A photovoltaic support fault diagnosis system based on an industrial internet platform

Abstract

The application discloses a photovoltaic support fault diagnosis system based on an industrial internet platform, relates to the technical field of industrial internet, and comprises a data acquisition module, which acquires real-time wind speed and vibration data of the photovoltaic support through sensors accessed by the industrial internet platform and obtains initial dynamic response characteristics of the structure in a key frequency range through modal analysis; a modeling and identification module, which, in the industrial internet platform, adopts a finite element method to construct a global model of the support according to the dynamic response characteristics and identifies key positions of damage-prone areas; the photovoltaic support fault diagnosis system based on the industrial internet platform improves the real-time performance, accuracy and engineering practicability of safety evaluation of the photovoltaic support structure and enhances the safety and stability of overall operation of the photovoltaic power station.

Description

Technical Field

[0001] This invention relates to the field of industrial internet technology, and specifically to a photovoltaic bracket fault diagnosis system based on an industrial internet platform. Background Technology

[0002] As a critical structure in photovoltaic (PV) power plants used to support PV modules, the structural safety of PV support systems directly impacts the long-term stable operation of the power plant. In actual operation, PV support systems are constantly exposed to a complex and changing natural environment, continuously bearing various external loads such as wind loads, snow loads, and temperature variations. Especially in large-span, large-scale PV power plants, damage to the support structure can easily trigger a chain reaction of risks, adversely affecting power generation efficiency and operational safety. Therefore, effective fault diagnosis and health assessment of PV support structures are crucial technical requirements for the operation and maintenance management of PV power plants.

[0003] Existing methods for diagnosing photovoltaic (PV) support faults primarily rely on periodic manual inspections, offline testing, or structural analysis based on single operating conditions. These methods typically depend on experience-based judgment or post-event analysis, making it difficult to reflect the true stress and dynamic response of the support under actual operating loads. This is especially true for issues like fatigue damage caused by wind loads and hidden cracks in connections, which are often difficult to detect early, exhibiting a significant lag. With the development of sensor and network communication technologies, some PV power plants have begun to introduce online monitoring methods, analyzing the support's condition by collecting operational data such as wind speed and vibration. However, existing online monitoring solutions often remain at the level of data acquisition or simple threshold alarms, lacking a systematic analysis of the overall structural dynamic characteristics, making it difficult to accurately assess the true structural state of the support under complex load conditions. Furthermore, the large amount of operational data collected is often not effectively integrated with the structural model, leading to significant uncertainty between monitoring results and the actual damage state. On the other hand, while structural analysis based on the finite element method can accurately describe the stress and deformation distribution of PV supports, engineering applications often require the creation of detailed models containing numerous local details, resulting in large computational scales and low computational efficiency. When the support structure is complex or requires frequent updates to load conditions, this type of analysis method struggles to meet the needs of online diagnosis and real-time assessment. Especially under sudden conditions such as strong winds, if structural response calculations cannot be completed within a short timeframe, it will be difficult to provide timely and effective support for operation and maintenance decisions. Furthermore, different parts of the photovoltaic support structure exhibit significantly different degrees of influence on the overall dynamic response and fault evolution; some areas are highly sensitive to typical damage modes, while the impact on other areas is relatively limited. Existing technologies generally lack a diagnostic mechanism that can effectively reduce computational complexity while ensuring analytical accuracy and focusing on the structural state of key components, making it difficult to balance computational efficiency with diagnostic accuracy. Summary of the Invention

[0004] The purpose of this invention is to provide a photovoltaic bracket fault diagnosis system based on an industrial internet platform, which solves the problems existing in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a photovoltaic support fault diagnosis system based on an industrial internet platform, comprising a data acquisition module, which acquires real-time wind speed and vibration data of the photovoltaic support through sensors connected to the industrial internet platform, and obtains the initial dynamic response characteristics of the structure within the key frequency range through modal analysis; a modeling and identification module, which constructs a global model of the support based on the dynamic response characteristics using the finite element method in the industrial internet platform, and identifies key parts of the damage-prone area; and a threshold determination module, which acquires detailed geometric parameters of the local complex structure of the identified key parts, and determines the threshold of these parts using a neural network algorithm in the industrial internet platform. The model simplification module extracts the vibration mode contribution of the secondary structure from the global model if the determined stress distribution threshold exceeds the preset threshold, and uses equivalent simplification techniques to reduce the computational complexity of the secondary structure; the consistency judgment module updates the dynamic response of the global model with the reduced computational complexity and judges whether the updated response is consistent with the initial dynamic response characteristics; the response calculation module obtains real-time load data and inputs it into the updated model if the judgment is consistent, to obtain the current stress and deformation distribution of the support; the damage identification module recalculates the vibration modes in the key frequency range using modal analysis based on the obtained stress and deformation distribution, and determines the potential damage location.

[0006] Preferably, the data acquisition module acquires real-time wind speed and vibration data of the photovoltaic support through sensors connected to the industrial internet platform. Modal analysis is used to obtain the initial dynamic response characteristics of the structure within the key frequency range. This includes acquiring a synchronous time-domain signal sequence containing wind load input and structural vibration output, generated by synchronizing wind speed time history values ​​and vibration acceleration values; generating a cross-power spectral density matrix based on the synchronous time-domain signal sequence, which is obtained by using a fast Fourier transform to retain complex data within a preset analysis bandwidth; performing singular value decomposition on the cross-power spectral density matrix to generate a structural modal data set, including natural frequencies, mode shape vectors, and damping ratio values; and mapping the structural modal data set to a preset frequency response function model to obtain the initial dynamic response characteristics of the photovoltaic support within the key frequency range.

[0007] Preferably, the modeling and identification module, within the industrial internet platform, constructs a global model of the support using the finite element method based on dynamic response characteristics, and identifies key parts of the damage-prone area. This includes acquiring real-time vibration signals of the support transmitted from the industrial internet platform, processing the real-time vibration signals using short-time Fourier transform to obtain time-frequency characteristics; mapping the support's geometric topology based on the time-frequency characteristics to generate finite element mesh elements; assembling the global stiffness matrix using the finite element mesh elements and solving for the nodal displacements; calculating the strain energy density based on the nodal displacements; generating a damage factor if the strain energy density exceeds a preset threshold; delineating the damage-prone area based on the spatial distribution gradient of the damage factor; and locating extreme points within the damage-prone area to identify key parts of the global model of the support.

[0008] Preferably, the threshold determination module, for the identified key parts, obtains detailed geometric parameters of their local complex structures and determines the stress distribution threshold of these parts in the industrial internet platform using a neural network algorithm. This includes obtaining detailed geometric parameters of the key parts in the industrial internet platform, including the curvature features and spatial coordinates of the reconstructed local structure surface; mapping the detailed geometric parameters into high-dimensional feature vectors and inputting them into a pre-trained convolutional neural network model; simulating the mechanical response features of the local structure through the convolutional neural network model; analyzing the mechanical response features to generate a predicted stress distribution map; and determining the stress distribution threshold of the key parts based on the peak region and gradient change of the predicted stress distribution map.

[0009] Preferably, the model simplification module, if the determined stress distribution threshold exceeds a preset threshold, extracts the vibration mode contribution of the secondary structure from the global model and uses equivalent simplification techniques to reduce the computational complexity of these structures. This includes acquiring nodal stress distribution data; if the maximum nodal stress value is greater than a preset limit, identifying the secondary structural entity; extracting the local stiffness matrix and local mass matrix of the secondary structural entity to calculate the vibration mode vector, and constructing a mode transformation matrix accordingly; using the mode transformation matrix to perform orthogonal transformation on the local stiffness matrix and local mass matrix to generate an equivalent stiffness matrix and an equivalent mass matrix; and assembling the equivalent stiffness matrix and equivalent mass matrix back into the global finite element model to reduce the computational complexity of the secondary structure.

[0010] Preferably, the consistency judgment module updates the dynamic response of the global model by reducing computational complexity and determines whether the updated response is consistent with the initial dynamic response features. This includes acquiring global model data to extract initial dynamic response features and constructing a sparse matrix representing the model structure to reduce computational complexity; performing low-dimensional space iteration based on the sparse matrix and incremental business data to generate an updated global model; inputting the test stimulus signal into the updated global model to output the updated dynamic response; calculating the response deviation value between the updated dynamic response and the initial dynamic response features; and determining that the updated response is consistent with the initial dynamic response features if the response deviation value is less than the feature drift threshold.

[0011] Preferably, the response calculation module, if the judgment is consistent, acquires real-time load data to update the model and obtains the current stress and deformation distribution of the support. This includes: if the monitored support operating state matches the preset consistency state, acquiring real-time load data stream; mapping the real-time load data stream to the geometric topology nodes of the finite element model to update the boundary conditions of the finite element model; assembling a global stiffness matrix based on the boundary conditions and support material properties, and solving the global stiffness matrix to obtain the node displacement vector; calculating strain field data based on the node displacement vector to obtain the current stress and deformation distribution of the support.

[0012] Preferably, the damage identification module, based on the obtained stress and deformation distribution, recalculates the vibration modes within the key frequency range using modal analysis and determines the potential damage location. This includes acquiring full-field stress distribution data and deformation distribution data, updating the geometric configuration using the deformation distribution data, and constructing a tangent stiffness matrix containing stress stiffening effects based on the full-field stress distribution data. It also extracts eigenvectors within the key frequency range from the tangent stiffness matrix as vibration mode shapes; calculates the element modal strain energy based on the vibration mode shapes and the tangent stiffness matrix; and if a local abrupt peak occurs in the element modal strain energy, it calculates damage location indicators based on the local abrupt peak to determine the potential damage location.

[0013] Preferably, it also includes a continuous monitoring module, which generates an online diagnostic report based on the determined potential damage location through an industrial internet platform, and cyclically inputs sensor data for the next cycle to maintain real-time monitoring. Specifically, it includes acquiring the potential damage location and local mutation peak value, establishing the spatial mapping coordinates of the potential damage location in a three-dimensional digital model and calculating the quantitative value of the damage degree; filling the spatial mapping coordinates and the quantitative value of the damage degree into the diagnostic report template to generate a structured online diagnostic report.

[0014] The continuous monitoring module generates an online diagnostic report based on the determined potential damage location through the industrial internet platform, and cyclically inputs sensor data for the next cycle to maintain real-time monitoring. It also parses the structured online diagnostic report. If the current structural health status label is non-failure, it constructs a standardized input vector sequence for the next time cycle. The standardized input vector sequence is loaded into the finite element analysis engine to update the boundary conditions to maintain real-time monitoring of the structural status.

[0015] As can be seen from the above technical solution, the present invention has the following beneficial effects: This photovoltaic (PV) support fault diagnosis system, based on an industrial internet platform, combines multi-source sensor data from the field with the computing and management capabilities of the industrial internet platform to achieve online perception and continuous monitoring of the PV support's operating status. Through modal analysis-based dynamic response feature extraction and global finite element modeling, the diagnostic process reflects the overall dynamic characteristics of the support under actual complex loads, avoiding the limitations of relying solely on experience or single-condition analysis. By identifying damage-prone areas and introducing stress threshold judgments and model equivalence simplification for key components, the system effectively reduces computational scale and time while maintaining diagnostic accuracy, improving real-time response capabilities under sudden conditions such as strong winds. Simultaneously, the platform architecture enables closed-loop processing of model updates, load input, and damage identification, continuously outputting reliable online diagnostic results. This provides timely and accurate decision-making support for PV power plant operation and maintenance personnel, thereby improving the real-time performance, accuracy, and engineering practicality of PV support structural safety assessments, and enhancing the overall safety and stability of the PV power plant. Attached Figure Description

[0016] Figure 1 This is a connection diagram of the photovoltaic bracket fault diagnosis system module of the present invention. Detailed Implementation

[0017] 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, and 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.

[0018] like Figure 1As shown, this invention provides a technical solution: a photovoltaic support fault diagnosis system based on an industrial internet platform, including a data acquisition module that acquires real-time wind speed and vibration data of the photovoltaic support through sensors connected to the industrial internet platform, and obtains the initial dynamic response characteristics of the structure within the key frequency range through modal analysis; a modeling and identification module that constructs a global model of the support using the finite element method based on the dynamic response characteristics in the industrial internet platform, and identifies key parts of the damage-prone area; a threshold determination module that acquires detailed geometric parameters of the local complex structure of the identified key parts, and determines the stress distribution threshold of these parts using a neural network algorithm in the industrial internet platform; and a model simplification module that, if the determined stress distribution threshold exceeds a preset value, performs a simplification process. The threshold is used to extract the vibration mode contribution of the secondary structure from the global model and to reduce the computational complexity of the secondary structure using an equivalent simplification technique. The consistency judgment module updates the dynamic response of the global model with the reduced computational complexity and judges whether the updated response is consistent with the initial dynamic response characteristics. The response calculation module, if consistent, obtains real-time load data and inputs it into the updated model to obtain the current stress and deformation distribution of the support. The damage identification module, based on the obtained stress and deformation distribution, uses modal analysis to recalculate the vibration modes in the key frequency range and determine the potential damage location. The continuous monitoring module, through the industrial internet platform, generates an online diagnostic report based on the determined potential damage location and cyclically inputs sensor data for the next cycle to maintain real-time monitoring.

