A multi-source supervised physical constraint visual monitoring method and system for rubber seismic isolation support

Abstract

The present application belongs to the technical field of performance detection of rubber seismic isolation support, and particularly relates to a multi-source supervised physical constraint visual monitoring method and system for rubber seismic isolation support, comprising the following steps: multi-source data acquisition; time alignment, space registration and supervised information construction; image quality discrimination, correction and standardization; multi-source multi-task teacher network training; physical constraint lightweight student network distillation; model compression, quantization and terminal adaptation; mobile or edge visual monitoring. The multi-source sensing information of the present application plays a role in the training stage, and only cameras and mobile or edge terminals can be retained in the deployment stage, reducing the need for long-term configuration of DIC, load, displacement or piezoelectric acquisition systems for each support. The load, displacement, DIC and piezoelectric information of the present application are no longer separated from image analysis, but participate in training through soft labels, spatial weights, teacher features and physical constraints, so that the model focuses on the areas related to the real deformation and damage of the support.

Description

Technical Field

[0001] This invention belongs to the field of performance testing technology for rubber seismic isolation bearings, specifically relating to a multi-source supervised physical constraint visual monitoring method and system for rubber seismic isolation bearings. Background Technology

[0002] Rubber seismic isolation bearings are widely used in building and bridge seismic isolation projects. However, under long-term vertical pressure and cyclic horizontal shear loads, they are prone to damage such as surface cracks, rubber bulging, interlaminar debonding, and abnormal buckling. Bearing damage directly reduces the seismic isolation performance of the structure and poses significant safety hazards. Therefore, routine damage and deformation quantitative monitoring of rubber seismic isolation bearings is a core requirement in the field of structural health monitoring. Currently, the industry's rubber seismic isolation bearing condition monitoring solutions mainly fall into four categories: image visual recognition monitoring, deformation measurement with dedicated equipment, built-in / external sensor monitoring, and intelligent detection based on neural networks. Existing related technical solutions have significant limitations. Based on patent database search results, the following are representative existing technical solutions: 1. Image-based rubber bearing defect identification technology Patent CN122176366A discloses a method for identifying bridge rubber bearing defects based on target detection and multi-label classification. This method relies on the original images of the bearing to complete the target location and defect classification, and uses image labels to achieve multi-defect identification, which can reduce interference from complex backgrounds to a certain extent. However, this method only uses manually labeled images as the sole supervisory data source, and does not incorporate mechanical data such as load, displacement, digital image correlation (DIC) strain field, and piezoelectric response that can be obtained during the experimental stage into the model training; at the same time, it does not establish a network to constrain the prediction results of the bearing's geometry and mechanical laws, and relies only on the "location-classification" two-level image recognition logic, which cannot simultaneously complete deformation quantification and damage credibility assessment.

[0003] 2. Support Deformation Measurement Technology Based on Dedicated Hardware Device Patent CN116659448A discloses an intelligent monitoring system for rubber bearing deformation, which uses a ring base, electric sliding rail, and other specialized mechanical structures to fit the sidewall boundary of the bearing and calculate the shear angle to quantify deformation. Patent CN111364623B uses a combination of a laser emitter, a dial, and a reflector to collect the three-dimensional displacement and rotation angle of the bearing. While these solutions can accurately obtain the geometric deformation parameters of the bearing, they suffer from high hardware dependence, requiring specialized motion, ranging, and laser sensing equipment, resulting in high on-site deployment costs. Furthermore, they can only quantify shear deformation and cannot simultaneously identify multiple types of surface damage such as cracks, debonding, and bulging, and are not suitable for lightweight detection scenarios using low-cost mobile terminals such as smartphones and tablets.

[0004] 3. Support mechanical monitoring technology based on built-in / external sensors Patent CN105442437B embeds fiber optic grating sensors inside the support, combined with pressure sensors to monitor shear deformation and vertical load. Patent CN121435103A integrates multiple time-series sensor data, including strain, displacement, vibration, and temperature, to achieve support status early warning. While these solutions can directly acquire the actual mechanical response of the support, they have two major drawbacks: First, fiber optic, piezoelectric, and displacement sensors need to be pre-installed during the support production or installation phase, making retrofitting existing supports difficult. Second, the sensor data is only used for status early warning and is not transferred to the visual inspection model as privileged monitoring information. This disconnects the mechanical monitoring and image inspection systems, preventing the use of high-precision mechanical data to optimize visual recognition accuracy.

[0005] 4. Knowledge distillation and physical information neural network detection technology Patent CN118096768A employs a teacher-student knowledge distillation structure to detect surface defects in steel plates. It improves defect recognition capabilities with few samples through model distillation. However, the detection object is steel structure plates, and the supervision information is only image defect labels. It does not introduce mechanical constraints for the large cyclic shear and large deformation conditions of rubber bearings. Patent CN121350732A constructs a physical information neural network using the overall mechanical equilibrium equation of the bridge as a physical constraint for stiffness damage identification of bridge components. However, this scheme is geared towards the macroscopic response of the entire bridge and does not construct cross-task consistency constraints for key points, contours, damage masks, and deformation regression of rubber bearing images. It also does not achieve a lightweight transfer reasoning mode of "multi-source sensor input during the training phase and only image input during the deployment phase".

[0006] Based on the publicly available technical solutions, the current intelligent monitoring technology for rubber seismic isolation bearings faces the following implementation challenges that are difficult to resolve: 1. Visual models rely on limited evidence for supervision, while human-labeled systems are prone to noise. Existing visual inspection solutions rely solely on manually labeled images to train the network. However, manual labeling makes it difficult to accurately distinguish the severity of damage and internal mechanical state, resulting in large labeling errors. Furthermore, objective mechanical data such as load and DIC strain field cannot be used to assist in correcting the labels, thus limiting the model's recognition accuracy.

[0007] 2. The mechanical sensing monitoring and visual inspection systems are disconnected. High-precision mechanical data such as load, displacement, DIC, and piezoelectricity are only used for laboratory condition analysis and are not converted into supervisory information that can be learned by visual networks. The multi-source high-precision data from the laboratory cannot empower low-cost visual inspection terminals on site.

