Intelligent identification system and method for ship structure failure mode based on multi-source field domain fusion

Abstract

This invention relates to an intelligent failure mode identification system and method for ship structures based on multi-source field fusion. The system includes a multi-source heterogeneous data acquisition module, a multi-modal data collaborative processing module, a virtual-real data fusion module, an intelligent failure mode identification module, and an online visualization and criterion generation module. It constructs a point-field fusion sensor network to synchronously acquire discrete measurement point and full-field digital image related data; it reconstructs the internal displacement field using a Bayesian inverse shell model and derives the continuous strain field using an improved sliding least squares method; and achieves soft replacement of virtual and real data through covariance cross-fusion and iterative nearest-point geometric registration. It utilizes a depth map convolutional network and a rule-based decision table to perform three-level failure mode identification and evolution tracking from stiffened plate to frame to overall structure; and renders the field response in real time and generates ultimate bearing capacity criteria. This system realizes the response field reconstruction and intelligent failure mode identification of the entire process from local buckling and plastic hinge formation to final collapse in ship structure ultimate strength tests.

Description

Technical Field

[0001] This invention relates to the field of ship structure model strength testing and failure mechanism analysis technology, and in particular to an intelligent identification system and method for ship structure failure modes based on multi-source field fusion. Background Technology

[0002] Ultimate strength model testing of ship structures is a benchmark method for verifying the safety of ship structures and revealing asymptotic failure mechanisms. Compared with numerical simulation, physical model testing can realistically reproduce the coupled effects of complex factors such as material nonlinearity, geometric nonlinearity, initial geometric defects, and welding residual stress. It accurately captures the complete failure path of the structure from local buckling and plastic hinge formation to overall collapse. The test data provides the final criteria for ship structure optimization design, code revision, and safety assessment. After the ultimate strength test of ship structure models enters the nonlinear stage, the model is in a critical failure state. Even a small disturbance of the load can cause the overall collapse or failure of the model structure. At this time, the response of the model structure exhibits severe spatial non-uniformity, and the strain field, displacement field, and deformation mode change rapidly, making the identification of failure modes of the model structure extremely difficult.

[0003] The existing technology has two main fundamental limitations: First, the model structure's response field information is insufficient. Traditional experimental methods rely on discrete point sensors such as resistance strain gauges and displacement gauges to collect data, and use macroscopic load-displacement curves to determine the ultimate strength of the structure. This method cannot obtain the full-field continuous response distribution of the model structure. As a result, key failure information in the nonlinear stage (such as the failure initiation location, buckling mode transition path, and plastic zone expansion) is missed, and it is even more impossible to capture the spatiotemporal coupling evolution law of multiple failure modes.

[0004] Secondly, the failure determination mechanism is lagging behind. Existing technologies generally adopt a post-experimental static interpretation mode, that is, after the test is completed, researchers identify the failure mode by visually observing the final damage morphology based on their personal experience and combining it with discrete measurement point data.

[0005] This mechanism has three major technical flaws: (1) Lack of temporal sequence: It is impossible to track and determine the dynamic process of failure initiation, expansion, and collapse in real time, resulting in the inability to trace the source of transient failure characteristics; (2) High subjectivity: Manual interpretation is easily affected by factors such as experience differences and visual limitations, which may lead to misjudgment and omission. (3) Unclear coupling mechanism: It is difficult to reveal the competitive triggering mechanism and dominant mode transformation criterion of multiple failure modes. The above defects make it impossible to accurately reveal the deep mechanism of the progressive failure process, which seriously affects the reliability of ship structural safety assessment and the effectiveness of optimization design.

[0006] Therefore, establishing an analytical method that integrates multi-source field test data and achieves complete reconstruction of structural response field and intelligent determination of failure modes through virtual and real data collaboration and deep learning algorithms has become a key technical problem that urgently needs to be solved in the field of ship structure testing technology. It has important engineering value and theoretical significance for improving the refinement and intelligence level of ship structure safety assessment. Summary of the Invention

[0007] To address the shortcomings of existing production technologies, this applicant provides an intelligent identification system and method for ship structure failure modes based on multi-source field fusion. This establishes a dynamic mapping relationship between structural response and failure modes, enabling response field reconstruction and intelligent identification of failure modes throughout the entire process of structural failure from local buckling and plastic hinge formation to final collapse during ultimate strength tests. This solves fundamental problems of traditional methods, such as reliance on discrete measurement points, delayed failure judgment, and inability to capture the dynamic evolution of multiple failure modes.

[0008] The technical solution adopted in this invention is as follows: A ship structure failure mode intelligent identification based on multi-source field fusion includes a multi-source heterogeneous data acquisition module, a multi-modal data collaborative processing module, a virtual-real data fusion module, a failure mode intelligent identification module, and an online visualization and criterion generation module connected in sequence. The multi-source heterogeneous data acquisition module is used to acquire full-process response data in the ultimate strength test of a ship structural model through a point-field fusion sensor network; The multimodal data collaborative processing module is used to solve, reconstruct, verify and fuse the measured data output by the multi-source heterogeneous data acquisition module to generate spatiotemporal continuous response field data of the failure key parts; The virtual-real data fusion module is used to generate a virtual-real fusion response field dataset of the entire field of the ship structure model by using virtual simulation data as a reference framework and fusing the measured field data output by the multimodal data collaborative processing module through geometric registration and confidence weighting. The failure mode intelligent identification module is used to receive the graph structure data output by the virtual and real data fusion module. Through the data-driven graph neural network and the rule-driven decision table, it performs three-level identification of the ship structure: "stiffening plate - plate frame - model as a whole", tracks the entire failure evolution process, and outputs the graded status, failure mode and text description. The online visualization and criterion generation module is used to receive the graded status, failure mode and text description output by the failure mode intelligent identification module, convert the time-series identification results into interactive graphic renderings in real time at the test site, and generate quantitative criteria that comply with the ship structure safety assessment specifications, thus realizing a closed loop from failure identification to strength determination.