[0019] This system uses an industrial internet platform as its data and computing core, and achieves online sensing of the photovoltaic support system's operational status through multi-source sensors. Real-time wind speed and vibration data reflect the coupling relationship between external loads and structural response. Modal analysis can extract dynamic characteristic parameters such as natural frequencies, mode shapes, and damping ratios of the photovoltaic support system within key frequency ranges, serving as an initial reference for the structural health status. Based on this, the modeling and identification module maps the support system's geometry, material parameters, and boundary conditions into a global computational model using the finite element method. It then identifies key areas prone to damage under long-term wind loads, vibration loads, or environmental influences by combining dynamic response characteristics. For these key areas, due to their complex structural morphology and significant stress concentration, the threshold determination module further acquires local geometric parameters and uses a neural network algorithm to establish a nonlinear mapping relationship between geometric features and stress distribution, thereby determining a stress distribution threshold that better reflects actual working conditions. When the local stress distribution threshold exceeds a preset safety threshold, the model simplification module analyzes the contribution of each structural unit to the overall vibration mode, effectively simplifying secondary structures with less impact on global dynamic behavior, reducing the computational scale while ensuring that the main dynamic characteristics are not distorted. The consistency assessment module verifies the rationality of the model simplification by comparing the degree of matching between the simplified model and the initial dynamic response characteristics within the key frequency range. Under the condition that consistency is achieved, the response calculation module inputs real-time load data into the updated model to obtain the stress and deformation distribution of the photovoltaic support under the current operating conditions. Subsequently, the damage identification module recalculates the vibration mode changes based on the new stress and deformation results, identifying potential damage locations through modal parameter shifts, thus achieving a refined diagnosis of the photovoltaic support structure's condition. The continuous monitoring module relies on the industrial internet platform to form a data closed loop, enabling the system to continuously track the health status of the support structure over a long period.

[0020] The data acquisition module acquires real-time wind speed and vibration data of the photovoltaic support through sensors connected to the industrial internet platform. Modal analysis is used to obtain the initial dynamic response characteristics of the structure within the key frequency range. This includes acquiring a synchronous time-domain signal sequence containing wind load input and structural vibration output, generated by synchronizing wind speed time history values ​​and vibration acceleration values; generating a cross-power spectral density matrix based on the synchronous time-domain signal sequence, which is obtained by using a fast Fourier transform to retain complex data within a preset analysis bandwidth; performing singular value decomposition on the cross-power spectral density matrix to generate a structural modal data set, including natural frequencies, mode shape vectors, and damping ratio values; and mapping the structural modal data set to a preset frequency response function model to obtain the initial dynamic response characteristics of the photovoltaic support within the key frequency range.

[0021] In this embodiment, the data acquisition module uses an industrial internet platform as the aggregation and scheduling carrier, organizing the sampling data from wind speed sensors and vibration acceleration sensors according to a unified time reference to form a synchronous time-domain signal sequence that simultaneously includes wind load input and structural vibration output. Specifically, both wind speed and vibration acceleration data are uniformly distributed sampling configurations from the platform and written into a timestamped data stream. The timestamps can be generated by the platform's time synchronization service or uniformly marked by the edge gateway after performing time synchronization on the data from each sensor. When there is a deviation in the sampling starting points of different data sources, alignment processing is used to resample the two data streams on the same time axis and fill in the missing points, ensuring that the wind speed value at any given time corresponds one-to-one with the vibration acceleration value at the same time. The alignment process typically includes: sorting and deduplicating the two timestamps; generating a standard time series based on the target sampling interval; performing interpolation and interval normalization on the original wind speed time history and vibration acceleration time history respectively; removing isolated outliers with abrupt changes or replacing them with the neighborhood median; performing detrending processing on sequences with DC drift; and, while ensuring that the physical meaning of the signal is not changed, smoothing filtering can be applied to high-frequency noise to improve the stability of subsequent frequency domain estimation. The key parameters involved in the above processing are jointly determined by engineering constraints and frequency domain analysis requirements: the sampling frequency is set according to the principle of at least twice the "upper limit of the key frequency range", and further considers anti-aliasing margin and sensor bandwidth, usually taking it as 2.5 times or higher than the upper limit of the key frequency range; the resampling interval is taken to be consistent with the sampling frequency; the allowable time alignment error threshold is determined according to the minimum period corresponding to the upper limit of the key frequency range, so that the alignment error does not exceed one-tenth of the minimum period, so as to avoid significant distortion of the phase relationship; the anomaly point judgment threshold can be determined according to the statistical distribution within the sliding window, for example, using the high quantile of the amplitude of vibration acceleration change within the window as the upper limit, so as to maintain adaptability under different wind conditions.

[0022] After obtaining the synchronous time-domain signal sequence, the system converts it into a cross-power spectral density matrix to characterize the frequency-domain coupling relationship between the wind load input and the outputs of multiple vibration measurement points. This calculation process is executed on the industrial internet platform side or edge computing node side, and is typically implemented using piecewise estimation and average noise reduction: First, the synchronous time-domain signal sequence is divided into multiple short-time data segments according to a preset segment length. Before calculation, a window function is applied to each segment to reduce spectral leakage, and overlap is set between adjacent data segments to improve the variance performance of the estimation. Then, a fast Fourier transform is performed on each segment to obtain complex frequency-domain data, and the cross-spectrum between the input and each output, as well as the cross-spectrum between each output, are calculated to form the cross-power spectral density matrix at the corresponding frequency points. Finally, the cross-spectrum results of all data segments are weighted and averaged to obtain a stable cross-power spectral density matrix, and only the complex data within the preset analysis bandwidth is retained to meet the analysis needs of the "critical frequency range". The parameters in the above steps are determined as follows: the preset analysis bandwidth is directly given by the key frequency range, usually based on the modal distribution of the support design, historical operating spectrum, and preliminary frequency sweep results. The lower limit is used to exclude low-frequency drift and slow-varying disturbances, while the upper limit is used to cover potential resonance regions. The segment length is determined by the desired frequency resolution; higher resolution results in a longer segment length, but this reduces time-varying adaptability. Therefore, the segment length is usually chosen so that the resolution is no greater than one percent of the key frequency range bandwidth, while also meeting online computation latency requirements. The overlap ratio is used to balance computational load and estimation stability, typically ranging from 0.5 to 0.75. Larger overlap results in better average performance but also higher computational load. Window function type... To control leakage and amplitude deviation, a window function that balances main lobe width and side lobe suppression is typically selected. Window function energy compensation is introduced during amplitude calibration to ensure spectral amplitude comparability. The average number of segments is determined by the stability of the on-site wind conditions and the allowable diagnostic update cycle. The more unstable the wind conditions, the fewer the average segments can be to maintain real-time performance; the more stable the wind conditions, the more the average segments can be to be to improve noise immunity. In addition, to avoid unreliable cross-spectral estimation when input energy is insufficient, an input validity threshold can be set. This threshold can be determined by the mean square level of the wind speed time history within the current cycle. When the mean square level is lower than the lower quantile of the historical normal interval, the weight of the results for that cycle is reduced or the output is delayed.

[0023] After generating the cross-power spectral density matrix, the system performs singular value decomposition (SVD) to obtain structural modal data sets, thereby extracting natural frequencies, mode shape vectors, and damping ratios. The specific implementation process is as follows: within a preset analysis bandwidth, the cross-power spectral density matrix is ​​read point-by-point according to frequency. Singular value decomposition is performed at each frequency point to obtain a sequence of singular values ​​sorted by energy contribution and the corresponding singular vectors. Then, the variation of singular values ​​with frequency is used as a characterization of modal significance. When a singular value exhibits a significant peak near a certain frequency and the peak shape meets the continuity requirement, the frequency corresponding to that peak is identified as a candidate natural frequency. Next, the singular vector corresponding to the main peak at that frequency point is taken as a candidate mode shape vector, and the mode shape vector is normalized to eliminate dimensional influences. Normalization can be achieved by using maximum component normalization or energy normalization to make the vibration modes comparable under different periods and operating conditions. The damping ratio is estimated near the candidate natural frequency based on the peak width or attenuation characteristics. In online implementation, the peak width estimation method is usually used: find the position where the amplitude drops to a certain proportion of the peak value on both sides of the candidate natural frequency, and convert the frequency difference on both sides into the damping ratio. When the peak width is unstable due to field noise, the time-domain attenuation estimation method can be used instead. That is, bandpass filtering is performed on the vibration response within the modal frequency band, the free attenuation segment is extracted and the attenuation rate of adjacent peaks is estimated to obtain the damping ratio. The thresholds and criteria involved in the above identification process need to be clearly defined: the threshold for candidate peak determination can be adaptively given by the background noise level of the singular value curve, for example, by using the median of the singular values ​​within the bandwidth plus a certain multiple of the dispersion as the threshold; those exceeding the threshold and having local maxima features are included in the candidate set; the continuity requirement is used to exclude isolated spikes, which can be achieved by requiring that the singular values ​​at several frequency points near the peak are all higher than the background level; the mode shape validity criterion is used to exclude false modes caused by local noise or loose installation, which can be achieved through the "mode shape consistency index", that is, comparing the similarity between the current period's candidate mode shape and the historical reference mode shape; if the similarity is lower than the historical reference mode shape, the candidate mode shape is considered to be valid. If a preset threshold is used, the mode will not be output. The similarity threshold can be determined by the similarity distribution of historical health data, and its lower quantile is taken as a conservative limit. The peak width ratio threshold of the damping estimation is used to define the intercept positions on both sides. It can be determined by the typical sharpness of the structural mode peak. Usually, a compromise is made between ensuring insensitivity to noise and sensitivity to peak shape changes. It is self-corrected on the platform side through historical data to ensure that the damping estimation stability of different site supports is consistent. When the system adopts the time-domain decay estimation method, the intercept threshold of the free decay segment can be triggered by the event of wind speed decrease or vibration energy sudden drop to ensure that the selected segment is as close as possible to the state without external excitation.

[0024] After extracting the natural frequencies, mode shape vectors, and damping ratios, the system maps the structural modal data set to a preset frequency response function model to obtain the initial dynamic response characteristics of the photovoltaic support within the key frequency range. The mapping is implemented as follows: The platform side pre-sets the structural form and parameter organization of the frequency response function model. This model is based on the concept of modal superposition, treating each identified mode as a contribution item and combining it with the corresponding mode shape vector and damping ratio to synthesize the responses at different frequency points. During mapping, the mode shape vectors are first scaled to match the sensor measurement point layout, and directional correction coefficients are introduced for the measurement point direction and installation direction. These directional correction coefficients can be determined based on the sensor installation calibration results. Subsequently, based on the candidate natural frequencies and damping ratios, the response contribution of each mode is calculated point-by-point within the key frequency range, and the overall frequency response is synthesized by measurement points. The synthesized results are then converted into a set of "initial dynamic response characteristics" defined by the platform. This set may include amplitude curves, phase change trends, peak frequencies, peak amplitudes, peak width characteristics, and relative response distributions between measurement points within the key frequency range. To ensure that the initial dynamic response characteristics can serve as a benchmark for subsequent consistency judgments and model updates, the system implements consistency and reliability constraints before output: when the input validity is insufficient, the modal peak value does not meet the continuity requirement, the mode shape consistency index is below the threshold, or the damping estimation fluctuation exceeds the threshold, the platform adopts a strategy of reducing the weight or delaying the update of the characteristics output for this period, and can mark the data for this period as requiring review to avoid the benchmark being contaminated. The determination principle of the above thresholds is based on the "statistical distribution of historical data during the healthy period": first, the distribution range of each indicator is established within the confirmed undamaged operating range, and then the threshold is taken as a boundary value that can cover most healthy samples and is sensitive to anomalies.

[0025] The modeling and identification module, within the industrial internet platform, constructs a global model of the support structure using the finite element method based on dynamic response characteristics, and identifies key components of the damage-prone areas. This includes acquiring real-time vibration signals of the support structure transmitted from the industrial internet platform, processing these signals using short-time Fourier transform to obtain time-frequency characteristics, mapping the support structure's geometric topology based on these time-frequency characteristics to generate finite element mesh elements, assembling the global stiffness matrix using these mesh elements, and solving for the nodal displacements. The module calculates the strain energy density based on the nodal displacements, and if the strain energy density exceeds a preset threshold, a damage factor is generated. The module delineates damage-prone areas based on the spatial distribution gradient of the damage factor, and locates extreme points within these areas to identify key components of the global model of the support structure.

[0026] In this modeling and identification module, the industrial internet platform first receives real-time vibration signals from the support structure. These signals are collected and uploaded by vibration sensors located at the support components or nodes. The platform performs unified time reference marking, missing point completion, and outlier suppression on the signals before they enter the time-frequency processing flow. The determination of sampling-related parameters follows the principle of "coverage of key frequency range and accessibility of online calculation": the sampling frequency is set based on the upper limit of the key frequency range, ensuring that the sampling frequency is not less than twice the upper limit of the key frequency range, while allowing margin based on the effective bandwidth of the sensors and the platform's computing resources; the data refresh cycle is determined by the platform's diagnostic output frequency, ensuring that at least one stable time-frequency feature update is completed within each refresh cycle; the window length for signal de-drift is determined by the time scale of the slow-changing trend of the support structure, ensuring that the slow-changing trend does not enter the effective frequency band of subsequent time-frequency features; the outlier judgment threshold is determined by the statistical distribution of vibration amplitude within the refresh cycle, preferably taking the high quantile boundary of the amplitude within the cycle as the upper limit, and identifying points exceeding the upper limit and with a duration shorter than the preset minimum duration as isolated outliers, thereby avoiding misleading time-frequency analysis due to occasional electromagnetic interference or communication jitter.

