector. The graph embedding vector comprehensively encodes the acoustic state, visual defect state, and spatial topological relationships of the pressure vessel at a specific moment. Therefore, the graph embedding vector contains information reflecting the structural health status.

[0026] In one embodiment of the present invention, the event detection module performs temporal accumulation of graph embedding vectors to identify potential damage events. Graph embedding vectors generated by the topology learning module at different time points are arranged in chronological order of their generation to form a graph embedding vector sequence. Each element in the graph embedding vector sequence represents the health status feature encoding of the pressure vessel at a specific sampling time. For each element in the graph embedding vector sequence, the cosine similarity between the graph embedding vector at the current time and the graph embedding vector at the previous adjacent time is calculated. The cosine similarity measures the difference in direction between the two vectors, and its value range is [-1, 1]. The closer the value is to 1, the more similar the system states represented by the two graph embedding vectors are. The cosine similarity of all adjacent element pairs in the graph embedding vector sequence is calculated sequentially to obtain a cosine similarity sequence corresponding to the length of the graph embedding vector sequence minus one. The cosine similarity sequence reflects the dynamic change trend of the pressure vessel structural health status at continuous time points.

[0027] In some embodiments, a sliding window statistical method is applied to the cosine similarity sequence to identify abrupt changes in the graph embedding vectors. A fixed-length sliding time window is defined, which moves progressively along the time axis along the cosine similarity sequence. Within each sliding time window, the arithmetic mean and standard deviation of all cosine similarity values ​​covered by the window are calculated. The arithmetic mean represents the average stability of state changes within that time period, and the standard deviation represents the fluctuation range of state changes. When the sliding window moves to a specific time point... Calculate the time point. The corresponding instantaneous cosine similarity value The instantaneous cosine similarity value The arithmetic mean calculated within the current sliding window and standard deviation Compare the values. Set a positive integer multiplier. When the conditions are met At that time, determine the point in time. The cosine similarity value at the marked time point is significantly lower than the recent average. These are potential mutation points, meaning that the structural health characteristics of the pressure vessel have undergone an abnormal change at this particular moment. The identified mutation points... Using a central time point as the starting point, we trace back a predetermined period of historical data and simultaneously extract a predetermined period of future data, which together constitute an analysis time period centered on the point of change.

[0028] From the raw data storage corresponding to this analysis period, the original acoustic emission time-domain signal and digital image data for that period are extracted. Energy entropy features and geometric moment features for the same time period are extracted from the feature fusion set. The extracted original acoustic emission time-domain signal, digital image, energy entropy features, and geometric moment features are packaged in chronological order and combined to form a structured damage event feature package. This damage event feature package fully encapsulates the multimodal raw data and high-level features for a period before and after the suspected damage event.

[0029] In one embodiment of the present invention, the causal assessment module takes a damage event feature package as input and drives a damage assessment model based on causal reasoning. Key multimodal feature change trends are extracted from the damage event feature package as input variables for causal reasoning. These input variables mainly include the change sequence of energy entropy calculated from acoustic emission signals within the event time window, and the change sequence of the geometric moments of the defect region calculated from digital images within the corresponding time window. Based on domain knowledge of pressure vessel failure mechanisms and statistical data from numerous historical damage cases, causal relationship hypotheses between multimodal features are pre-constructed, forming an initial causal graph. In this initial causal graph, nodes represent specific feature variables such as "energy entropy increase rate" and "geometric moment change," while directed edges represent the assumed causal relationship direction between variables. For example, a directed edge from the "energy entropy increase rate" node to the "geometric moment change" node indicates that the "energy entropy increase rate" is assumed to be a cause of the "geometric moment change."

