CN-121997282-A - Pressure vessel health evaluation system based on multi-mode data fusion

CN121997282ACN 121997282 ACN121997282 ACN 121997282ACN-121997282-A

Abstract

The invention discloses a pressure vessel health assessment system based on multi-mode data fusion, and relates to the technical field of pressure vessel safety detection. The method comprises the steps of acquiring acoustic emission time domain signals and weld digital images by a synchronous fusion module, generating a feature fusion set through space-time registration, inputting the feature fusion set into a graph neural network with specific nodes and side weights by a topology learning module, generating graph embedding vectors by graph convolution, accumulating graph embedding vectors by an event detection module, calculating cosine similarity recognition mutation points, extracting damage event feature packages, constructing a multi-mode feature causal graph by a causal evaluation module, analyzing direct and indirect influence factors, and outputting damage type labels and confidence. The system realizes multi-mode feature topology association learning and causal driving damage evaluation, and improves the accuracy of pressure vessel health detection.

Inventors

WANG ZEHAO
GUO YUNTENG
LI YAOMING
ZHANG KAIKAI
HOU YANYANG
ZHU YAKAI
LI JIAHAO

Assignees

中石化管道技术服务有限公司

Dates

Publication Date: 20260508
Application Date: 20260410

Claims (10)

1. A pressure vessel health assessment system based on multimodal data fusion, comprising: The synchronous fusion module is used for acquiring the acoustic emission time domain signal and the digital image of the weld joint region, constructing a multi-mode original data set, and executing time synchronization and space registration on the multi-mode original data set to generate a feature fusion set; The topological learning module is used for inputting the feature fusion set into a pre-constructed graph neural network, nodes of the graph neural network are composed of acoustic emission sensor positions and centers of defect areas in the digital image, edge weights are determined by acoustic emission signal arrival time differences and image space distances together, topological association among multi-mode features is learned through graph convolution operation, and graph embedding vectors containing structural health states are generated; the event detection module is used for carrying out time sequence accumulation on the graph embedding vectors, calculating cosine similarity between the graph embedding vectors under different time steps, identifying mutation points of the graph embedding vectors, taking time corresponding to the mutation points as trigger time of potential damage events, and extracting multi-mode features before and after the trigger time to form damage event feature packages; The causal evaluation module is used for driving a causal reasoning-based damage evaluation model by taking the damage event feature package as input, analyzing direct influence factors and indirect influence factors of acoustic feature changes on geometric feature evolution by constructing a causal graph among multi-mode features, and outputting an evaluation result containing damage type labels and confidence.
2. The pressure vessel health assessment system of claim 1, wherein the acquiring the acoustic emission time domain signal and the digital image of the weld region creates a multi-modality raw dataset, performing time synchronization and spatial registration on the multi-modality raw dataset, generating a feature fusion set, comprising: acquiring acoustic emission time domain signals through an acoustic emission sensor array arranged on the surface of the pressure vessel, and simultaneously acquiring digital images of a welding line area through a digital ray imaging system of the pressure vessel to form a multi-mode original data set containing acoustic features and visual features; Performing time synchronization and space registration on the multi-mode original data set to establish an associated index of a corresponding space position in the acoustic emission event and the digital image, decomposing an acoustic emission time domain signal through a wavelet packet to extract energy entropy characteristics, and morphological segmentation of the digital image to extract geometric moment characteristics of a defect edge to form a characteristic fusion set containing time-frequency domain acoustic characteristics and space domain geometric characteristics; The method for acquiring acoustic emission time-domain signals through an acoustic emission sensor array arranged on the surface of a pressure container, simultaneously acquiring digital images of a welding line area through a digital ray imaging system of the pressure container, forming a multi-mode original data set containing acoustic features and visual features, and comprises the following steps: Starting an acquisition mode of the acoustic emission sensor array, setting a sampling frequency to be a preset high-frequency value, continuously receiving an elastic wave signal generated in the operation process of the pressure vessel, and converting the elastic wave signal into a discrete voltage time domain signal serving as an acoustic emission time domain signal; synchronously starting a scanning mode of the digital ray imaging system, performing multi-angle exposure imaging on a welding line area of the pressure container, and obtaining a gray level image sequence containing the internal structure of the welding line; splicing and correcting the gray image sequence, eliminating geometric distortion caused by imaging angle deviation, and generating a digital image covering the whole welding seam area; and aligning the acoustic emission time domain signal with the digital image according to the time stamp, removing the data frame with the unmatched time stamp, retaining the acoustic emission time domain signal with the consistent time stamp and the digital image, and combining to form the multi-mode original data set containing the acoustic characteristic and the visual characteristic.