[0027] Subsequently, a short-time Fourier transform (SFT) is performed on the real-time vibration signal to obtain time-frequency features. This step is implemented by dividing the real-time vibration signal into several short-time segments, applying a window function to each segment to suppress spectral leakage, performing a frequency domain transform on the segment to obtain the energy distribution of the segment at each frequency, and finally stitching the results of each segment in chronological order to form a time-frequency distribution map. Time-frequency features for modeling and identification are extracted within a preset analysis bandwidth. The parameters involved in the short-time processing are determined as follows: the window length is used to balance time resolution and frequency resolution. The lower limit of the window length is based on the principle of covering at least a certain multiple of the corresponding period of the lower limit of the key frequency range, so as to make the main frequency features stably displayed. The upper limit of the window length is based on the principle of not exceeding a reasonable proportion of the refresh cycle, so as to ensure the real-time performance of online output. The window shift step size is used to control the continuity of adjacent time-frequency slices. The smaller the step size, the better the continuity, but the computational load increases. The platform can set the upper limit of the step size based on edge computing power and network latency, and appropriately reduce the step size when the wind conditions fluctuate drastically to improve the transient capture capability. The overlap ratio is determined by both the window length and the step size, preferably by making the window length and step size equal to the overlap ratio. Sufficient overlap between adjacent segments is ensured to smooth the dominant frequency trajectory without causing computational congestion. The window function type is determined by the sidelobe suppression requirements; when site noise is high or weak modes need to be highlighted, a window function with stronger sidelobe suppression is preferred. A uniformization process is introduced during the amplitude output stage to ensure comparability of energy levels between different periods. The preset analysis bandwidth is directly determined by the key frequency range; its lower limit is used to exclude low-frequency drift and slow-varying disturbances, while its upper limit covers sensitive frequency bands where the support structure may experience resonance or localized impacts. If the site has specific wind-induced frequency bands, the bandwidth boundaries are fine-tuned for specific conditions without excessively increasing computational load. The determination of the time-frequency characteristics aims to ensure they can be used for constrained structural dynamics description. This preferably includes the energy distribution of the key frequency band over time, the continuity of the dominant frequency trajectory, the frequency and duration of transient impacts, and the temporal location of energy surges within the key frequency band, thus providing a basis for subsequent geometric topology mapping and mesh refinement.

[0028] After obtaining the time-frequency characteristics, finite element mesh elements are generated by mapping the support geometry to these characteristics. This step is implemented as follows: First, the platform side establishes the support geometry data, including component types, component connection relationships, node spatial locations, constraint locations, and load transfer paths. The geometry can be formed by importing design data, as-built data, or on-site digital modeling data. Then, based on the time-frequency characteristics, the geometry is dynamically consistent with the data. For example, when energy in a certain frequency band is abnormally concentrated and exhibits continuous temporal characteristics, the corresponding sensor-neighboring components are marked as areas requiring improved representation accuracy. This marking result then guides mesh generation, performing local refinement on marked areas and applying conventional meshing to unmarked areas, thereby improving the spatial resolution of potentially abnormal areas while maintaining a controllable computational scale. The mesh generation parameters are determined as follows: The mesh element size is determined by both the component scale and the target spatial resolution. Ideally, the element size at high-risk locations such as connectors, hole neighborhoods, weld neighborhoods, and support intersections should be smaller than the component's characteristic size by a reasonable proportion to ensure that energy concentrations can be distinguished. For regions far from risk points and with stable contributions to overall dynamics, the element size is appropriately enlarged to reduce computational load. The element type is determined by the component's stress mode. Tension-compression dominated rod-type components, bending dominated beam-type components, and plate-type components are represented using elements matching their deformation modes. Furthermore, in critical connection areas… Element types with higher expressive power are used to avoid oversimplification of local stiffness; material parameters are determined by material grade, batch data, or platform material library, and correction coefficients can be introduced based on the effects of temperature, corrosion, or aging. The initial value of the correction coefficient can be given by the design margin, and subsequent calibration can be gradually performed based on the dynamic response characteristics monitored over a long period of time; connection and constraint parameters are determined by the foundation connection type, anchor bolt constraint state, and structural boundary conditions. When there are signs of loose bolts or foundation settlement on site, the constraint stiffness can be adjusted by the offset of dynamic response characteristics to make the model's response consistent with the observations in the key frequency range.

[0029] After the finite element mesh is generated, the global stiffness matrix is ​​assembled using the finite element mesh and the nodal displacements are obtained by solving. The assembly process is as follows: On the platform side, a global degree-of-freedom numbering system is established according to the element connection relationship. The stiffness contribution of each mesh element is mapped to the corresponding position in the global stiffness matrix according to its node number. At the same time, boundary constraints are applied to the global matrix in the form of constraint conditions to form a solvable global equation system. Then, the load input is determined according to the working conditions. The load can be composed of the equivalent wind load converted from wind speed, the component self-weight, and additional loads. The wind load is determined by wind speed data, the windward area of ​​the support, and the assumptions of wind direction and wind pressure distribution. On the platform side, the wind speed can be divided into several load levels and a corresponding load distribution template can be configured for each level to ensure the stability and consistency of online calculation. After the load and constraints are determined, a solution strategy suitable for the online scenario is used to calculate the nodal displacements. The selection of the solution strategy is determined by the degree-of-freedom scale and real-time requirements. When the model scale is large, sparse solution and iterative strategy are preferred, and a convergence criterion is set to control the calculation delay. The convergence threshold involved in the solution process is determined based on the principle that "the displacement update amplitude is relatively stable and does not affect subsequent energy calculations". A threshold can be determined on the baseline data to allow the fluctuation of the results to fall within an acceptable range, and it can be appropriately relaxed when computing resources are tight to ensure timely output.

[0030] After obtaining the nodal displacements, the strain energy density is calculated based on these displacements. This calculation process involves the platform mapping the nodal displacements to element deformations according to the element connection relationships, further obtaining the strain state within the element, combining this with material parameters to obtain the stress state, and finally converting the deformation energy of the element under the current deformation into the energy level per unit equivalent volume, thus forming an element-level strain energy density distribution. To avoid misjudgment of local spikes due to mesh discretization or measurement noise, the platform performs spatial smoothing on the strain energy density distribution. The smoothing range is determined by the mesh element size, preferably covering several adjacent layers of elements to suppress isolated spikes while preserving the true energy concentration bands. Simultaneously, a stability constraint is introduced in the time dimension: only when the strain energy density of a certain element remains high for several consecutive refresh cycles is it considered a reliable anomaly and included in the threshold comparison. The above-mentioned smoothing range and continuous cycle number parameters are determined as follows: the lower limit of the smoothing range is based on the principle of covering at least one spatial scale of the connection structure, and the continuous cycle number is determined by a trade-off between diagnostic sensitivity and false alarm tolerance. When it is desired to detect sudden loosening earlier, a smaller continuous cycle number is used, and when it is desired to pay more attention to fatigue accumulation and long-term degradation, a larger continuous cycle number is used.

[0031] After the strain energy density is calculated, the system compares it with a preset threshold. If the strain energy density exceeds the preset threshold, a damage factor is generated. The preset threshold is determined using a combination of "baseline period statistics and stratified operating condition correction": The platform first selects multiple operating cycles confirmed to be in a healthy state as the baseline period, obtains the baseline distribution of strain energy density for each unit using the same grid and calculation link as online, and establishes a normal fluctuation upper boundary for each unit; then, the preset threshold is set as this upper boundary and a safety margin is added to ensure that the threshold is neither too low, leading to widespread false alarms, nor too high, leading to missed early damage. Considering that different wind speed levels will cause different overall energy levels, the platform stratifies the baseline period data according to wind speed ranges or equivalent load levels, establishes a normal fluctuation upper boundary within each stratum, and automatically selects the corresponding threshold based on the current wind speed or load level during online operation, thereby achieving operating condition self-adaptation. If the site lacks sufficient baseline data, the platform can first use the strain energy density reference level calculated by the design model, superimposed with an empirical margin, as the initial threshold. During continuous monitoring, the threshold boundary is gradually converged using newly added health data. The frequency of threshold updates is determined by seasonal changes and structural stability. It is preferable to update the threshold only after the structural state is stable and the data volume reaches a preset scale to avoid threshold drift.

[0032] Damage factors are generated to quantify the degree of exceeding a threshold and improve the usability of spatial analysis. This is achieved by calculating the degree of exceedance of the strain energy density relative to the threshold for each element; the greater the exceedance, the higher the damage factor. To ensure comparability between different components and regions, the platform normalizes the damage factors to fall within a fixed range, and introduces a time-cumulative weight to give higher weight to persistent anomalies, thereby distinguishing between short-term noise spikes and stable degradation. The normalization range is uniformly set by the platform for cross-site alignment. The time-cumulative weight is determined by the diagnostic objective; when the objective is early fatigue identification, the weight of persistence is increased; when the objective is sudden loosening detection, the weight of instantaneous peak values ​​is increased and the cumulative effect is reduced. Furthermore, to avoid amplifying the degree of exceedance under low-load conditions, the system can introduce a load correction coefficient, which is determined by comparing the current load level with the baseline load level, ensuring that the damage factor output maintains a consistent interpretation scale across different load levels.

[0033] After obtaining the damage factors, damage-prone areas are delineated based on the spatial distribution gradient of the damage factors. This step is implemented as follows: the platform establishes spatial neighborhoods based on the adjacency relationships of grid cells, calculates the variation amplitude of the damage factors between each cell and its neighboring cells, and aggregates these amplitudes spatially to form a gradient distribution map. When the gradient exhibits continuous high-value bands or clustered high-value areas within a local region, it indicates that the anomaly has a clear spatial boundary and a concentration trend, and the system marks this region as a candidate damage-prone area. To ensure the delineation results are stable and usable, the platform sets regional screening thresholds and connectivity rules: the connectivity rules are determined by the grid topology, preferably using shared edges or shared nodes as connectivity criteria; the regional intensity threshold is determined by the upper boundary of the gradient distribution during the baseline period and superimposed with a margin, ensuring that areas exceeding this threshold have significant anomaly significance; the regional area threshold is determined by the grid scale and component scale, ensuring that the candidate region covers at least one inspectable structural unit range, thereby excluding small spot areas formed by isolated noise. All of the above thresholds can be set according to different operating conditions to ensure that under high wind conditions, the overall energy increase will not lead to a generally high gradient, resulting in incorrect delineation.

[0034] After delineating the damage-prone areas, extreme points are located within these areas to identify key components of the global scaffold model. The extreme point location is achieved by locally comparing damage factors within each candidate region and selecting the location where the damage factor reaches its local maximum as a candidate key point. Simultaneously, locations where the gradient reaches its local maximum can be further filtered to identify the steepest transition zone of the anomaly boundary, thus balancing the location requirements of both the "strongest anomaly point" and the "anomaly initiation boundary point." To avoid output redundancy caused by multiple adjacent elements having approximately maximum values, an extreme point merging radius is set on the platform side. Multiple candidate points with a distance smaller than the merging radius are merged into one key component, with the point having a higher damage factor or longer duration used as the representative point. The merging radius is determined by the mesh element size and connection construction scale, ensuring a one-to-one correspondence between the final key component and the actual component or node location. This facilitates subsequent modules in reading its geometric parameters and performing stress threshold determination and continuous diagnosis. To enhance the reliability of critical components, the platform can also introduce a consistency screening threshold, which requires critical components to remain in the candidate set for a certain number of consecutive refresh cycles. The number of consecutive cycles is determined by the false alarm tolerance. When there is strong on-site interference, a larger number of consecutive cycles is used to improve stability.

[0035] The threshold determination module acquires detailed geometric parameters of the local complex structure of the identified key parts and determines the stress distribution threshold of these parts in the industrial internet platform using a neural network algorithm. This includes acquiring detailed geometric parameters of the key parts in the industrial internet platform, including the curvature features and spatial coordinates of the reconstructed local structure surface; mapping the detailed geometric parameters into high-dimensional feature vectors and inputting them into a pre-trained convolutional neural network model; simulating the mechanical response characteristics of the local structure through the convolutional neural network model; analyzing the mechanical response characteristics to generate a predicted stress distribution map; and determining the stress distribution threshold of the key parts based on the peak region and gradient change of the predicted stress distribution map. In this threshold determination module, the key parts identified in claim 3 are first used as the processing objects. The industrial internet platform establishes a unique part identifier for each key part and retrieves the corresponding local structural geometric data from the platform's data warehouse accordingly. The sources of geometric data may include local geometric fragments obtained by trimming the design model, on-site point cloud measurement data, photogrammetric reconstruction data, or structural detail data supplemented during operation and maintenance. However, regardless of the data source, the platform converts it into a local surface representation under the same coordinate reference to ensure the comparability and traceability of subsequent curvature features and spatial coordinates. The coordinate reference is usually determined based on the global model coordinate system of the support, and a local coordinate system is established in the key parts to describe the details. The origin of the local coordinate system is preferably set at the geometric center of the key part or at the reference node of the force path, and the coordinate axis direction is preferably consistent with the axis and normal direction of the main component, so that the spatial coordinates obtained by reconstruction in different cycles can be directly compared. When there is an attitude deviation in the on-site data, the platform performs rigid body registration between the local structure and the reference geometry through an alignment strategy, so that the same structural details fall into a consistent spatial position range in different acquisition batches. The matching threshold involved in the registration strategy is determined by the point cloud density and the structural scale. Ideally, the average deviation after registration should not exceed a reasonable proportion of the local minimum feature size to avoid the curvature calculation being amplified by the registration error.

