Library health monitoring and early warning system based on heterogeneous multi-source data fusion

Abstract

The invention relates to the technical field of structural health monitoring, and particularly discloses a library body health monitoring and early warning system based on heterogeneous multi-source data fusion, and the system obtains library wall strain, environment wind load and three-dimensional deformation data in real time through a multi-source data collection module; high-risk, transitional and stable areas are divided by utilizing a space gridding layered mapping module and combining library body structure features, and risk sensitivity is improved through dynamic weight adjustment; the cross-frequency-domain association feature extraction module constructs a space-time-frequency domain feature matrix, fuses strain spectrum, wind load harmonic and deformation curvature data, and verifies feature credibility in combination with an adversarial generative network; the resonance risk dynamic identification module judges a resonance hot spot based on energy transfer path analysis and a modal cloud picture standing wave node line; and the closed-loop early warning module realizes graded early warning and strategy adaptive adjustment through encryption scanning, digital twin model iteration and dynamic threshold optimization.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of structural health monitoring, in particular to a library body health monitoring and early warning system based on heterogeneous multi-source data fusion. BACKGROUND

[0002] The fly ash library is a typical large powder library and an important facility and equipment in the thermal power industry, which is used for storing fly ash in the production process of the thermal power plant. The operation and management level of the fly ash library is of great significance to the safe production of the thermal power plant, the protection of the surrounding ecological environment, and the improvement of the solid waste recycling rate. Therefore, the research and development of intelligent equipment for the fly ash library will bring significant ecological, economic and social benefits to the thermal power industry, and will be conducive to the transformation and upgrading of the industry and the improvement of quality and efficiency.

[0003] The operation and maintenance of the fly ash library include library body structure health monitoring, fly ash weight and material level monitoring, fly ash hardening removal, fly ash metering and operation management, etc., which has a very important influence on the safe production of the thermal power plant. The production accidents caused by fly ash storage, forced shutdown of the boiler and stop of power generation due to fly ash library failure occur from time to time. Due to the poor internal working environment of the fly ash library (closed space, high dust, strong corrosion), the current operation and maintenance automation level of the fly ash library is low, and the operation is mainly manual, which has the problems of great safety hidden danger, long downtime, high maintenance cost, etc. In addition, the fly ash library also has the safety hidden danger of structural instability.

[0004] The prior art has the following disadvantages: When the library body is excited by strong wind, the vibration string sensor (10Hz sampling rate) captures the low-frequency strain main frequency (such as 2Hz), and the three-dimensional scanning robot (0.017Hz sampling rate) acquires the transient deformation point cloud data at the vibration peak moment, which cannot be accurately aligned with the high-frequency fluctuation signal (such as 4Hz) of the environmental wind speed instrument in the time-frequency domain. When the system forcibly associates the non-synchronous data, the random vibration harmonics will be misjudged as structural resonance due to phase shift and harmonic aliasing effect, resulting in false triggering of the safety protection mechanism. SUMMARY

[0005] The purpose of the present application is to provide a library body health monitoring and early warning system based on heterogeneous multi-source data fusion to solve the problems in the above background.

[0006] The purpose of the present application can be achieved by the following technical solutions: The library body health monitoring and early warning system based on heterogeneous multi-source data fusion comprises: A multi-source data acquisition module, which is used for real-time acquisition of library wall strain data, environmental wind load data and library body three-dimensional deformation data; A spatial gridding layered mapping module, which divides high-risk areas, transition areas and stable areas according to the structural characteristics of the library body, constructs a multi-dimensional spatial grid matrix, and maps the sensor positions to the grid nodes; A cross-frequency domain correlation feature extraction module, which fuses strain spectrum, wind load spectrum and deformation curvature data into a time-space-frequency domain feature matrix by dimensionality elevation, and extracts the coupling features of spatial position and frequency by constraint optimization decomposition; A resonance risk dynamic identification module, which locates the spatial-frequency energy focusing hotspots based on the energy transmission path analysis of the coupling features, and verifies the resonance risk in combination with the weak point parameters of the library body structure; A closed-loop early warning and feedback module, which triggers three-dimensional deformation encryption scanning and finite element model checking when the hotspot area energy intensity exceeds the dynamic threshold, outputs graded warning and dynamically adjusts the monitoring strategy.

[0007] As a further scheme of the present application, the construction of the multi-dimensional spatial grid matrix specifically includes the following steps: The circumferential weld distribution position and historical damage data of the library body welding structure are extracted, and the library wall is divided into high-risk areas, transition areas and stable areas along the height direction in combination with the library wall stress finite element cloud map; The high-risk area covers the library bottom ring foundation to the maximum bending moment acting height of the library wall, the transition area covers the bending moment attenuation zone to the structural deformation inflection point height, and the stable area covers the area above the inflection point; The high-risk area is equally divided into 12 sector-shaped ring zones along the circumferential direction, the transition area is equally divided into 8 sector-shaped ring zones, and the stable area is equally divided into 4 sector-shaped ring zones with the library body center axis as the reference; The high-risk area is divided into 6 layer domains according to the stress gradient, the transition area is divided into 4 layer domains, and the stable area is divided into 2 layer domains along the vertical direction, forming a multi-dimensional spatial grid matrix; The vibration string sensor coordinates are matched with the three-dimensional deformation scanning point cloud, each sensor is associated with the intersection grid node of the corresponding ring zone and layer domain through the nearest neighbor topology mapping, and the high-risk area node is given a 3 times weight coefficient, and the transition area is given a 2 times weight coefficient.

[0008] As a further scheme of the present application, the high-risk areas, transition areas and stable areas are divided according to the structural characteristics of the library body, a multi-dimensional spatial grid matrix is constructed, and the sensor positions are mapped to the grid nodes, which at least further include the following steps: Real-time micro-deformation data of the weld heat-affected zone are obtained, and when the deformation variance in a single ring zone exceeds the threshold, the corresponding ring zone is subdivided into 16 sub-ring zones, and virtual monitoring nodes are generated at the junctions of the sub-ring zones; Based on the thickness distribution of the tank wall steel plate, the environmental corrosion rate and the historical load spectrum, the structural risk entropy values of each grid node are calculated; The entropy value, real-time strain gradient and weld distance are input into the fuzzy decision maker, and a node weight correction factor is dynamically output, and the correction factor range is set to 1.0-5.0; When a certain grid node sensor fails, the adjacent node data is called and the corresponding node state is reconstructed through radial basis interpolation, and a three-dimensional scanning robot is driven to perform encryption scanning on the ring belt where the node is located, so that the missing signal is replaced by laser point cloud data.