[0009] Its further technical solution lies in: The multi-source heterogeneous data acquisition module includes a point sensor array unit, a field measurement unit, a spatiotemporal synchronization triggering unit, and a data encapsulation interface unit; The point-type sensor array unit consists of resistance strain sensors, displacement sensors and actuator load sensors, and is arranged along the intersection nodes of the hull plate stiffeners, the free edges of openings and the extreme points of the span. The field measurement unit consists of a digital image correlation measurement system and a 3D laser scanner; The spatiotemporal synchronization triggering unit uses the load step signal of the load sensor as the reference clock and synchronously drives the data acquisition actions of the point sensor array unit and the field measurement unit through the hardware triggering channel to achieve a time synchronization accuracy of better than 1ms for heterogeneous data sources. The data encapsulation interface unit is used to receive the discrete measurement point data stream of the point sensor array unit and the field data stream of the field measurement unit, perform data verification, outlier removal and format standardization, and encapsulate it into a unified data packet with spatiotemporal tags for output.

[0010] The multimodal data collaborative processing module includes a DIC data calculation unit, a displacement field reconstruction unit, a strain field deduction unit, and a multi-source data verification and fusion unit. The DIC data processing unit is used to establish the correlation function through the zero-mean normalized minimum distance sum of squares criterion, and to iteratively solve the shape function parameters using the inverse combination Gauss-Newton algorithm to achieve sub-pixel accuracy calculation of the displacement field of the outer surface of the structure. The dynamic boundary conditions of the region to be reconstructed are extracted from the global displacement field obtained by the calculation. The displacement field reconstruction unit is used to receive discrete strain time history data and the dynamic boundary conditions, construct the inverse shell element model of the region to be reconstructed, and solve the maximum a posteriori estimate under the Bayesian framework to obtain the displacement field of the key internal parts. The strain field derivation unit is used to receive discrete strain data and derive the continuous strain field using an improved sliding least squares method. The multi-source data verification and fusion unit is used to correct systematic errors in the displacement field of the digital image correlation solution by using the measured data of point sensors as a benchmark and adopting the covariance cross-fusion algorithm to adaptively allocate fusion weights according to the uncertainty of each source data.

[0011] The objective function for establishing the displacement field reconstruction element is:

[0012] Where u is the displacement vector of the node to be determined. ε meas u is the measured strain vector. BC B represents the boundary displacement extracted by DIC, B is the strain-displacement matrix, L is the boundary extraction operator, and W is the boundary displacement extracted by DIC. ε With W uThese are the observation confidence and boundary condition accuracy weight matrices, respectively. λ Let Φ(u) be the regularization factor; taking the partial derivative of Φ(u) with respect to u and setting it to zero, we obtain the equilibrium equation:

[0013] In the formula, K e =B T W ε B is the stiffness matrix assembled based on strain measurement points, K b =L T W u L is the boundary constraint matrix, F e =B T W ε ε meas F b =L T W u u BC .

[0014] The virtual-real data fusion module includes a geometric alignment unit and a feature mapping and fusion unit; The geometric alignment unit is used to perform multi-level registration of the virtual node set and the measured node set in three-dimensional space, including coarse alignment sub-units and fine registration sub-units using the iterative nearest point algorithm; The feature mapping and fusion unit is used to retrieve the K nearest Euclidean neighbors of each target node in the virtual node set from the registered measured node set, calculate the contribution weight based on the inverse distance weighting principle, map the measured displacement field and strain field to the virtual node according to the weight, and perform soft replacement based on the confidence level of the measured data. , where β is the confidence weight of the measured data.

[0015] The intelligent failure mode identification module includes a graph data normalization unit, a time sequence graph sampling unit, a buckling state identification unit, a failure mode discrimination unit, and a failure evolution tracking unit; The graph data standardization unit is used to transform the input graph structure data into an internal calculation format, fill in missing items caused by measurement blind spots with neighborhood mean, and perform Z-score standardization on features of different dimensions. The timing diagram sampling unit is used to select the interface node between the strip plate and the web plate of the stiffened plate of the model as a seed at each load step, and prioritizes to retain nodes in the strong deformation region where the displacement gradient is greater than a preset threshold, so that the node size of the sampled diagram is reduced to 15%~20% of the original diagram. The buckling state identification unit identifies nine types of buckling states of stiffened plates frame by frame based on a depth map convolutional network. The failure mode discrimination unit performs plate-frame level failure mode mapping based on buckling state combination and preset rules. When the states of multiple web plates are inconsistent, the dominant mode is determined according to the priority of "overall > shear > tilt > web plate > plate grid". The failure evolution tracking unit is used to scan and record the load step, mode type and location of the first failure of each board frame by frame, identify the earliest failed board as the initial failure source, track the spatial expansion and transformation of failure modes in subsequent load steps, and determine the stable final state when the failure modes are stable for five consecutive frames and the proportion of failed boards exceeds the threshold.

[0016] The online visualization and criterion generation module includes a field rendering engine unit, a criterion generation unit, and a threshold warning unit; The field rendering engine unit uses OpenGL to perform hardware-accelerated rendering of the ship structure response field, sends node displacements to the vertex shader to drive deformation animation, and inputs strain components to the fragment shader to generate continuous cloud maps. The criterion generation unit is used to extract the bearing capacity-displacement curve from the time series database, automatically detect the peak point or the point of 5% drop in the curve as the ultimate bearing capacity criterion, and calculate the stiffness degradation rate and safety margin. The threshold warning unit is used to preset three thresholds: when the proportion of failed plates exceeds 10%, a loading deceleration command is sent; when the stiffness degradation rate exceeds 30%, an audible and visual alarm is triggered; and when the overall buckling failure probability exceeds 0.8, an emergency alarm signal is sent.