[0030] In some embodiments, a Bayesian network learning method is employed to learn the structure and parameters of the initial causal graph. The edge weights in the initial causal graph are learned and corrected using a large amount of sample data from the damage event feature package. The edge weights characterize the strength and conditional probability of causal relationships. Through the learning process, edges that are not supported by the data can be removed, edges that strongly support causal relationships can be strengthened, and new causal relationships revealed by the data may be added, thereby determining a final causal graph structure that better fits the observed data. Based on the final causal graph structure, an intervention analysis algorithm is used to quantitatively analyze the causal effect of acoustic feature changes on geometric feature evolution. Intervention analysis assesses the strength of the causal effect by calculating the difference between the conditional probability distribution of the outcome variable (such as the change in geometric moments) when a causal variable is "intervened" (e.g., fixing the "rate of increase in energy entropy" to a specific value) and the marginal probability distribution when no intervention is implemented. Direct causal effect refers to the direct influence of the causal variable on the outcome variable, while indirect causal effect refers to the influence of the causal variable on the outcome variable by affecting other mediating variables. Through calculation, direct and indirect influencing factors that have a major impact on the evolution of geometric features can be identified. The identification of major influencing factors is based on the magnitude of their causal effect size. It can be done through the formula: in: Indicating in intervention variables Make it take the value Under the condition, outcome variable The probability distribution, and Representative variable The formula calculates two different intervention levels. from Become hour, The average difference in the distribution, i.e. the average causal effect.

[0031] Optionally, the initial causal graph can be constructed using expert knowledge manuals or by text mining and rule extraction from historical nondestructive testing reports and post-failure analysis reports. The Bayesian network learning algorithm can be based on a score-based search method, such as using the BIC scoring function and hill-climbing algorithm, or on a method based on conditional independence testing. In practice, the calculation of the "do" operator in intervention analysis typically involves finding a suitable set of covariates for adjustment using the "backdoor criterion" or "frontdoor criterion" on the causal graph, or by simulating probabilistic inference after model fitting. The value ranges of the main influencing factors can be divided into multiple levels, such as "low," "medium," and "high." Damage type determination rules can exist in the form of lookup tables or decision trees, mapping the value ranges of different influencing factors and the combination of causal effect strengths to specific damage type labels, such as "fatigue crack propagation," "corrosion thinning," and "welding defect cracking," and assigning a confidence value based on posterior probability calculation to each determination result, ultimately forming an evaluation result containing damage type labels and confidence levels. The evaluation results can be output in a structured format.

[0032] Understandably, damage assessment models based on causal reasoning go beyond traditional correlation analysis, aiming to reveal the inherent causal mechanisms between multimodal features. Extracting trend sequences from damage event feature packages as input enables the model to analyze dynamic evolution processes rather than static snapshots. The initial causal graph incorporates prior physical knowledge and engineering experience, providing a reasonable starting point for learning. Bayesian network learning methods utilize real-world data to verify and revise prior causal assumptions, ensuring the final causal graph possesses both physical interpretability and data-driven validity. Intervention analysis algorithms can quantitatively isolate the pure causal contributions of different features to damage evolution, helping to identify true driving factors rather than accompanying phenomena. Classification based on the values ​​and effect strengths of major influencing factors against pre-defined rules ensures that the assessment results not only provide the damage type but also quantify the degree of certainty through confidence measurements. The entire process transforms multimodal correlated data into damage assessment conclusions with causal interpretation.

[0033] Table 1: Examples of Initial Cause-Effect Graph Nodes and Partial Causal Relationship Assumptions

[0034] See Figure 4The figure presents the distribution of causal effect strength among multimodal features. Rows represent causal variables (acoustic and visual feature dimensions), columns represent outcome variables, and cell values ​​and colors jointly characterize the causal effect strength (colors gradually change from blue to red, corresponding to an increase in strength from 0 to 1). This quantifies the causal driving role of each feature variable in the damage evolution process: Strong causal effect identification: For example, the causal effect strength of "microcrack initiation" on "damage type" reaches 0.90, and the effect strength of "stress level mutation" on "damage type" is 0.85, indicating that these are the core driving factors for damage type determination; the effect strength of "energy entropy increase rate" on "geometric moment change" is 0.85, revealing the strong direct influence of acoustic feature changes on the evolution of geometric defects. Weak correlation and autocorrelation features: Most variables have a causal effect of 0 on themselves (e.g., "energy entropy increase rate" → "energy entropy increase rate"), consistent with the assumption that self-driven variables are meaningless in causal inference; "damage type," as the final evaluation label, generally has extremely low effect strength on other feature variables, reflecting its positioning as an output result. Multimodal interaction features: Significant cross-effects exist between acoustic features (such as "energy entropy increase rate" and "high-frequency energy ratio") and visual features (such as "geometric moment change" and "defect morphology change"). For example, the effect strength of "defect morphology change" on "energy entropy increase rate" is 0.80, verifying the necessity of multimodal data fusion in damage assessment and providing a quantitative basis for subsequent intervention analysis and damage type determination. Heatmaps are not only an intuitive mapping of the causal graph structure after Bayesian network learning, but also the input foundation for intervention analysis algorithms. By quantifying the pure causal contribution of each feature, key driving factors of damage evolution can be accurately identified, providing interpretable quantitative support for damage type determination and remaining life assessment of pressure vessels.

