Intelligent identification method for surface cracks in radiator vibration test

Abstract

The application discloses a kind of surface crack intelligent identification method in radiator vibration test, belong to image recognition technical field, it includes obtaining the vibration excitation parameter of vibration table real-time output;Determine image acquisition phase window based on vibration excitation parameter;Dynamic trigger shooting is executed to image acquisition phase window, and optimization image sequence is generated;Vibration perception preprocessing is carried out to optimization image sequence, and enhanced image is generated;From enhanced image, visual feature and physical response feature are extracted and fused, to obtain fusion feature set;Collaborative analysis is carried out to fusion feature set, and crack identification result and corresponding confidence score are output.The application adopts vibration phase adaptive control technology, combines multi-physical field feature fusion mechanism, can realize the accurate identification of microcrack under strong vibration environment, improve the detection capability of crack.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and in particular to an intelligent method for identifying surface cracks in radiator vibration tests. Background Technology

[0002] As a critical heat exchange component in vehicles and various engineering machinery, the structural reliability of the radiator directly affects the stable operation of the entire system. In product development and quality control, vibration testing is a core method for simulating the dynamic loads on the radiator under actual working conditions and evaluating its fatigue resistance and structural durability. During vibration, the initiation and propagation of surface cracks are the main forms leading to eventual failure. Therefore, real-time and accurate crack monitoring of the radiator surface during vibration testing is of great significance for preventing catastrophic failures and optimizing product design. This process falls under the category of non-destructive testing using computer vision technology.

[0003] Currently, machine vision-based techniques are partially employed for surface crack detection in vibration tests. Typically, this approach involves setting up an industrial camera near the test bench and capturing images of the radiator surface at fixed time intervals or after a set number of vibration cycles. After acquiring the images, traditional image processing algorithms, such as edge detection, thresholding, and morphological filtering, are used to process them in an attempt to highlight the crack outline. Finally, by setting geometric criteria such as length and aspect ratio, potential cracks are identified.

[0004] Existing technologies have several shortcomings in practical applications. First, in the image acquisition stage, taking pictures at fixed time intervals is highly unreliable. Because a crack under sinusoidal vibration load only reaches its maximum opening at the very brief moment of the displacement peak, it may be in a closed or semi-closed state for most of the remaining time, appearing as weak or even invisible features in the image. Fixed-period shooting greatly increases the likelihood of missing the optimal observation window, resulting in the acquisition of a large number of low-quality, low-information images, increasing the burden of data storage and processing, and directly affecting the crack detection rate. Second, in the feature recognition stage, relying solely on the geometric or texture features of static images makes it difficult to effectively distinguish real cracks from false defects such as surface scratches, welds, material textures, or oil stains. These false defects are morphologically very similar to early micro-cracks, easily leading to a large number of false alarms and misreports. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides an intelligent method for identifying surface cracks in radiator vibration tests. By employing vibration phase adaptive control technology and combining it with a multi-physics field feature fusion mechanism, it can accurately identify microcracks under strong vibration conditions, thereby improving crack detection capabilities.

[0006] The above objectives can be achieved through the following approach:

[0007] A method for intelligent identification of surface cracks in radiator vibration tests includes: acquiring vibration excitation parameters output in real time by a vibration table; calculating the vibration period and determining the image acquisition phase window based on these parameters; generating a phase-synchronized optimized image sequence by dynamically triggering image capture; performing vibration-sensing preprocessing on the image sequence to generate enhanced images; synchronously extracting the crack geometric evolution field and damage response tensor from the enhanced images, and fusing them into a feature set through physical constraints; performing fracture mechanics-guided spatiotemporal decoupling and dual-domain collaborative analysis on the fused feature set, and outputting crack identification results and confidence scores.

[0008] Optionally, determining the image acquisition phase window based on the vibration excitation parameters includes: calculating the vibration period based on the real-time frequency in the vibration excitation parameters; identifying the phase interval of the vibration displacement extreme value based on the vibration period; obtaining a potential acquisition phase point set based on the phase interval of the vibration displacement extreme value; and performing response timing optimization and synchronous dynamic compensation on the potential acquisition phase point set to determine the image acquisition phase window.

[0009] Optionally, obtaining the potential acquisition phase point set based on the phase interval of the vibration displacement extreme value includes: obtaining a crack propagation sensitive phase set based on the phase interval of the vibration displacement extreme value; and performing multimodal co-registration on the crack propagation sensitive phase set to obtain the potential acquisition phase point set.

[0010] Optionally, the step of performing vibration-sensing preprocessing on the optimized image sequence to generate an enhanced image includes: obtaining a vibration coupling feature tensor based on the optimized image sequence; and performing dynamic convolution filtering on the vibration coupling feature tensor to generate an enhanced image.

[0011] Optionally, the step of extracting visual features and physical response features from the enhanced image and fusing them to obtain a fused feature set includes: extracting visual features from the enhanced image to obtain a crack geometric evolution field; extracting physical response features from the enhanced image to obtain a damage response tensor; and performing multimodal recognition fusion on the crack geometric evolution field and the damage response tensor to obtain a fused feature set.

[0012] Optionally, the step of extracting physical response features from the enhanced image to obtain the damage response tensor includes: calculating a phase-analyzed pixel displacement field based on the enhanced image; extracting local anomalous strain regions from the phase-analyzed pixel displacement field to obtain a crack core strain field; quantifying the amplitude change rate and directional consistency of the crack core strain field to generate a micro-strain dynamic vector; and encoding the micro-strain dynamic vector using multi-physics coupling damage features to obtain the damage response tensor.

[0013] Optionally, the step of performing collaborative analysis on the fused feature set to output crack identification results and corresponding confidence scores includes: performing spatiotemporal feature decoupling of the fused feature set under physical constraints to obtain spatial structural components and time-varying dynamic components; performing dual-domain collaborative causal analysis on the spatial structural components and time-varying dynamic components to output crack identification results and corresponding confidence scores.

[0014] Optionally, the step of physically constraining the spatiotemporal feature decoupling of the fused feature set to obtain the spatial structure component and the time-varying dynamic component includes: obtaining a material constitutive perception feature tensor based on the fused feature set; separating the material constitutive perception feature tensor into a vibrational phase synchronization field to obtain a phase-aligned feature field; and performing fracture mechanics-guided spatiotemporal decoupling of the phase-aligned feature field to obtain the spatial structure component and the time-varying dynamic component.