[0017] The method for intelligent identification of ship structural failure modes based on multi-source field fusion includes the following steps: S1, Dynamic Synchronous Acquisition of Multi-Source Heterogeneous Data: A point-field fusion sensor network is constructed, and the time synchronization accuracy of heterogeneous data sources is better than 1ms through a spatiotemporal synchronization triggering unit. S2, Multimodal Data Collaborative Processing and Field Reconstruction: The surface displacement field is calculated using a subpixel algorithm based on digital image correlation data; the displacement field of key internal components is reconstructed using a Bayesian framework and inverse shell model by combining discrete strain data and dynamic boundary conditions; the continuous strain field is derived using an improved sliding least squares method; and multi-source data are adaptively weighted and fused using a covariance crossover algorithm. S3, Dynamic Fusion of Virtual and Real Response Fields: Based on a nonlinear finite element virtual model, the measured field data and the virtual model are geometrically aligned using an iterative nearest point algorithm. Based on K-nearest neighbor retrieval and inverse distance weighting, the measured displacement / strain field is mapped and soft replacement is performed to generate a full-field virtual-real fusion response dataset. S4, Intelligent Failure Mode Identification and Evolution Tracking: The fused graph structure data is standardized and compressed by sampling through a time-series graph; the buckling state of the stiffened plate is identified frame by frame using a depth graph convolutional network; the failure mode is mapped to the plate frame level based on preset rule logic; the initial failure source, spatial expansion path and mode transition event are recorded, and the overall failure mode is determined in combination with the load conditions. S5, online visualization and criterion generation: Real-time display of response field deformation animation and strain cloud map through hardware-accelerated rendering; automatic extraction of quantitative criteria such as ultimate bearing capacity and stiffness degradation rate; triggering of multi-level safety warnings; output of assessment reports that comply with ship specifications.

[0018] In S4, based on preset rule logic mapping to board-level failure modes, specifically including: Logical judgment is performed on the state of the strip and web within the same frame: If the plate exhibits localized buckling and the web shows no deformation, it is determined to be plate buckling; If the band plate buckles as a whole and the upper edge of the web plate buckles, it is determined to be overall buckling; If the middle of the web is bent but the bands are not deformed, it is determined to be web bending; If the upper edge of the web is flexed and accompanied by local flexion of the band plate, it is determined to be lateral flexion; If the strip exhibits shear buckling, it is determined to be shear buckling; When the states of multiple web plates are inconsistent, the dominant mode is determined according to the priority of "overall > shear > tilt > web plate > plate grid". S4 records the initial failure source, spatial expansion path, and mode transition events, specifically including: The load step, mode type, and location of the first failure of each board frame are recorded by scanning frame by frame, and the earliest failed board frame is identified as the initial failure source. Track the spatial expansion and transformation of failure modes in subsequent load steps, and record key events such as buckling mode transition, multi-mode competition and dominant mode transformation; When the failure mode is stable for five consecutive frames and the proportion of failed boards exceeds the threshold, it is determined to be a stable final state, forming an "initial-expansion-stable" timing chain.

[0019] In S5, multiple levels of security alerts are triggered, specifically including: Level 1 warning: When the proportion of failed plates exceeds 10%, a loading deceleration command is sent to the test loading system; Level 2 warning: When the stiffness degradation rate exceeds 30%, an audible and visual alarm is triggered; Level 3 warning: When the overall buckling failure probability exceeds 0.8, an emergency alarm signal is sent to the test loading system controller.

[0020] The beneficial effects of this invention are as follows: This invention has a compact and reasonable structure, is easy to operate, and... This invention represents a breakthrough in ship structural model failure mode identification, moving from "discrete point, post-event judgment" to "continuous, real-time intelligent full-field analysis." By constructing a "point-field fusion" sensor network, integrating discrete point sensors with field measurement technologies such as DIC and 3D scanning, this invention acquires spatiotemporally continuous full-field response data during ship structural ultimate strength tests. This overcomes the inherent flaw of traditional methods that rely on sparse measurement points, leading to the omission of crucial failure information, and provides a data foundation for accurately capturing the entire process of failure initiation and propagation. Through virtual-real fusion and intelligent recognition algorithms, failure judgment is transformed from post-test manual experience-based interpretation to real-time automatic identification during the loading process, significantly shortening the identification time and enabling in-situ, online tracking of the dynamic evolution of structural failure.

[0021] This invention improves the objectivity, accuracy, and traceability of failure mode identification. It employs a combination of data-driven depth graph convolutional networks and rule-driven logical judgment to achieve automatic identification and mapping of failure modes from stiffened plate buckling state to plate frame level and even the overall failure mode of the ship structure model. This method effectively eliminates the drawbacks of traditional manual interpretation, such as strong subjectivity and poor consistency, significantly improving identification accuracy. Simultaneously, the system fully records the spatiotemporal path of failure evolution (initial failure source, propagation process, mode transition), making the failure mechanism quantifiable and traceable, providing detailed data support for a deeper understanding of the progressive failure mechanism of structures.

[0022] This invention significantly enhances the integrity and reliability of response field reconstruction through multi-source information fusion and collaborative processing. It innovatively proposes a displacement field reconstruction method based on a Bayesian framework and an improved sliding least squares strain field derivation method. This method fuses discrete measurement point data with full-field measurement data under the constraints of a physical model, achieving accurate reconstruction of the displacement / strain fields of key components within the structure. Furthermore, through dynamic fusion technology of virtual and real data, high-confidence measured data is used to perform "soft replacement" correction on the finite element virtual simulation results, generating a more realistic and complete full-field response dataset, providing high-quality input for subsequent intelligent recognition.