[0036] After completing the geometric data aggregation and coordinate unification, the platform performs surface reconstruction on the complex local structures of key parts to form continuous and computable local surfaces. The specific implementation process of surface reconstruction is as follows: the original geometric data is cleaned to remove outliers, duplicates, and obvious measurement burrs; holes or boundary smoothing are performed on discontinuous areas caused by holes or occlusions to ensure that the surface topology meets the connectivity requirements for curvature calculation; smoothing is introduced on excessively rough surfaces to suppress noise, but the smoothing intensity is limited by the constraint of "not smoothing out real geometric abrupt changes". Therefore, the smoothing parameters are determined with reference to the minimum radius of structural details such as local real fillets, weld toe transitions, bent edges, and hole chamfers, and preferably a reasonable proportion of the smoothed minimum radius not less than the designed or measured minimum radius. Reconstruction resolution is a key parameter in this step, determined by the feature scale of the critical component. When the critical component contains small-scale geometric details such as holes, weld transitions, and stiffener roots, the reconstruction resolution is preferably set to a smaller scale than these details to ensure that curvature abrupt changes can be represented on the mesh or surface. When the critical component is dominated by large-scale bending or gradual transitions, the reconstruction resolution can be appropriately relaxed to reduce computational load. To balance online inference efficiency and threshold accuracy, the platform can establish a tiered resolution strategy for different types of critical components, such as selecting the corresponding reconstruction resolution based on component thickness level, connection type level, or historical stress concentration level, thus providing a stable selection rule for the parameter during engineering deployment.

[0037] After obtaining the reconstructed local structural surface, the platform calculates the curvature features and spatial coordinates in the detailed geometric parameters. The calculation of curvature features is based on discrete sampling points on the local surface. Specifically, the platform generates a set of sampling points on the reconstructed surface, with the spacing between sampling points determined by the reconstruction resolution. Adaptive encrypted sampling can be performed in areas with higher known risks. For each sampling point, the platform extracts local surface fragments in its neighborhood and performs local fitting. The neighborhood size is a key parameter for curvature calculation, and its value is determined by a trade-off between noise level and geometric detail scale. If the neighborhood is too small, the curvature will be sensitive to noise; if the neighborhood is too large, it will average out real abrupt changes. Therefore, the neighborhood size is preferably to cover at least several adjacent sampling points near the sampling point, while not exceeding a reasonable proportion of the minimum geometric abrupt change bandwidth. After the local fitting is completed, the platform extracts curvature-type characterizations from the fitting results and records the curvature-type characterizations together with the spatial coordinates of the sampling point in the local coordinate system, forming a "curvature feature and spatial coordinate" pair. To ensure that features from different key components can be input into the same model and maintain scale consistency, the platform performs scale normalization on spatial coordinates. The normalization scale benchmark can be the side length of the key component clipping box, the feature length of the main component of the key component, or the equivalent scale of the surface area of ​​the key component. Curvature features can also be range-clipping and normalized to avoid extreme noise points causing excessively large feature values ​​that could interfere with model inference. The determination of the above normalization boundaries is given by the platform's historical data statistics, preferably covering the vast majority of curvature distribution ranges under healthy conditions, and limiting the influence of abnormal peaks through clipping, thereby improving the stability of threshold inference. After acquiring the curvature features and spatial coordinates, the platform maps detailed geometric parameters into high-dimensional feature vectors and inputs them into a pre-trained convolutional neural network model. This mapping is achieved by organizing the discrete "curvature features and spatial coordinates" into a structured input acceptable to the model. The platform can employ regular rasterization, local patch aggregation, or voxelization, but regardless of the method, it adheres to the same principle: to express the location, intensity, and connectivity of local geometric abrupt changes as completely as possible within a fixed input size. The input size is a crucial parameter, its value determined by the fixed input requirements of the pre-trained model and fixed during model training. During online inference, the platform clips and scales key areas to fit within this fixed input size range, while maintaining consistent feature density through interpolation or resampling. The resampling density is determined by both the reconstruction resolution and the model input size. Ideally, each geometrically abrupt change in the input should cover at least a sufficient number of input units so that the convolutional neural network can recognize its spatial texture and boundary morphology. When the original sampling density is insufficient, the platform increases the density through interpolation, but interpolation is only used for density uniformity and does not change the relative position and intensity of the actual geometrical changes. To avoid excessive differences in feature scale between different key parts that could lead to model inference bias, the platform performs standardization on the input features. The standardization parameters are derived from statistics saved during the model pre-training phase and are directly called during online inference, thus ensuring consistency in data distribution between training and inference.

[0038] The pre-trained convolutional neural network model runs on the platform as an inference service. The inference process involves feeding high-dimensional feature vectors into the model, which then extracts the spatial relationships of local geometry layer by layer using a multi-layer convolution, nonlinear mapping, and feature aggregation structure. The final output is a feature result characterizing the mechanical response of the local structure. In engineering semantics, this mechanical response feature corresponds to "the possible location of stress concentration, the relative level of concentration intensity, the diffusion path of the concentration area, and the local high-gradient boundary." The platform treats this output as a simulation result of the mechanical response of the local structure under typical loads. The main operating parameters involved in the inference process include inference batch size, inference concurrency, and latency limit. The platform determines the concurrency strategy based on the site size and online refresh cycle, ensuring that all key components are inferred within each refresh cycle. When there are many key components, the platform prioritizes inferring key components with higher historical risk levels and updates low-risk key components at a lower frequency to reduce computational pressure without altering the technical approach of the claims. To ensure that the inference results can be used for threshold determination, the platform can also output an inference credibility index. The threshold for the credibility index is determined by the stability statistics of the model on the historical validation set. When the credibility is lower than the threshold, the platform will reduce the weight of the inference result or trigger the reacquisition of geometric data to avoid threshold drift caused by insufficient geometric reconstruction quality.

[0039] After obtaining the mechanical response characteristics, the platform analyzes these characteristics to generate a predicted stress distribution map. The analysis process involves reconstructing the model output onto a local surface grid or raster based on the surface coordinates of key components, ensuring each surface location corresponds to a predicted stress level, thus forming a visualized and computable stress distribution map. When the model output is a multi-scale feature, the platform first performs multi-scale fusion, uniformly superimposing the overall trend at the coarse scale with the local peaks at the fine scale, so that the final map reflects both the global stress trend and highlights the local concentration at geometric abrupt changes. The weight parameters involved in the fusion are determined by the calibration results during the model pre-training phase. During online use, these parameters are kept fixed or fixed levels are selected according to the key component category to avoid introducing unstable dynamic weights that could lead to inconsistent thresholds. The spatial resolution of the stress map is jointly determined by the aforementioned input organization method and reconstruction resolution. To maintain the engineering feasibility of the thresholds, the platform can limit the map resolution to a scale that "can locate specific structural details or inspectable locations," avoiding excessively fine resolution that would result in overly scattered threshold locations that are difficult to apply.

[0040] Based on the predicted stress distribution map, the platform determines the stress distribution threshold of key parts according to the peak region and gradient change. The determination of this threshold is divided into two levels: peak intensity threshold and peak region boundary threshold, which are combined and output on the platform side. First, the peak region is determined: the platform searches for local maximum regions in the stress distribution map and aggregates them into a peak region candidate set according to connectivity rules. The connectivity rules are determined by the adjacency relationship of the map grid, preferably using connectivity criteria of shared edges or shared nodes. To eliminate isolated points caused by noise spikes, the platform sets a minimum region area threshold, which is determined by the minimum identifiable structural scale of the key part, so that the peak region covers at least one actual structural detail range. Second, the gradient change boundary is determined: the platform calculates the steepness of stress change along the spatial direction outside the peak region and finds the boundary band with the most obvious transition from the "high stress area to the background area". The boundary band determination threshold is determined by healthy baseline data or model training calibration data, preferably ensuring that the identification result of the boundary band is stable and the positional deviation does not exceed the preset tolerance in the healthy state. The preset tolerance is determined by the reconstruction resolution, usually taken as several grid spacings, to ensure that the boundary identification is consistent in different acquisition batches. Finally, the stress distribution threshold was determined: the platform extracts representative stress levels within the peak region as a strength benchmark, and incorporates concentration correction based on the gradient steepness of the boundary zone. When the gradient is steeper, the threshold is appropriately increased to reflect a higher risk of stress concentration; when the gradient is gentler, the threshold is appropriately decreased to avoid missing large-scale moderate stress zones. The final threshold value also incorporates a safety margin, which follows the principle of "covering the upper boundary of healthy fluctuations and being sensitive to anomalies": during the healthy period or confirmed damage-free period, the platform repeatedly generates stress distribution maps for the critical location, statistically analyzes the normal fluctuation range of peak stress and gradient characteristics, and sets the threshold at a reasonable distance outside the upper boundary of normal fluctuations. This ensures that the threshold is not frequently triggered by normal wind fluctuations, while also responding promptly to stress increases at geometrically abrupt changes. When sufficient healthy data is lacking on-site, the platform can initially use a conservative threshold given by model training calibration data as an initial value, and gradually calibrate the margin with newly added healthy data during continuous monitoring, allowing the threshold to converge with the site characteristics. To ensure that the threshold can be directly called by subsequent modules, the platform stores the threshold in a structured manner as "key part identifier, peak area range description, threshold value, applicable working condition label, and update timestamp", and sets a minimum update interval for threshold updates. The update interval is determined by the geometric change rate of key parts and the maintenance frequency.

[0041] The model simplification module, if the determined stress distribution threshold exceeds a preset threshold, extracts the vibration mode contributions of secondary structures from the global model and uses equivalent simplification techniques to reduce the computational complexity of these structures. This includes acquiring nodal stress distribution data; if the maximum nodal stress value is greater than a preset limit, identifying the secondary structural entity; extracting the local stiffness matrix and local mass matrix of the secondary structural entity to calculate the vibration mode vector, and constructing a mode transformation matrix accordingly; using the mode transformation matrix to perform orthogonal transformation on the local stiffness matrix and local mass matrix to generate an equivalent stiffness matrix and equivalent mass matrix; and assembling the equivalent stiffness matrix and equivalent mass matrix back into the global finite element model to reduce the computational complexity of the secondary structure.

[0042] In this simplified processing module, the triggering prerequisite is that the threshold determination module has already given a stress distribution threshold for the critical parts, and this stress distribution threshold exceeds a preset threshold. The preset threshold here expresses the risk level where "stress concentration in the critical parts has reached a level requiring improved online calculation efficiency." Its determination is constrained by the structural design allowable level, the statistical upper boundary of the health period, and the operation and maintenance strategy: The platform preferably first provides a basic threshold range based on design specifications or enterprise internal control standards. Then, it uses confirmed, damage-free operational data to statistically analyze the peak level and fluctuation range of the predicted stress distribution in the critical parts. The preset threshold is set outside the upper boundary of this fluctuation range and a safety margin is added to ensure that normal wind fluctuations do not trigger simplification, while stable triggering occurs when stress concentration significantly increases. Simultaneously, the preset threshold can be set in layers according to wind speed level or equivalent load level. The layer boundaries are determined by historical wind speed distribution and the sensitive range of the support response to avoid frequent false triggering under high wind conditions or delayed triggering under low wind conditions. The trigger criteria are implemented on the platform side using a rule engine. The rule engine reads the stress distribution threshold corresponding to the key part identifier and its applicable working condition label. After matching the current working condition, it calls the corresponding preset threshold and completes the over-limit judgment. Once the over-limit judgment is completed, it immediately enters the secondary structure identification and equivalent simplification process.

[0043] After entering the simplified process, the module first acquires the nodal stress distribution data. This data is output by the response calculation module or the finite element solution process. The platform stores the stress values ​​of each node under the current working condition using a unified data structure, retaining the node number, spatial coordinates, and identifiers of the associated component and substructure to support subsequent structural entity identification. The acquisition process specifically includes: reading the stress results of each element from the global finite element model solution and converting them to the node level according to preset mapping rules; when the same node is shared by multiple elements, the platform can use a weighted aggregation strategy to generate representative stress values ​​for that node. The weighting weight is determined by the element volume or the correlation strength between the element and the node, thus avoiding unreasonable amplification of abnormal peaks in local elements at the node level; to improve the stability of the criteria, the platform introduces a standardization process for the nodal stress values, allowing direct comparison of nodal stresses from different refresh cycles, and can perform load normalization on the nodal stresses according to the working condition. The load normalization scale is determined by the current load level and the reference load level to ensure a consistent interpretation scale for "stress increments" under different wind conditions.

[0044] After obtaining the nodal stress distribution data, the module performs a judgment and identification process: "If the maximum nodal stress value is greater than a preset limit, then secondary structural entities are identified." The calculation of the maximum nodal stress value is implemented on the platform side in a traversal manner: all nodal stress values ​​involved in the calculation in this cycle are scanned, the maximum value is taken, and the corresponding node number, the component to which the node belongs, and the substructure to which it belongs are recorded. The preset limit is used to express that "the current global response has entered a high stress state, and it is necessary to reduce the order of the parts of the model that occupy a large degree of freedom to ensure subsequent real-time iterations." This limit is different from the aforementioned preset threshold. The preset threshold is a stress distribution threshold for key parts, used to trigger the simplification process, while the preset limit is a global maximum nodal stress, used to guide "which structures should be regarded as secondary structural entities and enter equivalent simplification." The principle for determining the preset limit is "bound to the global stress level and consistent with the risk level of key parts." Specifically, the platform first statistically analyzes the distribution range of the global maximum nodal stress in the health period data to obtain its normal fluctuation upper boundary. Then, combined with the trigger level of the stress distribution threshold of key parts, the preset limit is set at a reasonable distance outside the normal fluctuation upper boundary, with an added safety margin to ensure that secondary structure identification is only initiated when the global stress significantly increases. When the system adopts a layered operating condition approach, the preset limit is also set according to wind speed or load level, and the upper boundary and margin are determined separately within each layer to ensure a consistent trigger probability for judgment under different operating conditions. To avoid false judgments caused by transient spikes, the platform can introduce a time persistence criterion, requiring the maximum nodal stress to exceed the preset limit for several consecutive refresh cycles before entering the identification process. The number of consecutive cycles is determined by a trade-off between false alarm tolerance and real-time requirements, preferably improving judgment stability without significantly increasing response lag.