[0008] 3. Unable to adapt to operating conditions with missing data from multiple sources and inconsistent data quality. Data loss issues often occur in engineering sites and during testing, such as DIC speckle shedding, piezoelectric signal coupling failure, and short-term loss of displacement channels. Existing solutions directly fill missing data with zero values, which can easily lead to misjudgment of the support's mechanical state. There is no unified mechanism to distinguish between the three scenarios of "data not collected", "data of low quality", and "true zero mechanical response".

[0009] 4. General-purpose visual networks lack specific physical constraints for rubber supports. Existing detection networks do not incorporate the inherent mechanical properties of rubber bearings, such as shear strain, profile bulging, and load-displacement relationships. Under conditions of bearing obstruction, reflection, shooting angle deviation, and scarce samples, the model is prone to outputting erroneous identification results that contradict the physical properties of the bearings.

[0010] 5. High-precision multi-source models cannot be deployed in the field at low cost. The teacher network that integrates multi-sensor data and the DIC full-field strain calculation system have high computing power requirements and expensive equipment, making it impossible to configure them for every support in the long term; simply compressing and quantizing the network will lose key features of fine cracks and edge deformation, and cannot balance recognition accuracy with terminal computing power limitations.

[0011] 6. The monitoring output results lack verification criteria, and their reliability is uncontrollable. The existing solution only outputs the result of a single disease category, and cannot simultaneously output the image acquisition quality, damage spatial location, deformation quantification value, and model prediction uncertainty. On-site staff cannot judge whether the identification result is valid, and it is difficult to determine whether re-shooting or manual review is required.

[0012] 7. The solution has poor adaptability and a high implementation threshold. Most existing monitoring solutions are tied to specific cameras, dedicated testing equipment, and fixed detection thresholds. When the equipment purchased on-site or the testing conditions change, the entire monitoring solution cannot be directly reused, resulting in a narrow range of applications.

[0013] In summary, existing technologies cannot simultaneously meet the four major engineering requirements of utilizing multi-source high-precision data in the laboratory, deploying lightweight single images on-site, ensuring the physical rationality of recognition results, and verifying detection results. There is an urgent need to propose a visual damage monitoring solution that is suitable for the working conditions of rubber seismic isolation bearings and balances training accuracy with on-site deployment costs. Summary of the Invention

[0014] To address the technical problems existing in current industry-standard rubber seismic isolation bearing condition monitoring solutions, this invention provides a multi-source supervised physical constraint visual monitoring method and system for rubber seismic isolation bearings.

[0015] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method for multi-source monitoring of physical constraints of rubber seismic isolation bearings using visual monitoring, comprising the following steps: S1. Multi-source data acquisition: Images of the rubber seismic isolation bearings are acquired through the image acquisition module, and at least one type of mechanical data, including load P(t), horizontal displacement X(t), and vertical displacement Z(t), is acquired through the load displacement module. Optionally, the DIC displacement-strain field can be acquired through the DIC auxiliary module, and the piezoelectric excitation response can be acquired through the piezoelectric auxiliary module. Support geometric metadata, camera intrinsic and extrinsic parameters, target dimensions, and other metadata are recorded. The source mask and reliability module generates a source availability mask m for each type of data. s (t), reliability weight ω s (t), calculate the effective weights r of the samples. s (t)=m s (t)ω s (t); S2. Construction of temporal alignment, spatial registration, and supervisory information: Multi-source data synchronization and alignment are completed using hardware triggering or software timestamp matching. The DIC strain field is projected onto the region of interest in the image to generate a spatial distillation weight map W(t) as spatial soft supervision. Combined with manual annotation and geometric calculation, key points of the upper and lower boundaries of the support, support contour, visible damage mask, damage multi-label categories, continuous severity, and deformation labels are generated. When conclusions from different sources conflict, the original data and conflict markers are retained, and a simple majority vote is not used to generate a unique truth value. S3. Image quality discrimination, correction and standardization: Calculate the image sharpness, exposure, glare, occlusion ratio, viewpoint deviation and other quality vectors q(t), and perform scale and perspective correction on the image; Images whose overall quality does not reach the strong supervision threshold are not used for high confidence label training, but only for training the quality discrimination branch and hard sample branch; S4. Multi-source multi-task teacher network training: The support image, mechanical auxiliary data, and sample effective weights are input into the multi-source teacher network. Multi-source feature fusion is completed through gating weighting or cross attention. The training output includes key points, contour damage mask, damage category, continuous severity, deformation, and uncertainty multi-task representation. The training set, validation set, and test set are divided according to the specimen number. Adjacent frames of the same specimen are not split into different datasets to avoid information leakage. S5. Lightweight Student Network Distillation under Physical Constraints: A lightweight student network is constructed that only receives support images and support geometric metadata as input; soft label distillation and multi-scale feature distillation are performed through the knowledge distillation module to transfer knowledge from the multi-source teacher network; and at least one rubber support physical consistency loss is introduced in conjunction with the physical consistency constraint module to construct the total loss Ltotal of the student network to complete joint training; the physical consistency loss is selected from at least one of geometric consistency loss, strain consistency loss, temporal consistency loss, load-displacement mechanical consistency loss, and cross-task consistency loss. S6, Model Compression, Quantization and Terminal Adaptation: The compression and quantization module performs channel pruning and INT8 / FP16 numerical quantization on the lightweight student network, verifies the classification, segmentation and regression errors before and after quantization, and encapsulates and generates a deployment model package containing model weights, calibration parameters, quality thresholds and risk thresholds. S7. Mobile or edge visual monitoring: Mobile or edge terminals load and deploy model packages, collect single-frame images or short videos of the support to complete inference; the output status is determined based on the overall image quality and the difference in consistency between multi-task outputs: if the image quality is insufficient, the correction fails, or the task prediction conflicts, a reshoot or manual review is triggered; otherwise, the output includes traceable monitoring evidence containing the damaged area, deformation quantization value, image quality score, and model uncertainty.

[0016] In step S1, the DIC displacement-strain field only generates a spatial weight map W(t) to increase the distillation weight of the high-strain region, without directly equating the high-strain region of DIC with the true value of the crack; the piezoelectric response generates a state soft tag through time-domain framing and time-frequency transformation, and the soft tag weight is determined by the source mask and the reliability weight ω output by the reliability module. s (t) Control; when DIC and piezoelectric data are missing, the corresponding source availability mask m s (t)=0, this type of data does not participate in network fusion and loss calculation.

