Building structure health monitoring system based on multi-sensor fusion

Abstract

The invention discloses a building structure health monitoring system based on multi-sensor fusion, and relates to the technical field of building structure safety monitoring and sensor data fusion processing. Comprising a sensor network, a data acquisition module, a data processing center and a result output module, wherein the sensor network is arranged at key parts of a building structure and consists of various types of sensors, and the data acquisition module is used for collecting original monitoring data of the sensor network. According to the building structure health monitoring system based on multi-sensor fusion, the problem of data drift of the sensors caused by environmental fluctuation is effectively solved by constructing the virtual calibration twin body interacting with the physical sensor network in real time and establishing a multi-layer data verification mechanism. The system can automatically identify and correct the abnormal deviation of the sensor data in the online monitoring process, and ensures the accuracy and reliability of the output data through the dual guarantee of primary screening and secondary consensus verification, thereby improving the credibility of the judgment of the health state of the building structure.

Description

Technical Field

[0001] This invention relates to the field of building structure safety monitoring and sensor data fusion processing technology, specifically a building structure health monitoring system based on multi-sensor fusion. Background Technology

[0002] In the existing field of building structural health monitoring, multi-sensor fusion technology has been widely used to improve the comprehensiveness of condition assessment. However, in actual long-term deployment, this system has exposed a key and urgent practical technical problem: sensor data drifts slowly with natural fluctuations in ambient temperature and humidity. This drift is gradual and non-linear, causing the basic accuracy of the monitoring data to deteriorate over time. Existing solutions mostly rely on preset fixed calibration coefficients or periodic on-site manual calibration. These methods cannot respond in real time to continuously changing environmental conditions. As a result, during the window between two calibrations, the system's ability to identify key structural anomalies, such as microcrack propagation or minor plastic deformation, decreases significantly, and may even lead to misjudgments, posing potential risks to structural safety. To address this, a building structural health monitoring system based on multi-sensor fusion aims to solve the practical problem of inaccurate monitoring caused by sensor data drift due to environmental fluctuations. Summary of the Invention

[0003] The purpose of this invention is to provide a building structure health monitoring system based on multi-sensor fusion to solve the problems mentioned in the background art.

[0004] To address the aforementioned technical problems, the present invention provides the following technical solution: a building structure health monitoring system based on multi-sensor fusion, comprising a sensor network composed of various types of sensors arranged in key parts of the building structure, a data acquisition module, a data processing center, and a result output module. The data acquisition module is used to collect raw monitoring data from the sensor network. The data processing center performs fusion analysis on the collected data. The result output module is used for display and alarm. The data processing center further includes a virtual calibration twin construction module, a multi-layer data verification module, and a model self-updating module. During the system initialization phase, the virtual calibration twin construction module establishes an initial benchmark mapping relationship library describing the inherent physical correlation between the readings of each sensor based on the multi-sensor benchmark data of the building structure under no-load and environmentally stable conditions. This mapping relationship library is obtained through machine learning model training and is used to simulate the sensor data interaction mode of the building structure under healthy conditions. The multi-layer data verification module includes a primary drift screening unit and a secondary consensus verification unit. The primary drift screening unit receives real-time sensor data and environmental parameters, calculates the theoretical expected correlation by embedding a physical information neural network with simplified mechanical equations of the building structure, generates primary correction data and marks suspected drift points. The secondary consensus verification unit calls the initial benchmark mapping relationship library and combines it with the current reliable sensor data. It uses a graph signal processing algorithm to reconstruct the consensus estimate of the suspected measurement points. It compares the original readings, primary correction data and consensus estimate through an arbitration mechanism to output high-confidence secondary fine-calibration data. The model self-updating module feeds back the secondary fine-tuning data and the corresponding data quality confidence report to the physical information neural network, enabling the model to achieve online incremental learning and adaptive optimization. This system achieves dynamic correction and data quality purification of sensor data environment drift through parallel interaction between virtual calibration twins and physical sensor networks and multi-layer verification loops.