[0045] The identification process for secondary structural entities is based on the core criterion of "relatively small contribution to global dynamics but large occupation of degrees of freedom," which is achieved on the platform side through structural partitioning and contribution assessment. In specific implementation, the platform first uses the substructure partitioning information of the global model to divide the support into several structural entity units, such as beam segments, bracing segments, connector sets, secondary beam sets, or auxiliary component sets, and establishes its node set and element set for each structural entity. Subsequently, the stress statistical index and dynamic contribution index of each structural entity are calculated. The stress statistical index can include the maximum value, mean value, and high quantile value of the nodal stress within the entity, which is used to determine whether the entity is in a high stress state. The dynamic contribution index is used to determine the degree of influence of the entity on the global response within the key frequency range. The determination of dynamic contribution indicators can utilize existing modal information: the platform retrieves the modal contribution results of key frequency ranges saved during the modeling or consistency judgment phases, calculates the participation degree of each structural entity in these modes, and marks entities with low participation and whose stress statistics are not dominant as candidate secondary structures. If the platform has not yet formed a stable modal contribution history, structural function classification rules can be used as an initial strategy. For example, main beams, main columns, and main connection nodes on the main load-bearing path can be classified as primary structures, while components used to support auxiliary functions or far from the main load transfer path can be classified as secondary structures. The classification can be gradually corrected using modal contribution results during continuous monitoring. The final confirmation of candidate secondary structures must also meet the constraint of "not disrupting the consistency of key frequency ranges after replacement." Therefore, the platform sets an upper limit on the number of candidate structures or an upper limit on the degree of freedom ratio. The upper limit is determined by the online calculation delay target, making the benefits of order reduction quantifiable and controllable.

[0046] After identifying the secondary structural entity, the module extracts its local stiffness matrix and local mass matrix to calculate the vibration mode vector and construct the modal transformation matrix accordingly. The local matrix extraction process is as follows: the platform uses the node set and element set of the secondary structural entity as indices to segment the corresponding local sub-model data from the global finite element model. During segmentation, boundary nodes are marked. Boundary nodes are the interface nodes connecting the secondary and primary structures, and their processing determines the coupling accuracy with the global model after equivalent simplification. Therefore, the platform configures the degree-of-freedom retention strategy for boundary nodes as a key parameter. The principle for determining this parameter is "motion transmission at the interface must not be weakened." Typically, all degrees of freedom of the boundary nodes are retained as explicit degrees of freedom, enabling the equivalent model to correctly transmit forces and displacements. When computational resources are extremely limited, only degrees of freedom related to the main force direction can be retained, but the response consistency needs to be more rigorously verified in the consistency judgment module to offset the risks brought by this approximation.

[0047] After extracting the local stiffness matrix and local mass matrix, the platform calculates the vibration mode vectors on the local sub-model. The calculation process is as follows: the modal morphology of the substructure is solved in the local degree of freedom space, resulting in a set of mode vectors arranged from low to high frequency. To adapt to online scenarios, the platform does not need to retain all modes, but selects the number of modes to retain based on either the "number of modal cutoffs" or the "modal contribution coverage rate". The method for determining the number of modal cutoffs is as follows: using the upper limit of the key frequency range as a reference, low-order modes that can cover the main dynamic characteristics within the key frequency range are retained. At the same time, to improve robustness, several additional modes can be retained to cover boundary effects. If the modal contribution coverage rate strategy is adopted, the coverage level of the retained modes to the local mass participation is statistically analyzed. The coverage level threshold is determined by empirical and historical consistency results. Preferably, retention is stopped after the coverage level reaches a preset proportion to achieve a balance between accuracy and computational load. To ensure that the modal vectors are numerically usable for subsequent orthogonal transformations, the platform normalizes the modal vectors and tracks and corrects instability caused by possible modal exchanges or modal proximity. The correction is based on the similarity of modal morphology between adjacent periods. If the similarity is lower than a preset threshold, the modal order is rearranged or a more stable modal matching strategy is adopted. This similarity threshold is determined by the modal similarity distribution during the healthy period, and its lower quantile boundary is taken to avoid normal fluctuations being misjudged as instability.

[0048] The modal transformation matrix is ​​constructed based on the selected set of modal vectors. The platform combines the retained modal vectors column-wise to form transformation relationships, enabling the original degrees of freedom of local substructures to be mapped to a lower-dimensional modal coordinate space. The dimension of the transformation matrix is ​​jointly determined by the "number of original local degrees of freedom" and the "number of retained modes." A smaller number of retained modes results in a more significant order reduction, but weakens the ability to express high-frequency details. Therefore, the platform adopts an adaptive strategy for this parameter: when online latency pressure increases or critical parts have higher risks requiring faster iteration, the number of retained modes is preferentially reduced to improve speed; when consistency judgment finds that the simplified response deviation increases, the number of retained modes is increased to restore accuracy. To avoid introducing non-physical coupling into the transformation, the platform performs orthogonalization on the transformation matrix to meet the numerical requirements of orthogonal transformation. The orthogonalization process adopts a stable numerical flow and sets an orthogonality check threshold, which is determined by numerical stability requirements. It is preferable to keep the condition number of the equivalent matrix after transformation good to prevent divergence in subsequent solutions.

[0049] After constructing the modal transformation matrix, the module uses this matrix to perform orthogonal transformations on the local stiffness matrix and local mass matrix to generate equivalent stiffness and mass matrices. The pure textual calculation process for this step is as follows: the platform first maps the local stiffness matrix to modal coordinate space to obtain the stiffness expression in modal coordinates, and then maps the local mass matrix to modal coordinate space in the same way to obtain the mass expression in modal coordinates. During the mapping process, the platform simultaneously performs interface processing on the degrees of freedom retained at the boundary nodes, ensuring that the coupling relationship between the boundary degrees of freedom and modal coordinates is correctly expressed, thereby ensuring that the equivalent model can maintain consistent force and displacement transmission with the main structure at the interface. After the equivalent stiffness matrix and equivalent mass matrix are generated, the platform performs numerical consistency checks, including symmetry checks, positive definiteness checks, and dimensional consistency checks. Symmetry and positive definiteness are used to ensure the stability of subsequent global solutions. The check thresholds are determined by the numerical error tolerance, preferably allowing minor numerical asymmetry without changing the physical meaning, which can be eliminated through symmetry correction. If the check fails, the platform can automatically backtrack and increase the number of retained modes or adjust the orthogonalization intensity to restore the stability of the equivalent matrix. The above check and backtracking strategies are stored as online robustness parameters in the platform's strategy library. The strategy trigger thresholds are obtained from statistics of unstable cases that have occurred in historical operations, enabling the system to have continuous adaptive capabilities.

[0050] After obtaining the equivalent stiffness matrix and equivalent mass matrix, the module reassembles them into the global finite element model to reduce the computational complexity of the secondary structure. The assembly process is as follows: the platform locates the original set of degrees of freedom corresponding to the secondary structure entity and its boundary interface set of degrees of freedom in the global model, and replaces the high-dimensional matrix blocks of the original local substructure from the global matrix with equivalent matrix blocks; during the replacement, the boundary node numbers are kept consistent with the global numbers, allowing the main structure to couple with the equivalent substructure without modification; simultaneously, the platform updates the data index of the global model, including the degree of freedom mapping table, the matrix sparse structure index, and the solver preprocessing information, to ensure that subsequent solutions can be directly executed on the new model. To prevent the dynamic response of the global model from deviating from the initial dynamic response characteristics after replacement, the platform immediately outputs a "simplified label" after assembly and sends the updated model to the consistency judgment module. The consistency judgment module verifies the response consistency within the key frequency range. If the consistency does not meet the preset criteria, the platform triggers a recovery mechanism, increasing the number of retained modes or removing some structures from the secondary structure set, generating a new equivalent matrix, and reassembling until the consistency is met or the preset maximum number of adjustments is reached. The maximum number of adjustments is determined by the online latency budget, preferably providing a limited number of self-correction opportunities without affecting real-time performance.

[0051] The consistency judgment module updates the dynamic response of the global model by reducing computational complexity and determines whether the updated response is consistent with the initial dynamic response features. This includes acquiring global model data to extract initial dynamic response features and constructing a sparse matrix representing the model structure to reduce computational complexity; performing low-dimensional space iteration based on the sparse matrix and incremental business data to generate an updated global model; inputting the test stimulus signal into the updated global model to output the updated dynamic response; calculating the response deviation value between the updated dynamic response and the initial dynamic response features; and determining that the updated response is consistent with the initial dynamic response features if the response deviation value is less than the feature drift threshold.

[0052] In this implementation, the module first acquires global model data to extract initial dynamic response features and constructs a sparse matrix characterizing the model structure to reduce computational complexity. The global model data includes node information, element information, material information, constraint information, load template information, and a global matrix data structure for dynamic calculations. The initial dynamic response features are not generated temporarily in this step but originate from benchmark features formed by the data acquisition module and the modeling module during system startup or a healthy baseline period. The platform persistently stores these features as a "reference benchmark" and indexes them according to key frequency ranges, measurement point sets, and feature types for direct retrieval during consistency checks. The initial dynamic response features preferably include at least the following categories: a set of peak frequency locations within the key frequency range, a set of response amplitudes corresponding to the peaks, an overall description of the response curve's shape within the key frequency band, the relative response distribution of the dominant modes at each measurement point, and a description of the damping-related response width. The reason these features need to be stored in multiple dimensions is to avoid misjudgments caused by using only a single indicator. For example, the peak frequency may remain consistent, but changes in response amplitude or relative distribution could also indicate simplification and distortion. Corresponding to the initial dynamic response characteristics, the platform constructs a sparse matrix in this step to express the structural coupling relationships of the global model by storing only non-zero terms, thereby reducing the matrix storage and computational scale. The sparse matrix construction process is as follows: the platform first reads the node connection relationships and unit assembly relationships of the global model and establishes a degree-of-freedom numbering table; then, according to the assembly relationships, it maps the non-zero coupling terms that should originally be written into the global matrix to the sparse index structure, retaining only degree-of-freedom pairs with structural connections or constraint couplings; finally, it generates a sparse storage format compatible with the solver and simultaneously generates row and column index acceleration structures so that subsequent incremental updates only require access to the affected sub-blocks. Key parameters involved in the sparsification process include the sparsity threshold and retention rules. The sparsity threshold is used to handle small terms that are close to 0 in value but are generated due to numerical errors. The threshold is determined by the numerical stability requirements. It is preferred to base it on the magnitude distribution of matrix terms in the healthy period solution, and terms that are significantly lower than the principal magnitude by several orders of magnitude are regarded as negligible terms. The retention rule prioritizes structural connection. That is, as long as there is a real connection or real constraint, the corresponding coupling term must be retained, even if its value is small, so as to avoid non-physical disconnection introduced by sparsification.

[0053] Secondly, the module performs low-dimensional space iteration based on the sparse matrix and incremental business data to generate an updated global model. Incremental business data describes "the parts that have changed relative to the initial version or the previous version," including at least the equivalent stiffness and equivalent mass data blocks replaced by the model simplification module, load template corrections caused by changes in operating conditions, boundary condition corrections caused by changes in constraint states, and calibration results of material or connection parameters from the platform side. The process of acquiring incremental business data is as follows: the platform reads the replacement list output by the simplification module, which includes the identifiers of the replaced substructures, interface degree-of-freedom mappings, equivalent matrix block data, and replacement timestamps; simultaneously, it reads the operating condition data for the current period to form a load increment description; then it reads constraint monitoring or maintenance inputs to form a boundary increment description; finally, it merges these increments into a set of increments that can be used for matrix updates and maps them to the index space of the sparse matrix according to their scope of influence. The core of low-dimensional space iteration is "iterative updates only within the low-dimensional subspace affected by the increments," thereby avoiding full recalculation. The implementation process is as follows: The platform locates the affected degree-of-freedom set based on the incremental set, and expands this set and its adjacent coupled degrees of freedom into an iterative subspace. The expansion range is determined by the structural coupling strength and the sensitivity of the key frequency range, preferably ensuring that the iterative subspace covers the main coupling paths around the interface degrees of freedom. Subsequently, the platform performs a finite number of iterative updates within this iterative subspace. Each iteration includes updating the sub-block matrix, correcting the sub-block solution state, and calculating the convergence index. When the convergence index meets the preset conditions or the number of iterations reaches the upper limit, the iteration stops, and the sub-block update results are written back to the global sparse matrix, forming the updated global model version. The key parameters here include the iterative subspace dimension, the upper limit of the number of iterations, the convergence criterion, and the incremental weights.

[0054] The larger the dimension of the iterative subspace, the easier it is to satisfy consistency, but the computational cost increases. Therefore, its determination method is usually a tiered strategy: when the simplified replacement involves only a few minor structures and has few interface degrees of freedom, a smaller dimension is used; when the replacement involves multiple substructures or has dense interface degrees of freedom, a larger dimension is used. The upper limit of the number of iterations is determined by the online latency budget, preferably ensuring that subsequent test stimulus calculations can still be completed within one refresh cycle. The convergence criterion is used to determine whether the iteration has reached stability. The criterion can be based on the change amplitude of key response indicators within the subspace. The change amplitude threshold is determined by the numerical fluctuation range during the healthy period, so that numerical errors are not mistaken for non-convergence. The incremental weight is used to control the update amplitude when multiple increments coexist. The weight is determined based on the principle of "avoiding overshoot caused by an excessively large update". When the source of the increment is reliable and the amplitude is small, the weight can be increased. When the source of the increment is uncertain or the amplitude is large, the weight is decreased and the increment is gradually approximated through multiple iterations.