[0035] In one embodiment of the present invention, the model coupling module maps the evaluation results output by the causal evaluation module to the three-dimensional finite element model of the pressure vessel. The three-dimensional geometric model file and material property parameter file of the pressure vessel are read and imported into the three-dimensional finite element analysis software to establish a three-dimensional finite element model for calculation. Based on the damage type labels contained in the evaluation results, the system's preset failure mode library is called. The failure mode library stores typical load conditions and boundary conditions corresponding to different damage types. The load conditions and boundary conditions corresponding to the current damage type label are matched. These load conditions include internal pressure loads, temperature loads, and mechanical vibration loads. The location of the damage area is located in the three-dimensional finite element model. The location information comes from the defect space coordinates or coordinates transformed by the associated index carried in the evaluation results. Local mesh refinement is performed at the location of the damage area. Local mesh refinement improves the calculation accuracy of this area by increasing the element density. The three-dimensional finite element model is submitted for solution calculation to obtain the stress distribution cloud map and strain distribution cloud map of the pressure vessel under specified load and boundary conditions. The maximum stress value within the damage area is extracted from the stress distribution cloud map. Nominal stress value far from the damage area The ratio of the maximum stress value to the nominal stress value is calculated as the stress concentration factor. The equivalent plastic strain value within the damaged area is extracted from the strain distribution contour map. As an increment of plastic strain.

[0036] In some embodiments, the iterative correction module feeds back the stress concentration factor and the plastic strain increment to the damage assessment model based on causal reasoning. The calculated stress concentration factor... With plastic strain increment Two new observation variable nodes are added to the causal graph structure of the damage assessment model based on causal inference. The mutual information values ​​between the newly added stress concentration factor node, plastic strain increment node, and the existing feature variable nodes in the causal graph are calculated. These mutual information values ​​measure the degree of interdependence between the two random variables. Based on the calculated mutual information values, the weight coefficients of the edges connecting the new variable nodes and the existing feature variable nodes in the causal graph are adjusted, increasing the weight coefficients of edges with larger mutual information values ​​and decreasing the weight coefficients of edges with smaller mutual information values. Using the causal graph structure with adjusted weight coefficients, causal inference calculations are re-performed on the input data in the damage event feature package, updating the damage type labels and corresponding confidence scores, and outputting the adjusted assessment results that incorporate mechanical simulation feedback information.

[0037] In practice, the trajectory analysis module performs time-series analysis on the adjusted evaluation results output by the iterative correction module. It extracts the damage severity index from the adjusted evaluation results. The confidence score of the damage type in the assessment results Stress concentration factor calculated by the model coupling module The product is calculated as follows: Multiple injury severity indicators generated chronologically over a period of time are arranged to form a time series of injury severity. A linear regression algorithm is then applied to this time series to fit a straight line representing the trend of injury development. The slope of the fitted line is... This represents the upward trend slope of the injury severity index. A sliding window variance calculation method is applied to the injury severity time series, setting a fixed-length sliding window and calculating the variance of the injury severity index data within each window. The x-axis is time, and the y-axis is the injury severity index. Establish a coordinate system for the vertical axis, and calculate the upward trend slope. The data points are marked on the graph, and the fluctuation amplitude sequence is marked as error bands or points near the corresponding data points. The damage severity index data points at each time point are connected to form a complete damage evolution trajectory curve.