[0009] As a further scheme of the application, the construction process of the space-time-frequency domain feature matrix is: A correlation constraint model of strain main frequency components and environmental wind load harmonic components is established, and the strain main frequency amplitude and the wind load harmonic amplitude in the same ring belt position are forced to satisfy a monotonically increasing relationship; The strain spectrum, wind load spectrum and deformation curvature data are reorganized into a space-time-frequency domain feature matrix according to time slices, and a non-negative constraint decomposition is used to extract a coupling factor matrix of spatial grid nodes and frequencies; Based on the coupling factor matrix, the energy focusing coefficients of each ring belt region are calculated, and when the focusing coefficient of a certain ring belt in a specific frequency band exceeds a reference value and the deformation curvature suddenly changes synchronously, it is marked as a resonance risk hotspot.

[0010] As a further scheme of the application, the strain spectrum, wind load spectrum and deformation curvature data are dimensionally fused into a space-time-frequency domain feature matrix, and the coupling characteristics of spatial positions and frequencies are extracted through constraint optimization decomposition, which at least includes the following steps: The marginal spectrum entropy of the tank wall vibration signal is analyzed in real time, and when the entropy value is lower than a threshold value, the wind load spectrum analysis frequency band is contracted to within the strain main frequency ± 0.5 octave; A strain-deformation time sequence generative adversarial network is constructed, the generator generates virtual high-frequency deformation data under the condition of low-frequency deformation curvature, and the discriminator introduces a tank body vibration differential equation to constrain the network output; The Hausdorff distance between the generated deformation data and the measured deformation data is calculated, and the singular value decay rate of the coupling factor matrix is combined to output the reliability weight of the space-frequency domain feature; The energy focusing coefficient, deformation mutation gradient and feature reliability weight are fused, the resonance risk entropy value is calculated through the entropy weight method, and when the dynamic threshold value is exceeded, the encryption scanning verification is triggered.

[0011] As a further scheme of the application, the energy transmission path analysis based on the coupling characteristics locates the space-frequency energy focusing hotspot, including the following steps: Based on the stiffness attenuation coefficient of the spatial grid node and the ring weld continuity parameter, a directed topological network of tank body vibration energy transmission is constructed; Wherein the node stiffness attenuation coefficient is calculated by the thickness of the steel plate and the corrosion rate, and the weld continuity is evaluated by fitting the weld displacement gradient with the three-dimensional deformation data; Calculate the resonance energy flow density of each grid node in the target frequency band, which is the product of the strain spectrum amplitude, the wind load power spectrum density and the node transfer efficiency; When the energy flow density of three consecutive nodes in the same ring band exceeds 2 times the reference value, mark the corresponding ring band as an energy focusing hotspot; Call the natural frequency of the heat-affected zone of the weld in the finite element model to verify whether the hotspot frequency band falls within the interval of ±15% of the natural frequency; At the same time, detect whether the deformation curvature change rate of the hotspot ring band increases suddenly to more than 3 times the value at the previous time.

[0012] As a further scheme of the present application, the combination of library body structure weak point parameters verifies the resonance risk, comprising the following steps: Extract the instantaneous phase of the vibration string sensor in the hotspot ring band. When the phase coherence of 80% of the sensors in the ring band exceeds 85% and lasts for more than 5 seconds, it is determined that the structure is resonating; Fuse the energy focusing coefficient, the phase coherence and the deformation mutation gradient, and calculate the correlation entropy value of the three through nonlinear Granger causality test; When the correlation entropy value is lower than the threshold value and the phase coherence lasts, output the resonance risk level; Drive the three-dimensional scanning robot to perform millisecond-level laser vibration measurement on the hotspot ring band, and reconstruct the library wall vibration mode cloud diagram; If the cloud diagram shows that the standing wave nodal line passes through the ring weld, a first-level warning is triggered and a weld fatigue life evaluation report is generated.

[0013] As a further scheme of the present application, the output hierarchical warning and dynamic adjustment of the monitoring strategy specifically comprises: When the energy flow density of the energy focusing hotspot exceeds the dynamic threshold value; First-level response: start three-dimensional deformation encryption scanning, and the scanning frequency is increased to 5 times the regular frequency; Second-level response: call the finite element model to check the resonance frequency matching degree, and if the deviation exceeds 15%, activate the weld damage prediction; Third-level response: when the energy flow density does not decay for 3 minutes, drive the board consolidation cleaning robot to remove the material pressure in the hotspot area; Calculate the Hausdorff distance between the deformation curvature obtained by the encryption scanning and the finite element predicted deformation; If the distance value is less than the deformation tolerance threshold, the warning level is downgraded; If the distance value increases and the weld damage probability exceeds 60%, output a second-level warning; Train a hidden Markov model based on historical warning data; when the heat spot area energy flow density decline rate is lower than a preset value; automatically prolong the encryption scanning duration to twice the original plan and shrink the robot operation radius to within 3 meters.

[0014] As a further scheme of the present application: the output graded early warning and dynamic adjustment of the monitoring strategy, comprising the following steps: input the encryption scanning data, robot operation record and energy flow density change rate into the digital twin, dynamically correct the boundary constraint conditions and material constitutive parameters of the finite element model; when the secondary early warning lasts for 10 minutes without being eliminated; synchronously start the three-dimensional topography acquisition system to reconstruct the library wall topology model; and call the concretion cleaning robot to perform preventive vibration on the hot spot ring; statistically determine the structure relaxation coefficient after each early warning is eliminated, which is defined as the ratio of the energy flow density decay rate to the deformation recovery rate; when the relaxation coefficient is greater than 1.5 times the reference value for three times in succession; the dynamic threshold is raised by 20% and a structure stiffness strengthening scheme is generated.