[0055] Next, the module inputs the test excitation signal into the updated global model to output the updated dynamic response. The purpose of the test excitation signal is not to simulate all the details of real wind load, but to provide a repeatable and comparable excitation benchmark for consistency judgment, making the dynamic responses output by different model versions under the same excitation conditions comparable. The test excitation signal is generated on the platform side using a standardized strategy. The generation process is as follows: the excitation frequency band coverage is determined based on the key frequency range, ensuring that the excitation includes energy components within the key frequency range; the excitation amplitude is set based on the model's current damping level and response amplitude range, ensuring that the output response is neither overwhelmed by noise due to too small an amplitude nor enters a nonlinear state due to too large an amplitude; the excitation length is determined based on the refresh period, ensuring that the output response covers sufficient time or frequency sampling points to extract stable features. Key parameters of the test excitation signal include frequency band range, amplitude level, excitation length, and input location or input method.

[0056] The frequency band range is directly determined by the key frequency range, and a certain margin can be extended at both the upper and lower boundaries to cover response changes near the boundaries. The size of the margin is determined by the sensitivity of the initial dynamic response characteristics near the boundaries. The amplitude level can be calibrated according to the typical response amplitude during the healthy period to ensure that the output response is within the identifiable range. The excitation length is determined by both the frequency resolution requirement and the online latency; the longer the length, the higher the resolution but the slower the calculation. The input position or input method is determined by the model excitation interface. It is possible to apply equivalent excitation at representative nodes or to cover the main force path with multi-point excitation. The selection principle is "sufficient excitation capability for the dominant modes of the key frequency band." After completing the excitation input, the platform calls the updated global model to perform dynamic response calculation and outputs updated dynamic response data. The dynamic response data is organized with the same caliber as the initial dynamic response characteristics, including the response amplitude curve, main peak position, and relative response distribution within the key frequency range, so as to facilitate comparison with the same caliber.

[0057] Subsequently, the module calculates the response deviation between the updated dynamic response and the initial dynamic response features, and uses a feature drift threshold for consistency determination. The response deviation is not calculated as a single difference, but rather as a comprehensive measure of multiple feature dimensions. The calculation process is as follows: the platform first aligns the frequency axes and measurement point sets of the two sets of features to ensure a one-to-one correspondence between the comparison objects; then, it calculates the main peak frequency offset, main peak amplitude offset, key frequency band curve morphology difference, and relative response distribution difference, and combines these differences into a response deviation value according to preset weights. The weight parameters reflect the contribution of different features to consistency, and are determined based on the principle of "higher weight for features that have a greater impact on subsequent stress and damage localization": for example, if subsequent diagnosis is extremely sensitive to the resonance frequency, the weight of the main peak frequency offset is increased; if subsequent diagnosis relies more on modal morphology consistency, the weight of the relative response distribution difference is increased. The weights can be fixed in the initial stage of system deployment and calibrated based on misjudgment cases and consistency verification results during long-term operation. However, the calibration process is controlled by version management to ensure that the weights do not change arbitrarily within the same version cycle, thereby maintaining the interpretability of the judgment. The characteristic drift threshold is used to define the boundary of "acceptable deviation". Its determination method also adopts health period statistics and working condition stratification correction: Under the condition that the structure is healthy and no simplification replacement is performed, the platform outputs dynamic response of the global model in multiple cycles with the same test stimulus, calculates the natural fluctuation deviation distribution between these cycles, and sets the characteristic drift threshold outside the upper boundary of the distribution and adds a safety margin so that normal fluctuations will not be judged as inconsistent. When there are obvious differences in working conditions, the platform performs stratified statistics on the deviation distribution according to wind speed level or load level, and sets drift thresholds in each stratum to avoid misjudgment caused by the natural increase of deviation under high wind conditions. To further suppress occasional numerical fluctuations, the platform can introduce a continuous judgment strategy, which requires that the response deviation value be less than the feature drift threshold in a number of consecutive judgments before outputting a "consistent" conclusion. The number of consecutive judgments is determined by a trade-off between the tolerance for misjudgment and the requirements for real-time performance. When an "inconsistent" conclusion is output, the platform refines the reasons for inconsistency into sources of offset, such as excessive shift in the main peak frequency or excessive difference in the relative response distribution, and feeds them back to the model simplification module to adjust the number of modes to be retained, adjust the dimension of the iterative subspace, or backtrack some replacements to form a closed-loop optimization.

[0058] The response calculation module, if the judgment is consistent, acquires real-time load data to update the model, obtaining the current stress and deformation distribution of the support. This includes: if the monitored support operating state matches the preset consistency state, acquiring real-time load data streams; mapping the real-time load data streams to the geometric topology nodes of the finite element model to update the boundary conditions of the finite element model; assembling the global stiffness matrix based on the boundary conditions and support material properties, and solving the global stiffness matrix to obtain the node displacement vectors; calculating the strain field data based on the node displacement vectors to obtain the current stress and deformation distribution of the support.

[0059] In this response calculation module, the prerequisite for entering the calculation process is that the consistency judgment module has output a judgment conclusion that "the updated response is consistent with the initial dynamic response characteristics." To ensure the executability of this prerequisite on the platform side, the system converts the consistency judgment result into a machine-readable status flag, and combines this status flag with the model version identifier, key frequency range label, test stimulus verification timestamp, and data quality status to form a "preset consistency state." The so-called "preset consistency state" is not an abstract description, but a set of joint conditions configurable on the platform side, used to constrain real-time load-driven stress and deformation calculations to only be performed when the model is reliable and the data is trustworthy. The parameters included in this status and their determination methods are as follows: The model version consistency parameter confirms that the model version called for real-time calculation is completely consistent with the model version that has passed the consistency check. This parameter is automatically generated and forcibly bound by the platform version management. The response deviation compliance parameter confirms that the response deviation value in the previous round of consistency judgment does not exceed the feature drift threshold. This parameter is output by the consistency judgment module. The data quality compliance parameter confirms that the sensor data in this cycle meets the minimum validity requirements. This parameter is determined based on the data missing ratio, outlier ratio, and time alignment error. The missing ratio threshold and outlier ratio threshold are obtained from the health period data and a margin is added to ensure that normal communication jitter does not cause frequent interruptions to the calculation. The load condition label matching parameter confirms that the current wind speed level or equivalent load level is consistent with the model's applicable load condition label. This parameter is determined by the load stratification strategy, thereby avoiding inputting loads that are not compatible with the load condition into the model, which would amplify the error. The above joint conditions are implemented on the platform side using a rule engine. The rule engine verifies each item and outputs a final result of "match" or "not match". If the monitored support operation status matches the preset consistency status, the real-time load acquisition and solution process is entered; if it does not match, stress and deformation distribution are not output in this cycle, or a degradation strategy is adopted, such as maintaining the previous reliable result and prompting that the data or model needs to be reviewed.

[0060] After confirming the operational status is consistent, the system collects real-time load data streams. The collection of real-time load data streams is not limited to wind speed alone; rather, the data is organized with the goal of "forming the load description required for finite element boundary conditions." Therefore, it can include wind speed, wind direction, gust index, temperature, component attitude, and auxiliary wind field information from meteorological stations or edge stations. The collection process is completed by the data access layer of the industrial internet platform: the platform configures the sampling frequency and reporting cycle for each load data source and synchronizes the execution time of different data sources to ensure that all load quantities have a consistent time reference within the same calculation cycle. The sampling frequency and reporting cycle are determined as follows: the reporting cycle is determined by the online diagnostic refresh cycle, ensuring that at least one complete set of load data is generated in each refresh cycle; the sampling frequency is determined by the rate of wind change and the structural response sensitivity. When there are significant gusts or rapid changes in wind direction at the site, the sampling frequency is increased to avoid the load being averaged by low-frequency sampling; when computing resources are limited, the platform can perform short-window aggregation output of the load data. The aggregation method can adopt a method that combines mean and peak values, preserving both the average load level and the short-term peak risk. Load data quality control also involves thresholds: for example, the wind speed validity threshold is used to determine whether the wind speed data is within the effective measurement range of the sensor, and the threshold is determined by the sensor range; the gust identification threshold is used to determine whether short-term fluctuations constitute gust events, and the threshold is determined by the upper boundary of the healthy period wind speed fluctuation statistics and the addition of margin, so that real sudden fluctuations are identified while normal perturbations are not amplified; the temperature correction trigger threshold is used to determine whether the material parameters need to be corrected for temperature, and the threshold is determined by the sensitive range of material properties changing with temperature and the temperature distribution of the site.

[0061] After the real-time load data stream is acquired, the system maps the real-time load data stream to the geometric topology nodes of the finite element model to update the boundary conditions of the finite element model. This mapping is based on the "one-to-one correspondence between the load action area and the set of structural nodes". The platform pre-establishes a load mapping table, which includes at least: the action surface or action component corresponding to each type of load, the set of nodes or elements of the action surface or component in the finite element mesh, the load allocation weights, and the direction conversion rules related to the wind direction. The specific implementation process of the mapping is as follows: The platform first reads the wind direction and component attitude of the current cycle to determine the load action direction and the windward surface; then, the wind speed and gust index are converted into equivalent load intensity. The conversion process is completed by the platform load template. The parameters of the load template are determined by the site structure dimensions, windward area, and empirical wind pressure distribution rules, and can be calibrated based on the measured response during long-term operation; then, the platform allocates the equivalent load to the nodes or elements according to the weights based on the load mapping table. The weights are determined in a way that ensures the load is continuous in space and proportional to the load-bearing area of ​​the component, avoiding the concentration of loads on a small number of nodes, which would lead to numerical instability. For boundary condition updates, the platform updates not only external loads but also constraint states, such as the equivalent stiffness parameters of anchor bolt constraints, support constraints, or connection constraints. The triggering conditions for constraint updates can come from maintenance inputs or monitored changes in connection states. The trigger threshold is determined by connection-related response anomaly indicators, ensuring that constraint changes are reflected in the model in a timely manner. To prevent excessive instantaneous changes caused by load mapping, the platform can introduce load change limiting parameters. These limiting parameters are determined by the upper boundary of historical load change rates. When the load change in the current cycle exceeds the limit, the platform performs a smooth load transition or requires data verification to avoid abnormal data directly driving unreliable displacements in the finite element solution.

[0062] After the boundary conditions are updated, the system assembles the global stiffness matrix based on the boundary conditions and the material properties of the support, and solves the global stiffness matrix to obtain the nodal displacement vectors. The assembly process is as follows: the platform reads the updated finite element model version, including node numbers, degree-of-freedom numbers, element connection relationships, material properties, and the set of boundary conditions; then, it calculates the contribution of each element to the overall stiffness and assembles it into the global stiffness matrix according to its degree-of-freedom number. During the assembly process, constraint conditions are simultaneously written into the constraint processing structure of the global matrix to ensure that constrained degrees of freedom are correctly handled in the solution; and nodal loads or element loads are summarized into a global load vector, making it consistent with the global degree-of-freedom number. Material properties include elastic-related parameters and density-related parameters. Elastic-related parameters are used for static or quasi-static displacement solutions, while density-related parameters are used when inertial effects need to be considered or dynamic response solutions are required. Material properties are determined primarily using design material data and batch data. If temperature correction is enabled on the platform, material properties are corrected based on the temperature of the current cycle, with the temperature correction coefficient determined by the material library or historical calibration results. If aging correction is enabled, material or connection stiffness is empirically corrected based on long-term monitored dynamic response drift trends, and the correction process is included in version management to ensure traceability. When solving for nodal displacement vectors, the platform selects a solution strategy to meet online latency requirements. The selection parameters for the solution strategy include solver type, iteration upper limit, and convergence threshold: direct solution can be used when the model size is small or the latency budget is sufficient; when the model size is large, sparse iterative solution is preferred, and preprocessing is configured to accelerate convergence. The iteration limit is determined by the available computation time within the refresh cycle, and the convergence threshold is determined by the stability requirements of the displacement update amplitude. It is preferably set within the range of numerical fluctuations when repeatedly solving the same working condition during the healthy period, so that numerical errors do not significantly affect subsequent strain and stress calculations. When convergence fails to meet the standard but reaches the iteration limit, the platform can output a degradation flag and mark the result of that cycle as low confidence, or trigger a rollback strategy that increases the number of retained modes and re-verifies consistency, in order to prevent non-converged results from entering the damage identification link.

[0063] After obtaining the nodal displacement vectors, the system calculates the strain field data based on these vectors, thereby obtaining the current stress and deformation distribution of the support. This calculation process is implemented on the platform side using elements as the basic objects: the platform projects the nodal displacements onto the deformation description within the element according to the element connection relationship, obtaining the deformation distribution of the element under the current load; then, the deformation is converted into strain field data, and the strain field data is converted into stress field data according to the material properties, thus forming the stress distribution of the support. The stress distribution can be output at the nodal level or the element level, with the output level determined by the input requirements of the subsequent damage identification module. If alignment with the geometric details of critical parts is required, it is preferable to output stress results with higher spatial resolution in the critical part areas. The deformation distribution is directly formed by the nodal displacements. The platform can interpolate the nodal displacements on the structural surface to generate a continuous deformation morphology description for presentation in the online diagnostic report. Key parameters involved in this stage include the result smoothing range, peak identification threshold, and output resolution. The result smoothing range is used to suppress isolated spikes caused by mesh discretization. Its determination is constrained by not weakening the true stress concentration, and preferably covers a reasonable proportion of several adjacent layers but does not exceed the minimum geometric change scale. The peak identification threshold is used to identify stress concentration areas. The threshold can be linked with the stress distribution threshold output by the threshold determination module, or the high quantile boundary of the stress statistics in this period can be used as an auxiliary threshold to maintain cross-period comparability and have the ability to adapt to the current period. The output resolution is jointly determined by the mesh scale and the report presentation requirements. Under the premise of meeting the spatial accuracy of subsequent damage localization, it avoids excessively high output resolution, which would lead to excessive storage and transmission burden on the platform.