[0023] This invention forms a complete technical closed loop from intelligent identification to engineering criteria, demonstrating strong practicality. Going beyond identification, it further integrates online visualization, automatic criterion generation, and multi-level safety early warning functions. The system can render failure evolution animations in real time and automatically extract core quantitative criteria conforming to shipbuilding standards, such as ultimate bearing capacity and stiffness degradation rate, directly serving engineering safety assessments. The three-level early warning mechanism can also be linked with the test loading system, effectively improving the safety of the test process. The final standardized assessment report greatly improves the efficiency and standardization of test data processing. Attached Figure Description

[0024] Figure 1 This is a diagram illustrating the overall architecture of the intelligent identification system for ship structural failure modes based on multi-source field fusion, as described in this invention.

[0025] Figure 2 This is a flowchart of the multimodal data collaborative processing of the present invention.

[0026] Figure 3 This is a schematic diagram of the multi-source heterogeneous data acquisition layout of the present invention.

[0027] Figure 4 This is a block diagram of the intelligent failure mode identification algorithm of the present invention.

[0028] Figure 5a This is a schematic diagram of a typical failure mode of the stiffened plate of the present invention (overall buckling).

[0029] Figure 5b This is a schematic diagram of a typical failure mode of the stiffened plate of the present invention (inter-stiffener buckling).

[0030] Figure 5c This is a schematic diagram of a typical failure mode of the stiffened plate of the present invention (beam-column buckling).

[0031] Figure 5d This is a schematic diagram of a typical failure mode of the stiffened plate of the present invention (web buckling).

[0032] Figure 5e This is a schematic diagram of a typical failure mode of the stiffened plate of the present invention (lateral buckling of the stiffener).

[0033] Figure 5f This is a schematic diagram of a typical failure mode of the stiffened plate of the present invention (shear buckling).

[0034] Figure 6 This is a screenshot of the visualization and criterion generation interface for this invention. Detailed Implementation

[0035] The specific embodiments of the present invention will now be described with reference to the accompanying drawings.

[0036] Please see Figure 1 The intelligent identification system for ship structure failure modes based on multi-source field fusion provided in this embodiment has the following specific technical solutions: including a multi-source heterogeneous data acquisition module, a multi-modal data collaborative processing module, a virtual and real data fusion module, an intelligent identification module for failure modes, and an online visualization and criterion generation module.

[0037] Among them, the multi-source heterogeneous data acquisition module refers to Figure 1 and Figure 3 The system acquires full-process response data of the ultimate strength test of the ship structure model through a point-field fusion sensor network.

[0038] The point-type sensor array unit is arranged along the intersection nodes of the hull plate stiffeners, the free edges of openings, and the extreme points of the span. This array consists of several resistance strain sensors, displacement sensors, and actuator load sensors, forming a measurement point layout covering the critical path of structural failure. The resistance strain sensors are foil strain gauges with a range of not less than 20,000 με and a sampling frequency of more than 20 Hz to capture strain data throughout the entire process from elastic to plastic stages. The displacement sensors adopt a redundant configuration of contact displacement gauges and non-contact laser displacement sensors, with a measurement accuracy better than ±0.1% of the full scale. The load sensor error is not greater than ±0.5%FS and serves as a global time-series reference source.

[0039] The field measurement unit consists of a digital image correlation (DIC) measurement system and a 3D laser scanner. The DIC measurement system employs a binocular stereo vision system with an industrial camera resolution of at least 5 megapixels and a sampling frequency of at least 20Hz. It captures the global displacement and strain fields through random speckle patterns (3-5mm in diameter, with a coverage of at least 60%) on the structural surface, achieving a spatial resolution of at least 5mm to distinguish the partial buckling modes of the plate structure. The 3D laser scanner has a measurement accuracy better than ±0.05mm and is used to acquire the initial geometry before the experiment and the geometry of the model after the experiment. Its point cloud registration accuracy is better than 1 / 50th of the short side length of the smallest plate grid in the structure.

[0040] The spatiotemporal synchronization triggering unit uses the load step signal of the load sensor as the reference clock. It synchronously drives the data acquisition actions of the point sensor array and the field measurement unit through a hardware triggering channel (such as a BNC cable branch). It also marks the acquired data with a unified timestamp, achieving a time synchronization accuracy of better than 1ms for heterogeneous data sources.

[0041] The data encapsulation interface unit receives the discrete measurement point data stream from the point sensor array and the field data stream from the field measurement unit, and performs data verification (such as range checking, 3D model checking, etc.). σ Outlier removal and format standardization are implemented, and the data is encapsulated into a unified data packet with spatiotemporal tags for output.

[0042] Among them, the multimodal data collaborative processing module solves, reconstructs, verifies and fuses the measured data output by the aforementioned acquisition module to generate spatiotemporal continuous response field data of the critical failure location.

[0043] The DIC data processing unit receives the raw speckle image sequence captured by the field measurement unit and establishes a correlation function using the zero-mean normalized minimum sum of squared distances (ZNSSD) criterion.

[0044] in, I and I0 represents the grayscale value of the deformed sub-region and the reference sub-region, respectively; p is the shape function parameter vector; and u(x;p) is the displacement shape function. The shape function parameters are iteratively solved using the inverse combined Gaussian-Newton algorithm, minimizing the correlation function to achieve sub-pixel accuracy in calculating the displacement field of the outer surface of the structure. Dynamic boundary conditions of the region to be reconstructed are extracted from the calculated global displacement field and encapsulated in the output using the load step as the time label.

[0045] The displacement field reconstruction unit receives the discrete strain time history data measured by the point sensor array and the aforementioned dynamic boundary conditions, constructs the inverse shell element model of the region to be reconstructed, solves the maximum a posteriori estimate within a Bayesian framework, and establishes the objective function:

[0046] Where u is the displacement vector of the node to be determined. ε meas u is the measured strain vector. BC B represents the boundary displacement extracted by DIC, B is the strain-displacement matrix, L is the boundary extraction operator, and W is the boundary displacement extracted by DIC. ε With W u These are the observation confidence and boundary condition accuracy weight matrices, respectively. λ Let be the regularization factor. Taking the partial derivative of Φ(u) with respect to u and setting it to zero, we obtain the equilibrium equation:

[0047] In the formula, K e =B T W ε B is the stiffness matrix assembled based on strain measurement points, K b =L T W u L is the boundary constraint matrix, F e =B T W ε ε meas F b =L T W u u BC The time-varying stiffness matrix is ​​updated and solved in real time under each load step, and the reconstruction error is controlled within 5%.