[0005] Furthermore, the virtual calibration twin construction module performs the following steps when establishing the initial benchmark mapping relationship library: Under the baseline conditions of no external load on the building structure and stable ambient temperature and humidity, raw data from all sensors are collected simultaneously for multiple consecutive cycles. Extract the time-domain and frequency-domain features of each sensor data, and calculate the correlation coefficient, covariance matrix, and transfer function between different sensor data. The above features and relational parameters are input into a deep spatiotemporal graph neural network for training. The network takes environmental temperature and humidity data as input and the physical constraint relationship between sensor readings as the hidden layer supervision signal, and learns a digital twin model that can predict the expected readings of each sensor under given environmental conditions. The trained digital twin model and its parameters are stored as an initial benchmark mapping relationship library, which is used for real-time data comparison and drift detection during the online monitoring phase.

[0006] Furthermore, the physical information neural network in the primary drift screening unit adopts a deep learning architecture constrained by partial differential equations. The forward propagation process of this network integrates the constitutive equations of the building structure and the spatial topological relationship of the sensor layout. The network input layer receives real-time sensor data vectors and environmental temperature and humidity vectors. The hidden layer learns the nonlinear mapping between environmental parameters and sensor readings through residual connections. The output layer generates the theoretical expected values ​​of each sensor. The difference analysis submodule calculates the residual sequence between the original readings and the theoretical expected values, and performs time-series smoothness and correlation analysis on the residuals. If the sliding variance of the residuals exceeds the preset threshold or shows a strong correlation with environmental changes, the data point is marked as a suspected drift point. The compensation generation submodule generates dynamic compensation coefficients for marked suspected drift points based on the magnitude and direction of the residuals, performs linear or nonlinear transformations on the original data, and generates first-level correction data.

[0007] Furthermore, the execution process of the secondary consensus verification unit includes: The graph signal processing algorithm uses the sensor network topology as a graph structure and reliable sensor data as node signals. It reconstructs the consensus estimate of the suspected measurement point in the current state through graph Laplace regularization. The drift arbitrator uses a majority decision mechanism to compare the Euclidean distance and cosine similarity of the included angle between the original readings of the suspected point, the first-level correction data, and the consensus estimate. If the similarity between the first-level correction data and the consensus estimate is higher than the similarity between the two and the original reading, then the point is determined to be a true drift, and the first-level correction data is adopted as the output. If the original reading has the highest similarity to the consensus estimate, then the point is determined to be likely to reflect the actual structural change, the original reading is retained, and a structural anomaly warning is triggered. The arbitration result, along with the confidence score, is recorded in the data quality confidence report.

[0008] Furthermore, the online incremental learning process of the model self-updating module is specifically as follows: At fixed time windows, data samples with confidence levels higher than a threshold are selected from the secondary refined data to form a micro-batch training set. Using these high-confidence samples as supervision signals, the gradient descent method is used to fine-tune the weight parameters of the physical information neural network. The adjustment range is controlled by the learning rate and the confidence weighting factor. Meanwhile, the key relation parameters in the initial baseline mapping relation library are updated by moving average based on the latest data distribution, enabling the digital twin model to track the long-term performance evolution of the building structure. The model version management submodule retains historical model parameters. If the validation performance degrades over several consecutive periods, it will automatically roll back to the previous stable version.

[0009] Furthermore, the various types of sensors include strain sensors, acceleration sensors, temperature sensors, and humidity sensors, which are fixed to the beams, columns, nodes, and foundations of the building structure in a spatially distributed layout.

[0010] Furthermore, the data acquisition module employs multi-channel synchronous sampling technology to ensure that the timestamps of data from different sensors are aligned, and the sampling frequency is adaptively adjusted according to the dynamic response characteristics of the structure.

[0011] Furthermore, the result output module provides a graphical interface that displays the secondary calibration data, data quality confidence report, and structural health status assessment results in real time. When a verified structural abnormality is detected, a warning message is issued through an audible and visual alarm and a remote communication interface.

[0012] Furthermore, the drift arbitrator also incorporates the sensor's historical reliability score as a weighting factor during the arbitration process. The historical reliability score is dynamically updated based on the frequency at which the sensor has been marked as drifting in the past and the accuracy of the arbitration results.

[0013] Furthermore, the model self-update module also includes a model health monitoring submodule, which periodically injects test data of known drift patterns into the simulation environment to verify the model correction performance. If the correction error continues to exceed the tolerance range, the model reconstruction process is triggered.