[0064] The damage identification module, based on the obtained stress and deformation distributions, recalculates the vibration modes within the critical frequency range using modal analysis and determines potential damage locations. This includes acquiring full-field stress and deformation distribution data, updating the geometric configuration using the deformation distribution data, and constructing a tangent stiffness matrix incorporating stress stiffening effects based on the full-field stress distribution data. Eigenvectors within the critical frequency range are extracted from the tangent stiffness matrix as vibration mode shapes. The module calculates the element modal strain energy based on the vibration mode shapes and the tangent stiffness matrix. If a local abrupt peak occurs in the element modal strain energy, damage location indicators are calculated based on this peak to determine the potential damage location.

[0065] In this embodiment, the module first acquires full-field stress distribution data and deformation distribution data. It then uses the deformation distribution data to update the geometric configuration and constructs a tangent stiffness matrix incorporating stress stiffening effects, based on the full-field stress distribution data. The acquisition of full-field stress distribution data and deformation distribution data is handled by the platform's data layer. This data layer performs consistency processing on the results from the finite element solution, ensuring that stress and deformation can be directly correlated within the same mesh and node numbering system. The deformation distribution data is used as follows: the platform reads the displacement value of each node under the current load and superimposes this displacement value onto the original node coordinates, thereby generating an updated set of node coordinates. This updated set of node coordinates constitutes the "updated geometric configuration." To avoid unreasonable local wrinkles or peaks in the geometric configuration due to numerical noise, the platform introduces geometric smoothing parameters and geometric update limiting parameters during geometric updates. The geometric smoothing parameter is used to spatially smooth the displacement field. Its determination principle is to "suppress discrete noise without smoothing the true deformation gradient." Ideally, the smoothing effect should cover a reasonable proportion of adjacent elements but not exceed the minimum geometric abrupt change scale. The geometric update limiting parameter is used to limit the update amplitude of node coordinates within a single period. The limit is determined based on the upper boundary of displacement increments in historical health data, with an added margin, ensuring that normal operating condition fluctuations do not trigger the limit, while abnormal jumps are suppressed and marked for review. If the platform determines that the current deformation distribution contains obvious anomalies, it can use neighborhood interpolation to replace the anomalies or trigger a re-solution, thereby ensuring the reliability of the geometric configuration update.

[0066] After the geometric configuration is updated, the platform constructs a tangent stiffness matrix based on the global stress distribution data to express the equivalent stiffness of the structure under the current stress state. The construction of the tangent stiffness matrix in engineering implementation adopts a method of "superimposing material stiffness contribution and stress-related contribution": the platform first recalculates the geometric and connectivity relationships of the elements according to the updated geometric configuration to ensure that the element stiffness contribution is consistent with the current configuration; then, it reads the global stress distribution and maps the stress state of each element to the stiffness correction term of that element, so that the influence of the element on the overall stiffness under high stress can be reflected in the tangent stiffness matrix; finally, the tangent stiffness contributions of each element are assembled according to the global degree of freedom number to form a global tangent stiffness matrix. Key parameters involved in this process include stress mapping aperture, stress smoothing range, and stiffening effect activation threshold. The stress mapping caliber specifies the input method for stress from elements to stiffness correction. The platform can use representative values ​​of element stress as input. The representative values ​​can be determined by the average value or high quantile value of element stress to avoid unreasonable amplification of local peaks in stiffness correction. The stress smoothing range is used to suppress the influence of stress dispersion noise on stiffness correction. Its determination is consistent with geometric smoothing and is constrained by the minimum geometric abrupt change scale. The stiffness effect activation threshold is used to avoid introducing unnecessary stiffness correction calculations when the stress level is low, thereby saving computational resources. This threshold is preferably determined based on the stress level distribution during the healthy period. When the stress level is below the median of normal fluctuations, stiffness correction can be disabled or weakened. When the stress level is close to or exceeds the upper boundary of the healthy period, stiffness correction is enabled to improve the accuracy of state representation. The final values ​​of the above thresholds must be consistent with the site condition stratification strategy and set according to wind speed level or load level to ensure that the construction caliber of the tangent stiffness matrix is ​​consistent under different conditions.

[0067] Secondly, the module extracts eigenvectors from the tangent stiffness matrix that fall within the critical frequency range as vibration mode shapes. To accomplish this step, the platform establishes a modal solving task between the tangent stiffness matrix and the mass correlation matrix, and employs a screening strategy applicable to the critical frequency range, extracting only the modal results corresponding to the critical frequency range to avoid computational burden caused by solving the entire frequency band. The modal extraction process is as follows: the platform first reads the critical frequency range boundary and converts it into modal screening conditions; then, it selects a solution strategy to generate a candidate mode set, which is arranged in ascending order of frequency; subsequently, based on the screening conditions, it locates the modes falling within the critical frequency range in the candidate set and outputs their corresponding eigenvectors as vibration mode shapes. Key parameters for this step include the upper limit of the number of extracted modes, the modal screening boundary margin, the solution convergence threshold, and the modal stability criterion. The upper limit for the number of extracted modes is used to control the computational load. Its determination is based on the width and modal density of the critical frequency range. Ideally, all dominant modes within the critical frequency range should be covered, with a small number of additional boundary modes retained to improve robustness. The modal screening boundary margin is used to prevent the omission of boundary modes due to numerical errors. The margin size can be determined based on the natural fluctuation range of the modal frequencies during the healthy period. The solution convergence threshold and modal stability criterion are used to ensure that the extracted mode shapes can be used for energy calculations. The modal stability criterion can be achieved through the similarity of modes in adjacent periods. The threshold is determined by the similarity distribution during the healthy period and the quantile boundary is taken to ensure that normal fluctuations are not misjudged as modal instability. If the modal stability does not meet the requirements, the platform can improve the solution accuracy or increase the number of extracted modes to avoid positioning errors caused by missed mode extraction or mode swapping within the critical frequency range.

[0068] Secondly, the module calculates the modal strain energy of the elements based on the vibration mode shape and tangent stiffness matrix, and uses the local abrupt peak of the modal strain energy as a damage trigger signal to further calculate damage location indicators to determine the potential damage location. The calculation of the modal strain energy of the elements is performed on the platform side on an element-by-element basis. Specifically, for each mode shape falling within the critical frequency range, the platform first maps the components of the mode shape in the global degree of freedom space to the degree of freedom set of each element to obtain the displacement shape of the element in that mode; then, it combines the displacement shape with the corresponding element contribution in the tangent stiffness matrix to obtain the energy contribution value of the element in that mode; finally, it summarizes the energy contributions of the same element in multiple target modes to form the modal strain energy index of the element. To improve numerical stability, the platform performs normalization processing on the modal strain energy of the elements. The normalization scale benchmark can be the total amount or high quantile value of the modal strain energy of all elements in that period, so that the energy distribution under different periods and different working conditions is comparable. Key parameters in this stage include the participating mode set, mode weights, energy smoothing range, and local mutation threshold. The participating mode set is obtained by screening key frequency ranges and can be further prioritized according to the contribution of modes to the measurement point response. Mode weights are used to reflect the differences in damage sensitivity of different modes. The weights can be determined based on the energy concentration of each mode on known vulnerable sites in the health period history. If historical information is lacking, equal weights can be used and gradually calibrated during continuous monitoring. The energy smoothing range is used to suppress isolated energy spikes caused by discrete noise, and its determination is constrained by the grid scale and the minimum geometric mutation scale. The local mutation threshold is used to determine whether the modal strain energy of a certain unit constitutes a "local mutation peak". The determination of this threshold adopts a joint strategy of "neighborhood comparison and health baseline statistics". That is, the platform calculates the degree of difference between the energy level of each unit and its neighboring units, and statistically analyzes the normal distribution of this degree of difference in the health period data. The threshold is set outside the upper boundary of the normal distribution and a margin is added so that normal energy fluctuations will not trigger mutation judgment, while damage-induced energy concentration can trigger it. The neighborhood range is a necessary parameter for this thresholding strategy. The neighborhood range is determined by the grid adjacency relationship. It is preferable to cover several layers of adjacent cells that are directly connected to the cell to make the contrast background representative.

[0069] When a local abrupt peak in the modal strain energy of a unit is detected, the platform calculates damage location indicators based on the local abrupt peak to determine the potential damage location. The calculation process for the damage location indicators is as follows: the platform generates a peak intensity indicator, a peak relative background indicator, and a peak persistence indicator for each abrupt peak unit. The peak intensity indicator reflects the absolute level of the abrupt peak, the peak relative background indicator reflects the prominence of the abrupt peak relative to the surrounding background, and the peak persistence indicator reflects whether the abrupt change persists over several consecutive calculation cycles. Then, the above indicators are combined into a location indicator according to preset weights. The higher the location indicator, the more likely there is potential damage at that location. The weight parameters are determined based on the principle of "reducing false alarms and improving repairability": when the field noise is high, the persistence weight is increased so that short-term spikes are less likely to be judged as damage; when earlier warning is needed, the relative background weight is increased so that minor but structurally concentrated spikes can also be detected. To avoid output redundancy caused by multiple adjacent units in the same damage area having high location indicators simultaneously, the platform introduces a region merging parameter, merging peak units that are less than the merging radius into a potential damage area, and outputting the unit with the highest location indicator or its geometric center as the representative location of that area. The merging radius is determined by the grid cell size and component connection scale, ensuring that the output location corresponds to the actual inspectable structural details. Simultaneously, the platform sets a threshold for the location index output. Only when the location index exceeds this threshold is it reported as a potential damage location. The output threshold is also determined by the upper boundary of the healthy period location index distribution and includes a margin. It can be set in layers according to working conditions to maintain consistent judgment sensitivity under different load levels. If a potential damage location appears continuously over multiple cycles and the location index shows an upward trend, the platform can upgrade its risk level and make it a key focus of the continuous monitoring module's online reporting.

[0070] The continuous monitoring module, through the industrial internet platform, generates an online diagnostic report based on the identified potential damage locations and continuously inputs sensor data for the next cycle to maintain real-time monitoring. This includes acquiring potential damage locations and local abrupt peak values, establishing spatial mapping coordinates of potential damage locations in a 3D digital model, and calculating quantitative values ​​of damage severity. The spatial mapping coordinates and quantitative values ​​of damage severity are then filled into the diagnostic report template to generate a structured online diagnostic report. The structured online diagnostic report is parsed, and if the current structural health status label is non-failure, a standardized input vector sequence for the next time cycle is constructed. This standardized input vector sequence is then loaded into the finite element analysis engine to update boundary conditions, thereby maintaining real-time monitoring of the structural status.

[0071] In this implementation, the module first acquires the potential damage location and local mutation peak value, establishes the spatial mapping coordinates of the potential damage location in the 3D digital model, and calculates the quantitative value of the damage degree. Acquiring the potential damage location is not simply reading a single point, but rather reading the set of positioning results output by the damage identification module. This set includes at least the grid index information of the potential damage location, the corresponding component identifier, the positioning index value, and the local mutation peak value used to characterize the anomaly intensity. The platform converts the grid index information into spatial coordinates in the 3D digital model. The conversion process is completed by a "grid-to-model" mapping table. This mapping table is established during the system modeling phase and maintained synchronously with model version updates, ensuring that the same component can still be accurately located under different cycles and different model versions. The establishment process of the spatial mapping coordinates specifically includes: first, determining the coordinate system type used in the 3D digital model, preferably using a global coordinate system consistent with the global finite element model for cross-module consistency; then, using the geometric center coordinates of the node or element to which the potential damage location belongs as the basic coordinates; when the potential damage location is output in the form of a region, the platform aggregates the coordinates of multiple elements within the region. The aggregation method can use the center coordinates weighted by the positioning index, making the output coordinates closer to the strongest anomaly. The parameters involved in the above aggregation include the region merging radius and the coordinate aggregation weight. The region merging radius is determined by the grid cell size and the component connection scale, so that the aggregated coordinates still correspond to the actual inspectable location. The coordinate aggregation weight is preferably based on the positioning index or the local abrupt change peak, so that higher risk points contribute more to the final coordinates.

[0072] The damage severity quantification value is used to convert "local mutation peak values" into comparable and gradable risk quantification results. The calculation process is implemented on the platform side through a combination of standardization and grading. Specifically, the platform first performs benchmarking on the local mutation peak values. The benchmark comes from healthy period statistics or benchmark period model output. Under healthy conditions, the platform statistically analyzes the distribution of local mutation peak values ​​within the same component and the same critical frequency range to obtain the normal fluctuation range. Then, the local mutation peak value of the current period is compared with the normal fluctuation range to obtain the degree of exceedance, and this degree of exceedance is converted into a damage severity quantification value. To avoid incomparability in quantification due to differences in geometry and stress among different components, the platform normalizes the quantification values ​​to ensure they fall within a unified dimension. It can also introduce zoning correction coefficients based on component type or structural zone. These zoning correction coefficients are determined by the differences in peak levels across different zones in historical healthy data, used to eliminate inherent energy level differences between structural zones. The thresholds involved in this stage include the upper boundary threshold of normal fluctuations and grading thresholds. The upper boundary threshold for normal fluctuations is determined by the upper boundary of the peak distribution during the healthy period, plus a margin, to prevent fluctuations under normal operating conditions from being misjudged as damage. The grading threshold is used to map the quantified damage level to a risk level. The determination of the grading threshold is based on the operation and maintenance strategy and fault tolerance objectives, preferably including at least a concern threshold, an early warning threshold, and a severe threshold. The concern threshold is slightly higher than the upper boundary of normal fluctuations for early warning; the early warning threshold corresponds to the risk level requiring scheduled inspection; and the severe threshold corresponds to the risk level requiring immediate action or shutdown for inspection. If there are significant differences in site operating conditions, the grading threshold can be set in layers according to wind speed or load levels, ensuring that risk labels have consistent meanings under different operating conditions.