[0017] The total loss of a lightweight student network satisfies the following calculation formula: L total =L task +λ KD L KD +λ phys L phys +λ q L quality The distillation loss LKD fusion spatial weight map W(t) applies element-wise alignment constraints to the multi-scale features of teachers and students; the physical consistency loss is a soft constraint, and the loss weight is dynamically and adaptively adjusted with the effective weight rs(t) of the sample to avoid physical assumptions suppressing the visible real damage features in the image.

[0018] When collecting short videos in step S7, robust temporal aggregation is performed to output the stable deformation mean, the maximum risk frame, and the damage superposition evidence annotation map. The mobile or edge terminal stores the standardized inspection record locally or uploads it to the backend system. The inspection record includes the original image, evidence map, support number, collection time, model version, image quality score, and review status mark.

[0019] The rubber seismic isolation bearings include natural rubber bearings, lead-core rubber bearings, and high-damping laminated rubber bearings; the mobile or edge terminal can be any one of a mobile phone, industrial handheld tablet, edge computing box, or industrial camera with built-in computing unit; during the training phase, the equipment and the field terminal are physically separated, and field deployment does not require continuous connection to load, DIC, or piezoelectric acquisition equipment.

[0020] A multi-source supervised physical constraint visual monitoring system for rubber seismic isolation bearings includes: an image acquisition module, a load displacement module, a DIC auxiliary module, a piezoelectric auxiliary module, a training and constraint module, a model encapsulation module, and a movement or edge monitoring module; The image acquisition module is used to acquire single-frame images, short videos, and camera calibration parameters of the support; the load displacement module is used to acquire time-series data of load, horizontal displacement, and vertical displacement; the DIC auxiliary module is used to output the surface displacement field, strain field, and data-related quality indicators of the support; the piezoelectric auxiliary module is used to output the piezoelectric excitation response and time-frequency domain state characterization. The training and constraint module integrates the source mask and reliability module, multi-source teacher network, knowledge distillation module, physical consistency constraint module, and lightweight student network to complete the entire process of multi-source data temporal and spatial registration, multi-source soft supervision construction, multi-source teacher network training, and physical constraint distillation training. The model encapsulation module integrates compression and quantization modules, which perform pruning and quantization transformation on the trained lightweight student network and encapsulate it to generate a standardized deployment model package. The mobile or edge monitoring module is deployed on a mobile or edge terminal in the field and only calls the image acquisition module to acquire images to complete offline lightweight inference. The training and constraint module and the model packaging module are deployed on the workstation / server, while the mobile or edge monitoring module is deployed on the field terminal. The two only complete data interaction through offline deployment of the model package. During the field monitoring stage, there is no need to connect to the load displacement module, DIC auxiliary module, and piezoelectric auxiliary module in real time.

[0021] The image acquisition module uses a mobile phone, a global shutter industrial camera, a high-speed camera, or a dual-target fixed camera; the load displacement module includes a testing machine data interface, a displacement meter, and a laser rangefinder; the DIC auxiliary module and the piezoelectric auxiliary module are optional supporting units. Without these two types of auxiliary modules, the system training process can be fully run by relying solely on the image and load / displacement data.

[0022] The physical consistency constraint module within the training and constraint module incorporates five independent loss calculation branches: geometric, strain, temporal, mechanical, and cross-task. The loss weight of each branch is determined by the effective sample weight r output by the source mask and reliability module. s (t) Real-time dynamic adjustment enables logical self-verification of multi-task prediction results.

[0023] The mobile or edge monitoring module has a built-in image quality discrimination unit, a scale perspective correction unit, a multi-task consistency verification unit, a local storage unit, and a cloud data upload interface. When the image clarity is insufficient, the support correction fails, or the multi-task prediction output conflicts, the module automatically outputs a prompt for reshoot or manual review, without forcing the output of a deterministic risk judgment conclusion.

[0024] The system is also equipped with loading test auxiliary equipment, rubber seismic isolation bearing specimens, lighting and shading components, and multi-source acquisition and synchronization units. The loading test auxiliary equipment applies graded vertical pressure and cyclic horizontal shear displacement to the rubber seismic isolation bearing specimens. The lighting and shading components ensure the imaging quality of the speckle target area of ​​the bearing. The multi-source acquisition and synchronization unit synchronizes the image, load, DIC, and piezoelectric acquisition timing.

[0025] Compared with the prior art, the beneficial effects of this invention are: 1. Reduced dependence on field equipment: The multi-source sensing information of this invention plays a role in the training phase, and only cameras and mobile or edge terminals are needed in the deployment phase, reducing the need for long-term configuration of DIC, load, displacement or piezoelectric acquisition systems for each support.

[0026] 2. Improve the mechanical relevance of labels and features: The load, displacement, DIC and piezoelectric information of this invention are no longer separated from image analysis, but participate in training through soft labels, spatial weights, teacher features and physical constraints, so that the model focuses on the region related to the actual deformation and damage of the support.

[0027] 3. Improve robustness under missing source conditions: The source mask and reliability weight of this invention distinguish between "not collected", "low quality" and "true zero response", avoiding erroneous supervision caused by simple splicing, and allowing subsequent devices to be added or removed gradually.

[0028] 4. Reduce physically unreasonable output: The geometric, strain, temporal, mechanical and cross-task constraints of this invention enable mutual verification of damage type, spatial location and deformation amount, reducing the risk of misjudgment caused by background texture, glare or local occlusion alone.

[0029] 5. Balancing accuracy and deployment efficiency: The teacher network of this invention utilizes multi-source information to obtain a strong representation, and then transfers the knowledge related to the deployment task to a lightweight student network; compression and quantization are used to verify the errors of multiple tasks separately, making it suitable for mobile or edge computing resources.

[0030] 6. Output verifiable evidence: The system of this invention simultaneously outputs the damage location, deformation amount, confidence level, image quality, and task consistency, and explicitly triggers re-shooting or manual verification when evidence is insufficient, rather than forcibly outputting high-risk conclusions. Attached Figure Description

[0031] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0032] The structures, proportions, sizes, etc. illustrated in this specification are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed herein, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0033] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a diagram of the teacher-student and physical constraint architecture of the present invention; Figure 3 This is a system composition block diagram of the present invention; Figure 4 This is a flowchart of the mobile or edge-end visual monitoring process of the present invention; Figure 5 This is a schematic diagram of the proposed test and data acquisition setup for the present invention.