[0014] This invention provides a building structure health monitoring system based on multi-sensor fusion. It has the following beneficial effects: This building structural health monitoring system, based on multi-sensor fusion, effectively solves the problem of data drift caused by environmental fluctuations by constructing a virtual calibration twin that interacts in real time with the physical sensor network and establishing a multi-layered data verification mechanism. The system can automatically identify and correct abnormal deviations in sensor data during online monitoring. Through primary screening and secondary consensus verification, it ensures the accuracy and reliability of the output data, thereby improving the credibility of the assessment of the building's structural health status.

[0015] This building structural health monitoring system, based on multi-sensor fusion, possesses self-learning and continuous optimization capabilities. Through a model self-updating mechanism, the digital twin model can adapt to the long-term performance evolution of the building structure. This feature not only reduces reliance on external manual calibration and maintenance costs but also enables timely identification of real structural changes and sensor malfunctions, providing stable and reliable technical support for building structural safety early warning and long-term performance evaluation. Attached Figure Description

[0016] Figure 1 This is a system architecture diagram of a building structure health monitoring system based on multi-sensor fusion according to the present invention; Figure 2 This is a model self-updating closed-loop diagram of a building structure health monitoring system based on multi-sensor fusion according to the present invention; Figure 3 This is a drift arbitration decision tree diagram for a building structure health monitoring system based on multi-sensor fusion according to 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] Please see Figures 1 to 3 The present invention provides a technical solution: a building structure health monitoring system based on multi-sensor fusion, comprising a sensor network composed of various types of sensors arranged in key parts of the building structure, a data acquisition module, a data processing center, and a result output module. The data acquisition module is used to collect the raw monitoring data of the sensor network, the data processing center performs fusion analysis on the collected data, and the result output module is used for display and alarm. The data processing center further includes a virtual calibration twin construction module, a multi-layer data verification module, and a model self-updating module. During the system initialization phase, the virtual calibration twin building module establishes an initial benchmark mapping relationship library describing the inherent physical correlation between the readings of each sensor based on the multi-sensor benchmark data of the building structure under no-load and environmentally stable conditions. This mapping relationship library is obtained through machine learning model training and is used to simulate the sensor data interaction mode of the building structure under healthy conditions. The multi-layer data verification module includes a primary drift screening unit and a secondary consensus verification unit. The primary drift screening unit receives real-time sensor data and environmental parameters, calculates the theoretical expected correlation by embedding a physical information neural network that simplifies the mechanical equations of the building structure, generates first-level correction data and marks suspected drift points. The secondary consensus verification unit calls the initial benchmark mapping relationship library and combines it with the current reliable sensor data. It uses a graph signal processing algorithm to reconstruct the consensus estimate of the suspected measurement points. It compares the original readings, first-level correction data and consensus estimate through an arbitration mechanism to output high-confidence second-level fine-calibration data. The model self-update module feeds back the secondary fine-tuning data and the corresponding data quality confidence report to the physical information neural network, enabling the model to achieve online incremental learning and adaptive optimization. This system achieves dynamic correction and data quality purification of sensor data environment drift through parallel interaction between virtual calibration twins and physical sensor networks and multi-layer verification loops.

[0019] It should be further explained that dynamic correction of sensor data environmental drift is achieved by constructing a virtual calibration twin that interacts in real time with the physical sensor network. During the initialization phase, the system first controls the data acquisition module to synchronously collect raw data from all sensors for multiple consecutive sampling periods under specific baseline conditions where the building structure is under no external load and the ambient temperature and humidity remain stable. Subsequently, the virtual calibration twin construction module uses a deep spatiotemporal graph neural network to train on this baseline data. The input to this network is the ambient temperature and humidity data. The hidden layer uses simplified constitutive equations describing the mechanical behavior of the building structure as physical constraints, while the output layer learns to generate a physical correlation model that should be satisfied between the readings of each sensor. This process ultimately establishes an initial baseline mapping relationship library that can characterize the intrinsic correlation laws of sensors under structurally healthy conditions.

[0020] During online monitoring, the multi-layer data verification module begins operation: its primary drift screening unit inputs real-time sensor data and environmental parameters into a trained physical information neural network. This network calculates the theoretical expected values ​​of each sensor based on learned physical constraints, compares the residuals between the theoretical values ​​and actual readings, and intelligently identifies suspected drift points by combining the temporal characteristics of the residuals with their environmental correlation, generating preliminary first-level correction data. Subsequently, the secondary consensus verification unit is activated. It calls the initial benchmark mapping relation library and uses the currently unmarked reliable sensor data as a benchmark to reconstruct the spatial consistency estimate of suspected measurement points on the sensor network topology using graph signal processing algorithms. The crucial drift arbitrator then performs a three-way comparison of the original readings, first-level correction data, and consensus estimates, making a voting decision based on preset similarity measurement rules, and finally outputs high-confidence second-level fine-calibration data that has undergone secondary verification.