[0015] Optionally, the step of dynamically triggering the image acquisition phase window to generate an optimized image sequence includes: dynamically gating the image acquisition phase window and performing real-time phase error compensation to obtain a phase-synchronized image frame set; and performing cross-period feature alignment and spatiotemporal fusion reconstruction on the phase-synchronized image frame set to generate an optimized image sequence.

[0016] Based on the same inventive concept, this invention also provides an intelligent surface crack recognition system for radiator vibration testing. The system includes: a vibration parameter acquisition module for acquiring vibration excitation parameters output in real time by a vibration table; a phase window decision module for determining an image acquisition phase window based on the vibration excitation parameters; a dynamic shooting control module for dynamically triggering shooting within the image acquisition phase window to generate an optimized image sequence; a vibration sensing preprocessing module for performing vibration sensing preprocessing on the optimized image sequence to generate an enhanced image; a multimodal recognition fusion module for extracting and fusing visual features and physical response features from the enhanced image to obtain a fused feature set; and a crack recognition decision module for performing collaborative analysis on the fused feature set and outputting crack recognition results and corresponding confidence scores.

[0017] Compared with the prior art, the present invention has the following advantages:

[0018] This invention precisely determines the image acquisition phase window with the most significant crack features by real-time analysis of vibration excitation parameters and performs synchronous dynamic triggering for image capture. This method abandons the inefficient traditional fixed-frequency or random shooting mode, ensuring that each frame of image has extremely high information value from the data acquisition source. This not only significantly improves the original signal-to-noise ratio of crack features but also significantly reduces data redundancy and the computational load of subsequent processing, thereby improving the overall operating efficiency of the identification system.

[0019] This invention proposes a crack identification strategy that integrates visual and physical response features. Instead of relying solely on the static geometry of the crack, this method innovatively introduces the mechanical behavior of the crack under vibration loads as a key criterion. By analyzing the local strain field and other physical responses in the crack region, a dual verification mechanism is established. This deep fusion of multimodal features can effectively distinguish between real, dynamic cracks that threaten structural safety and static, harmless surface scratches or oil stains, fundamentally improving the accuracy and reliability of the identification.

[0020] The vibration sensing preprocessing and physical constraint-based collaborative analysis method employed in this invention significantly enhance the system's ability to suppress complex backgrounds and noise. Through dynamic convolutional filtering, dynamic crack signals with the same frequency as the vibration can be extracted from time-series images, effectively filtering out static texture interference. Subsequent collaborative analysis and causal inference further base recognition decisions on an understanding of physical laws, making the recognition process more robust and interpretable, avoiding the misjudgments easily caused by traditional methods that rely solely on apparent features.

[0021] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating a method for intelligent identification of surface cracks in a radiator vibration test according to an embodiment of the present invention.

[0024] Figure 2 This is a multimodal signal cooperative registration diagram according to an embodiment of the present invention.

[0025] Figure 3 This is a vibration coupling feature tensor diagram according to an embodiment of the present invention.

[0026] Figure 4 This is a diagram showing the crack geometry evolution field extraction results according to an embodiment of the present invention.

[0027] Figure 5 This is a diagram showing the strain field analysis results of the crack core in an embodiment of the present invention.

[0028] Figure 6 This is a confidence score curve for crack identification according to an embodiment of the present invention.

[0029] Figure 7 This is a schematic diagram of the structure of an intelligent surface crack recognition system for radiator vibration testing according to an embodiment of the present invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Reference Figure 1 One embodiment of the present invention proposes an intelligent identification method for surface cracks in radiator vibration tests. It adopts vibration phase adaptive control technology and combines a multi-physics field feature fusion mechanism to achieve accurate identification of microcracks under strong vibration environment and improve crack detection capability.

[0032] The method described in this embodiment specifically includes:

[0033] Obtain the vibration excitation parameters output in real time by the vibration table;

[0034] Based on the vibration excitation parameters, the image acquisition phase window is determined;

[0035] Dynamically triggering image capture is performed on the image acquisition phase window to generate an optimized image sequence;

[0036] The optimized image sequence is subjected to vibration sensing preprocessing to generate enhanced images;

[0037] Visual features and physical response features are extracted from the enhanced image and fused to obtain a fused feature set;

[0038] The fused feature set is subjected to collaborative analysis to output crack identification results and corresponding confidence scores.

[0039] Specifically, the method first acquires the physical excitation parameters of the vibration table in real time to fundamentally understand the dynamic environment of the radiator. Based on this physical insight, the method proactively predicts and determines the vibration phase with the most significant crack characteristics, forming an optimal image acquisition phase window. Precise dynamic triggering is then performed based on this window to obtain a highly information-density optimized image sequence. Subsequently, the method uses the vibration information contained in the sequence to perform perceptual preprocessing on the images, solidifying and enhancing the dynamic crack response characteristics. Building on this, the method is no longer limited to a single visual representation but extracts in parallel visual features describing the crack geometry and physical response features characterizing its mechanical behavior. By deeply fusing these two different but intrinsically related features, a comprehensive feature set with both morphological and mechanical connotations is formed. Finally, through collaborative analysis of this fused feature set, a judgment on the presence or absence of a crack is derived, and the credibility of this judgment is quantified.

[0040] Optionally, determining the image acquisition phase window based on the vibration excitation parameters includes:

[0041] The vibration period is calculated based on the real-time frequency in the vibration excitation parameters.

[0042] Based on the vibration period, the phase interval of the vibration displacement extreme value is identified;

[0043] Based on the phase interval of the extreme values ​​of the vibration displacement, a potential set of acquisition phase points is obtained;

[0044] The potential acquisition phase point set is optimized for response timing and dynamically compensated for synchronization to determine the image acquisition phase window.