[0034] Among them: 201 is the image acquisition module, 202 is the load displacement module, 203 is the DIC auxiliary module, 204 is the piezoelectric auxiliary module, 205 is the training and constraint module, 206 is the model packaging module, 207 is the motion or edge detection module, 301 is the multi-source teacher network, 302 is the source mask and reliability module, 303 is the knowledge distillation module, 304 is the physical consistency constraint module, 305 is the lightweight student network, 306 is the compression and quantization module, 307 is the deployment model package, 501 is the loading test auxiliary equipment, 502 is the rubber seismic isolation bearing specimen, 503 is the camera and lens, 504 is the lighting and shading components, 506 is the multi-source acquisition and synchronization unit, and 507 is the mobile or edge terminal. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. These descriptions are only for further illustrating the features and advantages of the present invention, and not for limiting the claims of the present invention. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0036] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0037] This embodiment provides a multi-source supervised physical constraint visual monitoring method for rubber seismic isolation bearings, such as... Figure 1 As shown, it includes the following steps: S1. Obtain a dataset with support images as the main data and at least one non-visual mechanical data as auxiliary supervision, and establish a correspondence between sample time or loading state. S2. Construct source availability labels and / or reliability weights for different sources; S3. Train a multi-source teacher network using images and auxiliary supervision to output damage and deformation representations; S4. Construct and deploy a lightweight student network with less input than the teacher network, where the student network receives at least the support image; S5. Train the student network jointly using teacher-student distillation loss and at least one rubber support physical consistency loss. S6. Compress or convert the student network into a mobile or edge-executable model; S7. During deployment, damage monitoring results are obtained using images, and the output, re-encoder, or verification status is determined based on image quality or multi-task consistency.

[0038] DIC, piezoelectric sensors, high-speed cameras, binocular cameras, specific backbone networks, specific quantization formats, and specific experimental equipment models are all optional technical features. This invention can also be implemented using only images with load / displacement, without requiring the availability of DIC or piezoelectric devices as a prerequisite.

[0039] Furthermore, S1: Multi-source data acquisition S11. Record metadata such as support type, diameter or side length, total rubber layer thickness, number of rubber layers, initial shape, camera internal and external parameters, lens focal length, shooting distance, marker size, and specimen number.

[0040] S12. Position mobile phones, industrial cameras, or high-speed cameras at fixed or repeatable locations, and select single-frame, continuous video, timed sampling, or loading-state triggered sampling according to the loading speed. For static and low-speed loading, prioritize the use of global shutter industrial cameras or cameras with short exposure control; high-speed cameras are used for transient crack propagation or rapid instability and are not considered essential equipment.

[0041] S13. Obtain at least one of the load P(t), horizontal displacement X(t), and vertical displacement Z(t). This can be obtained through visual calculations using a testing machine interface, displacement gauge, laser rangefinder, scale, or target.

[0042] S14. Optionally, the DIC displacement field and strain field E(x,y,t) can be obtained, and the speckle region or measurement region can be mapped to the original image coordinates through calibration relationships.

[0043] S15. Optionally, the piezoelectric excitation and response a(t) can be obtained, and the piezoelectric characterization corresponding to the loading state can be obtained through time-domain framing, frequency-domain transformation, time-frequency transformation or learning-based coding.

[0044] S16. Generate a source availability mask m and a reliability weight ω for each source record based on availability, synchronization error, signal-to-noise ratio, occlusion rate, correlation coefficient, or other quality metrics.

[0045] Furthermore, S2: Construction of temporal alignment, spatial registration, and supervisory information. S21. Prioritize hardware-triggered synchronization; if hardware synchronization is not possible, use unified timestamps, loading step start and end points, peak events, or correlation matching for software alignment.

[0046] S22. Perform interpolation or time window aggregation on low-frequency load and displacement signals, but set boundaries at loading direction switching, peak values, unloading start points, and pause states to avoid smoothing across boundaries.

[0047] S23. Project the DIC strain field onto the region of interest in the image to generate a continuous spatial weight map. The high-strain region increases the supervision weight of the damage candidate region, but is not directly equivalent to the crack ground truth.

[0048] S24. Generate state soft labels or teacher features based on the correlation coefficient, band energy, arrival time, amplitude attenuation, or learned characterization of the piezoelectric response, and control their weights by coupling quality.

[0049] S25. Generate key points of upper / lower connection boundaries, support contours, visible damage masks, damage multi-label categories, continuous severity, and deformation labels through manual annotation, geometric calculation, and multi-source assistance.

[0050] S26. When conclusions from different sources conflict, the original data, conflict markers, and their respective reliability should be preserved, and a simple majority vote should not be used to forcibly generate a unique truth value.

[0051] Furthermore, S3: Image quality assessment, correction, and standardization. S31. Calculate quality indicators such as sharpness, exposure, glare, motion blur, occlusion ratio, target integrity, viewpoint deviation, and scale effectiveness to form a quality vector q.

[0052] S32. Images whose quality does not meet the strong supervision threshold are not used for training high-confidence labels, but can be used to train the quality discrimination branch, hard sample branch, or retake decision branch.

[0053] S33. Geometric enhancement simultaneously transforms key points, masks, and DIC projection maps; optical enhancements such as brightness and contrast only affect the image and do not change load, displacement, or strain labels.

[0054] S34. After correcting the multi-view images separately, map them to a unified support coordinate system; under single-view conditions, use the target, known support dimensions, or geometric features of the upper and lower connecting plates to complete scale and perspective correction.

[0055] Furthermore, such as Figure 2 As shown, S4: Multi-source, multi-task teacher network training S41. The visual encoder extracts features at at least two scales to account for overall shearing morphology, edge bulging, and fine cracks.

[0056] S42. Load / displacement encoders handle current values, increments, loading directions, cycle numbers, and peak positions; DIC encoders handle displacement gradients and strain concentrations; piezoelectric encoders handle time-domain, frequency-domain, or time-frequency characteristics.

[0057] S43. Multi-source fusion is performed using splicing, gated weighting, cross-attention, or feature modulation, and the source availability mask m and reliability ω are explicitly input into the fusion module. Missing sources are not masqueraded as all-zero measurements.

[0058] S44. The multi-task output of the multi-source teacher network 301 includes at least two of the following: key points, contours or damage masks, damage categories, continuous severity, horizontal or vertical deformation, and uncertainty.