[0021] Meanwhile, the model self-updating module continuously collects these high-confidence data and their quality reports, using them as samples for incremental learning. It fine-tunes the parameters of the physical information neural network using a gradient descent algorithm, ensuring that key parameters in the initial benchmark mapping relation library can track the long-term performance evolution of the structure, thus forming a closed-loop calibration system with self-optimization capabilities. This technical solution, through parallel interaction between the virtual twin and the physical system, progressive confirmation via multi-layered verification mechanisms, and continuous model evolution, fully achieves the technical effects of online identification, correction, and system adaptation of sensor data drift.

[0022] When establishing the initial benchmark mapping relationship library, the virtual calibration twin construction module performs the following steps: Under the baseline conditions of no external load on the building structure and stable ambient temperature and humidity, raw data from all sensors are collected simultaneously for multiple consecutive cycles. Extract the time-domain and frequency-domain features of each sensor data, and calculate the correlation coefficient, covariance matrix, and transfer function between different sensor data. The above features and relational parameters are input into a deep spatiotemporal graph neural network for training. The network takes environmental temperature and humidity data as input and the physical constraint relationship between sensor readings as the hidden layer supervision signal, and learns a digital twin model that can predict the expected readings of each sensor under given environmental conditions. The trained digital twin model and its parameters are stored as an initial benchmark mapping relationship library, which is used for real-time data comparison and drift detection during the online monitoring phase.

[0023] It should be further explained that the detailed process of establishing the initial benchmark mapping relationship library by the virtual calibration twin construction module is defined. This process is the foundation for the system to achieve high-precision drift correction. In the benchmark establishment phase of system initialization, it is first necessary to ensure that the building structure is under specific conditions of no external load and highly stable ambient temperature and humidity. At this time, the data acquisition module synchronously triggers all sensors to collect data for multiple consecutive cycles to obtain a standard dataset that can represent the intrinsic health state of the structure. The next step is feature extraction and relationship modeling. The collected multi-source sensor data is preprocessed, including calculating the statistical characteristics of each sensor reading in the time domain, such as mean, variance, and peak factor. In the frequency domain, the dominant frequency components and power spectral density are extracted through fast Fourier transform. Furthermore, the correlation characteristics between different sensor data streams are analyzed. By calculating the Pearson correlation coefficient, constructing the covariance matrix, and estimating the transfer function between sensor nodes, the inherent physical constraint relationships between the sensor readings are quantified. These extracted features and relational parameters are systematically organized into training samples and input into a deep spatiotemporal graph neural network for model training. The special design of this network is that its graph structure accurately corresponds to the actual spatial layout of the sensors on the building structure. The node attributes contain sensor type and location information, and the edge weights are initialized to the physical correlation strength derived from the principles of structural mechanics. During the training process, the network uses real-time collected environmental temperature and humidity data as conditional input and the physical relationship that should be satisfied between sensor readings as implicit supervision signals. The network parameters are optimized through the backpropagation algorithm, and finally a digital twin model that can accurately simulate the correlation between the expected readings of each sensor under any given environmental conditions is obtained. The trained model and all its parameters are stored in a complete initial benchmark mapping library. This library not only contains the static mapping relationships of sensor data, but also encapsulates the complex dynamic interaction patterns between environmental parameters and physical constraints, providing an accurate benchmark for real-time data comparison and intelligent drift detection during the online monitoring phase.

[0024] The physical information neural network in the primary drift screening unit adopts a deep learning architecture constrained by partial differential equations. The forward propagation process of this network integrates the constitutive equations of the building structure and the spatial topological relationship of the sensor layout. The network input layer receives real-time sensor data vectors and environmental temperature and humidity vectors. The hidden layer learns the nonlinear mapping between environmental parameters and sensor readings through residual connections. The output layer generates the theoretical expected values ​​of each sensor. The difference analysis submodule calculates the residual sequence between the original readings and the theoretical expected values, and performs time-series smoothness and correlation analysis on the residuals. If the sliding variance of the residuals exceeds the preset threshold or shows a strong correlation with environmental changes, the data point is marked as a suspected drift point. The compensation generation submodule generates dynamic compensation coefficients for marked suspected drift points based on the magnitude and direction of the residuals, performs linear or nonlinear transformations on the original data, and generates first-level correction data.