[0045] Specifically, the system first continuously acquires vibration excitation parameters from the control system of the vibration test bench, with the core parameter being the real-time changing excitation frequency. The instantaneous vibration period of the current vibration load on the radiator can be calculated using this real-time frequency. This calculation follows a physical formula: the vibration period is equal to the reciprocal of the real-time frequency. The formula is as follows:

[0046] ,

[0047] in, This represents the calculated vibration period. This represents the real-time frequency obtained from the vibration excitation parameters. The vibration period is then obtained. The next step of the method is to identify the phase intervals of the extreme vibration displacements. Within a complete vibration cycle, the maximization of crack opening on the radiator surface occurs when the displacement reaches its positive or negative peak; the phases at these moments provide the best observation opportunities. Therefore, the system calibrates the phase intervals containing these peak displacements based on the vibration cycle. To further improve the accuracy of phase localization, the method uses the identified phase intervals of the extreme vibration displacements as a basis and employs multimodal co-registration technology, such as multimodal signal co-registration... Figure 2 As shown, the phases most sensitive to crack propagation are precisely located, resulting in a set of crack propagation-sensitive phases. The co-registration here integrates data from other measurement modes, such as strain gauges or acoustic emission sensors, ensuring that the identified phases represent the peak points of the actual stress and strain of the structure, rather than theoretical peak points of the driving signal. Thus, the system obtains a potentially optimal set of acquisition phase points composed of multiple high-value phase moments. Finally, considering the inherent system delays between issuing the trigger command and the camera completing the exposure, such as signal transmission and shutter response time, the system performs response timing optimization and synchronous dynamic compensation on this point set. This optimization process pre-calculates and compensates for system delays, while dynamic compensation continuously adjusts the trigger lead based on real-time frequency changes, ultimately determining a precisely synchronized and optimized time window—the optimal image acquisition phase window—to guide subsequent image acquisition.

[0048] Optionally, obtaining the potential acquisition phase point set based on the phase interval of the vibration displacement extreme value includes:

[0049] Based on the phase interval of the extreme values ​​of the vibration displacement, a crack propagation sensitive phase set is obtained;

[0050] Multimodal co-registration is performed on the crack propagation sensitive phase set to obtain a potential acquisition phase point set.

[0051] Specifically, the system first needs to filter out the crack propagation sensitive phase set from the phase interval. This step is based on the fact that there may be a slight phase difference between the overall displacement peak value and the peak value of local stress and strain at the crack tip. The opening and closing of the crack is directly determined by the stress field in its tip region. Therefore, it is necessary to further locate those precise phases that can induce the maximum tensile stress and thus make the crack characteristics most significant within the phase interval of the vibration displacement extreme values. These phases together constitute the crack propagation sensitive phase set, representing the theoretically most favorable set of moments for crack observation. Subsequently, in order to accurately align the theoretically optimal phases with the actual response in the physical world, the method performs multimodal co-registration on the crack propagation sensitive phase set. Here, "multimodal" refers to the simultaneous introduction of other physical sensors in addition to visual sensors, such as strain gauges or acoustic emission sensors attached to key locations on the heat sink. During the vibration test, the system precisely aligns each candidate phase in the crack propagation sensitive phase set with the real-time strain peak value measured by the strain gauge or the energy release peak value captured by the acoustic emission sensor in time. The goal of co-registration is to find one or more phase points where the response signals from multiple different physical sensors simultaneously reach their peak or exhibit the strongest correlation. The optimal synchronization phase can be determined using the following co-response function:

[0052] ,

[0053] in, These are the synchronization phase points obtained through registration. It is a candidate phase in the crack propagation sensitive phase set. It is the first Modal sensors in phase Normalized signal strength at time, These are the confidence weights for the corresponding modes. It is a normalization function that solves the problem of incomparable dimensions between different sensors, and performs preliminary normalization on all single modes. The function is uniformly mapped to the dimensionless interval [0,1] to achieve comparability of cross-modal signals. By solving this function, the phase point that makes the multi-source physical evidence reach its strongest value can be found. This point or these points constitute the highly reliable potential optimal acquisition phase point set of the final output.

[0054] Optionally, the step of performing vibration-sensing preprocessing on the optimized image sequence to generate an enhanced image includes:

[0055] Based on the optimized image sequence, the vibration coupling feature tensor is obtained;

[0056] The vibration coupling feature tensor is subjected to dynamic convolution filtering to generate an enhanced image.

[0057] Specifically, the system first constructs a vibration coupling feature tensor from the acquired optimized image sequence, i.e., multiple frames of images continuously captured within the image acquisition phase window. The vibration coupling feature tensor is as follows: Figure 3 As shown. The vibration coupling feature tensor is a three-dimensional structure, and its dimension is defined as follows. ,in These are the height and width of the image, respectively. The time dimension (corresponding to the number of phase frames within the vibration period, with a value of 8) 32 frames (positively correlated with vibration frequency). Number of image channels (grayscale image) Color images =3). This tensor is generated by stacking optimized image sequences according to the acquisition time sequence, preserving the dynamic changes of the crack within the vibration period. The dynamic convolutional filtering uses a three-dimensional spatiotemporal convolution kernel with The structure, The dimension in the width direction is pixels, with dimensions in the height direction as The pixel dimensions and parameters are designed as follows: The value can be 3 or 5 pixels. When the fin spacing on the heatsink surface is ≥0.5mm, When the fin spacing is <0.5mm (dense structure), To accommodate the spatial scale of minute cracks. Time dimension size. The value is 3 5 frames. Based on vibration period. Calculations show that the following conditions are met:

[0058] ，

[0059] The proportionality coefficient is set to 0.1. This refers to the total number of frames within a single cycle, ensuring coverage of the dynamic process of a crack opening and closing completely in one operation. Input / output channels. The feature representation capability is enhanced through channel expansion. Convolutional kernel parameters are jointly determined through "physical constraints + data training": Physical prior initialization: The time dimension weights are set according to the crack opening and closing rules, assigning high weights to the extreme phases of vibration displacement (the moment of maximum crack opening / closing) (center frame weight 0.3, adjacent frames 0.2, edge frames 0.1); the spatial dimension is initialized using a Laplacian kernel to strengthen edge gradient features and highlight the crack outline. Training optimization: 500 sets of radiator vibration crack samples (covering 0.05...) are used. A 3D convolutional neural network was constructed using a dataset of 1mm microcracks (different materials and load conditions). The signal-to-noise ratio (SNR) of the dynamic response in the crack region was used as the optimization objective, and the Adam optimizer (learning rate) was employed. The parameters were fine-tuned (after 30 iterations). Vibration phase labels were introduced as supervision during training to constrain the consistency between the temporal features and the physical vibration laws. Image sequences. Includes spatial coordinates and discrete time or phase index By stacking these two-dimensional images along the time axis, a three-dimensional data volume is formed. This is the vibration coupling feature tensor. This tensor not only preserves the spatial information of each frame of the image, but more importantly, it encodes the subtle morphological changes of the crack under different vibration phases, such as opening and closing, as a sequence of pixel intensity changes in the tensor over time, thus coupling vibration information with image features. Next, the method performs dynamic convolution filtering on this vibration coupling feature tensor. Unlike traditional two-dimensional convolution for processing static images, dynamic convolution filtering uses a three-dimensional spatiotemporal convolution kernel. Perform operations on this tensor. The convolution kernel has a spatial dimension of... Capture the morphological features of cracks, and simultaneously in the time dimension. The kernel is designed to identify specific patterns of change. For example, it is optimized to produce a high response to high-frequency alternating brightness changes in pixels over time (corresponding to the rapid opening and closing of cracks), while being insensitive to static backgrounds or slow changes in illumination. The filtering process can be represented as:

[0060] ,

[0061] in, It is the filtered feature tensor. It is the input vibration coupling feature tensor. It is a three-dimensional dynamic convolution kernel. The output of this operation... A significantly enhanced signal is generated in the crack region. Finally, in order to generate a single enhanced image... The system processes the filtered three-dimensional tensor. Projection aggregation is performed along the time dimension, for example, by using the time maximum value projection, which takes each spatial pixel. At all points in time The maximum response value is used as the final pixel value. In this way, even if the crack is not clear in some frames, as long as it is successfully enhanced in any frame of the sequence, its features will be preserved and highlighted in the final enhanced image.

[0062] Optionally, the step of extracting visual features and physical response features from the enhanced image and fusing them to obtain a fused feature set includes:

[0063] Visual features are extracted from the enhanced image to obtain the crack geometric evolution field;

[0064] Physical response features are extracted from the enhanced image to obtain the damage response tensor;

[0065] The crack geometric evolution field and the damage response tensor are fused using multimodal recognition to obtain a fused feature set.

[0066] Specifically, the system first extracts visual features from the enhanced image to generate a crack geometric evolution field. The crack geometric evolution field extraction result is as follows: Figure 4 As shown, the system utilizes deep learning segmentation models or advanced image processing algorithms to accurately identify and delineate the crack contour in the enhanced image and its source optimized image sequence. Subsequently, the changes in this contour over time are quantitatively analyzed to calculate the variations of key geometric parameters such as crack length, width, and tip curvature radius over time. These time-varying geometric parameters collectively constitute a vector field distributed in space and time, namely the crack geometric evolution field. Next, the method extracts physical response features in parallel from the same enhanced image and source sequence to obtain the damage response tensor. This step goes beyond observing the surface morphology and instead analyzes the mechanical behavior of the structure under vibration. By employing digital image correlation (DIC) or optical flow methods, the system calculates the displacement trajectory of each pixel in the image during vibration. Based on this displacement field, the strain field distribution on the material surface can be further obtained by solving its spatial gradient. The presence of a crack causes strain concentration in its tip region; this localized high-strain region is a direct physical representation of structural damage. The system constructs the time-varying strain field and the derived stress state information into a multi-dimensional tensor, namely the damage response tensor, which accurately describes the mechanical response of the crack region. Finally, multi-modal recognition fusion is performed on the obtained crack geometric evolution field and damage response tensor to generate a final fused feature set. This fusion is a deep integration at the feature level. The system concatenates the morphological and evolutionary feature vectors extracted from the crack geometric evolution field with the mechanical response feature vectors extracted from the damage response tensor, or uses a more complex weighted fusion strategy to generate a unified and higher-dimensional combined feature vector for each suspected crack region. Each element in this fused feature set simultaneously encodes the crack's "appearance" and "behavior," i.e., its static geometric properties and dynamic mechanical properties.

[0067] Optionally, extracting physical response features from the enhanced image to obtain the damage response tensor includes:

[0068] Based on the enhanced image, the phase-analyzed pixel displacement field is calculated;

[0069] The local abnormal strain region of the phase-analyzed pixel displacement field is extracted to obtain the crack core strain field;

[0070] The amplitude change rate and directional consistency of the strain field at the crack core are quantified to generate a micro-strain dynamic vector;

[0071] The damage characteristic is encoded by multi-physics coupling of the micro-strain dynamic vector to obtain the damage response tensor.

[0072] Specifically, the system first calculates the phase-resolved pixel displacement field based on the enhanced image obtained from the vibration sensing preprocessing step and the underlying original optimized image sequence. Here, "phase resolution" means determining the precise displacement of each pixel on the heat sink surface at each key phase of the vibration cycle. This calculation is achieved using digital image correlation (DIC) technology. The system uses the first frame or reference frame in the optimized image sequence as a baseline and precisely tracks its spatial movement by matching the texture patterns of tiny pixel regions in subsequent frames, thus resolving the displacement field for each pixel. In different phases A displacement vector is generated, and this displacement vector field is the phase-resolved pixel displacement field. Next, the method extracts local anomalous strain regions from this displacement field to obtain the crack core strain field. The analysis results of the crack core strain field are as follows: Figure 5 As shown. Strain is the spatial rate of change of displacement, a direct physical quantity measuring local deformation of a material. The system calculates the strain tensor of each pixel by spatially differentiating the phase-analyzed pixel displacement field. The crack tip is a typical region of stress-strain concentration, with strain values ​​much higher than those in the matrix of the material. Therefore, the system identifies these high-strain regions by setting a strain threshold or using an anomaly detection algorithm; the set of these regions is defined as the crack core strain field. This process can be described by the following equation:

[0073] ,

[0074] in, At the pixel and phase The strain tensor at the point It is the displacement vector in the phase-analyzed pixel displacement field obtained in the previous step. It is a spatial gradient operator. The transpose of the tensor is used. After determining the strain field at the crack core, the system further quantifies the dynamic characteristics of this strain field to generate a micro-strain dynamics vector. This step aims to capture the dynamic behavior of the crack "breathing" under vibrational loads. The system analyzes each point in the crack core strain field along the time or phase dimension, calculating two key indicators: the rate of change of amplitude, i.e., how quickly the strain magnitude changes over time; and the consistency of direction, i.e., whether the principal strain direction remains stable within the vibration period. These two dynamic indicators are combined into a vector, namely the micro-strain dynamics vector, which describes the dynamic fingerprint of the local strain in the crack region. Finally, this vector is encoded using multi-physics coupling damage features to generate the final damage response tensor. This encoding process associates the micro-strain dynamics vector with prior physical knowledge such as the constitutive relation of the material and fatigue damage models, transforming it into a structured feature representation that can be efficiently understood by machine learning models, namely the damage response tensor. Multiphysics coupling damage feature encoding is a process of quantitatively associating microstrain dynamic vectors with physical knowledge such as material constitutive relations and fatigue damage models. Specifically, it includes the following steps: physical parameter mapping based on material constitutive relations (e.g., Hooke's law for metallic materials), transforming strain parameters in the microstrain dynamic vectors into stress parameters. For the linear elastic stage, the stress tensor... With strain tensor The mapping relationship is as follows:

[0075] ,

[0076] in, The elastic modulus of the material (determined by the material of the heat sink, such as 69 GPa for aluminum alloy). Poisson's ratio (0.33 for aluminum alloys). The trace of the strain tensor This is a unit tensor. This step transforms the "rate of change of amplitude" in the micro-strain dynamics vector into the "rate of change of stress amplitude," and "directional consistency" into "principal stress direction consistency." Fatigue damage quantification: Combining a fatigue damage model (such as Miner's linear cumulative damage rule), the cumulative damage value in the crack core region is calculated. For cyclic vibration loading, the damage accumulation formula is:

[0077] ,

[0078] in, Cumulative damage value (It was determined to be fatigue failure at that time). Current stress level The number of loops below, The fatigue life corresponding to this stress level (obtained from the material's SN curve). Feature dimension fusion: The final form of the damage response tensor is defined as... ,in: The spatial coordinates of the crack core region; The time dimension corresponds to the phase point of the vibration period, which is consistent with the time index of the optimized image sequence. For physical feature dimension ( Corresponding strain amplitude, Corresponding stress amplitude, Corresponding to the principal direction angle, (Corresponding cumulative damage value). The encoding process is achieved through tensor concatenation: the stress and damage features obtained in the previous steps are combined with the geometric features (direction, amplitude) in the original micro-strain dynamic vector according to dimension. Stack them to form a four-dimensional damage response tensor.

[0079] Optionally, the step of performing collaborative analysis on the fused feature set to output crack identification results and corresponding confidence scores includes:

[0080] The spatiotemporal features of the fused feature set are physically constrained to decouple them, resulting in spatial structure components and time-varying dynamic components.

[0081] A dual-domain collaborative causal analysis is performed on the spatial structural components and the time-varying dynamic components to output crack identification results and corresponding confidence scores.

[0082] Specifically, the system first needs to physically constrain the spatiotemporal features of the fused feature set to decouple them, aiming to separate two independent parts describing the static geometry of the crack and its dynamic evolution behavior. This decoupling process is constrained by the physical behavior of the material, ensuring that the separated components have clear physical meaning. Through this step, the originally highly coupled spatiotemporal features are decomposed into two orthogonal sub-feature sets: one is a spatial structure component that does not change or changes slowly over time, which mainly encodes the static properties of the crack, such as the average geometry, length, and path; the other is a time-varying dynamic component that changes rapidly over time, which mainly encodes the dynamic response characteristics of the crack, such as opening and closing during the vibration cycle and stress cycling at the tip. After successfully decoupling the spatial structure component and the time-varying dynamic component, the system performs a two-domain co-causal analysis on these two components. The two-domain co-causal analysis includes the following steps: first, heterogeneous data processing: the static spatial components (such as crack length and location) are discretized according to the monitoring period and transformed into pseudo-time series data of the same length as the dynamic time series. Then, the Granger test is applied: the aligned spatial pseudo-time series is processed... and dynamic timing Construct an autoregressive model:

[0083] ,

[0084] express Time-varying dynamic components at any given moment (such as physical quantities that evolve over time, such as stress cycling at the crack tip and strain changes). This represents the constant term of the model, which is the intercept of the regression equation; This represents the influence of historical information of the time-varying dynamic components on the current value, where yes The time-varying dynamic components at time t, These are the coefficients corresponding to historical terms, reflecting the temporal correlation of the dynamic components themselves; This represents the influence of historical information of spatial structure components on time-varying dynamic components, where yes Spatial structural components at any given time (pseudo-time series data that has been discretized over time, such as historical values ​​of parameters like crack length and location). These are the coefficients of the corresponding spatial terms, used to determine the causal driving effect of spatial features on dynamic features; This represents the error term, indicating random perturbations that the model cannot explain. (The rest of the text appears to be a fragment and requires further context for accurate translation.) Test and judgment Significance, verification right The causality. Finally, proceed with... Quantification:

[0085] ，

[0086] in To test the statistic, This is the critical value. The maximum value in the sample is denoted as [0,1], reflecting the causal strength. Here, "dual domain" refers to the spatial and temporal domains. Co-occurrence analysis no longer simply inputs the two components as independent evidence into a classifier, but rather explores the causal relationship between them. For example, the analysis model verifies whether a specific crack geometry described by the spatial structural component leads to a specific dynamic response observed by the time-varying dynamic component. This causal inference can be achieved using Granger causality tests or more complex structural causal models, the core of which is determining whether changes in spatial characteristics can predict or "cause" changes in temporal characteristics. Dual-domain co-occurrence causal analysis model It can be formalized as:

[0087] ,

[0088] in, Represents spatial structural components. Represents time-varying dynamic components, output This is a quantitative indicator of causal relationship strength. The model will only award a high score if and only if a significant, physically consistent causal relationship exists between the two components. Finally, the system integrates the results of the causal analysis with the strength of the two components themselves, and outputs the final crack identification result—that is, whether a crack exists—through a decision function or a pre-trained classification model. Simultaneously, the system will also generate a quantitative confidence score based on the strength of the causal analysis and the significance of the features. The crack identification confidence score curve is shown below. Figure 6 As shown, the score reflects the system's level of confidence in the identification result.