[0038] The model coupling module effectively bridges the gap between data-driven causal assessment and physics-based mechanical simulation. Stress concentration factor and plastic strain increment are key parameters for quantifying the local mechanical effects of damage. The iterative correction module measures the statistical dependence of new observed variables on existing characteristic variables through mutual information and adjusts the weights of the causal graph accordingly, enabling the causal inference model to absorb mechanical simulation results and achieve self-optimization of the assessment results. The trajectory analysis module transforms discrete assessment points into continuous evolutionary trends. The upward trend slope quantitatively describes the speed of damage development, while the fluctuation amplitude reflects the instability of the development process. The damage evolution trajectory curve visually demonstrates the dynamic behavior of damage over time. The life assessment module upgrades from state assessment to life prediction by comparing the actual evolution trajectory with the theoretical design life curve. The degree of deviation provides a quantitative basis for classifying health levels, and the final assessment report integrates multi-source information and multi-level analysis results.

[0039] Optionally, in the model coupling module, local mesh refinement can be performed automatically based on error estimation, or manually set according to a preset mesh size classification strategy. The load cases in the failure mode library can be static or time-varying load spectra. In the iterative correction module, mutual information can be calculated using a histogram-based method or a k-nearest neighbor-based algorithm. When adjusting weighting coefficients, linear interpolation or a nonlinear mapping method based on the sigmoid function can be used. In the trajectory analysis module, the linear regression algorithm can use the least squares method. The length of the sliding window can differ from the window length in the event detection module to accommodate different timescales required for trend analysis and fluctuation analysis. The health level classification in the life assessment module can set multiple thresholds based on the degree of deviation. The assessment report can include a damage evolution trajectory curve, a list of key parameters, health level and health index, and a summary of maintenance recommendations.

[0040] See Figure 5 In the mechanical simulation analysis of pressure vessel damage evolution, the stress concentration factor and the increment of plastic strain are the core time-series indicators characterizing the mechanical effects of local damage. Together, they constitute the damage evolution trend curve of stage 2. Specifically, the solid line represents the stress concentration factor, with an initial value of approximately 1.2. It shows a continuous upward trend with the advancement of time steps, accompanied by small fluctuations, and finally climbs to approximately 2.7 at 100 hours, reflecting the gradual intensification of stress concentration and the dynamic characteristics of local stress redistribution in the damaged area. The dashed line represents the increment of plastic strain, with an initial value of approximately 0.001. It also increases monotonically with the advancement of time steps, finally reaching approximately 0.021 at 100 hours, reflecting the cumulative process of plastic deformation in the damaged area. The trends of the two curves are highly synchronized, both showing a stable growth trend, indicating a significant positive correlation between the increase in stress concentration and the accumulation of plastic strain. This pattern provides a direct mechanical basis for the iterative correction module to correct the weights of the causal graph and the trajectory analysis module to construct the damage evolution trajectory. At the parameter level, the time step is in hours, covering an observation period of 0 to 100 hours. The fluctuation range of the stress concentration factor reflects the mechanical response disturbance caused by load fluctuation or local crack propagation, while the smooth growth of the plastic strain increment reflects the gradual accumulation characteristics of the material's plastic deformation.

[0041] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A pressure vessel health assessment system based on multimodal data fusion, characterized in that, include: The synchronous fusion module is used to acquire acoustic emission time-domain signals and digital images of the weld area, construct a multimodal raw dataset, perform time synchronization and spatial registration on the multimodal raw dataset, and generate a feature fusion set. The topology learning module is used to input the feature fusion set into a pre-constructed graph neural network. The nodes of the graph neural network are composed of the location of the acoustic emission sensor and the center of the defect region in the digital image. The edge weights are determined by the arrival time difference of the acoustic emission signal and the spatial distance of the image. The topological association between multimodal features is learned through graph convolution operations to generate a graph embedding vector containing the structural health status. The event detection module is used to perform temporal accumulation on the graph embedding vectors, calculate the cosine similarity between graph embedding vectors at different time steps, identify abrupt change points in the graph embedding vectors, take the time corresponding to the abrupt change point as the trigger time of the potential damage event, and extract multimodal features before and after the trigger time to form a damage event feature package. The causal assessment module is used to drive a causal reasoning-based damage assessment model by taking the damage event feature package as input. The causal reasoning-based damage assessment model constructs a causal graph between multimodal features, analyzes the direct and indirect influencing factors of acoustic feature changes on geometric feature evolution, and outputs assessment results including damage type labels and confidence levels.

