CN-121980522-A - Intelligent evaluation method and system for welding structure performance based on multi-mode large model

CN121980522ACN 121980522 ACN121980522 ACN 121980522ACN-121980522-A

Abstract

The invention relates to a welding structure performance intelligent evaluation method and a system based on a multi-mode large model, which belong to the technical field of welding, wherein the system comprises a data acquisition and access layer for acquiring multi-mode data and writing in a unified traceability mark; the method comprises the steps of preprocessing, space-time calibration and alignment of a calibration alignment layer, outputting a multi-mode sample after alignment, obtaining fusion characterization by a multi-mode large model layer through quality perception gating fusion, carrying out uncertainty quantification and out-of-distribution detection, carrying out constraint calibration by combining knowledge retrieval by a mechanism constraint and state inference layer, outputting crack states, health indexes, residual life intervals and key evidence indexes, outputting confidence and risk levels by an interpretable and decision support layer, triggering a rechecking strategy and generating an audit report, and carrying out reflux updating of a model version by a maintenance verification result by a closed-loop learning and digital standing account layer. The invention can realize space-time alignment, quality perception fusion, mechanism constraint reasoning and closed-loop updating, and output accurate state estimation and interpretable evidence.

Inventors

SUN CUIMIN
DONG JIAWEI
CUI SHUWAN
PAN XIUBIN
HAN LEIGANG
XUE BIN
WANG CHENGBO
YANG QIYE
HU QICHENG
GUAN WEI

Assignees

广西大学
广西科技大学

Dates

Publication Date: 20260505
Application Date: 20260323

Claims (10)