[0073] After completing the calculation of spatial mapping coordinates and damage degree quantification, the module fills the spatial mapping coordinates and damage degree quantification values ​​into the diagnostic report template, generating a structured online diagnostic report. The diagnostic report template is defined on the platform side using a fixed set of fields, which includes at least the monitoring batch identifier, site identifier, support identifier, model version identifier, operating condition label, spatial mapping coordinates of potential damage locations, corresponding component identifier, local mutation peak value, damage degree quantification value, risk level, structural health status label, recommended treatment actions, and a data quality summary. The filling process is implemented as follows: the platform writes each potential damage location as a record into the report body, sorting them by risk level, prioritizing the display of the highest risk locations; when the number of potential damage locations exceeds a preset limit, the platform retains the first few highest-risk records and merges the remaining records into a statistical summary to avoid excessively long reports affecting online readability. The maximum number of records is determined by the maintenance reading load and the platform's display capacity, ideally controlling the length without omitting high-risk points. The "Recommended Actions" in the report are generated by the rules engine. The rules engine reads the risk level, persistence indicators, and historical trends, and outputs inspection suggestions, encrypted monitoring suggestions, or alarm suggestions. The threshold for persistence is determined by the number of consecutive occurrences. The value of the number of consecutive occurrences is determined by a trade-off between false alarm tolerance and the timeliness requirements for risk handling, prioritizing not triggering strong actions for short-term spikes and triggering higher-level actions for persistent anomalies. The structured output format of the report is uniformly defined by the platform to ensure subsequent parsing and cross-system integration capabilities.

[0074] After generating a structured online diagnostic report, the module parses the report. If the current structural health status label is "non-failure," it constructs a standardized input vector sequence for the next time period. The report parsing is implemented as follows: the platform reads and verifies the report fields, confirming that the report version and field integrity meet the parsing rules; then it reads the structural health status label and enters the branch logic. The determination of the structural health status label is not subjective but is inferred by the platform based on risk level, quantitative damage degree, trend indicators, and consistency status. The determination process is as follows: when the highest risk level is below the warning threshold and the location result does not meet the persistence threshold, the health status label is set to "non-failure"; when the warning threshold is reached or the persistence threshold is met, the health status label can be upgraded to "needs attention" or "warning" status; when the severe threshold is reached or multiple high-risk points appear with a rapid upward trend, the health status label can be set to "failure risk" or "failure status." The thresholds involved here include trend determination thresholds and multi-point aggregation thresholds. The trend determination threshold is determined by the upper boundary of the incremental value of the damage degree quantification within a continuous period and is used to identify abnormal acceleration. The multi-point aggregation threshold is determined by the number of high-risk points within the same structural partition and is used to identify systemic problems or cascading degradation. If the health status label is a non-failure state, a standardized input vector sequence for the next period is constructed. The standardized input vector sequence is used to convert the sensor data, operating condition information, and necessary diagnostic context of the next period into an input set that the finite element analysis engine can directly consume. The construction process specifically includes: the platform obtains the raw sensor data of the next period from the data acquisition module and performs denoising, drift removal, missing data completion, and time alignment on multi-dimensional data such as wind speed, vibration, and temperature; then, it performs scale unification according to preset standardization rules to ensure that data of different dimensions fall into a unified numerical range. The parameters of the standardization rules come from health period statistics or engineering configurations, including the mean benchmark, fluctuation scale, and clipping boundary; when the data exceeds the clipping boundary, the platform clips it and marks it as abnormal to avoid extreme abnormal inputs that lead to unstable solutions. The standardized input vector sequence may also include load condition labels, load levels, model version identifiers, and a summary of risks from the previous period. This enables the finite element analysis engine to select the correct load template and parameter level when updating boundary conditions. The update frequency of standardized parameters is determined by seasonal variations and site operational stability. Ideally, updates should be periodic while ensuring stability to avoid standardization distortion caused by long-term drift.

[0075] Finally, the standardized input vector sequence is loaded into the finite element analysis engine to update the boundary conditions, maintaining real-time monitoring of the structural state. This loading process is implemented on the platform side via interface calls: the platform maps the standardized input vector sequence to the input ports of the finite element analysis engine according to preset fields. The input ports include load input ports and boundary condition input ports. The load input port is used to generate the equivalent load for the next cycle, and the boundary condition input port is used to update the constraint state and connection parameters. After loading, the finite element analysis engine selects the corresponding load template based on the input vector sequence and generates node or element load assignments. Then, it updates the boundary conditions and enters the solution process for the next cycle, thus realizing a closed loop of "sensor data input—model update—response calculation—damage identification—report output—next cycle input construction". To ensure the continuity of the closed loop, the platform sets input validity thresholds and engine loading success thresholds: the input validity threshold is determined by the missing proportion, outlier proportion, and the decidability of the working condition label. When it is below the threshold, automatic updates are paused and a prompt is displayed that the data needs to be reviewed. The engine loading success threshold is determined by the interface return status and key field verification. If loading fails, a retry or rollback to the previous reliable boundary conditions is triggered. The values ​​of the above thresholds are determined based on the platform's operational stability requirements and false alarm tolerance, so that the system can run continuously and avoid outputting unreliable results when data or interfaces are abnormal.

[0076] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A photovoltaic bracket fault diagnosis system based on an industrial internet platform, characterized in that, include: The data acquisition module obtains real-time wind speed and vibration data of the photovoltaic support through sensors connected to the industrial internet platform, and obtains the initial dynamic response characteristics of the structure in the key frequency range through modal analysis. The modeling and identification module, within the industrial internet platform, constructs a global model of the support structure using the finite element method based on dynamic response characteristics, and identifies key areas prone to damage. The threshold determination module acquires detailed geometric parameters of the local complex structure of the identified key parts, and determines the stress distribution threshold of these parts through a neural network algorithm in the industrial internet platform. The model simplification module extracts the vibration mode contribution of the secondary structure from the global model if the determined stress distribution threshold exceeds the preset threshold, and uses equivalent simplification technology to reduce the computational complexity of the secondary structure. The consistency judgment module updates the dynamic response of the global model with the reduced computational complexity and determines whether the updated response is consistent with the initial dynamic response characteristics. If the judgment is consistent, the response calculation module will obtain real-time load data and input it into the updated model to obtain the current stress and deformation distribution of the support. The damage identification module uses modal analysis to recalculate the vibration modes within the critical frequency range based on the obtained stress and deformation distribution, and determines the location of potential damage.

2. The photovoltaic bracket fault diagnosis system based on an industrial internet platform according to claim 1, characterized in that: The data acquisition module acquires real-time wind speed and vibration data of the photovoltaic support through sensors connected to the industrial internet platform, and obtains the initial dynamic response characteristics of the structure within the key frequency range through modal analysis, including: Obtain a synchronized time-domain signal sequence containing wind load input and structural vibration output. The synchronized time-domain signal sequence is generated by synchronizing wind speed time history values ​​and vibration acceleration values. A cross-power spectral density matrix is ​​generated based on the synchronous time-domain signal sequence. The cross-power spectral density matrix is ​​then obtained by preserving the complex data within the preset analysis bandwidth through a fast Fourier transform. Singular value decomposition is performed on the cross power spectral density matrix to generate a structural modal data set, which includes natural frequencies, mode shape vectors, and damping ratio values. By mapping the structural modal data set to a preset frequency response function model, the initial dynamic response characteristics of the photovoltaic support within the key frequency range are obtained.

3. The photovoltaic bracket fault diagnosis system based on an industrial internet platform according to claim 1, characterized in that: The modeling and identification module, within the industrial internet platform, constructs a global model of the support using the finite element method based on dynamic response characteristics, and identifies key components of the damage-prone area, including: The real-time vibration signal of the support transmitted by the industrial internet platform is obtained, and the time-frequency characteristics are obtained by processing the real-time vibration signal using short-time Fourier transform. Finite element mesh elements are generated based on the time-frequency characteristic mapping of the support geometry and topology. The global stiffness matrix is ​​assembled using the finite element mesh elements and the nodal displacements are obtained by solving the solution. The strain energy density is calculated based on the nodal displacement. If the strain energy density exceeds a preset threshold, a damage factor is generated. Based on the spatial distribution gradient of damage factors, damage-prone areas are delineated, and extreme points are located within these areas to identify key components of the global stent model.

4. The photovoltaic bracket fault diagnosis system based on an industrial internet platform according to claim 1, characterized in that: The threshold determination module, for the identified key parts, obtains detailed geometric parameters of their local complex structures, and determines the stress distribution thresholds of these parts in the industrial internet platform using a neural network algorithm, including: Obtain detailed geometric parameters of key components in the industrial internet platform, including curvature characteristics and spatial coordinates after reconstructing the local structural surface; Detailed geometric parameters are mapped to high-dimensional feature vectors and input into a pre-trained convolutional neural network model, which then simulates the mechanical response characteristics of the local structure. The mechanical response characteristics are analyzed to generate a predicted stress distribution map. The stress distribution threshold of key parts is determined based on the peak region and gradient change of the predicted stress distribution map.

5. A photovoltaic bracket fault diagnosis system based on an industrial internet platform according to claim 1, characterized in that: If the determined stress distribution threshold exceeds a preset threshold, the model simplification module extracts the vibration mode contributions of secondary structures from the global model and uses equivalent simplification techniques to reduce the computational complexity of these structures, including: Acquire nodal stress distribution data; if the maximum nodal stress value is greater than a preset limit, then identify the secondary structural entity. The local stiffness matrix and local mass matrix of the secondary structural entities are extracted to calculate the vibration mode vector, and the mode transformation matrix is ​​constructed accordingly. The local stiffness matrix and local mass matrix are orthogonally transformed using the modal transformation matrix to generate equivalent stiffness matrix and equivalent mass matrix; The equivalent stiffness matrix and equivalent mass matrix are assembled back into the global finite element model to reduce the computational complexity of the secondary structure.

6. The photovoltaic bracket fault diagnosis system based on an industrial internet platform according to claim 1, characterized in that: The consistency judgment module updates the dynamic response of the global model with reduced computational complexity and determines whether the updated response is consistent with the initial dynamic response characteristics, including: We acquire global model data to extract initial dynamic response features and construct a sparse matrix representing the model structure to reduce computational complexity. The updated global model is generated by iterating in a low-dimensional space based on sparse matrices and incremental business data. The test stimulus signal is input into the updated global model to output the updated dynamic response; Calculate the response deviation between the updated dynamic response and the initial dynamic response characteristics. If the response deviation is less than the characteristic drift threshold, the updated response is determined to be consistent with the initial dynamic response characteristics.

7. The photovoltaic bracket fault diagnosis system based on an industrial internet platform according to claim 1, characterized in that: If the response calculation module determines that the values ​​match, it obtains real-time load data and inputs it into the updated model to obtain the current stress and deformation distribution of the support, including: If the monitored stent operating status matches the preset consistency status, then collect the real-time load data stream; The real-time load data stream is mapped to the geometric topology nodes of the finite element model to update the boundary conditions of the finite element model. Assemble the global stiffness matrix based on the boundary conditions and support material properties, and solve the global stiffness matrix to obtain the nodal displacement vectors. The strain field data is calculated based on the nodal displacement vectors, thereby obtaining the current stress distribution and deformation distribution of the support.

8. The photovoltaic bracket fault diagnosis system based on an industrial internet platform according to claim 1, characterized in that: The damage identification module, based on the obtained stress and deformation distribution, recalculates the vibration modes within the critical frequency range using modal analysis, and determines the potential damage locations, including: Acquire full-field stress distribution data and deformation distribution data, update the geometric configuration using deformation distribution data, and construct a tangent stiffness matrix that includes stress stiffening effect by combining full-field stress distribution data; Eigenvectors located within the key frequency range are extracted from the tangent stiffness matrix as vibration mode shapes. The modal strain energy of the element is calculated based on the vibration mode shape and tangent stiffness matrix. If a local abrupt peak occurs in the modal strain energy of the element, the damage location index is calculated based on the local abrupt peak to determine the potential damage location.

9. A photovoltaic bracket fault diagnosis system based on an industrial internet platform according to claim 1, characterized in that, It also includes a continuous monitoring module, which generates online diagnostic reports based on the identified potential damage locations through an industrial internet platform, and cyclically inputs sensor data for the next cycle to maintain real-time monitoring. Specifically, it includes: Obtain the potential damage location and local mutation peak value, establish the spatial mapping coordinates of the potential damage location in the three-dimensional digital model, and calculate the quantitative value of the damage degree. The spatial mapping coordinates and the quantified values ​​of the damage level are filled into the diagnostic report template to generate a structured online diagnostic report.

10. A photovoltaic bracket fault diagnosis system based on an industrial internet platform according to claim 9, characterized in that: The continuous monitoring module, through the industrial internet platform, generates an online diagnostic report based on the determined potential damage location, and cyclically inputs sensor data for the next cycle to maintain real-time monitoring. It also includes: Parse the structured online diagnostic report. If the current structural health status label is non-failure, construct a standardized input vector sequence for the next time period. The standardized input vector sequence is loaded into the finite element analysis engine to update the boundary conditions in order to maintain real-time monitoring of the structural state.

Images