3. The pressure vessel health assessment system of claim 2, wherein said time synchronizing and spatial registering of said multi-modal raw dataset to establish an associated index of acoustic emission events and corresponding spatial locations in the digital image comprises: Selecting a unified clock reference from the multi-mode original data set, and calibrating clock signals of all acoustic emission sensors and clock signals of a digital ray imaging system to the clock reference; establishing a Cartesian coordinate system in the digital image, and calibrating the conversion relation between the pixel coordinates and the actual physical coordinates of the digital image by taking the geometric center of the pressure container as an origin and the trend of the welding seam as the coordinate axis direction; Identifying the arrival time and the arrival sensor number of the acoustic emission event in the acoustic emission time domain signal through a peak detection algorithm, and calculating acoustic emission source space coordinates of the acoustic emission event by combining the physical installation coordinates of the acoustic emission sensor; Mapping the acoustic emission source space coordinate into a Cartesian coordinate system of the digital image through a nearest neighbor matching algorithm, finding a nearest pixel point as an associated pixel point, establishing a corresponding relation between an acoustic emission event and the associated pixel point, and generating an associated index.
4. The system for evaluating health of a pressure vessel based on multi-modal data fusion according to claim 3, wherein the extracting energy entropy features from the acoustic emission time domain signal after wavelet packet decomposition, extracting geometric moment features of defect edges from the digital image after morphological segmentation, forming a feature fusion set comprising time-frequency domain acoustic features and spatial domain geometric features, comprises: Selecting a preset wavelet basis function and a decomposition layer number, and carrying out wavelet packet decomposition on the acoustic emission time domain signal to obtain energy spectrums of different frequency bands; Calculating the ratio of energy of each frequency band to total energy, constructing an energy distribution vector, and calculating the shannon entropy of the energy distribution vector as an energy entropy characteristic; carrying out gray level processing and binarization processing on the digital image, and removing image noise by adopting a morphological filter combining open operation and closed operation; identifying a defect region in the image through a connected region marking algorithm, extracting the contour edge of the defect region, and calculating the second-order geometric moment of the contour edge as a geometric moment feature; and arranging the energy entropy features and the geometric moment features according to the sequence of the associated indexes, and combining to form a feature fusion set containing the time-frequency domain acoustic features and the spatial domain geometric features.
5. The system of claim 1, wherein the feature fusion set is input to a pre-constructed graph neural network, nodes of the graph neural network are composed of acoustic emission sensor positions and centers of defect areas in digital images, edge weights are determined by acoustic emission signal arrival time differences and image space distances together, topological relations among the multi-modal features are learned through a graph convolution operation, and graph embedding vectors containing structural health states are generated, and the method comprises the steps of: initializing an adjacency matrix of a graph neural network, and taking coordinates of the position of an acoustic emission sensor and the center of a defect area as a node set of the graph; For each node pair, calculating the arrival time difference of the acoustic emission signal from one sensor to the other sensor as time weight, calculating the pixel distance of the two nodes in the digital image as space weight, normalizing the time weight and the space weight, and adding to obtain edge weight; Respectively mapping the energy entropy features and the geometric moment features in the feature fusion set to corresponding nodes to serve as initial feature vectors of the nodes; Performing multi-layer graph rolling operation, wherein each layer of graph rolling operation performs weighted summation and nonlinear activation on the initial feature vector of the node and the feature vector of the neighbor node, and updates the feature vector of the node; and after all graph convolution operations are completed, performing global pooling operation on the feature vectors of all nodes to obtain a graph embedded vector with fixed dimension, wherein the graph embedded vector contains structural health state information.
6. The system for evaluating health of a pressure vessel based on multi-modal data fusion according to claim 1, wherein the time sequence accumulation is performed on the graph embedding vectors, cosine similarity between the graph embedding vectors at different time steps is calculated to identify mutation points of the graph embedding vectors, time corresponding to the mutation points is taken as trigger time of a potential damage event, and multi-modal features before and after the trigger time are extracted to form a damage event feature packet, comprising: Arranging the graph embedding vectors in time sequence to form a graph embedding vector sequence; for each element in the embedded vector sequence of the graph, calculating the cosine similarity between the element and the previous element to obtain a cosine similarity sequence; Applying sliding window statistics to the cosine similarity sequence, calculating the mean value and standard deviation of the cosine similarity in the window, and marking the corresponding moment as a mutation point when the cosine similarity at a certain moment is smaller than the mean value minus a plurality of times of standard deviation; taking the abrupt change point as a center, taking time periods with preset time lengths forwards and backwards respectively, extracting acoustic emission time domain signals and digital images in the corresponding time periods from the multi-mode original data set, and simultaneously extracting energy entropy features and geometric moment features in the corresponding time periods from the feature fusion set, and combining to form a damage event feature packet.