1. An intelligent evaluation system for welding structure performance based on a multi-mode large model is characterized by comprising: The data acquisition and access layer is used for acquiring an on-line monitoring signal, a nondestructive testing image, a visual image or an infrared image and a working condition text or a history text, outputting an original multi-mode data packet and metadata thereof, writing a unified traceability identification in the metadata, wherein the unified traceability identification is used as a primary key associated with cross-layer data and at least comprises a structure identification, a welding seam identification and an acquisition time stamp; The system comprises a preprocessing and calibrating alignment layer, a time synchronization calibration and space positioning calibration and ROI mapping, wherein the preprocessing and calibration alignment layer is connected with the data acquisition and access layer and is used for receiving an original multi-mode data packet and metadata thereof, denoising, segmenting, outlier processing and missing complementation are carried out on an on-line monitoring signal, a nondestructive testing image, a visual image or an infrared image in the original multi-mode data packet, quality evaluation is carried out, modal quality indexes of signal to noise ratio, definition and missing rate are calculated, the time synchronization calibration and space positioning calibration and the ROI mapping are carried out, mapping is carried out on a signal segment and a welding seam region of interest or a welding toe region of interest, and a detection pose and a structural coordinate system, an aligned multi-mode sample is output, and the aligned multi-mode sample carries a uniform tracing mark; The multi-mode large model layer is connected with the preprocessing and calibration alignment layer and is used for receiving the aligned multi-mode samples, respectively encoding the multi-mode samples by utilizing an encoder set to obtain unified embedded representation, and also is used for adaptively adjusting the contribution weight of each mode through quality perception gating fusion according to the calculated mode quality index and executing cross-mode attention fusion to obtain fusion representation; the system comprises a multi-mode large model layer, a mechanism constraint and state push layer, a constraint calibration result, a state quantity estimation and residual life interval prediction based on the constraint calibration result, a crack existence probability, a crack size or damage index, a health index, a residual life interval and a development trend, a working condition attribution and trend analysis, a working condition attribution result and a key evidence index, wherein the multi-mode large model layer is connected with the mechanism constraint and state push layer and is used for receiving fusion characterization, uncertainty indexes and out-of-distribution detection results, searching related knowledge segments from a mechanism model library, a standard threshold library and a historical case library, and carrying out constraint and calibration on reasoning output of the fusion characterization by combining the uncertainty indexes and the out-of-distribution detection results; The system comprises an interpretable and decision-making support layer, a mechanism constraint and state inference layer, a working condition attribution result, a failure index, a residual life interval and a development trend, wherein the working condition attribution result, the key evidence index, the uncertainty index, the out-of-distribution detection result, a confidence coefficient, an interval and out-of-distribution judgment are output, a recheck strategy or a manual recheck strategy is triggered according to the out-of-distribution judgment, an evidence thermodynamic diagram or a key fragment is generated, and a report and audit record are formed; The system comprises a closed-loop learning and digital ledger layer, a multi-mode large model layer, a version management and computer maintenance management system interface or an enterprise asset management interface, a model version and a threshold version, wherein the closed-loop learning and digital ledger layer is connected with the interpretable and decision support layer and is used for receiving the risk level, the maintenance suggestion or the recheck suggestion and the audit report, acquiring a maintenance verification result or a recheck result backflow corresponding to the maintenance suggestion or the recheck suggestion output by the decision support layer, which can be interpreted as a supervision signal, executing incremental update, recording the model version, the data version and the threshold version, returning the updated model version to the multi-mode large model layer, returning the updated threshold version to the mechanism constraint and state push layer for subsequent evaluation, and combining the version management and computer maintenance management system interface or the enterprise asset management interface to construct the digital ledger.
2. The system of claim 1, wherein a first data interface is arranged between the data acquisition and access layer and the preprocessing and calibration alignment layer, the first data interface is used for transmitting an original multi-mode data packet and metadata thereof, and the metadata at least comprises a sampling rate, a time stamp, a pose or a coordinate, equipment parameters and a working condition text, and contains a unified traceability identifier; A second data interface is arranged between the preprocessing and calibration alignment layer and the multi-mode large model layer, and is used for transmitting aligned multi-mode samples, wherein the aligned multi-mode samples at least comprise time alignment parameters, space calibration matrix or interested region coordinates, missing complement masks and quality index vectors, and contain uniform traceability marks; A third data interface is arranged between the multi-mode large model layer and the mechanism constraint and state push layer, and is used for transmitting fusion characterization, initial state prediction results and uncertainty and distribution external detection results and contains the unified traceability identification; A fourth data interface is arranged between the mechanism constraint and state push layer and the interpretable and decision support layer, and is used for transmitting crack state quantity or damage state quantity, health index, residual life interval, trend and working condition attribution and key evidence index, and contains uniform traceability marks, wherein the key evidence index is used for indicating corresponding time slices, image frames, ROI coordinates and matched standard or mechanism item numbers, so that the interpretable and decision support layer generates evidence thermodynamic diagrams, key slices and risk classification reports; a fifth data interface is arranged between the interpretable and decision support layer and the closed-loop learning and digital ledger layer, and is used for transmitting risk level, maintenance advice or recheck advice and audit records, wherein the audit records at least comprise unified traceability identification, data version, model version, threshold version, quality index and uncertainty description.
3. The system of claim 2, wherein the time alignment parameters include at least a time drift correction amount that is a scalar, a segmented scalar sequence, or a time-varying function parameter, and is stored in association with a uniform trace-back identifier, and wherein the alignment parameters including the time drift correction amount are output when generating the assessment report.
4. The system of claim 2, wherein the spatial calibration matrix or the region of interest coordinates transmitted by the second data interface are used to establish a correspondence between the signal segments, the image frames, and the weld locations, the correspondence being used to interpret spatial back-projection and audit trail when the decision support layer generates an evidence thermodynamic diagram.
5. The system of claim 1, wherein the unified trace back identifier further comprises a detection pose, an acquisition channel number and/or a working condition label, the multi-mode large model layer comprises an inference calibration module, a feedback interface is arranged between the inference calibration module and the mechanism constraint and state inference layer and is used for transmitting constraint verification results and deviation amounts of the mechanism constraint and the state inference layer to the inference calibration module, and the encoder set comprises a signal encoder, an image encoder and a text encoder.
6. The system of claim 1, wherein the closed loop learning and digitizing ledger layer transmits the updated model version to a multi-modal large model layer and the updated threshold version to a mechanism constraint and status inference layer.
7. The intelligent evaluation method for the performance of the welding structure based on the multi-mode large model is characterized by comprising the following steps of: S1, multi-source data acquisition and tracing, namely acquiring an on-line monitoring signal, a nondestructive testing image, a visual image or an infrared image and a working condition text or a history text, obtaining an original multi-mode data packet and metadata thereof, writing a unified tracing identifier in the metadata, wherein the unified tracing identifier is used as a primary key associated with cross-layer data and at least comprises a structure identifier, a welding seam identifier and an acquisition timestamp; S2, preprocessing and space-time calibration alignment, namely denoising, segmenting, outlier processing and missing complementation are carried out on-line monitoring signals, nondestructive testing images, visual images or infrared images in the original multi-mode data packet in the step S1, modal quality indexes of signal-to-noise ratio, sharpness and missing rate are calculated, time synchronization calibration and space positioning calibration are carried out, mapping is carried out on signal segments and weld joint regions of interest or weld toe regions of interest, and detection pose and a structural coordinate system, so that aligned multi-mode samples are obtained, and the aligned multi-mode samples carry uniform traceability marks; S3, multi-mode large model reasoning, namely respectively utilizing a signal encoder, an image encoder and a text encoder to encode the aligned multi-mode samples in the step S2 to obtain unified embedded representation, adaptively adjusting the contribution weights of all modes through quality perception gating fusion according to the calculated modal quality indexes, and executing cross-mode attention fusion to obtain fusion characterization; S4, mechanism constraint and state inference, namely retrieving relevant knowledge segments from a mechanism model library, a standard threshold library and a historical case library, carrying out constraint and calibration on the inference output of the fusion characterization in the step S3 by combining the uncertainty index and the out-of-distribution detection result in the step S3, and outputting constraint calibration results; s5, interpretation and decision support, namely outputting confidence coefficient, interval and out-of-distribution judgment, triggering rechecking or manual rechecking strategies according to the out-of-distribution judgment, generating an evidence thermodynamic diagram or key fragments, forming reports and audit records, integrating the crack size or damage index, health index, residual life interval, development trend, working condition attribution result and the output confidence coefficient and out-of-distribution judgment in the step S4, calculating risk score, mapping risk grade, and outputting maintenance advice or rechecking advice and audit report; S6, performing closed-loop learning and digitizing the ledger, namely acquiring a maintenance verification result or a review result reflux corresponding to the maintenance suggestion or the review suggestion output in the step S5 as a supervision signal, executing incremental updating, recording a model version, a data version and a threshold version, transmitting the updated model version back to a multi-mode large model layer corresponding to the step S3, transmitting the updated threshold version back to a mechanism constraint and state push layer corresponding to the step S4, and constructing the digitizing the ledger by combining version management and a computer maintenance management system interface or an enterprise asset management interface; the steps S1 to S6 are sequentially performed, and the output of the previous step is used as input data, parameters or constraints of the subsequent step.
8. The method of claim 7, wherein in step S2, when the unified trace back identifier is incomplete, the corresponding original multi-mode data packet is marked as to-be-checked or to-be-complemented.
9. The method of claim 7, wherein the condition for triggering the review or manual review strategy in step S5 includes at least one of a confidence level being lower than a first threshold, an out-of-distribution determination being out of distribution, a risk level not being lower than a second risk level, a critical modality loss rate exceeding a second threshold, a timestamp drift exceeding a preset threshold, and a mechanism constraint verification failing.
10. The method of claim 9, wherein the first threshold is 0.75, the second risk level is R2, and the second threshold is 20%.