2. The pressure vessel health assessment system based on multimodal data fusion according to claim 1, characterized in that, The process involves acquiring acoustic emission time-domain signals and digital images of the weld area, constructing a multimodal raw dataset, performing time synchronization and spatial registration on the multimodal raw dataset, and generating a feature fusion set, including: Acoustic emission time-domain signals are acquired by an array of acoustic emission sensors arranged on the surface of the pressure vessel, and digital images of the weld area are acquired by a digital X-ray imaging system of the pressure vessel, forming a multimodal raw dataset containing acoustic and visual features. The multimodal raw dataset is time-synchronized and spatially registered to establish an association index between acoustic emission events and corresponding spatial locations in digital images. The acoustic emission time-domain signal is decomposed by wavelet packet to extract energy entropy features, and the digital image is morphologically segmented to extract geometric moment features of defect edges, forming a feature fusion set that includes time-frequency domain acoustic features and spatial domain geometric features. The process involves acquiring acoustic emission time-domain signals through an array of acoustic emission sensors arranged on the surface of the pressure vessel, and simultaneously acquiring digital images of the weld area through a digital X-ray imaging system of the pressure vessel, forming a multimodal raw dataset containing both acoustic and visual features, including: The acoustic emission sensor array is activated in acquisition mode, and the sampling frequency is set to a preset high-frequency value. The elastic wave signal generated during the operation of the pressure vessel is continuously received, and the elastic wave signal is converted into a discrete voltage time-domain signal as the acoustic emission time-domain signal. The scanning mode of the digital X-ray imaging system is activated simultaneously to perform multi-angle exposure imaging of the weld area of ​​the pressure vessel and obtain a grayscale image sequence containing the internal structure of the weld. The grayscale image sequence is stitched and corrected to eliminate geometric distortion caused by imaging angle deviation and generate a digital image covering the entire weld area. The acoustic emission time-domain signal and digital image are aligned according to timestamps. Data frames with mismatched timestamps are removed, and the acoustic emission time-domain signal and digital image with consistent timestamps are retained to form a multimodal original dataset containing acoustic and visual features.

3. The pressure vessel health assessment system based on multimodal data fusion according to claim 2, characterized in that, The step of performing time synchronization and spatial registration on the multimodal raw dataset to establish an association index between acoustic emission events and corresponding spatial locations in digital images includes: A unified clock reference is selected in the multimodal raw dataset, and the clock signals of all acoustic emission sensors and the clock signal of the digital X-ray imaging system are calibrated to the clock reference. A Cartesian coordinate system is established in the digital image, with the geometric center of the pressure vessel as the origin and the weld direction as the coordinate axis direction, to calibrate the conversion relationship between the pixel coordinates of the digital image and the actual physical coordinates. In the acoustic emission time domain signal, the arrival time and sensor number of the acoustic emission event are identified by the peak detection algorithm. Combined with the physical installation coordinates of the acoustic emission sensor, the spatial coordinates of the acoustic emission source of the acoustic emission event are calculated. By using the nearest neighbor matching algorithm, the spatial coordinates of the acoustic emission source are mapped to the Cartesian coordinate system of the digital image. The nearest pixel is found as the associated pixel, and the correspondence between the acoustic emission event and the associated pixel is established to generate an association index.