[0059] S45. Divide the training set, validation set, and test set according to the specimen number to avoid information leakage caused by adjacent frames of the same specimen being assigned to different sets.

[0060] Furthermore, S5: Lightweight Student Network Distillation under Physical Constraints S51. Construct a lightweight student network that takes only visual images and optional support geometry metadata as input. The student network consists of a lightweight backbone, a lightweight multi-scale feature fusion layer, and a multi-task prediction head.

[0061] S52. Perform soft label distillation on the class probabilities of the multi-source teacher network 301 and the lightweight student network 305; align multi-scale features through a mapping layer and perform feature distillation; further sample relation, spatial relation, or task relation distillation can be performed.

[0062] S53. Adjust the distillation space weights based on the DIC strain weight map, crack candidate region, support edge, and key point neighborhood to prioritize the retention of support-specific features within the limited student capacity.

[0063] S54. Add at least one of geometric consistency, DIC strain consistency, temporal consistency, load-displacement or residual deformation consistency, and cross-task consistency to the 305 training loss of the lightweight student network.

[0064] S55. Independent masks are used for loading direction switching, unloading, occlusion, missing source and low quality samples; the physical constraints output by the physical consistency constraint module 304 are soft constraints, and their weights are controlled by data reliability and training phase to avoid erroneous physical assumptions suppressing real visible damage.

[0065] Support geometric characterization When T r When available, the nominal shear strain γ(t) and nominal vertical compressive strain ε of the support are... v (t) can be expressed as: γ(t) = X(t) / Tᵣ (1) εᵥ(t) = Z(t) / Tᵣ (2) The bulging ratio b(t) and damage area ratio ρd(t) obtained from the contour prediction can be expressed as: b(t) = [D max [(t) - D0] / D0(3) ρd(t) = Ad(t) / max[Aroi(t), ε] (4) Where ε is a positive number to prevent the denominator from being zero. When the support area is incomplete or A_roi(t) is lower than the effective threshold, equation (4) is not calculated, and the process is switched to re-shooting or manual review.

[0066] Joint training objectives The total network loss for students is: Ltotal = Ltask + λKD LKD + λphys Lphys + λq Lquality (5) The multi-tasking loss can be expressed as: Ltask = λseg Lseg + λkp Lkp + λcls Lcls + λreg Lreg + λunc Lunc (6) Distillation loss can be expressed as: LKD = αLlogit + βΣl ‖W(t) ⊙ [φl(FTl) - FSl]‖1 + χLrelation (7) In equation (7), φ_l represents the feature size and channel mapping, W(t) is formed by the DIC weight, the damage candidate region, the support edge and the source reliability, and is resampled to the corresponding spatial size before participating in the feature distillation of the l-th layer; ⊙ represents element-wise multiplication.

[0067] The physical consistency loss can be expressed as: Lphys = λg Lgeo + λe Lstrain + λt Ltemporal + λm Lmechanics + λcLcross (8) Among them, Lgeo constrains the consistency of keypoint displacement, contour change, and deformation regression values; Lstrain maintains a soft correlation between visible damage probability and high strain regions; Ltemporal constrains adjacent frame increments and short-term stability; Lmechanics constrains the output to be compatible with loading stages, load-displacement relationships, or residual displacements; Lcross constrains the logical relationships between classification, segmentation, keypoint, and regression tasks. t is the time corresponding to the image frame or loading state; P(t) is the vertical or synthetic load; X(t) is the horizontal relative displacement of the support; Tr is the total rubber layer thickness of the support; D0 is the feature diameter / width under no-load or baseline conditions; D_max(t) is the current maximum feature diameter / width of the contour; Ad(t) is the current predicted damage area; A_roi(t) is the current effective area of ​​the support; m s (t) is the availability mask for source s; ω s (t) represents the reliability weight of source s; q(t) represents the image quality vector or overall quality score; pT, pS represent the class probability vectors of teachers and students; FTl, FSl represent the teacher and student features at the l-th layer; l represents the feature layer index; W(t) represents the spatial distillation or strain soft weight map; r s(t) represents the effective weight of source s; φl(・) represents the feature size and channel mapping of the l-th layer; ε is a positive number to prevent the area ratio from having a zero denominator; Ltotal represents the total loss of the student network; Ltask represents the multi-task supervision loss; LKD represents the teacher-student distillation loss; Lphys represents the physical consistency loss; Lquality represents the quality discrimination or quality weighting loss; Lseg, Lkp, Lcls, Lreg, and Lunc represent the segmentation, keypoint, classification, regression, and uncertainty losses; Llogit and Lrelation represent the soft label and relation distillation losses; Lgeo, Lstrain, Ltemporal, Lmechanics, and Lcross represent the geometric, strain, temporal, mechanical, and cross-task consistency losses; α, β, and χ represent the weights of the distillation components; and λ* represents the weights of each loss term.

[0068] Source reliability control The effective weight for source s is defined as r s (t)=m s (t)ω s (t). m when the source is missing. s (t)=0, the source does not participate in fusion and the corresponding loss; when the source exists but its quality is low, 0<ω s When (t) < 1, its effect continuously weakens. When all auxiliary sources are unavailable, the teacher or student degenerates into visually supervised branch training, but retains geometric and cross-task constraints.

[0069] like Figure 3 As shown, the system includes: Image acquisition module 201 is used to acquire single-frame images or videos and camera parameters; The load displacement module 202 is used to acquire at least one of load, horizontal displacement and vertical displacement; DIC auxiliary module 203 is used to output the displacement field, strain field and data quality of the support surface; The piezoelectric auxiliary module 204 is used to output excitation response and state characterization; The training and constraint module 205 integrates the source masking and reliability module 302, the multi-source teacher network 301, the knowledge distillation module 303, the physical consistency constraint module 304, and the lightweight student network 305, which are used to align multi-source data, train the multi-source teacher network 301, construct reliability and train the lightweight student network 305 under physical constraints. The model encapsulation module 206 integrates the compression and quantization module 306, which is used for pruning, quantization, operator transformation and encapsulation of the model, calibration parameters, category dictionary and threshold, and generates the deployment model package 307. The mobile or edge monitoring module 207, deployed on the mobile or edge terminal 507, is used for quality judgment, correction, inference, consistency verification, result display and recording.