[0015] The present application has the following beneficial effects: (1) The present application fuses heterogeneous multi-source data (library wall strain, environmental wind load, three-dimensional deformation, etc.), combines dynamic spatial gridding modeling and cross-frequency domain feature extraction technology, constructs a time-space-frequency domain feature matrix, accurately identifies resonance risk hot spots. At the same time, the closed-loop early warning module realizes real-time response and strategy adjustment to high-risk areas through encryption scanning, digital twin model iteration and dynamic threshold optimization, reduces the false alarm rate and improves the timeliness of early warning, ensures the efficient management and control of the safety of the library structure.

[0016] (2) Through dynamic weight correction of spatial grid nodes (based on fuzzy decision maker and structure risk entropy value), sensor fault tolerance (radial basis interpolation and laser scanning compensation) and strategy optimization driven by hidden Markov model, the system can maintain stable operation under complex working conditions. In addition, the collaborative updating mechanism of digital twin and finite element model enables the system to have continuous learning ability, adapt to environmental changes and structure degradation trend, prolong the service life of equipment and reduce maintenance cost. BRIEF DESCRIPTION OF DRAWINGS

[0017] The present application will be further described below in conjunction with the drawings.

[0018] Figure 1 is a block diagram of the system of the present application. DETAILED DESCRIPTION

[0019] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0020] See also Figure 1 As shown, the present invention is a warehouse health monitoring and early warning system based on heterogeneous multi-source data fusion, including: A multi-source data acquisition module, which is used to obtain real-time data on the strain of the storage wall, environmental wind load data, and three-dimensional deformation data of the storage body; A spatial grid layered mapping module, which divides the reservoir into high-risk areas, transition areas, and stable areas based on the reservoir structure characteristics, constructs a multi-dimensional spatial grid matrix, and maps sensor locations to grid nodes; A cross-frequency domain correlation feature extraction module, which performs dimension-upgrading and fusion of strain spectrum, wind load spectrum, and deformation curvature data into a time-space-frequency domain feature matrix, and extracts the coupling characteristics of spatial position and frequency through constrained optimization decomposition; A resonance risk dynamic identification module, which locates the energy focus hotspots in the spatial-frequency domain based on the energy transfer path analysis of the coupling characteristics and verifies the resonance risk in combination with the weak point parameters of the warehouse structure; A closed-loop early warning and feedback module triggers three-dimensional deformation encryption scanning and finite element model verification when the energy intensity of the hot spot area exceeds a dynamic threshold, outputs a graded early warning and dynamically adjusts the monitoring strategy.

[0021] In the multi-source data acquisition module, 172 vibrating-wire strain sensors are installed on the silo wall surface, including 112 horizontally and 60 vertically. The sensors are arranged according to a gradient distribution principle: six rings of 12 sensors are equally spaced in each ring from 0 to 7.2 meters above the silo wall; four rings of 8 sensors are installed in the middle from 7.2 to 15 meters above the silo wall; and two rings of 4 sensors are installed in each ring from the top above 15 meters above the silo wall. All sensors are connected via armored cables to six 32-channel vibrating-wire acquisition boxes, housed in waterproof protective casings and secured to welded clips 1.5 meters above the silo wall. The data sampling rate is 10 Hz, and sensor temperature is recorded simultaneously during acquisition for strain temperature compensation. The compensation coefficient is determined by laboratory calibration to be 0.5 microstrain per degree Celsius.

[0022] A three-dimensional ultrasonic anemometer was installed at the center of the reservoir roof, measuring 3 meters above the highest point. The anemometer has a sampling rate of 20 Hz, a measurement range of 0-60 m / s, and an accuracy of ±0.1 m / s. To eliminate turbulent interference from the reservoir roof, the anemometer is equipped with a fairing and a dynamic yaw correction mechanism: if the wind direction changes by more than 30 degrees within 10 seconds, the motor automatically activates to rotate the anemometer into the wind. Wind load spectrum analysis uses a Hanning window-weighted fast Fourier transform with a window length of 1024 sampling points and a 50% overlap.

[0023] A high-precision track system is laid circumferentially along the warehouse wall, equipped with a laser 3D scanning robot. The total track length is 94.2 meters (based on a 30-meter warehouse diameter), and the robot's positioning accuracy is ±0.1 mm. Scanning uses line laser triangulation, forming a dot array with a 0.5 mm spacing on the warehouse wall surface. A single full warehouse scan takes 60 seconds. The deformation curvature calculation process is as follows: first, the local surface of the warehouse wall is extracted from the point cloud data, and the surface equation is fitted using the moving least squares method. The maximum eigenvalue of the curvature tensor is then calculated as the deformation curvature indicator. Scans are performed once an hour in normal monitoring mode and once a minute in early warning mode.

[0024] A three-dimensional coordinate system was established with the center of the reservoir floor as the origin, with the Z axis pointing vertically upward. All sensor clocks were synchronized via a GPS timing module, with a time error of less than 1 millisecond. The spatial registration method employed involved pre-setting 12 reference spheres on the reservoir wall. Upon each startup, the 3D scanning robot first scanned the target sphere coordinates and mapped the vibrating wire sensor positions to the target sphere coordinate system. The position mapping residual error was kept within 2 mm. Ambient wind load data was aligned with strain data using timestamps.

[0025] Strain data preprocessing includes third-order Butterworth low-pass filtering (cutoff frequency 5Hz) and temperature drift compensation. Wind load data preprocessing uses median filtering to eliminate impulse noise and converts the wind pressure distribution to the reservoir coordinate system (based on a pre-calibrated wind pressure coefficient matrix using computational fluid dynamics). Three-dimensional deformation data is aligned with the historical point cloud using an iterative closest point algorithm to extract the deformation displacement field. The processed data is stored in a time-sliced ​​structure, with each slice containing four types of data: timestamp, spatial grid code, strain spectrum amplitude, wind load power spectrum density, and deformation curvature.

[0026] Automatic diagnostics are performed every 24 hours. The vibrating wire sensor measures the string's damping coefficient using an excitation coil, flagging a fault if the damping changes by more than 15% from the initial value. The anemometer uses acoustic time-of-flight cross-validation, triggering a calibration procedure if the deviation exceeds 5%. The scanning robot verifies its accuracy by measuring the diameter of a standard sphere, automatically adjusting its optical calibration parameters if the error exceeds 0.2 mm. Diagnostic results are updated in real time to the monitoring database.