[0089] Optionally, the step of physically constraining the spatiotemporal feature decoupling of the fused feature set to obtain spatial structure components and time-varying dynamic components includes:

[0090] Based on the fused feature set, the material constitutive awareness feature tensor is obtained;

[0091] Vibrational phase synchronization field separation is performed on the constitutive sensing feature tensor of the material to obtain a phase-aligned feature field;

[0092] The phase-aligned feature field is decoupled in a fracture mechanics-guided spatiotemporal manner to obtain the spatial structural component and the time-varying dynamic component.

[0093] Specifically, this method details the spatiotemporal feature decoupling of physical constraints to obtain the specific implementation path for spatial structural components and time-varying dynamic components. The entire process begins with the fused feature set obtained in the previous step. First, based on this fused feature set, the system generates a material constitutive-aware feature tensor. The core of this step is to combine the fused features with the constitutive relation of the heat sink material. The constitutive relation is a physical law describing the relationship between stress and strain in a material; for example, for metallic materials, it follows Hooke's law within the linear elastic range. By substituting the strain information contained in the fused feature set into the constitutive model of the material, the system can deduce the corresponding stress field. In this way, the original feature tensor is given a deeper physical meaning, upgrading from a simple geometric and kinematic description to a material constitutive-aware feature tensor that can reflect the internal mechanical state of the material. Next, the method performs vibration phase synchronization field separation on the material constitutive-aware feature tensor to obtain a phase-aligned feature field. Since image acquisition is performed within a phase window, different frames correspond to different small phases of the vibration period. To eliminate the effects of phase jitter and clearly analyze the evolution of features within a complete cycle, the system needs to align feature data from different cycles. Using the synchronization signal output from the vibration table, the system precisely maps each feature data point to the corresponding phase of the standard vibration cycle. This process is similar to phase-locked amplification of a chaotic pulse signal, resulting in a phase-aligned feature field where the evolution of features is strictly organized according to the standardized vibration phase, eliminating the effects of random time delays and frequency drift. Finally, the system performs fracture mechanics-guided spatiotemporal decoupling on this phase-aligned feature field, ultimately obtaining the desired spatial structural components and time-varying dynamic components. This decoupling process is guided by fracture mechanics theory. For example, according to fracture mechanics, the average stress intensity factor of the crack determines its propagation trend, a relatively slowly varying quantity that can be classified as a spatial structural component; while the stress cycle range at the crack tip determines the accumulation rate of fatigue damage, a quantity that changes rapidly with the phase and can be classified as a time-varying dynamic component. Decoupling can be achieved through matrix decomposition techniques such as Principal Component Analysis (PCA) or Independent Component Analysis (ICA). Fracture mechanics-guided spatiotemporal decoupling achieves constraints by modifying the objective function and selecting basis functions. Modifications to the objective function include adding fracture mechanics constraint terms to the reconstruction error terms of PCA / ICA, such as those based on the Paris formula.

[0094] ,

[0095] These are weighting coefficients. This represents the crack propagation rate. From space basis calculate, These are material constants;

[0096] Based on Griffith's theory:

[0097] ,

[0098] These are the weighting coefficients. For strain energy release rate, This is the critical value. By filtering the basis functions, we have: retaining those that satisfy... (spatial domain) and and The basis functions exhibit a power-law relationship (time domain), and invalid components are eliminated. The choice of the decomposition basis is not purely based on statistical data, but is constrained by the physical model of fracture mechanics, ensuring that each principal component has a clear physical correspondence. Decoupling can be expressed as:

[0099] ,

[0100] in, It is the input phase-aligned feature field. The decoupling operator, which is guided by fracture mechanics, outputs spatial structure components. and time-varying dynamic components Two independent feature sets.

[0101] Optionally, the step of dynamically triggering image capture within the image acquisition phase window to generate an optimized image sequence includes:

[0102] Dynamic gating triggering and real-time phase error compensation are applied to the image acquisition phase window to obtain a phase-synchronized image frame set;

[0103] The phase-synchronized image frame set is subjected to cross-cycle feature alignment and spatiotemporal fusion reconstruction to generate an optimized image sequence.

[0104] Specifically, the system first needs to perform dynamic gating triggering and real-time phase error compensation for the image acquisition phase window. Specifically, the system generates a high-level gating signal based on the start and end phases of the window. When the real-time phase of the vibration table enters the window, the gating signal opens, allowing the high-speed camera to continuously acquire images; when the phase leaves the window, the gating closes, and acquisition pauses. More importantly, due to the response delay of the mechanical and electrical systems, there will be a slight phase error between the actual captured phase and the commanded phase. The system uses a feedback loop to monitor the actual phase of the captured images in real time and compares it with the target phase to calculate the phase error. Then, through a PID controller or similar control algorithm, the advance of the next trigger signal is dynamically adjusted to compensate for this error in real time, ensuring that each acquired image frame falls as accurately as possible on the target phase. The output of this process is a set of images acquired within the optimal phase window over one or more vibration cycles, i.e., a phase-synchronized image frame set. After obtaining the phase-synchronized image frame set, the method further performs cross-cycle feature alignment and spatiotemporal fusion reconstruction to generate the final optimized image sequence. Vibration tests typically last for thousands or even millions of cycles, and analyzing images from a single cycle may be insufficient to capture the slow evolution of cumulative crack damage. Therefore, the system acquires phase-synchronized image frames from multiple vibration cycles spaced far apart. However, due to the potential for minute macroscopic displacements or attitude changes in the radiator during prolonged vibration, image frames acquired from different cycles may not be perfectly aligned spatially. To address this, the system employs an image registration algorithm, using the image frames from the first acquisition cycle as a baseline, and performing rigid or non-rigid transformations on the image frames from all subsequent cycles to eliminate the influence of macroscopic displacements and achieve cross-cycle feature alignment. After alignment, the system fuses and reconstructs these data in the spatiotemporal domain. For example, averaging or weighted averaging multiple frames of the same phase after alignment effectively suppresses random noise, enhances the signal-to-noise ratio, and preserves the subtle but continuous morphological changes caused by fatigue damage accumulation in the crack. This fusion and reconstruction process ultimately outputs a highly purified, feature-enhanced image sequence that reflects the long-term evolution trend of the crack—the optimized image sequence.