[0070] The training and constraint module 205 can be located on a workstation or server, while the mobile or edge monitoring module 207 is located on a field terminal; the two are connected via a model package 307 rather than real-time multi-source data. Therefore, the training equipment and the field terminal can be physically separated.

[0071] Furthermore, S6: Model compression, quantization, and terminal adaptation. S61. Select pruning targets based on channel importance, feature redundancy, and actual measured time consumption at the target terminal, retaining key points, edges, and layers related to minor damage.

[0072] S62. Use post-training quantization or quantization-aware training to convert weights and / or activations to INT8, FP16, or other formats supported by the terminal; verify the classification probability, keypoint error, mask consistency, and deformable regression error before and after quantization.

[0073] S63. Convert the student network into a terminal executable format, which may be a general exchange format, a mobile inference framework format, or a chip-specific format. The format name is not a necessary limitation.

[0074] S64. Encapsulate the model weights, camera calibration parameters, applicable support types, geometric parameters, category dictionary, quality threshold, risk threshold, and version verification information into a model package.

[0075] Furthermore, such as Figure 4 As shown, S7: Visual monitoring of mobile or edge devices. S71. Load deployment model package 307 and check model version, applicable support type and file integrity.

[0076] S72. Acquire a single frame image or short video, calculate the quality vector q; if the image quality is below the threshold, stop risk assessment and prompt for reshoot.

[0077] S73. Position the main body of the support, the upper and lower connecting boundaries and the target, and perform dimensional, posture and perspective corrections; if reliable correction cannot be achieved, output "requires verification".

[0078] S74. Perform lightweight student network 305 inference to obtain key points, contours or damage masks, class probabilities, continuity severity, deformation, and uncertainty.

[0079] S75. Check the consistency between key point displacement, contour bulging, deformation regression, and damage segmentation; reduce confidence or trigger reshoot / manual review when output conflicts exceed the threshold.

[0080] S76. Perform robust temporal aggregation on short videos, output stable deformation, maximum risk frame, representative evidence map and abnormal frame list to form a traceable inspection record.

[0081] The parameters in the table below are used to illustrate the scope of implementation. The formal values ​​are determined by testing and end-user capabilities and are not intended to limit the claims.

[0082]

[0083] Example 1: Low-cost basic closed loop This embodiment is used to verify the basic closed loop of S1 to S7 before all equipment has been purchased, and is not aimed at obtaining the final engineering accuracy.

[0084] (1) Select small rubber blocks, small rubber supports, waste supports or controllable deformation models, and set scale targets on the side surface; use mobile phones or ordinary industrial cameras to collect images or short videos at fixed positions.

[0085] (2) Obtain horizontal displacement reference values ​​through small loading devices, displacement stages, rulers or visual targets; if there is no load channel for the time being, use the image plus displacement as the minimum multi-source training data.

[0086] (3) Set at least two of the following states: normal, shear deformation, local bulging, surface crack, or artificially controlled defect. Artificial defects should be marked separately from real failure samples to avoid mistaking simulated defects for real fatigue or debonding.

[0087] (4) Establish the correspondence between image frames and displacement states, and mark the key points of the upper and lower boundaries, support contours, visible damage areas and state categories.

[0088] (5) Train the image and add displacement to the multi-source teacher network 301, and then distill it into the visual lightweight student network 305; at least shear strain geometry, key point-contour-regression consistency and short-term stability constraints are adopted.

[0089] (6) Convert the lightweight student network 305 to an Android mobile terminal or edge computing terminal to verify whether the shooting, quality judgment, correction, reasoning, evidence overlay and record export are connected.

[0090] Example 2: Industrial Camera and DIC Enhancement Solution (1) An industrial camera that supports external triggering, short exposure, and fixed focal length lenses shall be selected; during quasi-static loading, acquisition shall be triggered based on displacement increment or loading step, and a high-speed camera may be used when cracks propagate rapidly. The specific camera model shall be registered after procurement, and the model shall not be limited to an essential feature of this invention.

[0091] (2) Set up stable lighting and shading structures to ensure that the side surface of the support, the upper and lower connecting boundaries, and the speckle area are within the effective depth of field. The DIC camera and the monitoring camera can be the same camera or different cameras that have been spatially calibrated.

[0092] (3) Simultaneously acquire images, loads, and displacements; the DIC auxiliary module 203 outputs the displacement field, principal strain or shear strain field, and correlation coefficients. Only regions with correlation coefficients meeting the threshold are used to form spatial soft weights.

[0093] (4) The multi-source teacher network 301 consists of a visual encoder, a load displacement encoder, and a DIC encoder. The fusion module reads the source mask and the mask and weights output by the reliability module 302, and outputs key points, contour / damage mask, damage type, and deformation amount.

[0094] (5) The lightweight student network 305 only receives monitoring images and support geometry metadata. During training, it is compared with pure vision student, distillation student, distillation plus physical constraint student, and quantization student to distinguish the contribution of each technical feature.

[0095] (6) It is recommended to randomly divide the dataset by specimen number rather than image frame; consecutive frames of the same specimen should not be included in the training set and test set at the same time.

[0096] Example 3: Full-scale multi-source integrated test of LNR600 This embodiment covers the full-scale route in existing test concepts. The specimen can be a natural rubber seismic isolation bearing LNR600, and the official parameters should be verified in conjunction with the testing machine's capabilities, safety assessment, and applicable standards.

[0097] (1) Vertical loading can be increased from 0 to 15 MPa in stages. It is recommended to use 2.5 MPa as a loading stage. After reaching the target vertical pressure, maintain stability and check the image, displacement and acquisition channel quality.

[0098] (2) Under the condition of maintaining a vertical pressure of 15 MPa, the horizontal shear can be increased from 0 to +250% of the nominal shear strain. It is recommended to use 50% as a state level, then return to 0 and load in the -250% direction; increase the sampling density during the direction switching, peak and unloading stages.

[0099] (3) It is proposed to use 3 specimens, with each specimen repeated 3 times as the initial design. The actual number will be adjusted according to the budget, specimen repeatability and statistical analysis. Training, verification and testing will be grouped by specimen.

[0100] (4) Arrange DIC speckle or target on the side surface of rubber isolation bearing specimen 502 to obtain the full field displacement and strain; at the same time, record the test machine load, horizontal displacement and vertical displacement. If the DIC speckle falls off or fails, the corresponding area is disabled by source masking.