[0027] When the wind speed exceeds 20 m / s, the sampling rate of the vibrating wire is increased to 20 Hz, and the anti-wind stabilization mode of the scanning robot is started (hydraulic leg lock track). When rain and snow weather is detected, the vibrating wire sensor protection cover starts electric heating and dehumidification, and the anemometer enables ultrasonic self-cleaning function. When the ambient temperature is lower than-10℃, the temperature control module in the acquisition box maintains the temperature in the box at 5±2℃.

[0028] In the spatial grid layer mapping module, according to the circumferential weld distribution position of the warehouse body welding structure and the historical damage detection report, combined with the stress finite element analysis result of the warehouse wall, the warehouse wall is divided into high-risk area, transition area and stable area along the height direction. The division range of the high-risk area is from the surface of the ring foundation at the bottom of the warehouse to the maximum bending moment acting height of the warehouse wall. The height is determined by finite element analysis, and the typical value is 25%-30% of the total height of the warehouse body. The transition area covers the maximum bending moment acting height to the structure deformation inflection point height. The inflection point is determined by long-term deformation monitoring data, which shows the position where the curvature change rate significantly decreases. The stable area is above the deformation inflection point to the top of the warehouse. In actual engineering, for a 30-meter ash warehouse, the high-risk area is usually set to 0-7.2 meters, the transition area is 7.2-15 meters, and the stable area is above 15 meters.

[0029] Taking the center axis of the warehouse body as the reference, the high-risk area is equally divided into 12 sector rings along the circumferential direction, each corresponding to a 30-degree central angle; the transition area is equally divided into 8 sector rings, each corresponding to a 45-degree central angle; and the stable area is equally divided into 4 sector rings, each corresponding to a 90-degree central angle. When divided vertically, the high-risk area is divided into 6 layers according to the stress gradient, with each layer being about 1.2 meters high; the transition area is divided into 4 layers, with each layer being about 1.95 meters high; and the stable area is divided into 2 layers, with each layer being about 7.5 meters high. Thus, a multi-dimensional spatial grid matrix composed of ring bands and layers is formed, in which the high-risk area contains 72 grid nodes (12 ring bands x 6 layers), the transition area contains 32 nodes (8 ring bands x 4 layers), and the stable area contains 8 nodes (4 ring bands x 2 layers).

[0030] The installation coordinates of the vibrating wire sensors are registered with the warehouse wall point cloud data obtained by the three-dimensional laser scanning, and each sensor is positioned to the corresponding grid node through the nearest neighbor topology mapping algorithm. During the mapping process, for the sensors near the boundaries of the ring bands or layer domains, the attribution determination is made according to the distance ratio between their actual positions and the centers of the nodes. After the mapping is completed, the weight coefficients are assigned according to the risk levels of the areas: the weight of the grid nodes in the high-risk area is set to 3.0, the weight of the nodes in the transition area is set to 2.0, and the weight of the nodes in the stable area remains the default weight 1.0. The weight coefficient will directly affect the contribution calculation in the subsequent data analysis.

[0031] Real-time micro-deformation data of the weld heat-affected zone is monitored. When the system detects that the deformation variance in a certain ring belt exceeds the set threshold (usually 2 times the standard deviation of the average deformation) for 3 consecutive times, the ring belt is automatically subdivided into 16 sub-rings, each corresponding to a 22.5-degree central angle. Virtual monitoring nodes are generated at the junction of the sub-rings, and the initial data of the virtual monitoring nodes is generated by bilinear interpolation of the data of the adjacent actual sensors. The subdivided grid improves the monitoring resolution of the local area, and is particularly suitable for abnormal working conditions such as weld cracking or accelerated corrosion.

[0032] For each grid node, the structural risk entropy value is calculated by integrating the measured thickness of the library wall steel plate, the monitored data of the environmental corrosion rate, and the analysis results of the historical load spectrum. The thickness of the steel plate is obtained from the construction acceptance report, the corrosion rate is obtained by back calculation based on the surface potential detection of the library wall, and the historical load spectrum is obtained from long-term strain monitoring statistics. The entropy value is calculated using the information entropy theory, which reflects the uncertainty of the structural state of the node location. The higher the entropy value, the greater the risk. The calculation result is normalized to a relative value in the range of 0-1, which is used for subsequent dynamic adjustment of the weight.

[0033] The structural risk entropy value, real-time strain gradient (ratio of strain difference of adjacent nodes to distance), and distance to the nearest weld are input into the fuzzy decision system. The system has five levels of fuzzy rules, such as "if the entropy value is high and the strain gradient is large, then the weight is greatly increased". After fuzzy reasoning and defuzzification, the dynamic weight correction factor of each node is output, with a correction range limited to 1.0-5.0. This process is executed every 10 minutes to ensure that the weight distribution responds to changes in the structural state in a timely manner.

[0034] When a sensor corresponding to a certain grid node fails, the real-time monitoring data of the adjacent 8 nodes is first called to reconstruct the state data of the failed node using a radial basis function interpolation algorithm. During the interpolation process, the interpolation weight is allocated according to the spatial distance and weight coefficient of each adjacent node to the failed node. At the same time, a three-dimensional scanning robot is controlled to perform intensive scanning of the ring belt where the failed node is located, with the scanning frequency increased to 3 times that of the normal mode. After curvature calculation on the obtained high-density point cloud data, the output of the failed sensor is replaced to form a fault-tolerant mechanism with multiple sources of data complementing each other.

[0035] The real-time state of the spatial grid matrix is displayed through a three-dimensional visualization interface, in which different ring belts and layers are rendered with gradient colors, and the color depth corresponds to the risk level. Manual interaction is supported to adjust the grid parameters, including temporarily adding virtual nodes and manually modifying weight coefficients. All adjustments are recorded in the database for subsequent analysis of the impact of grid optimization on monitoring effectiveness.

[0036] In the cross-frequency domain correlation feature extraction module, based on the dynamic characteristics of the library body structure, a physical correlation model of the strain main frequency component and the environmental wind load harmonic component is established. For the position of each spatial grid node in the ring band, the strain main frequency amplitude and the wind load harmonic amplitude are forced to satisfy the monotonic increasing relationship. In specific implementation, the strain main frequency (usually the peak frequency in the range of 0.5-5 Hz) of each ring band is obtained through the vibration string sensor, and the corresponding wind load harmonic component (usually an integer multiple of the strain main frequency) recorded by the anemograph is extracted. The system automatically checks the amplitude variation trend of the two, and if the strain amplitude of a certain ring band increases while the wind load amplitude decreases, which is an abnormal situation, it is determined that the data of the ring band is abnormal and the review program is started. The constraint model effectively avoids false correlation of non-wind-induced vibration.