[0048] The strain field derivation unit receives discrete strain data and uses an improved sliding least squares method to derive the continuous strain field. A compactly supported domain weight function is constructed:

[0049] in, d i =||x x i || represents the distance from the point to be determined to the measuring point.R To determine the radius of influence (which is 1.5 times the average distance between adjacent measuring points), the coefficients of the strain field basis functions are obtained by solving a local weighted least squares problem at each point to be determined.

[0050] The multi-source data verification and fusion unit uses the measured data from the displacement sensors in the point sensor array as a benchmark to correct systematic errors in the displacement field calculated by DIC. It employs a covariance cross-fusion algorithm, adaptively allocating fusion weights according to the uncertainty of each source data.

[0051] Where P1 and P2 are the covariance matrices of the point sensor and the DIC data, respectively. ω Point-based data is weighted (not less than 0.6), and DIC data weight is not higher than 0.4. The mean and standard deviation of the system error of the DIC displacement field are calculated based on the measured values ​​of the displacement sensor, and the DIC data is corrected frame by frame.

[0052] The data encapsulation interface unit encapsulates the displacement field and strain field data output by the DIC data calculation unit, displacement field reconstruction unit, and strain field deduction unit into a unified data packet with spatiotemporal tags, confidence level markers, Jacobian matrix information, and load step index.

[0053] The virtual-real data fusion module uses virtual simulation data as a baseline framework and fuses the measured field data output by the multimodal data collaborative processing module through geometric registration and confidence weighting to generate a virtual-real fusion response field dataset for the entire field of the ship structure model.

[0054] Before the experiment, the basic framework generation unit constructs a nonlinear finite element model of the ship structure, integrating the initial geometry, the actual material stress-strain curves, the measured plate thickness, and the residual stress distribution of the welds. Through incremental loading calculations, it obtains the full-field response data (node ​​coordinates, element topology, node displacements, element strains, and stresses) for each load step, serving as virtual reference data for virtual-real integration. The finite element calculation results are converted into a graph data structure, where finite element mesh nodes correspond to graph nodes, element edges correspond to graph edges, and node displacements and element strains are stored as graph node attributes.

[0055] The data extraction and indexing unit selects the set of node numbers to be fused from the virtual benchmark data according to the fusion range defined in the fusion rule file, extracts the set of node numbers for the corresponding region from the measured data, and establishes a dual-number list index mapping.

[0056] The geometric alignment unit performs multi-level registration of the virtual node set and the measured node set in 3D space. In the coarse alignment stage, the centroid coordinates of the two node sets are calculated separately. A rigid body translation transformation is used to make the two centroids coincide. Then, a scaling factor is calculated based on the bounding box scale, and a uniform scaling transformation is performed on the measured point cloud to eliminate scale differences. In the fine registration stage, the Iterative Closest Point (ICP) algorithm is used. Iteratively, nearest neighbor search and least squares are performed to solve for the optimal rotation matrix and translation vector until the root mean square error change rate between adjacent iterations is less than 0.1%, obtaining a measured node set that is precisely aligned with the virtual node coordinate system.

[0057] The feature mapping and fusion unit searches for the nearest Euclidean distance for each target node in the virtual node set within the registered measured node set. K =8 neighboring nodes, contribution weights are calculated based on the inverse distance weighting principle:

[0058] in, d k For the first k The distances between each neighboring node and the target node are calculated. The measured displacement and strain fields are weighted and mapped to virtual nodes, and soft replacement is performed based on the confidence level of the measured data.

[0059] Among them, u meas The measured displacement after interpolation. β The confidence weights for the measured data are set at 0.6–0.8 for high confidence and 0.3–0.4 for low confidence, with high-confidence measured data being used first to correct the virtual results.

[0060] The integrity verification unit performs a topology consistency check on the fused full-field response data: if a certain element contains both fused and unfused nodes, the continuity of boundary physical quantities is ensured by element shape function interpolation; it verifies whether the strain-displacement satisfies the geometric equation, marks elements with excessive residuals and triggers local reconstruction.

[0061] The failure mode intelligent identification module receives graph structure data output by the virtual and real data fusion module. Through data-driven graph neural network and rule-driven decision table, it performs three-level identification of ship structure: stiffening plate - plate frame - model as a whole, and tracks the entire failure evolution process.

[0062] The graph data standardization unit transforms the input graph structure data into an internal calculation format, verifies whether each node feature vector contains three-dimensional coordinates, three-axis displacements, strain components, component type, and load step information, completes missing items due to measurement blind spots using neighborhood mean, performs Z-score standardization on features of different dimensions, verifies the graph edge connection topology, marks isolated nodes detached from the main structure as low confidence, preserves the mechanical transfer paths at the plate-web junction and model connections, and maintains the continuity of the structural space.

[0063] The timing diagram sampling unit selects the junction node of the strip plate and web plate of the stiffened plate of the model as a seed at each load step. It prioritizes retaining nodes in strong deformation areas with displacement gradients greater than 0.1 mm / mm, and selects representative nodes in weak deformation areas according to the inverse distance weight, so that the node scale of the sampled diagram is reduced to 15%–20% of the original diagram, and the buckling initiation position and plastic hinge extension path are fully preserved.

[0064] The buckling state identification unit identifies buckling states frame-by-frame based on a depth map convolutional network. It aggregates displacement patterns and strain states of neighboring nodes through multi-layer message passing, extracts feature vectors of stiffened plate members, and identifies nine states: local buckling, overall buckling, combined buckling, shear buckling, and essentially no deformation in the stiffened plate strip; and intermediate buckling, upper edge buckling, arcuate buckling, and essentially no deformation in the stiffened plate web. A classifier output probability exceeding 0.7 is considered a valid identification; otherwise, it is marked as an uncertain state.