[0025] Furthermore, the specific logic for the compensation generation submodule to generate dynamic compensation coefficients and first-level correction data is as follows, to ensure the reproducibility of the data correction process: In the linear compensation scenario, applicable when the residual sequence exhibits a stable linear change, and the difference analysis submodule determines that the temporal smoothness of the residual sequence satisfies "the fluctuation of the residual change rate within 5 consecutive sampling periods is ≤5%", a linear compensation model is adopted. The dynamic compensation coefficient k is calculated as follows: Extract the historical residual data {e1,e2,...,e} from the past 30 normal sampling periods of the sensor, i.e., the data not marked as suspected drift. 30}, and the corresponding environmental parameters, namely the changes in temperature and humidity: {ΔT1,ΔH1,...,ΔT 30 ,ΔH 30}; The least squares method was used to linearly fit the residuals and environmental changes, and the fitting equation was obtained as e=a·ΔT+b·ΔH+c, where a is the temperature influence coefficient, b is the humidity influence coefficient, and c is the basic residual offset. The dynamic compensation coefficient k is determined by the current residual e. curr The predicted residual e of the fitted equation pred The deviation is determined by the formula: ; where the denominator is increased by 10 -6 To avoid division by zero, the value of k is limited to [0.8, 1.2], and the boundary value is used when the value exceeds this range. Level 1 calibration data x cal1 The formula for calculating x is: cal1 =x raw ·k+(e pred -e curr ), where x rawThis is the sensor's raw reading.

[0026] In nonlinear compensation scenarios, this is suitable for residual sequences exhibiting nonlinear fluctuations. When the sliding variance of the residual sequence exceeds a preset threshold, the threshold is determined by three times the standard deviation of historical normal data, i.e., σ. th =3σ normal When the correlation coefficient R is non-linear with environmental changes, the correlation coefficient R 2 <0.7, a quadratic polynomial compensation model is adopted: Based on the current ambient temperature and humidity (T) curr H curr As input, based on the "environment-residual" sample set {(T) stored in the initial baseline mapping relation library. i H i ,e i Construct a quadratic polynomial fitting model: e | i=1,2,...,n} (n≥50). pred =p0+p1T+p2H+p3T 2 +p4TH+p5H 2 ; Where p0-p5 are polynomial coefficients, the error function is minimized by solving the Newton-Raphson method. get; The dynamic compensation coefficient adopts a piecewise function form: when|e curr -e pred |≤2σ normal hour, ; When 2σ normal <|e curr -e pred |≤4σ normal hour, ; The lower limit of k is set to 0.6 to avoid overcorrection; Level 1 calibration data x cal1 The formula for calculating x is: cal1 =x raw +(e pred -e curr )·k 2 .

[0027] It should be further explained that the physical information neural network adopts a special design that integrates physical laws and data-driven approaches. Its network structure consists of an input layer, multiple hidden layers, and an output layer. The hidden layers ensure effective gradient propagation through residual connections, and during the forward propagation of the network, simplified constitutive equations describing the mechanical behavior of building structures are embedded as physical constraints. These equations are usually expressed as partial differential equations about stress, strain, and ambient temperature. Their discretized numerical models are encoded into the network loss function as regularization terms.

[0028] During operation, the input layer receives multi-dimensional input features composed of real-time sensor data vectors and synchronously acquired ambient temperature and humidity vectors; the hidden layer learns the complex nonlinear mapping relationship between environmental parameters and sensor readings through activation functions and weight matrices. The output of each layer is not only passed to the next layer, but also fused with the output of the previous layer through skip connections, thereby enhancing the model's ability to represent features; the output layer finally generates a sequence of theoretical expected values ​​for each sensor under the current environmental conditions.

[0029] Subsequently, the difference analysis submodule is activated. It calculates the residual sequence between the original sensor readings and the theoretical expected values, and performs in-depth time-series characteristic analysis on the residual sequence, including calculating its moving average and moving variance within a certain time window, and analyzing the correlation between the residual amplitude and the environmental temperature and humidity change curves. If the residual moving variance of a certain sensor data point continues to exceed the preset threshold calculated based on historical stable data, or if its residual sequence shows a definite strong correlation pattern with the changes in environmental parameters, then the data point is marked as a suspected drift point.