[0037] The preprocessed strain spectrum, wind load spectrum and deformation curvature data are reorganized according to 5-minute time slices. In each slice, the strain spectrum takes the amplitude spectrum of the 0.1-10 Hz frequency band at each spatial grid node, the wind load spectrum takes the power spectral density of the 0.5-20 Hz frequency band, and the deformation curvature takes the rate of change in the time period. The first dimension of the three-dimensional feature matrix corresponds to the spatial grid node number, the second dimension is the frequency component, and the third dimension contains three types of data channels. When filling the matrix, the data of the frequency points not directly measured is smoothed by cubic spline interpolation to ensure the integrity of the matrix.

[0038] Non-negative constraint decomposition is performed on the space-time-frequency domain feature matrix to extract the spatial grid node factor matrix and the frequency factor matrix. Special constraint conditions are set during the decomposition process: nodes in the same ring band share the fundamental frequency component in the frequency factor matrix. After solving by the iterative optimization algorithm, the core tensor reflecting the space-frequency coupling characteristics is obtained. Based on the tensor, the energy focusing coefficient of each ring band is calculated. The calculation method is: the coupling strength in a certain frequency band (such as 2-4 Hz) is spatially integrated in the ring band, and then divided by the total area of the ring band. When the energy focusing coefficient of a certain ring band exceeds 1.8 times of the historical statistical benchmark value, and the deformation curvature rate of change in the same period exceeds 3 times of the average value, it is marked as a potential resonance risk hotspot.

[0039] The marginal spectrum entropy of the library wall vibration signal is calculated in real time, which is used to quantify the concentration degree of vibration energy distribution. When the spectrum entropy value is lower than the set threshold value (indicating that the energy is concentrated in a few frequency bands), the wind load spectrum analysis bandwidth is automatically contracted: taking the current strain main frequency as the center, only the wind load components within ±0.5 octave range are retained to participate in the subsequent analysis. For example, when the strain main frequency is detected as 2 Hz, the wind load analysis frequency band is limited to 1.5-3 Hz. This mechanism effectively suppresses the interference of broadband noise and improves the signal-to-noise ratio of feature extraction. The degree of frequency band contraction is dynamically adjusted according to the spectrum entropy value, forming an adaptive analysis window.

[0040] An adversarial neural network is constructed, which contains a generator and a discriminator. The generator takes the low-frequency deformation curvature (from regular scans with 1 per hour) and strain spectrum as input, and outputs the predicted high-frequency deformation field (equivalent to virtual scan data with 6 per minute). The discriminator not only judges the authenticity of the data, but also introduces the library body vibration control equation as a physical constraint: check whether the generated data satisfies the vibration modal relationship determined by the structural stiffness matrix and mass matrix. During network training, a transfer learning strategy is adopted: first pre-train on small-scale calibration test data, and then gradually adapt to the actual ash library characteristics through online learning. The generated virtual data is labeled with a confidence index for subsequent fusion.

[0041] The Hausdorff distance between the calculated deformation data and the measured deformation data is calculated, which reflects the maximum difference in spatial distribution. At the same time, the singular value decay rate of the coupling factor matrix is analyzed to evaluate the stability of feature extraction. These two indicators are input into the fuzzy reasoning system to output the credibility weight of the spatial-frequency domain features, with a weight range of 0-1. For feature components with a credibility less than 0.6, the system automatically triggers data reacquisition or decomposition parameter adjustment processes to ensure the reliability of the analysis results.

[0042] The entropy weight method is used to fuse the energy focusing coefficient, deformation mutation gradient and feature credibility weight three indicators. First, normalize each indicator, then determine the objective weight coefficient according to its information entropy. The calculated resonance risk entropy value is quantized to 0-100 points, and when it exceeds the dynamic threshold (initially set to 75 points), the system performs a three-level response: the first level is encryption scanning verification (upgraded to 1 per minute), the second level is real-time checking of the finite element model, and the third level is early warning information push. The dynamic threshold is automatically adjusted according to the historical false alarm rate, and if there is no false alarm for 3 consecutive times, the threshold is lowered by 5 points to improve sensitivity.

[0043] For the annulus marked as a resonance hotspot, the system automatically retrieves the deformation data of the last 3 encrypted scans, and verifies the risk persistence through the curvature change trend. At the same time, feature extraction parameters (such as the number of layers of non-negative decomposition, frequency band contraction ratio, etc.) are recorded to the knowledge base, and when a real resonance is verified, the parameter settings are optimized for subsequent analysis. All feature extraction cases are periodically reviewed and evaluated every month to update the baseline values and threshold parameters, forming a self-optimizing closed-loop system.

[0044] In the dynamic identification module for resonance risk, a directed topological network is constructed based on the mechanical properties of spatial grid nodes. The node stiffness attenuation coefficient is calculated by subtracting corrosion loss from the measured steel plate thickness, and the corrosion rate is updated monthly based on potential detection data. Weld continuity parameters are extracted from three-dimensional deformation data: five monitoring points are taken on each side of the weld, and the angle of the displacement gradient vector is calculated. An angle less than 15 degrees is considered to be good continuity. Network edge weights are set to the energy transfer efficiency between adjacent nodes, and their value depends on the node spacing and the integrity rating of the connecting weld. After the network is constructed, accessibility analysis is performed to identify the main energy transfer paths and potential blockage points.

[0045] For each grid node, the resonant energy flux density is calculated within the target frequency band (usually 1-10 Hz). The strain spectrum amplitude is taken from the temperature-compensated data of the vibrating wire sensor, the wind load power spectrum density is derived from the turbulence-corrected results of the ultrasonic anemometer, and the node transfer efficiency is derived from the topological network analysis. During the calculation, each parameter is first normalized, and then the geometric mean of the three is taken as the energy flux density indicator. The system continuously tracks the energy flux density distribution of each ring belt. When it is detected that the energy flux density of three consecutive nodes in the same ring belt exceeds twice the historical baseline value of the ring belt and lasts for more than three sampling periods, the ring belt is marked as an energy focus hotspot.