[0065] The failure mode discrimination unit performs frame-level failure mode mapping based on buckling state combinations and preset rules. It logically judges the state of the strip and web within the same frame: local buckling of the stiffened strip with no web deformation → frame buckling; overall buckling of the strip and buckling at the upper edge of the web → overall buckling; buckling in the middle of the web with no strip deformation → web buckling; buckling at the upper edge of the web accompanied by local buckling of the strip → tilt buckling; shear buckling of the strip → shear buckling. When multiple web states are inconsistent, the dominant mode is determined according to the priority "overall > shear > tilt > web > frame".

[0066] The failure evolution tracking unit scans and records the load step, mode type, and location of the first failure of each plate frame frame by frame, and identifies the earliest failed plate frame as the initial failure source; it tracks the spatial expansion and transformation of failure modes in subsequent load steps, and records key events such as buckling mode transition, multi-mode competition, and dominant mode transformation; when the failure modes are stable for five consecutive frames and the proportion of failed plates exceeds the threshold, it is determined to be a stable final state, forming an "initial-expansion-stable" time sequence chain.

[0067] The overall failure determination unit determines six overall modes based on the load combination of the central arch / sag bending moment, torque and surface pressure, combined with the spatial distribution of plate frame failure modes: bottom compression buckling failure, deck compression buckling failure, warping failure, shear failure, bottom diagonal buckling failure or deck diagonal buckling failure; if the proportion of failed plates is less than 5%, it is determined as not failed.

[0068] The text description generation unit generates a structured natural language description based on a four-segment template of "load condition - initial failure - propagation process - final distribution", embedding quantitative indicators such as failure propagation rate, load-bearing capacity reduction rate, and failure plate proportion.

[0069] The online visualization and criterion generation module transforms the time-series identification results into interactive graphic renderings of the test site in real time, and generates quantitative criteria that comply with ship structural safety assessment standards.

[0070] The field rendering engine unit uses OpenGL to perform hardware-accelerated rendering of the ship structure response field. It sends the node displacement to the vertex shader to drive the deformation animation, and inputs the strain components to the fragment shader to generate a continuous cloud map (tensile strain orange-red, compressive strain cyan-blue). The buckling node starts to blink every 0.5 seconds. It supports functions such as panel group display and section dragging, with a rendering frame rate of 30fps and a latency of less than 200ms.

[0071] The failure evolution animation unit generates keyframe animations according to the load step: the failed racks adopt a dual effect of transparency gradient (0→0.7) and color deepening (gray→red), the unfailed areas remain semi-transparent gray, and the failure expansion boundary of the current load step is marked with a white highlight line; the animation synchronously overlays a digital dashboard (current load, number of failed racks / total, load capacity percentage) and a four-segment failure history text.

[0072] The criterion generation unit extracts load-displacement curves from the time-series database and automatically detects peak points or points with a 5% drop in the curves as ultimate bearing capacity criteria. It calculates the stiffness degradation index from the buckling state matrix, calculates the safety margin from the plate frame failure modes, calculates the proportion of shear failure for torque conditions, and calculates the buckling depth of the compression zone for bending moment conditions, generating a core criterion table that includes ultimate load, stiffness degradation rate, safety margin, and failure mode proportion.

[0073] The threshold warning unit has three preset thresholds: Level 1 warning (yellow) sends a loading deceleration command to the test loading system when the proportion of failed plates exceeds 10%; Level 2 warning (orange) triggers an audible and visual alarm when the stiffness degradation rate exceeds 30%; Level 3 warning (red) sends an emergency alarm signal to the test loading system controller when the overall buckling failure probability exceeds 0.8.

[0074] The report automatically generates basic information of the unit integration test, key frame screenshots of the failure evolution animation, ultimate bearing capacity and stiffness degradation curves, plate frame failure mode distribution cloud map, overall failure mode judgment conclusion, four-segment failure history text description and core criterion summary table, generating an assessment report that conforms to shipbuilding industry standards.

[0075] A method for intelligent identification of ship structural failure modes based on multi-source field fusion includes the following steps: S1. Dynamic Synchronous Acquisition of Multi-Source Heterogeneous Data: Construct a point-field fusion sensor network, and achieve time synchronization accuracy of better than 1ms for heterogeneous data sources through a spatiotemporal synchronization trigger unit. Use data encapsulation interface to verify data format standardization.

[0076] S2. Multimodal Data Collaborative Processing and Field Reconstruction: Based on DIC data, the surface displacement field is solved using a sub-pixel algorithm; combining discrete strain data and dynamic boundary conditions, the displacement field of key components inside the model is reconstructed using a Bayesian framework and an inverse shell model; the continuous strain field is derived using an improved sliding least squares method; and multi-source data are adaptively weighted and fused using a covariance crossover algorithm.

[0077] S3. Dynamic Fusion of Virtual and Real Response Fields: Based on a nonlinear finite element virtual model, the measured field data is geometrically aligned with the virtual model using the ICP algorithm; the measured displacement / strain field is mapped based on K-nearest neighbor search and inverse distance weighting, and soft replacement is performed to generate a full-field virtual and real fused response dataset.

[0078] S4. Intelligent Failure Mode Identification and Evolution Tracking: The fused graph structure data is standardized and compressed by sampling through a time-series graph; the buckling state of the stiffened plate is identified frame by frame using a depth graph convolutional network; the failure mode is mapped to the plate frame level based on preset rule logic; the initial failure source, spatial expansion path and mode transition event are recorded, and the overall failure mode is determined in combination with the load conditions.

[0079] S5. Online Visualization and Criterion Generation: Real-time display of response field deformation animation and strain cloud map through OpenGL hardware acceleration rendering; automatic extraction of quantitative criteria such as ultimate bearing capacity and stiffness degradation rate; triggering multi-level safety warnings; outputting assessment reports that comply with ship specifications.