[0101] (5) Optional piezoelectric auxiliary module 204 can be configured to build piezoelectric excitation and response channels. The sweep frequency range can be from 1kHz to 100kHz, and the sampling frequency is not less than 1MHz. The proposed data acquisition device model is NIUSB-6366, and the proposed power amplifier model is ChipMingTian C47. The above models are only optional implementation examples. Before purchasing, the number of channels, synchronization method, output range and sampling capability must be verified.

[0102] (6) The multi-source teacher network 301 is trained by combining images, load displacement, DIC and piezoelectric representations; the lightweight student network 305 is only input with image and geometric metadata during the deployment phase. Piezoelectric features are only used to construct state soft labels or teacher intermediate features, and the mobile terminal does not need to access the piezoelectric auxiliary module 204.

[0103] (7) The test record shall include at least the original data, synchronization information, source quality, annotation version, model version, training configuration, quantization configuration and terminal test record, so that the agent can select the implementation data that can be written into the specification from the real results.

[0104] The accompanying test setup for this embodiment is as follows: Figure 5 As shown, it includes a loading test auxiliary device 501, a rubber seismic isolation bearing specimen 502, a camera and lens 503, an illumination and shading assembly 504, a multi-source acquisition and synchronization unit 506, and a mobile or edge terminal 507. The loading test auxiliary device 501 applies graded vertical pressure and cyclic horizontal shear displacement to the rubber seismic isolation bearing specimen 502. The illumination and shading assembly 504 ensures the imaging quality of the speckle target area of ​​the bearing. The multi-source acquisition and synchronization unit 506 unifies and synchronizes the image, load, DIC, and piezoelectric acquisition timing.

[0105] Example 4: On-site inspection of mobile or edge devices (1) The mobile or edge terminal 507 first selects the support type or reads the support identifier, and calls the corresponding deployment model package 307 and built-in geometric parameters.

[0106] (2) On-site personnel take a single frame image or a short video of 2 to 5 seconds on the side of the support through the image acquisition module 201; the terminal provides real-time prompts on whether the target is complete, whether it is overexposed, whether it is blurry, and whether the shooting angle can be corrected.

[0107] (3) After the image quality is qualified, the Lightweight Student Network 305 inference outputs the candidate region of damage, category probability, severity, horizontal / vertical deformation estimate and uncertainty; when the short video is input, it outputs the temporal stability statistics and the maximum risk frame.

[0108] (4) When the segmentation output is obviously damaged but the classification is normal, the key point deformation and the difference between the regression value are too large, or the image quality is not up to standard, the motion or edge monitoring module 207 does not directly output a definitive risk conclusion, but pops up a prompt for reshoot or manual review.

[0109] (5) The inspection record includes the original image, damage superimposed evidence image, support number, acquisition time, model version, image quality score, quantification result and review status mark; the record can be stored locally on the mobile or edge terminal 507, or uploaded to the background management system.

[0110] The above description only illustrates the preferred embodiments of the present invention. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention, and all such changes should be included within the protection scope of the present invention.

Claims

1. A method for multi-source monitoring of physical constraints and visual monitoring of rubber seismic isolation bearings, characterized in that, Includes the following steps: S1. Multi-source data acquisition: Images of the rubber seismic isolation bearing are acquired through the image acquisition module (201), and at least one of the mechanical data of load P(t), horizontal displacement X(t), and vertical displacement Z(t) is acquired through the load displacement module (202). The DIC displacement strain field can be acquired through the DIC auxiliary module (203), and the piezoelectric excitation response can be acquired through the piezoelectric auxiliary module (204). Support geometric metadata, camera internal and external parameters, target size and other metadata are recorded. The source mask and reliability module (302) generates a source availability mask m for each type of data. s (t), reliability weight ω s (t), calculate the effective weights r of the samples. s (t)=m s (t)ω s (t); S2. Construction of temporal alignment, spatial registration, and supervisory information: Multi-source data synchronization and alignment are completed using hardware triggering or software timestamp matching. The DIC strain field is projected onto the region of interest in the image to generate a spatial distillation weight map W(t) as spatial soft supervision. Combined with manual annotation and geometric calculation, key points of the upper and lower boundaries of the support, support contour, visible damage mask, damage multi-label categories, continuous severity, and deformation labels are generated. When conclusions from different sources conflict, the original data and conflict markers are retained, and a simple majority vote is not used to generate a unique truth value. S3. Image quality judgment, correction and standardization: Calculate the image sharpness, exposure, glare, occlusion ratio, viewpoint deviation and other quality vectors q(t), and perform scale and perspective correction on the image; Images whose overall quality does not reach the strong supervision threshold are not used for high-confidence label training, but are only used to train the quality discrimination branch and the hard sample branch. S4. Multi-source multi-task teacher network training: The support image, mechanical auxiliary data, and sample effective weights are input into the multi-source teacher network (301). Multi-source feature fusion is completed through gating weighting or cross attention. The training outputs key points, contour damage mask, damage category, continuous severity, deformation, and uncertainty multi-task representation. The training set, validation set, and test set are divided according to the specimen number. Adjacent frames of the same specimen are not split into different datasets to avoid information leakage. S5. Lightweight student network distillation under physical constraints: Construct a lightweight student network (305) that only receives support images and support geometric metadata as input; perform soft label distillation and multi-scale feature distillation through the knowledge distillation module (303) to transfer knowledge from the multi-source teacher network (301), and combine it with the physical consistency constraint module (304) to introduce at least one rubber support physical consistency loss to construct the student network total loss Ltotal to complete joint training; the physical consistency loss is selected from at least one of geometric consistency loss, strain consistency loss, temporal consistency loss, load-displacement mechanical consistency loss, and cross-task consistency loss; S6. Model compression, quantization and terminal adaptation: The compression and quantization module (306) performs channel pruning and INT8 / FP16 numerical quantization on the lightweight student network (305), verifies the classification, segmentation and regression errors before and after quantization, and encapsulates and generates a deployment model package (307) containing model weights, calibration parameters, quality thresholds and risk thresholds. S7. Mobile or edge visual monitoring: The mobile or edge terminal (507) loads and deploys the model package (307), collects single-frame images or short videos of the support to complete inference; the output status is determined based on the overall image quality and the difference in consistency between multi-task outputs: if the image quality is insufficient, the correction fails, or the task prediction conflicts, a reshoot or manual review is triggered; otherwise, the output includes traceable monitoring evidence containing the damaged area, deformation quantization value, image quality score, and model uncertainty.