[0046] The natural frequency parameters corresponding to the hotspot ring are extracted from a finite element model database, which has been calibrated through modal testing. The hotspot frequency band monitored in real time is compared with the model's natural frequency to verify that it falls within the ±15% range. For the weld's heat-affected zone, additional consideration is given to the material degradation coefficient, extending the natural frequency range to -20%. Once the match is verified, the system automatically retrieves historical fatigue cumulative damage data for the weld to supplement the risk level assessment.

[0047] Within the ring marked as a hotspot, instantaneous phase data is extracted from all vibrating wire sensors. The instantaneous phase angle of each sensor signal is calculated using a Hilbert transform, and the phase consistency of the sensors within the ring is assessed using the circular statistics method. Structural resonance is identified when the phase difference of more than 80% of the sensors remains within ±15 degrees (corresponding to 85% coherence) for more than 5 seconds. To improve robustness, the system excludes anomalous phase data caused by sensor failure.

[0048] A three-element time series containing the energy focusing coefficient, phase coherence index and deformation mutation gradient is constructed. The non-linear correlation between parameters is analyzed by an improved Granger causality test method: the time series is divided into 10-second sliding windows, and the conditional entropy and transfer entropy are calculated in each window. Finally, the correlation entropy value is synthesized by the information geometry method, and the smaller the value, the stronger the deterministic relationship between parameters. The dynamic entropy threshold is set, and when the measured entropy value is lower than the threshold and the phase coherence continues to meet the conditions, the resonance risk level is determined as "confirmed" state.

[0049] The three-dimensional scanning robot is controlled to perform millisecond-level laser vibration measurement on the hot spot ring. The sampling frequency is set to 1 kHz, and the duration is 10 seconds. The obtained vibration displacement data is reconstructed by a modal decomposition algorithm to obtain the vibration cloud of the library wall. The standing wave characteristics in the cloud are analyzed, and the intersection position of the standing wave nodal line (vibration amplitude zero point connecting line) and the circumferential weld is checked. When the nodal line crosses the weld and the strain gradient at the intersection point exceeds the allowable value, it is determined as a high-risk state. The system automatically generates a special evaluation report containing vibration modal animation and weld stress concentration coefficient.

[0050] According to the continuous monitoring results, the risk level is dynamically adjusted: if the energy flow density continues to rise while the phase coherence remains, the risk level is increased by one level every 5 minutes; if the deformation rate slows down or the entropy value rises, the downgrade evaluation process is started. The risk level is divided into four levels: observation (blue), warning (yellow), emergency (orange) and danger (red), each corresponding to a different response plan. All level changes are recorded with time stamp and decision basis to form a complete risk evolution file.

[0051] For cases confirmed as resonance, the system calls the weld fatigue analysis module. Based on the vibration amplitude, frequency and material S-N curve, the fatigue damage increment caused by this resonance event is calculated. Combined with the historical cumulative damage data, the remaining fatigue life is predicted. When the predicted life is less than 3 months, a special inspection work order is triggered and pushed to the maintenance management system. The evaluation report contains key parameters such as crack initiation probability and critical crack size suggestion value.

[0052] In the closed-loop early warning and feedback module, the system monitors the energy flow density indicators of each ring in real time. When the energy flow density of the hotspot area exceeds the dynamic threshold (initial value set to 2.5 times the historical average of the ring), the hierarchical response mechanism is started. The first level response is to increase the three-dimensional deformation scanning frequency: in the normal mode, the scanning is increased from once every hour to once every 12 minutes (5 times frequency), and the scanning range is focused on the hotspot ring and its adjacent two rings. The second level response calls the finite element model for real-time checking. When the deviation between the measured resonance frequency and the model predicted value exceeds 15%, the weld damage probability prediction algorithm based on fracture mechanics is automatically activated. The third level response is triggered when the energy flow density continues to increase for 3 minutes without attenuation trend. The control board cleaning robot moves to the position directly below the hotspot ring and performs directional vibration work to loosen the accumulated material. The vibration intensity is adjusted according to the energy flow density value.

[0053] The deformation curvature data obtained by the encryption scanning is spatially matched with the finite element prediction value. The improved Hausdorff distance algorithm is used to calculate the difference between the two: first, the scanning point cloud and the predicted surface are meshed, and the maximum value of the curvature difference in each mesh element is calculated. The system presets the deformation tolerance threshold as 0.15 rad / m. When the measured distance is less than this value, it is determined as a false alarm and the early warning level is downgraded; when the distance value continues to increase and the weld damage probability model output value exceeds 60%, it is upgraded to the second early warning state. During the verification process, the system automatically marks the areas with significant differences, which are used to guide the local modification of the finite element model.

[0054] Based on the historical early warning case library, a hidden Markov model is trained, which includes three hidden states (risk increase, risk stability, risk decrease) and six observation variables (energy flow density, deformation curvature, etc.). When the energy flow density of the hotspot area decreases at a rate lower than the preset value (usually 15% per hour), the model outputs a strategy adjustment suggestion: extend the encryption scanning duration from the default 30 minutes to 60 minutes, and shrink the control cleaning robot's working radius from 5 meters to 3 meters, forming a more intensive monitoring-disposal closed loop. The strategy adjustment parameters are automatically updated every week to ensure adaptation to the time-varying characteristics of the library structure performance.

[0055] A gray library digital twin is constructed, with the core being a parameterized finite element model. After each encryption scanning is completed, the deformation field data, robot vibration operation records (including vibration position, intensity, duration) and energy flow density change curve are input into the twin. The system adjusts the model parameters through the back propagation algorithm: the boundary constraint conditions are corrected according to the measured support reaction force data, and the material constitutive parameters are optimized according to the stress-strain relationship curve. The modified model is immediately used for the next round of early warning analysis, forming a continuously evolving simulation environment. When the model version is updated, the historical copies are preserved for backtracking and comparative analysis.