[0080] The following table shows the failure mode discrimination criteria:

[0081] The above description is an explanation of the present invention and not a limitation thereof. The scope of the present invention is defined by the claims. Within the scope of protection of the present invention, any form of modification may be made.

Claims

1. An intelligent identification method for ship structural failure modes based on multi-source field fusion, characterized in that: It includes a multi-source heterogeneous data acquisition module, a multi-modal data collaborative processing module, a virtual-real data fusion module, a failure mode intelligent identification module, and an online visualization and criterion generation module, which are connected in sequence. The multi-source heterogeneous data acquisition module is used to acquire full-process response data in the ultimate strength test of a ship structural model through a point-field fusion sensor network; The multimodal data collaborative processing module is used to solve, reconstruct, verify and fuse the measured data output by the multi-source heterogeneous data acquisition module to generate spatiotemporal continuous response field data of the failure key parts; The virtual-real data fusion module is used to generate a virtual-real fusion response field dataset of the entire field of the ship structure model by using virtual simulation data as a reference framework and fusing the measured field data output by the multimodal data collaborative processing module through geometric registration and confidence weighting. The failure mode intelligent identification module is used to receive the graph structure data output by the virtual and real data fusion module. Through the data-driven graph neural network and the rule-driven decision table, it performs three-level identification of the ship structure: "stiffening plate - plate frame - model as a whole", tracks the entire failure evolution process, and outputs the graded status, failure mode and text description. The online visualization and criterion generation module is used to receive the graded status, failure mode and text description output by the failure mode intelligent identification module, convert the time-series identification results into interactive graphic renderings in real time at the test site, and generate quantitative criteria that comply with the ship structure safety assessment specifications, thus realizing a closed loop from failure identification to strength determination.

2. The intelligent identification system for ship structural failure modes based on multi-source field fusion as described in claim 1, characterized in that: The multi-source heterogeneous data acquisition module includes a point sensor array unit, a field measurement unit, a spatiotemporal synchronization triggering unit, and a data encapsulation interface unit; The point-type sensor array unit consists of resistance strain sensors, displacement sensors and actuator load sensors, and is arranged along the intersection nodes of the hull plate stiffeners, the free edges of openings and the extreme points of the span. The field measurement unit consists of a digital image correlation measurement system and a 3D laser scanner; The spatiotemporal synchronization triggering unit uses the load step signal of the load sensor as the reference clock and synchronously drives the data acquisition actions of the point sensor array unit and the field measurement unit through the hardware triggering channel to achieve a time synchronization accuracy of better than 1ms for heterogeneous data sources. The data encapsulation interface unit is used to receive the discrete measurement point data stream of the point sensor array unit and the field data stream of the field measurement unit, perform data verification, outlier removal and format standardization, and encapsulate it into a unified data packet with spatiotemporal tags for output.

3. The intelligent identification system for ship structural failure modes based on multi-source field fusion as described in claim 1, characterized in that: The multimodal data collaborative processing module includes a DIC data calculation unit, a displacement field reconstruction unit, a strain field deduction unit, and a multi-source data verification and fusion unit. The DIC data processing unit is used to establish the correlation function through the zero-mean normalized minimum distance sum of squares criterion, and to iteratively solve the shape function parameters using the inverse combination Gauss-Newton algorithm to achieve sub-pixel accuracy calculation of the displacement field of the outer surface of the structure. The dynamic boundary conditions of the region to be reconstructed are extracted from the global displacement field obtained by the calculation. The displacement field reconstruction unit is used to receive discrete strain time history data and the dynamic boundary conditions, construct the inverse shell element model of the region to be reconstructed, and solve the maximum a posteriori estimate under the Bayesian framework to obtain the displacement field of the key internal parts. The strain field derivation unit is used to receive discrete strain data and derive the continuous strain field using an improved sliding least squares method. The multi-source data verification and fusion unit is used to correct systematic errors in the displacement field of the digital image correlation solution by using the measured data of point sensors as a benchmark and adopting the covariance cross-fusion algorithm to adaptively allocate fusion weights according to the uncertainty of each source data.

4. The intelligent identification system for ship structural failure modes based on multi-source field fusion as described in claim 3, characterized in that: The objective function for establishing the displacement field reconstruction element is: Where u is the displacement vector of the node to be determined. ε meas u is the measured strain vector. BC B represents the boundary displacement extracted by DIC, B is the strain-displacement matrix, L is the boundary extraction operator, and W is the boundary displacement extracted by DIC. ε With W u These are the observation confidence and boundary condition accuracy weight matrices, respectively. λ Let Φ(u) be the regularization factor; taking the partial derivative of Φ(u) with respect to u and setting it to zero, we obtain the equilibrium equation: In the formula, K e =B T W ε B is the stiffness matrix assembled based on strain measurement points, K b =L T W u L is the boundary constraint matrix, F e =B T W ε ε meas F b =L T W u u BC .

5. The intelligent identification system for ship structural failure modes based on multi-source field fusion as described in claim 1, characterized in that: The virtual-real data fusion module includes a geometric alignment unit and a feature mapping and fusion unit; The geometric alignment unit is used to perform multi-level registration of the virtual node set and the measured node set in three-dimensional space, including coarse alignment sub-units and fine registration sub-units using the iterative nearest point algorithm; The feature mapping and fusion unit is used to retrieve the K nearest Euclidean neighbors of each target node in the virtual node set from the registered measured node set, calculate the contribution weight based on the inverse distance weighting principle, map the measured displacement field and strain field to the virtual node according to the weight, and perform soft replacement based on the confidence level of the measured data. , where β is the confidence weight of the measured data.