2. The method for multi-source monitoring of physical constraints and visual monitoring of rubber seismic isolation bearings according to claim 1, characterized in that: In step S1, the DIC displacement strain field only generates a spatial weight map W(t) to increase the distillation weight of the high strain region, and does not directly equate the high strain region of DIC with the true value of the crack. The piezoelectric response is time-domain framed and time-frequency transformed to generate a state soft tag. The soft tag weight is the reliability weight ω output by the source mask and the reliability module (302). s (t) Control; when DIC and piezoelectric data are missing, the corresponding source availability mask m s (t)=0, this type of data does not participate in network fusion and loss calculation.

3. The method for multi-source monitoring of physical constraints and visual monitoring of rubber seismic isolation bearings according to claim 1, characterized in that, The total loss of the lightweight student network (305) satisfies the calculation formula: L total =L task +λ KD L KD +λ phys L phys +λ q L quality The distillation loss LKD fusion spatial weight map W(t) applies element-wise alignment constraints to the multi-scale features of teachers and students; the physical consistency loss is a soft constraint, and the loss weight is dynamically and adaptively adjusted with the effective weight rs(t) of the sample to avoid physical assumptions suppressing the visible real damage features in the image.

4. The method for multi-source monitoring of physical constraints and visual monitoring of rubber seismic isolation bearings according to claim 1, characterized in that, When collecting short videos in step S7, robust temporal aggregation is performed to output the stable deformation mean, the maximum risk frame, and the damage superposition evidence annotation map; the mobile or edge terminal (507) stores the standardized inspection record locally or uploads it to the background system. The inspection record includes the original image, evidence map, support number, collection time, model version, image quality score, and review status mark.

5. The method for multi-source monitoring of physical constraints and visual monitoring of rubber seismic isolation bearings according to claim 1, characterized in that, The rubber vibration isolation bearings include natural rubber bearings, lead-core rubber bearings, and high-damping laminated rubber bearings; the mobile or edge terminal (507) is any one of a mobile phone, industrial handheld tablet, edge computing box, or industrial camera built-in computing unit; during the training phase, the equipment and the field terminal are physically separated, and the field deployment does not require continuous access to load, DIC, and piezoelectric acquisition equipment.

6. A multi-source monitoring system for physical constraints of rubber seismic isolation bearings, characterized in that, include: Image acquisition module (201), load displacement module (202), DIC auxiliary module (203), piezoelectric auxiliary module (204), training and constraint module (205), model encapsulation module (206), and motion or edge detection module (207). The image acquisition module (201) is used to acquire single-frame images, short videos, and camera calibration parameters of the support; the load displacement module (202) is used to acquire time-series data of load, horizontal displacement, and vertical displacement; the DIC auxiliary module (203) is used to output the surface displacement field, strain field, and data-related quality indicators of the support; the piezoelectric auxiliary module (204) is used to output the piezoelectric excitation response and time-frequency domain state characterization. The training and constraint module (205) integrates the source mask and reliability module (302), the multi-source teacher network (301), the knowledge distillation module (303), the physical consistency constraint module (304), and the lightweight student network (305) to complete the entire process of multi-source data temporal and spatial registration, multi-source soft supervision construction, multi-source teacher network training, and physical constraint distillation training. The model encapsulation module (206) integrates the compression and quantization module (306), performs pruning and quantization transformation on the trained lightweight student network (305), and encapsulates it to generate a standardized deployment model package (307). The mobile or edge monitoring module (207) is deployed on the field mobile or edge terminal (507), and only calls the image acquisition module (201) to acquire images to complete offline lightweight inference; The training and constraint module (205) and the model packaging module (206) are deployed on the workstation / server, and the mobile or edge monitoring module (207) is deployed on the field terminal. The two only complete data interaction through offline deployment of the model package (307). During the field monitoring stage, there is no need to connect to the load displacement module (202), the DIC auxiliary module (203), and the piezoelectric auxiliary module (204) in real time.

7. The multi-source monitoring physical constraint visual monitoring system for rubber seismic isolation bearings according to claim 6, characterized in that: The image acquisition module (201) uses a mobile phone, a global shutter industrial camera, a high-speed camera, or a dual-target positioning camera; the load displacement module (202) includes a testing machine data interface, a displacement meter, and a laser rangefinder; the DIC auxiliary module (203) and the piezoelectric auxiliary module (204) are optional supporting units. Without the two types of auxiliary modules, the system training process can be fully run by relying solely on the image and load / displacement data.

8. The multi-source monitoring physical constraint visual monitoring system for rubber seismic isolation bearings according to claim 6, characterized in that: The physical consistency constraint module (304) inside the training and constraint module (205) has five independent loss calculation branches: geometry, strain, time series, mechanics, and cross-task. The loss weight of each branch is the effective weight r of the sample output by the source mask and reliability module (302). s (t) Real-time dynamic adjustment enables logical self-verification of multi-task prediction results.

9. A multi-source monitoring physical constraint visual monitoring system for rubber seismic isolation bearings according to claim 6, characterized in that: The motion or edge monitoring module (207) has a built-in image quality discrimination unit, scale perspective correction unit, multi-task consistency verification unit, local storage unit, and cloud data upload interface. When the image clarity is insufficient, the support correction fails, or the multi-task prediction output conflicts, the module automatically outputs a prompt for reshoot or manual review, without forcing the output of a deterministic risk judgment conclusion.

10. A multi-source monitoring system for physical constraints of rubber seismic isolation bearings according to claim 6, characterized in that: The system is also equipped with loading test auxiliary equipment (501), rubber seismic isolation bearing specimen (502), lighting and shading components (504), and multi-source acquisition and synchronization unit (506); the loading test auxiliary equipment (501) applies graded vertical pressure and cyclic horizontal shear displacement to the rubber seismic isolation bearing specimen (502), the lighting and shading components (504) ensure the imaging quality of the speckle target area of ​​the bearing, and the multi-source acquisition and synchronization unit (506) synchronizes the image, load, DIC, and piezoelectric acquisition sequence.

Images