[0056] When the second-level warning state continues for 10 minutes without being lifted, the cross-system collaborative disposal process is initiated. First, the three-dimensional morphology acquisition system is called to perform a full-circumferential scan of the hot spot ring belt, and a topological model with millimeter-level accuracy is reconstructed, with special attention paid to the surface microcracks in the weld area. At the same time, the plate-hardening cleaning robot is instructed to switch to preventive vibration mode: vibration points are arranged at intervals of 2 meters below the hot spot ring belt, each point is vibrated for 30 seconds, and the vibration frequency is set to 80% of the natural frequency of the ring belt to avoid resonance. The operating data of the two systems are shared in real time, the morphology data is used to verify the vibration effect, and the vibration records are used to optimize the operating parameters.

[0057] After each warning is lifted, the system calculates the structural relaxation coefficient as a performance evaluation indicator. This coefficient is defined as the ratio of the energy flux density decay rate to the deformation recovery rate, reflecting the structural damping characteristics. The calculation uses monitoring data within 30 minutes after the warning is lifted, and the rate of change is calculated through linear regression. The coefficient baseline value is statistically derived from the initial 100 days of stable operation data (typical values ​​are 1.2-1.8). When the relaxation coefficient exceeds 1.5 times the baseline value for three consecutive warning events, the structural stiffness is determined to have degraded. The system automatically and permanently increases the dynamic threshold by 20% to enhance conservatism, and simultaneously generates a reinforcement plan including the installation location of the reinforcement ring and the steel plate.

[0058] A comprehensive monthly assessment of the early warning system is conducted, analyzing metrics such as false alarm rate, missed alarm rate, and average response time, and analyzing the correlation between warning levels and subsequent damage development. The assessment results are used to optimize three parameters: adjusting the trigger thresholds for each level of response, revising the material parameter update strategy for the finite element model, and improving the vibration operation procedures for the cleaning robot. All optimization solutions must be verified through digital twin simulation before deployment to ensure system stability.

[0059] The present invention works by acquiring multi-source data (storage wall strain, environmental wind load, and three-dimensional deformation), dynamic spatial gridding modeling (demarcation of high-risk, transitional, and stable regions and optimization of node weights), cross-frequency domain feature extraction (construction of a space-time-frequency matrix and adversarial generation verification), multi-dimensional identification of resonance risk (energy focusing coefficient, phase coherence, and standing wave nodal lines in modal cloud maps), and closed-loop strategy optimization (digital twin iterative model, hierarchical early warning, and robotic response). This system achieves high-precision health monitoring and dynamic risk management of storage structures. The system innovatively integrates physical modeling with data-driven approaches, improving the accuracy of resonance risk identification and the efficiency of early warning responses through dynamic weight adjustment, energy transfer path analysis, and adaptive threshold optimization.

[0060] The above is a detailed description of an embodiment of the present invention. However, the content described is only a preferred embodiment of the present invention and should not be considered to limit the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. The warehouse health monitoring and early warning system based on heterogeneous multi-source data fusion is characterized by: include: A multi-source data acquisition module, which is used to obtain real-time data on the strain of the silo wall, environmental wind load data, and three-dimensional deformation data of the silo body; A spatial grid layered mapping module, which divides the reservoir into high-risk areas, transition areas, and stable areas based on the reservoir structure characteristics, constructs a multi-dimensional spatial grid matrix, and maps sensor locations to grid nodes; A cross-frequency domain correlation feature extraction module, which performs dimension-upgrading and fusion of strain spectrum, wind load spectrum, and deformation curvature data into a time-space-frequency domain feature matrix, and extracts coupling features of spatial position and frequency through constrained optimization decomposition; A resonance risk dynamic identification module, which locates the energy focus hotspots in the space-frequency domain based on the energy transfer path analysis of the coupling characteristics and verifies the resonance risk in combination with the weak point parameters of the warehouse structure; A closed-loop early warning and feedback module triggers three-dimensional deformation encryption scanning and finite element model verification when the energy intensity of the hot spot area exceeds a dynamic threshold, outputs a graded early warning and dynamically adjusts the monitoring strategy.

2. The warehouse health monitoring and early warning system based on heterogeneous multi-source data fusion according to claim 1 is characterized in that: The multidimensional space grid matrix construction specifically includes the following steps: The distribution location and historical damage data of the circumferential welds of the silo's welded structure were extracted, and combined with the finite element cloud map of the silo's wall stress, the silo's wall was divided into high-risk areas, transition areas, and stable areas along the height direction. The high-risk area covers the height from the annular foundation at the bottom of the reservoir to the maximum bending moment of the reservoir wall, the transition area covers the bending moment attenuation zone to the height of the structural deformation inflection point, and the stable area covers the area above the inflection point. Taking the central axis of the reservoir as the benchmark, the high-risk area is divided into 12 sector-shaped ring zones, the transition area is divided into 8 sector-shaped ring zones, and the stable area is divided into 4 sector-shaped ring zones. The high-risk area is divided into 6 layers according to the stress gradient vertically, the transition area is divided into 4 layers, and the stable area is divided into 2 layers, forming a multi-dimensional space grid matrix. The coordinates of the vibrating string sensor are aligned with the three-dimensional deformation scanning point cloud. Through nearest neighbor topological mapping, each sensor is associated with the intersection grid node of the corresponding ring belt and layer domain, and the nodes in the high-risk area are given a weight coefficient of 3 times and the nodes in the transition area are given a weight coefficient of 2 times.

3. The warehouse health monitoring and early warning system based on heterogeneous multi-source data fusion according to claim 1 is characterized in that: The method of dividing the high-risk area, transition area and stable area according to the structural characteristics of the reservoir, constructing a multi-dimensional space grid matrix, and mapping the sensor positions to the grid nodes includes at least the following steps: The micro-deformation data of the heat-affected zone of the weld is acquired in real time. When the deformation variance within a single ring band exceeds a threshold, the ring band is subdivided into 16 sub-ring bands, and virtual monitoring nodes are generated at the intersection of the sub-ring bands. Based on the thickness distribution of the warehouse wall steel plate, the environmental corrosion rate and the historical load spectrum, the structural risk entropy value of each grid node is calculated; The three parameters of entropy value, real-time strain gradient and weld distance are input into the fuzzy decision maker, and the node weight correction factor is dynamically output. The correction factor range is set to 1.0-5.0; When a grid node sensor fails, the data of adjacent nodes is called and the node status is reconstructed through radial basis interpolation. At the same time, the 3D scanning robot is driven to perform encrypted scanning of the ring where the node is located, replacing the missing signal with laser point cloud data.