7. The system of claim 1, wherein the damage event feature pack is taken as an input to drive a causal reasoning-based damage assessment model, and the causal reasoning-based damage assessment model analyzes direct influence factors and indirect influence factors of acoustic feature changes on geometric feature evolution by constructing a causal graph among multi-modal features, and outputs an assessment result comprising damage type labels and confidence, and the method comprises: extracting the energy entropy change trend of the acoustic emission signal and the geometric moment change trend of the digital image from the damage event feature packet as input variables of causal reasoning; Based on field knowledge and historical data statistics, constructing a causal relation assumption among multi-mode features to form an initial causal diagram, wherein nodes represent feature variables and directed edges represent causal relations; Adopting a Bayesian network learning method, and utilizing data in a damage event feature packet to learn and correct edge weights in an initial causal graph to determine a final causal graph structure; in the final causal graph, calculating the direct causal effect and the indirect causal effect of the acoustic feature change on the geometrical feature evolution through an intervention analysis algorithm, and identifying main influence factors; And according to the value interval of the main influence factor and the strength of the causal effect, comparing with a preset damage type judgment rule, outputting a damage type label and a corresponding confidence value to form an evaluation result.
8. The multi-modal data fusion based pressure vessel health assessment system of claim 1, further comprising: the model coupling module is used for mapping the evaluation result into a three-dimensional finite element model of the pressure container, activating corresponding load working conditions and boundary conditions in the three-dimensional finite element model according to failure modes corresponding to the damage type labels, and calculating to obtain stress concentration coefficients and plastic strain increments of a damaged area; The iteration correction module is used for feeding back the stress concentration coefficient and the plastic strain increment to the damage evaluation model based on causal reasoning, correcting the weight coefficient of the influence factor in the causal graph and re-outputting the adjusted evaluation result; the track analysis module is used for carrying out time sequence analysis on the adjusted evaluation result, extracting the rising trend slope and the fluctuation amplitude of the damage severity index in the evaluation result and constructing a damage evolution track curve; The life evaluation module is used for comparing the damage evolution track curve with the pressure vessel design life curve, identifying the deviation degree of the damage evolution track curve and the pressure vessel design life curve, classifying the health grade according to the deviation degree, and generating an evaluation report containing the health grade and the health index; Mapping the evaluation result to a three-dimensional finite element model of the pressure vessel, activating corresponding load working conditions and boundary conditions in the three-dimensional finite element model according to failure modes corresponding to damage type labels, and calculating to obtain stress concentration coefficients and plastic strain increments of damaged areas, wherein the method comprises the following steps: reading a three-dimensional geometric model and material attribute parameters of the pressure vessel, and importing the three-dimensional geometric model and the material attribute parameters into a three-dimensional finite element model; According to the damage type label in the evaluation result, a preset failure mode library is called, and corresponding load working conditions and boundary conditions are matched, wherein the load working conditions comprise internal pressure load, temperature load and mechanical vibration load; Positioning the position of a damaged area in the three-dimensional finite element model, and refining grid division at the position to improve the calculation accuracy; Submitting solution calculation to obtain a stress distribution cloud image and a strain distribution cloud image of the damaged area; extracting a maximum stress value and a nominal stress value of a damaged area from the stress distribution cloud chart, and calculating the ratio of the maximum stress value to the nominal stress value as a stress concentration coefficient; and extracting equivalent plastic strain values of the damaged areas from the strain distribution cloud chart as plastic strain increment.
9. The system of claim 8, wherein the feedback of the stress concentration coefficient and the plastic strain increment to the causal reasoning-based damage assessment model, the correction of the weight coefficient of the influence factor in the causal graph, and the re-output of the adjusted assessment result comprise: Taking the stress concentration coefficient and the plastic strain increment as new observation variables, and adding the new observation variables into a causal graph of the damage evaluation model based on causal reasoning; calculating mutual information values between stress concentration coefficients, plastic strain increments and original characteristic variables, and measuring correlation strength between the stress concentration coefficients and the plastic strain increments; according to the size of the mutual information value, adjusting the weight coefficient of the corresponding edge in the causal graph, increasing the weight of the edge with strong correlation and reducing the weight of the edge with weak correlation; And (3) using the causal graph after the weight coefficient is corrected to carry out reasoning calculation on the data in the damage event feature package again, updating the damage type label and the confidence coefficient value, and outputting an adjusted evaluation result.
10. The system for evaluating health of a pressure vessel based on multi-modal data fusion according to claim 9, wherein the time-series analysis is performed on the adjusted evaluation result, the rising trend slope and the fluctuation range of the damage severity index in the evaluation result are extracted, and a damage evolution track curve is constructed, and the system comprises: Extracting a numerical value of a damage severity index from the adjusted evaluation result, wherein the damage severity index is obtained by calculating a product of the confidence numerical value and the stress concentration coefficient; arranging the damage severity indexes in time sequence to form a damage severity time sequence; a linear regression algorithm is applied to the damage degree time sequence, a straight line is obtained through fitting, and the slope of the straight line is the slope of the rising trend; A sliding window variance calculating method is applied to the damage degree time sequence, variance values in each window are calculated, and the set of all variance values is the fluctuation amplitude; and marking the ascending trend slope and the fluctuation amplitude in a coordinate system by taking time as an abscissa and the damage severity index as an ordinate, and connecting all data points to form a damage evolution track curve.