4. The pressure vessel health assessment system based on multimodal data fusion according to claim 3, characterized in that, The process involves extracting energy entropy features from the acoustic emission time-domain signal via wavelet packet decomposition, and extracting geometric moment features of defect edges from the digital image via morphological segmentation, forming a feature fusion set that includes both time-frequency domain acoustic features and spatial domain geometric features. By selecting preset wavelet basis functions and decomposition levels, wavelet packet decomposition is performed on the acoustic emission time-domain signal to obtain the energy spectrum of different frequency bands; Calculate the ratio of energy in each frequency band to the total energy, construct an energy distribution vector, and calculate the Shannon entropy of the energy distribution vector as an energy entropy feature; Digital images are subjected to grayscale and binarization processing, and image noise is removed by a combination of opening and closing morphological filters. Defect regions in an image are identified using a connected component labeling algorithm. The contour edges of the defect regions are extracted, and the second-order geometric moments of the contour edges are calculated as geometric moment features. The energy entropy features and geometric moment features are arranged in the order of their associated indices and combined to form a feature fusion set that includes time-frequency domain acoustic features and spatial domain geometric features.

5. A pressure vessel health assessment system based on multimodal data fusion according to claim 1, characterized in that, The fused feature set is input into a pre-constructed graph neural network. The nodes of the graph neural network are formed by the locations of the acoustic emission sensors and the centers of defect regions in the digital image. The edge weights are determined by the arrival time difference of the acoustic emission signals and the spatial distance of the image. Topological relationships between multimodal features are learned through graph convolution operations to generate graph embedding vectors containing structural health status, including: Initialize the adjacency matrix of the graph neural network, and use the coordinates of the acoustic emission sensor position and the center of the defect region as the node set of the graph; For each node pair, the arrival time difference of the acoustic emission signal from one sensor to another is calculated as the time weight, and the pixel distance between the two nodes in the digital image is calculated as the spatial weight. The time weight and spatial weight are normalized and then added together to obtain the edge weight. The energy entropy features and geometric moment features in the feature fusion set are mapped to the corresponding nodes, which serve as the initial feature vectors of the nodes. Perform multi-layer graph convolution operations. Each layer of graph convolution operations performs a weighted summation and non-linear activation on the initial feature vector of a node and the feature vectors of its neighboring nodes to update the node's feature vector. After all graph convolution operations are completed, the feature vectors of all nodes are globally pooled to obtain a fixed-dimensional graph embedding vector, which contains structural health information.

6. The pressure vessel health assessment system based on multimodal data fusion according to claim 1, characterized in that, The graph embedding vectors are accumulated over time, and the cosine similarity between graph embedding vectors at different time steps is calculated to identify abrupt change points in the graph embedding vectors. The time corresponding to the abrupt change point is taken as the trigger time of the potential damage event. Multimodal features before and after the trigger time are extracted to form a damage event feature package, including: Arrange the graph embedding vectors in chronological order to form a graph embedding vector sequence; For each element in the graph embedding vector sequence, calculate its cosine similarity with the previous element to obtain the cosine similarity sequence; Apply sliding window statistics to the cosine similarity sequence, calculate the mean and standard deviation of the cosine similarity within the window, and mark the corresponding time point as a change point when the cosine similarity at a certain time is less than the mean minus a certain number of times the standard deviation. Centered on the mutation point, time periods of preset length are selected both forward and backward. Acoustic emission time-domain signals and digital images within the corresponding time periods are extracted from the multimodal raw dataset. At the same time, energy entropy features and geometric moment features within the corresponding time periods are extracted from the feature fusion set and combined to form a damage event feature package.

7. A pressure vessel health assessment system based on multimodal data fusion according to claim 1, characterized in that, Using the damage event feature package as input, a damage assessment model based on causal reasoning is driven. This model constructs a causal graph among multimodal features, analyzes the direct and indirect influencing factors of acoustic feature changes on geometric feature evolution, and outputs assessment results including damage type labels and confidence levels, including: The energy entropy change trend of acoustic emission signals and the geometric moment change trend of digital images are extracted from the feature package of damage events and used as input variables for causal inference. Based on domain knowledge and historical data statistics, we construct causal relationship hypotheses among multimodal features and form an initial causal graph, where nodes represent feature variables and directed edges represent causal relationships. A Bayesian network learning method is used to learn and correct the edge weights in the initial causal graph using data from the damage event feature package, thereby determining the final causal graph structure. In the final causal diagram, the direct and indirect causal effects of acoustic feature changes on geometric feature evolution are calculated using intervention analysis algorithms, and the main influencing factors are identified. Based on the value range of the main influencing factors and the intensity of the causal effect, and in accordance with the preset damage type determination rules, the damage type label and the corresponding confidence level value are output to form the evaluation result.