4. The warehouse health monitoring and early warning system based on heterogeneous multi-source data fusion according to claim 1 is characterized in that: The construction process of the time-space-frequency domain feature matrix is ​​as follows: A correlation constraint model between the strain main frequency component and the environmental wind load harmonic component is established to force the strain main frequency amplitude and the wind load harmonic amplitude within the same ring position to satisfy a monotonically increasing relationship. The strain spectrum, wind load spectrum and deformation curvature data are reorganized into a time-space-frequency domain feature matrix by time slice, and the coupling factor matrix between the spatial grid nodes and the frequency is extracted by using non-negative constraint decomposition. The energy focusing coefficient of each ring zone area is calculated based on the coupling factor matrix. When the focusing coefficient of a ring zone in a specific frequency band exceeds the baseline value and the deformation curvature changes synchronously, it is marked as a resonance risk hotspot.

5. The warehouse health monitoring and early warning system based on heterogeneous multi-source data fusion according to claim 1 is characterized in that: The method of dimensionalizing the strain spectrum, wind load spectrum, and deformation curvature data into a time-space-frequency domain feature matrix and extracting the coupling characteristics of spatial position and frequency through constrained optimization decomposition includes at least the following steps: Real-time analysis of the marginal spectrum entropy of the reservoir wall vibration signal. When the entropy value is lower than the threshold, the wind load spectrum analysis band is narrowed to within ±0.5 octave of the main strain frequency. A strain-deformation time series generative adversarial network is constructed. The generator generates virtual high-frequency deformation data based on the low-frequency deformation curvature, and the discriminator introduces the vibration differential equation of the storage body to constrain the network output. The Hausdorff distance between the generated deformation data and the measured deformation data is calculated, and the credibility weight of the spatial-frequency domain feature is output by combining the singular value decay rate of the coupling factor matrix. The energy focusing coefficient, deformation mutation gradient and feature credibility weight are integrated to calculate the resonance risk entropy value through the entropy weight method. When the dynamic threshold is exceeded, the encrypted scanning verification is triggered.

6. The warehouse health monitoring and early warning system based on heterogeneous multi-source data fusion according to claim 1 is characterized in that: The energy transfer path analysis based on coupling characteristics and positioning of space-frequency domain energy focusing hotspots include the following steps: Based on the stiffness attenuation coefficient of spatial grid nodes and the continuity parameters of circumferential welds, a directed topological network for vibration energy transfer of the warehouse is constructed. The node stiffness attenuation coefficient is calculated by the steel plate thickness and corrosion rate, and the weld continuity is evaluated by fitting the weld displacement gradient with 3D deformation data. Calculate the resonant energy flux density of each grid node within the target frequency band, where the energy flux density is the product of the strain spectrum amplitude, the wind load power spectrum density, and the node transfer efficiency; When the energy flux density of three consecutive nodes in the same ring belt exceeds 2 times the baseline value, the ring belt is marked as an energy focus hotspot; Call the natural frequency of the heat-affected zone of the weld in the finite element model to verify whether the hot spot frequency band falls within the range of ±15% of the natural frequency; At the same time, check whether the rate of change of the hotspot ring deformation curvature suddenly increases to more than three times that of the previous moment.

7. The warehouse health monitoring and early warning system based on heterogeneous multi-source data fusion according to claim 1 is characterized in that: The method of verifying the resonance risk by combining the weak point parameters of the warehouse structure includes the following steps: Extract the instantaneous phase of the vibrating wire sensor within the hotspot ring. When the phase coherence of 80% of the sensors within the ring exceeds 85% and lasts for more than 5 seconds, it is determined to be structural resonance. The energy focusing coefficient, phase coherence and deformation mutation gradient are integrated and the correlation entropy of the three is calculated through nonlinear Granger causality test. When the correlation entropy value is lower than the threshold and the phase coherence persists, the resonance risk level is output; Drive the 3D scanning robot to perform millisecond-level laser vibration measurement on the hotspot ring and reconstruct the vibration mode cloud map of the warehouse wall; If the cloud map shows that the standing wave node line passes through the circumferential weld, a level 1 warning is triggered and a weld fatigue life assessment report is generated.

8. The warehouse health monitoring and early warning system based on heterogeneous multi-source data fusion according to claim 1 is characterized in that: The output of graded warnings and dynamic adjustment of monitoring strategies specifically include: When the energy flux density of the energy focusing hotspot exceeds the dynamic threshold; Level 1 response: Initiate 3D deformation encryption scanning, increasing the scanning frequency to 5 times the normal frequency; Secondary response: The finite element model is used to check the resonant frequency matching. If the deviation exceeds 15%, weld damage prediction is activated. Level 3 response: When the energy flux density does not decay for 3 minutes, the compaction cleaning robot is driven to release the material pressure in the hot spot area; The deformation curvature obtained by the encrypted scan is compared with the finite element predicted deformation to calculate the Hausdorff distance; If the distance value is less than the deformation tolerance threshold, the warning level is downgraded; If the distance value increases and the probability of weld damage exceeds 60%, a second-level warning is output; Training Hidden Markov Model based on historical warning data; When the energy flux density decrease rate in the hot spot area is lower than the preset value; Automatically extend the encryption scan time to twice the original plan and shrink the robot's operating radius to within 3 meters.

9. The warehouse health monitoring and early warning system based on heterogeneous multi-source data fusion according to claim 1 is characterized in that: The output of graded warnings and dynamic adjustment of monitoring strategies include the following steps: Input encrypted scanning data, robot operation records, and energy flux density change rates into the digital twin to dynamically modify the boundary constraints and material constitutive parameters of the finite element model; When the Level 2 warning lasts for 10 minutes; Synchronously start the 3D topology acquisition system to reconstruct the reservoir wall topology model; The compaction cleaning robot is called upon to perform preventive vibration on the hot spot ring; The structural relaxation coefficient after each warning is lifted is calculated, which is defined as the ratio of the energy flux density decay rate to the deformation recovery rate; When the relaxation coefficient is greater than 1.5 times the baseline value for three consecutive times; Increase the dynamic threshold by 20% and generate a structural stiffness enhancement plan.

Images