Description

Pressure vessel health evaluation system based on multi-mode data fusion Technical Field The invention belongs to the technical field of pressure vessel safety detection, and particularly relates to a pressure vessel health assessment system based on multi-mode data fusion. Background In the existing pressure vessel health evaluation field, a single-mode detection means is mostly adopted, characteristic extraction and anomaly discrimination of a time domain signal are emphasized by acoustic emission detection, geometric form identification of a focus weld joint region defect is detected by digital image detection, a part of multi-mode detection scheme only carries out simple characteristic splicing processing on acoustic signals and image data, and a synchronous fusion link only completes basic data integration operation. The single-mode detection means cannot simultaneously characterize dynamic acoustic response and static geometric form of pressure vessel damage, the multi-mode scheme of simple characteristic splicing does not establish spatial association of a detection unit and a defect area, and structural layer association attributes of multi-mode characteristics cannot be mined. The existing damage event detection mostly adopts a fixed threshold judgment mode, the accumulated similarity calculation is not carried out on the sequence feature vector, and the triggering time of the potential damage is difficult to accurately capture. The existing damage assessment model realizes classification by means of feature correlation analysis, cannot distinguish the action relation among different modal features, and the damage assessment result lacks the distinguishing basis of an internal logic level. The graph neural network defined by specific nodes and edge weights is required to be constructed to realize multi-mode feature topology association learning, and the acting factors of acoustic features and geometric features are required to be analyzed through a causal graph to complete damage assessment. Disclosure of Invention The present invention aims to solve at least one of the technical problems existing in the prior art; therefore, the invention provides a pressure vessel health evaluation system based on multi-mode data fusion, which comprises: The synchronous fusion module is used for acquiring the acoustic emission time domain signal and the digital image of the weld joint region, constructing a multi-mode original data set, and executing time synchronization and space registration on the multi-mode original data set to generate a feature fusion set; The topological learning module is used for inputting the feature fusion set into a pre-constructed graph neural network, nodes of the graph neural network are composed of acoustic emission sensor positions and centers of defect areas in the digital image, edge weights are determined by acoustic emission signal arrival time differences and image space distances together, topological association among multi-mode features is learned through graph convolution operation, and graph embedding vectors containing structural health states are generated; the event detection module is used for carrying out time sequence accumulation on the graph embedding vectors, calculating cosine similarity between the graph embedding vectors under different time steps, identifying mutation points of the graph embedding vectors, taking time corresponding to the mutation points as trigger time of potential damage events, and extracting multi-mode features before and after the trigger time to form damage event feature packages; The causal evaluation module is used for driving a causal reasoning-based damage evaluation model by taking the damage event feature package as input, analyzing direct influence factors and indirect influence factors of acoustic feature changes on geometric feature evolution by constructing a causal graph among multi-mode features, and outputting an evaluation result containing damage type labels and confidence. Further, the acquiring the acoustic emission time domain signal and the digital image of the weld region, constructing a multi-mode original data set, performing time synchronization and spatial registration on the multi-mode original data set, and generating a feature fusion set, including: acquiring acoustic emission time domain signals through an acoustic emission sensor array arranged on the surface of the pressure vessel, and simultaneously acquiring digital images of a welding line area through a digital ray imaging system of the pressure vessel to form a multi-mode original data set containing acoustic features and visual features; Performing time synchronization and space registration on the multi-mode original data set to establish an associated index of a corresponding space position in the acoustic emission event and the digital image, decomposing an acoustic emission time domain signal through a wavelet packet to extract energy entropy ch