6. The intelligent identification system for ship structural failure modes based on multi-source field fusion as described in claim 1, characterized in that: The intelligent failure mode identification module includes a graph data normalization unit, a time sequence graph sampling unit, a buckling state identification unit, a failure mode discrimination unit, and a failure evolution tracking unit; The graph data standardization unit is used to transform the input graph structure data into an internal calculation format, fill in missing items caused by measurement blind spots with neighborhood mean, and perform Z-score standardization on features of different dimensions. The timing diagram sampling unit is used to select the interface node between the strip plate and the web plate of the stiffened plate of the model as a seed at each load step, and prioritizes to retain nodes in the strong deformation region where the displacement gradient is greater than a preset threshold, so that the node size of the sampled diagram is reduced to 15%~20% of the original diagram. The buckling state identification unit identifies nine types of buckling states of stiffened plates frame by frame based on a depth map convolutional network. The failure mode discrimination unit performs plate-frame level failure mode mapping based on buckling state combination and preset rules. When the states of multiple web plates are inconsistent, the dominant mode is determined according to the priority of "overall > shear > tilt > web plate > plate grid". The failure evolution tracking unit is used to scan and record the load step, mode type and location of the first failure of each board frame by frame, identify the earliest failed board as the initial failure source, track the spatial expansion and transformation of failure modes in subsequent load steps, and determine the stable final state when the failure modes are stable for five consecutive frames and the proportion of failed boards exceeds the threshold.

7. The intelligent identification system for ship structural failure modes based on multi-source field fusion as described in claim 1, characterized in that: The online visualization and criterion generation module includes a field rendering engine unit, a criterion generation unit, and a threshold warning unit; The field rendering engine unit uses OpenGL to perform hardware-accelerated rendering of the ship structure response field, sends node displacements to the vertex shader to drive deformation animation, and inputs strain components to the fragment shader to generate continuous cloud maps. The criterion generation unit is used to extract the bearing capacity-displacement curve from the time series database, automatically detect the peak point or the point of 5% drop in the curve as the ultimate bearing capacity criterion, and calculate the stiffness degradation rate and safety margin. The threshold warning unit is used to preset three thresholds: when the proportion of failed plates exceeds 10%, a loading deceleration command is sent; when the stiffness degradation rate exceeds 30%, an audible and visual alarm is triggered; and when the overall buckling failure probability exceeds 0.8, an emergency alarm signal is sent.

8. A method for an intelligent identification system for ship structural failure modes based on multi-source field fusion as described in claim 1, characterized in that: The following steps are included: S1, Dynamic Synchronous Acquisition of Multi-Source Heterogeneous Data: A point-field fusion sensor network is constructed, and the time synchronization accuracy of heterogeneous data sources is better than 1ms through a spatiotemporal synchronization triggering unit. S2, Multimodal Data Collaborative Processing and Field Reconstruction: The surface displacement field is calculated using a subpixel algorithm based on digital image correlation data; the displacement field of key internal components is reconstructed using a Bayesian framework and inverse shell model by combining discrete strain data and dynamic boundary conditions; the continuous strain field is derived using an improved sliding least squares method; and multi-source data are adaptively weighted and fused using a covariance crossover algorithm. S3, Dynamic Fusion of Virtual and Real Response Fields: Using a nonlinear finite element virtual model as a benchmark, the measured field data and the virtual model are geometrically aligned through an iterative nearest-point algorithm; Based on K-nearest neighbor retrieval and inverse distance weight mapping of the measured displacement / strain field, soft replacement is performed to generate a full-field virtual-real fusion response dataset; S4, Intelligent Failure Mode Identification and Evolution Tracking: The fused graph structure data is standardized and compressed by temporal graph sampling; the buckling state of the stiffened plate is identified frame by frame using a depth graph convolutional network. Based on preset rule logic mapping to board-level failure modes; Record the initial failure source, spatial expansion path, and mode transition events, and determine the overall failure mode in conjunction with the load conditions; S5, online visualization and criterion generation: Real-time display of response field deformation animation and strain cloud map through hardware-accelerated rendering; automatic extraction of quantitative criteria such as ultimate bearing capacity and stiffness degradation rate; triggering multi-level safety warnings; Output an assessment report that complies with ship specifications.

9. The method of the intelligent identification system for ship structural failure modes based on multi-source field fusion as described in claim 8, characterized in that: In S4, based on preset rule logic mapping to board-level failure modes, specifically including: Logical judgment is performed on the state of the strip and web within the same frame: If the plate exhibits localized buckling and the web shows no deformation, it is determined to be plate buckling; If the band plate buckles as a whole and the upper edge of the web plate buckles, it is determined to be overall buckling; If the middle of the web is bent but the bands are not deformed, it is determined to be web bending; If the upper edge of the web is flexed and accompanied by local flexion of the band plate, it is determined to be lateral flexion; If the strip exhibits shear buckling, it is determined to be shear buckling; When the states of multiple web plates are inconsistent, the dominant mode is determined according to the priority of "overall > shear > tilt > web plate > plate grid". S4 records the initial failure source, spatial expansion path, and mode transition events, specifically including: The load step, mode type, and location of the first failure of each board frame are recorded by scanning frame by frame, and the earliest failed board frame is identified as the initial failure source. Track the spatial expansion and transformation of failure modes in subsequent load steps, and record key events such as buckling mode transition, multi-mode competition and dominant mode transformation; When the failure mode is stable for five consecutive frames and the proportion of failed boards exceeds the threshold, it is determined to be a stable final state, forming an "initial-expansion-stable" timing chain.

10. The method of the intelligent identification system for ship structural failure modes based on multi-source field fusion as described in claim 8, characterized in that: In S5, multiple levels of security alerts are triggered, specifically including: Level 1 warning: When the proportion of failed plates exceeds 10%, a loading deceleration command is sent to the test loading system; Level 2 warning: When the stiffness degradation rate exceeds 30%, an audible and visual alarm is triggered; Level 3 warning: When the overall buckling failure probability exceeds 0.8, an emergency alarm signal is sent to the test loading system controller.

Images