8. A pressure vessel health assessment system based on multimodal data fusion according to claim 1, characterized in that, The system also includes: The model coupling module is used to map the evaluation results to the three-dimensional finite element model of the pressure vessel. According to the failure mode corresponding to the damage type label, the corresponding load condition and boundary condition are activated in the three-dimensional finite element model to calculate the stress concentration factor and plastic strain increment of the damaged area. The iterative correction module is used to feed back the stress concentration factor and the plastic strain increment to the damage assessment model based on causal reasoning, correct the weight coefficients of the influencing factors in the causal graph, and re-output the adjusted assessment results. The trajectory analysis module is used to perform time series analysis on the adjusted assessment results, extract the upward trend slope and fluctuation amplitude of the damage severity index in the assessment results, and construct the damage evolution trajectory curve. The life assessment module is used to compare the damage evolution trajectory curve with the pressure vessel design life curve, identify the degree of deviation between the damage evolution trajectory curve and the pressure vessel design life curve, classify the health level according to the degree of deviation, and generate an assessment report containing the health level and health index. The evaluation results are mapped to a three-dimensional finite element model of the pressure vessel. Based on the failure mode corresponding to the damage type label, the corresponding load conditions and boundary conditions are activated in the three-dimensional finite element model to calculate the stress concentration factor and plastic strain increment of the damaged region, including: Read the three-dimensional geometric model and material property parameters of the pressure vessel and import them into the three-dimensional finite element model; Based on the damage type label in the assessment results, a preset failure mode library is called to match the corresponding load conditions and boundary conditions. The load conditions include internal pressure load, temperature load and mechanical vibration load. In the three-dimensional finite element model, the location of the damaged area is located, and the mesh is refined at the location to improve the calculation accuracy; Submit the solution calculation to obtain the stress distribution cloud map and strain distribution cloud map of the damaged area; Extract the maximum stress value and the nominal stress value of the damaged area from the stress distribution cloud map, and calculate the ratio of the maximum stress value to the nominal stress value as the stress concentration factor; extract the equivalent plastic strain value of the damaged area from the strain distribution cloud map as the plastic strain increment.

9. A pressure vessel health assessment system based on multimodal data fusion according to claim 8, characterized in that, The step of feeding back the stress concentration factor and plastic strain increment to the damage assessment model based on causal reasoning, correcting the weight coefficients of the influencing factors in the causal graph, and re-outputting the adjusted assessment results includes: The stress concentration factor and the plastic strain increment are added as new observed variables to the causal graph of the damage assessment model based on causal reasoning. Calculate the mutual information values ​​between the stress concentration factor and the plastic strain increment and the original characteristic variables to measure the strength of their correlation. Based on the mutual information value, adjust the weight coefficient of the corresponding edge in the causal graph, increasing the weight of edges with strong correlation and decreasing the weight of edges with weak correlation. Using the causal graph with corrected weighting coefficients, the data in the damage event feature package is recalculated to update the damage type label and confidence value, and the adjusted evaluation results are output.

10. A pressure vessel health assessment system based on multimodal data fusion according to claim 9, characterized in that, The adjusted assessment results are subjected to time series analysis to extract the upward trend slope and fluctuation amplitude of the damage severity index in the assessment results, and to construct a damage evolution trajectory curve, including: The damage severity index is extracted from the adjusted assessment results. The damage severity index is calculated by multiplying the confidence level value and the stress concentration factor. Arrange the severity of injury indicators in chronological order to form a time series of injury severity. A linear regression algorithm is applied to the time series of damage severity to fit a straight line, and the slope of the straight line is the slope of the upward trend. The sliding window variance calculation method is applied to the time series of damage severity to calculate the variance value within each window. The set of all variance values ​​is the fluctuation amplitude. Using time as the horizontal axis and the severity of damage as the vertical axis, the upward trend slope and fluctuation amplitude are plotted on the coordinate system, and the data points are connected to form a damage evolution trajectory curve.

Images