Description

Intelligent evaluation method and system for welding structure performance based on multi-mode large model Technical Field The invention belongs to the field of in-service health monitoring, nondestructive testing and life assessment of a welded structure, and particularly relates to an intelligent evaluation method and system for performance of the welded structure based on a multi-mode large model. Background The welding structure is widely applied to various industrial equipment, and the weld toe, the weld root and the heat affected zone of the welding structure are easy to generate fatigue cracks, defect expansion or local failure under the action of complex load spectrum, corrosion environment and temperature fluctuation, so that the operation safety of the equipment is directly affected. The existing performance evaluation technology has obvious defects: 1. Depending on single-mode data, only based on health monitoring of signals such as strain/vibration/acoustic emission or only based on defect interpretation of nondestructive detection images such as ultrasonic/ray, false alarm missing report is easy to generate when noise is enhanced, working condition is transferred or detection conditions are changed, and cross-mode evidence fusion and interpretability are lacking; 2. Although the multi-mode fusion technology is applied, the problems of inconsistent sampling frequency, time stamp drift, difficulty in unifying detection pose and weld joint coordinates, data loss, large quality fluctuation and the like exist in field data, so that the fusion result is difficult to reproduce stably; 3. The traditional model lacks uncertainty quantification, out-of-distribution identification and review strategy linkage mechanism, engineering decision lacks confidence and risk basis, and practical application requirements are difficult to meet. Therefore, an integrated evaluation method capable of realizing multi-mode data alignment, quality perception fusion, mechanism constraint reasoning and uncertainty quantification is needed, and the problems of poor stability, insufficient interpretability and high decision risk in the prior art are solved. Disclosure of Invention Aiming at the problems in the prior art, the invention provides an intelligent evaluation method and an intelligent evaluation system for the performance of a welding structure based on a multi-mode large model, which aim to deeply fuse and intelligently process multi-source heterogeneous data comprising on-line monitoring signals, nondestructive testing images, visual or infrared images, working condition texts and the like so as to solve the problem that the evaluation accuracy is insufficient under the conditions of multi-source heterogeneous, quality fluctuation and working condition transfer in the prior art. In order to achieve the above object, the present invention is specifically as follows: an intelligent evaluation system for welding structure performance based on a multi-mode large model, comprising: The data acquisition and access layer is used for acquiring an on-line monitoring signal, a nondestructive testing image, a visual image or an infrared image and a working condition text or a history text, outputting an original multi-mode data packet and metadata thereof, writing a unified traceability identification in the metadata, wherein the unified traceability identification is used as a primary key associated with cross-layer data and at least comprises a structure identification, a welding seam identification and an acquisition time stamp; The system comprises a preprocessing and calibrating alignment layer, a time synchronization calibration and space positioning calibration and ROI mapping, wherein the preprocessing and calibration alignment layer is connected with the data acquisition and access layer and is used for receiving an original multi-mode data packet and metadata thereof, denoising, segmenting, outlier processing and missing complementation are carried out on an on-line monitoring signal, a nondestructive testing image, a visual image or an infrared image in the original multi-mode data packet, quality evaluation is carried out, modal quality indexes of signal to noise ratio, definition and missing rate are calculated, the time synchronization calibration and space positioning calibration and the ROI mapping are carried out, mapping is carried out on a signal segment and a welding seam region of interest or a welding toe region of interest, and a detection pose and a structural coordinate system, an aligned multi-mode sample is output, and the aligned multi-mode sample carries a uniform tracing mark; The multi-mode large model layer is connected with the preprocessing and calibration alignment layer and is used for receiving the aligned multi-mode samples, respectively encoding the multi-mode samples by utilizing an encoder set to obtain unified embedded representation, and also is used for adaptively adjusting the cont