[0105] Based on the same inventive concept, such as Figure 7 As shown, the present invention also provides an intelligent surface crack identification system for radiator vibration testing, the system comprising:

[0106] The vibration parameter acquisition module is used to acquire the vibration excitation parameters output in real time by the vibration table;

[0107] The phase window decision module is used to determine the image acquisition phase window based on the vibration excitation parameters;

[0108] The dynamic shooting control module is used to dynamically trigger shooting within the image acquisition phase window and generate an optimized image sequence.

[0109] A vibration sensing preprocessing module is used to perform vibration sensing preprocessing on the optimized image sequence to generate an enhanced image;

[0110] A multimodal recognition fusion module is used to extract visual features and physical response features from the enhanced image and fuse them to obtain a fused feature set;

[0111] The crack identification decision module is used to perform collaborative analysis on the fused feature set and output crack identification results and corresponding confidence scores.

[0112] To verify the feasibility of this invention in practice, it was applied to the quality control of radiator products at an automotive parts manufacturer. This manufacturer needed to conduct high-frequency vibration fatigue tests on its new aluminum radiators to ensure their structural reliability during long-term use. The core challenge was to detect minute surface cracks generated during vibration in real time and accurately. Currently, the manufacturer's detection methods mainly relied on manual penetrant testing or fixed-interval camera photography after the vibration test. However, manual inspection could not achieve process monitoring and was costly, inefficient, and highly subjective; while fixed-interval photography generated massive amounts of redundant data and was prone to missing the critical moment when cracks reached maximum opening under specific vibration phases, leading to missed detections and misjudgments. The manufacturer hoped to use the method of this invention to achieve automated and intelligent crack identification and monitoring throughout the entire vibration testing process. In this embodiment, the manufacturer installed the radiator under test on an electromagnetic vibration test bench and deployed the intelligent identification system described in this invention. The system acquires vibration excitation parameters, such as real-time frequency and amplitude, in real time through the vibration table controller. Based on these parameters, the system intelligently determines the optimal image acquisition phase window and controls a high-speed industrial camera to dynamically trigger shooting. The acquired optimized image sequence undergoes vibration sensing preprocessing, multimodal feature fusion, and co-causal analysis to finally output crack identification results and confidence levels. To verify the beneficial effects of this invention, the manufacturer selected a batch of radiator samples for comparative testing. The experimental group used the method of this invention, while the control group used the traditional fixed-time interval shooting method (100 frames per second). The vibration conditions of the test were a sinusoidal frequency sweep between 50Hz and 200Hz, lasting for 1 million cycles. In the step of determining the optimal acquisition phase window, the system acquired a real-time vibration table frequency of 150Hz, from which the vibration period was calculated. =1 / 150≈6.67ms. The system identified the phase interval where the theoretical displacement extreme value was located. To further refine the positioning, the system used miniature strain gauges attached to the radiator branch pipe connection for multimodal co-registration. It was found that when the strain gauge reading reached its peak, the actual phase lagged slightly behind the theoretical peak. By solving the co-response function, the system determined the optimal image acquisition phase window that maximized the crack opening and dynamically compensated for the camera trigger command. In the image acquisition and preprocessing stage, the experimental group system only triggered the camera to capture images within the optimal phase window of each vibration cycle. After every 10,000 cycles, the acquired phase-synchronized image frame set was aligned and fused across cycles to generate an optimized image sequence. This sequence was then constructed as a vibration coupling feature tensor and an enhanced image was generated through dynamic convolution filtering. The results showed that a tiny crack with a length of about 0.5mm was significantly highlighted in the enhanced image, while it was almost invisible in the single-frame original image of the control group. In the feature extraction and fusion step, the system extracts the geometric evolution field of the crack from the enhanced image (such as the process of the crack length increasing from 0.5 mm to 0.8 mm), and uses digital image correlation (DIC) technology to calculate that the crack tip region contains up to 1500... The periodic strain concentration forms a damage response tensor. Through physical constraint feature field coupling, the system confirms a high degree of consistency between the visual morphology and the physical response, forming a highly reliable fusion feature set. In the final co-analysis step, the system decouples the fusion features into a spatial structural component (the average geometry of the crack) and a time-varying dynamic component (the cyclic characteristics of the stress at the crack tip). Dual-domain co-causal analysis confirms that this specific geometry is the direct cause of the dynamic mechanical response. Therefore, the system outputs "fatigue crack identified" with a 98.5% confidence score. In contrast, the control group's analysis algorithm misreported a surface scratch as a crack due to its low signal-to-noise ratio. The data shows that this invention has significant advantages in terms of accuracy, efficiency, and reliability in crack identification. The experimental group's system achieved early and accurate identification of micro-cracks and effectively eliminated the interference of false defects, while the control group experienced missed detections and false alarms. The method of this invention provides timely and reliable data support for the fatigue life assessment and structural optimization of radiators.

[0113] Table 1 Comparison of Crack Detection Performance

[0114] Comparison items Experimental group (this invention) Control group (traditional method) Total data volume (GB) 65 800 Time to first detection of 0.5mm crack 25 45 Final detection rate (%) 99.5 82.0 False alarm rate (%) 0.5 15

[0115] Table 2 Comparison of Crack Quantitative Analysis Accuracy

[0116] Actual crack length (mm) Actual group measurement (mm) Control group measurements (mm) 0.52 0.55 Not detected 1.13 1.10 1.35 2.45 2.41 2.70 Measurement mean relative error (%) 3.2 20.5

[0117] Table 3 Comparison of System Analysis Efficiency and Reliability

[0118] Comparison items Experimental group (this invention) Control group (traditional method) Time taken per analysis (s) 5.2 1.8 Total analysis time (h) for the complete test cycle 1.5 12.0 Confidence score of results Provided (average 98%) Not provided Automation level Fully automatic Semi-automatic (requires manual review)

[0119] As can be seen from the data in Tables 1 to 3 above, the experimental group, after applying the method of this invention, showed significantly better performance in all aspects compared to the control group. In Table 1, the data acquisition volume of this invention is only 8.1% of that of the traditional method, greatly saving storage and computing resources. Simultaneously, for early microcracks of 0.5 mm, the detection time of this invention is advanced by 200,000 cycles, achieving a final detection rate of 99.5% with an extremely low false alarm rate, demonstrating its high sensitivity and high reliability. Table 2 shows the accuracy of this invention in crack size quantification. The average relative error of the experimental group was only 3.2%, far lower than the 20.5% of the control group. This is attributed to feature extraction and fusion based on physical response, which can more accurately reflect the true state of the crack. In Table 3, although the calculation of a single analysis of this invention is more complex and takes slightly longer, due to its efficient data acquisition strategy, the total analysis time throughout the entire experimental cycle is only 1.5 hours, far less than the 12 hours required by the control group to process massive amounts of data. More importantly, this invention can output a quantitative confidence score, providing a reliable basis for engineering decision-making and achieving true full-process automation.

[0120] It should be noted that the electrical connections between the various units described above do not necessarily represent direct or indirect connections. Any indirect connection method can be applied to the embodiments of the present invention as long as it achieves the purpose of the present invention. The above descriptions are merely exemplary embodiments of the present invention and should not be construed as limiting the scope of the present invention.

[0121] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

Claims

1. A method for intelligent identification of surface cracks in radiator vibration testing, characterized in that, The method includes: Obtain the vibration excitation parameters output in real time by the vibration table; Based on the vibration excitation parameters, the optimal image acquisition phase window is determined, including: calculating the vibration period based on the real-time frequency in the vibration excitation parameters; identifying the phase interval of the vibration displacement extreme value based on the vibration period; obtaining the potential optimal acquisition phase point set based on the phase interval of the vibration displacement extreme value; and performing response timing optimization and synchronous dynamic compensation on the potential optimal acquisition phase point set to determine the optimal image acquisition phase window. Dynamically trigger shooting is performed within the optimal image acquisition phase window to generate an optimized image sequence; The optimized image sequence is subjected to vibration sensing preprocessing to generate enhanced images; Visual features and physical response features are extracted from the enhanced image and fused to obtain a fused feature set, including: extracting visual features from the enhanced image to obtain a crack geometric evolution field; extracting physical response features from the enhanced image to obtain a damage response tensor; and performing multimodal recognition fusion on the crack geometric evolution field and the damage response tensor to obtain a fused feature set. The fused feature set is subjected to collaborative analysis to output crack identification results and corresponding confidence scores.

2. The intelligent identification method for surface cracks in a radiator vibration test according to claim 1, characterized in that, The potential optimal set of acquisition phase points obtained based on the phase interval of the extreme value of the vibration displacement includes: Based on the phase interval of the extreme values ​​of the vibration displacement, a crack propagation sensitive phase set is obtained; Multimodal co-registration is performed on the crack propagation sensitive phase set to obtain the potential optimal acquisition phase point set.

3. The intelligent surface crack identification method in radiator vibration test according to claim 1, characterized in that, The step of performing vibration-sensing preprocessing on the optimized image sequence to generate enhanced images includes: Based on the optimized image sequence, the vibration coupling feature tensor is obtained; The vibration coupling feature tensor is subjected to dynamic convolution filtering to generate an enhanced image.

4. The intelligent identification method for surface cracks in a radiator vibration test according to claim 1, characterized in that, The step of extracting physical response features from the enhanced image to obtain the damage response tensor includes: Based on the enhanced image, the phase-analyzed pixel displacement field is calculated; The local abnormal strain region of the phase-analyzed pixel displacement field is extracted to obtain the crack core strain field; The amplitude change rate and directional consistency of the strain field at the crack core are quantified to generate a micro-strain dynamic vector; The damage characteristic is encoded by multi-physics coupling of the micro-strain dynamic vector to obtain the damage response tensor.

5. The intelligent identification method for surface cracks in a radiator vibration test according to claim 1, characterized in that, The collaborative analysis of the fused feature set, outputting crack identification results and corresponding confidence scores, includes: The spatiotemporal features of the fused feature set are physically constrained to decouple them, resulting in spatial structure components and time-varying dynamic components. A dual-domain collaborative causal analysis is performed on the spatial structural components and the time-varying dynamic components to output crack identification results and corresponding confidence scores.

6. The intelligent identification method for surface cracks in a radiator vibration test according to claim 5, characterized in that, The process of physically constraining the spatiotemporal feature decoupling of the fused feature set to obtain spatial structure components and time-varying dynamic components includes: Based on the fused feature set, the material constitutive awareness feature tensor is obtained; Vibrational phase synchronization field separation is performed on the constitutive sensing feature tensor of the material to obtain a phase-aligned feature field; The phase-aligned feature field is decoupled in a fracture mechanics-guided spatiotemporal manner to obtain the spatial structural component and the time-varying dynamic component.

7. The intelligent identification method for surface cracks in a radiator vibration test according to claim 1, characterized in that, The step of dynamically triggering image capture within the optimal image acquisition phase window to generate an optimized image sequence includes: Dynamic gating triggering and real-time phase error compensation are applied to the optimal image acquisition phase window to obtain a phase-synchronized image frame set; The phase-synchronized image frame set is subjected to cross-cycle feature alignment and spatiotemporal fusion reconstruction to generate an optimized image sequence.

8. A surface crack intelligent identification system for radiator vibration testing, applied to a surface crack intelligent identification method for radiator vibration testing as described in any one of claims 1-7, characterized in that, The system includes: The vibration parameter acquisition module is used to acquire the vibration excitation parameters output in real time by the vibration table; The phase window decision module is used to determine the optimal image acquisition phase window based on the vibration excitation parameters, including: calculating the vibration period based on the real-time frequency in the vibration excitation parameters; identifying the phase interval of the vibration displacement extreme value based on the vibration period; obtaining the potential optimal acquisition phase point set based on the phase interval of the vibration displacement extreme value; and performing response timing optimization and synchronous dynamic compensation on the potential optimal acquisition phase point set to determine the optimal image acquisition phase window. The dynamic shooting control module is used to dynamically trigger shooting within the optimal image acquisition phase window to generate an optimized image sequence; A vibration sensing preprocessing module is used to perform vibration sensing preprocessing on the optimized image sequence to generate an enhanced image; A multimodal recognition and fusion module is used to extract visual features and physical response features from the enhanced image and fuse them to obtain a fused feature set, including: extracting visual features from the enhanced image to obtain a crack geometric evolution field; extracting physical response features from the enhanced image to obtain a damage response tensor; and performing multimodal recognition and fusion on the crack geometric evolution field and the damage response tensor to obtain a fused feature set. The crack identification decision module is used to perform collaborative analysis on the fused feature set and output crack identification results and corresponding confidence scores.

Images