[0030] Finally, the compensation generation submodule, for the marked suspected drift points, uses a preset linear compensation algorithm or a lookup table-based nonlinear mapping function to generate a dynamically changing compensation coefficient based on the magnitude, direction, and time evolution trend of the residuals. This coefficient is then applied to the original sensor readings to generate primary calibration data that has undergone preliminary environmental compensation, providing high-quality input for subsequent secondary consensus verification.

[0031] The execution process of the second consensus verification unit includes: The graph signal processing algorithm uses the sensor network topology as a graph structure and reliable sensor data as node signals. It reconstructs the consensus estimate of the suspected measurement point in the current state through graph Laplace regularization. The drift arbitrator uses a majority decision mechanism to compare the Euclidean distance and cosine similarity of the included angle between the original readings of the suspected point, the first-level correction data, and the consensus estimate. To quantify the credibility of arbitration results, the drift arbitrator, when implementing a majority decision mechanism, needs to calculate a confidence score using the following method and determine the data quality level based on a preset threshold: The core metrics are calculated as follows: Euclidean distance normalization: Let the original reading be x. raw The first-level calibration data is x cal1 The consensus estimate is x. cons The Euclidean distances between the three are: d1 = |x raw -x cal1 |,d2=|x raw -x cons |,d3=|x cal1 -x cons |; The distance is normalized to eliminate the influence of dimensions: For i=1,2,3, after normalization ∈[0,1].

[0032] Cosine similarity of the included angle: Treating the three as three-dimensional vectors: , , Environmental parameters are incorporated to ensure relevance; similarity is calculated as follows: , , ;sim similarity i ∈[-1,1], taking the absolute value yields abs(sim i )∈[0,1].

[0033] The confidence score formula includes the following: The confidence score S is calculated using a weighted summation method, with weight coefficients determined based on historical arbitration accuracy. This was validated through a trial run of 1000 arbitration cases. The similarity weight ω1 = 0.6 and the normalized distance weight ω2 = 0.4 were set. If the arbitration result is "adopt the first-level correction data", then sim3 is the largest and If the minimum is: ; If the arbitration result is "Retain the original reading and issue a warning" (sim2 maximum and (Minimum), then: ; Scoring threshold setting: Confidence scoring threshold S th The "3σ principle" is used to determine: After system initialization, the confidence scores {S1, S2, ..., S} of the first 100 valid arbitrations are collected. 100}, calculate the mean μ S and standard deviation σ S; Set a high confidence threshold S th1 =μ S +0.5σ S Used to filter samples for model self-updating, with a low confidence threshold S. th2 =μ S -0.5σ S When S th2 At that time, a manual review reminder will be triggered; The threshold is updated every 30 days based on the latest arbitration data, using the following formula: To ensure a smooth transition and avoid excessive threshold fluctuations.

[0034] If the similarity between the first-level correction data and the consensus estimate is higher than the similarity between the two and the original reading, then the point is determined to be a true drift, and the first-level correction data is adopted as the output. If the original reading has the highest similarity to the consensus estimate, then the point is determined to be likely to reflect the actual structural change, the original reading is retained, and a structural anomaly warning is triggered. The arbitration result, along with the confidence score, is recorded in the data quality confidence report.

[0035] It should be further explained that the secondary consensus verification unit first initiates a graph signal processing algorithm. This algorithm abstracts the sensor network into a graph structure, where nodes represent individual sensors, edges represent the spatial and physical relationships between sensors, and the weights of the edges are initialized using the association strength learned from the initial benchmark mapping relation library. The algorithm uses the currently unmarked reliable sensor data as the observation signals of known nodes in the graph. By solving an optimization problem based on graph Laplacian regularization, it reconstructs and estimates the signal values ​​of suspected drifting points by utilizing the spatial consistency of the known node signals, thereby obtaining the consensus estimate value for each suspected point.

[0036] Subsequently, the drift arbitrator begins operation, employing a multi-index fusion majority decision mechanism: First, it calculates the Euclidean distance between the raw readings, the first-level correction data, and the consensus estimate to measure the absolute magnitude of the numerical differences; simultaneously, it calculates the cosine similarity of the angles between them to assess the consistency of data change trends; based on this, the arbitrator also considers the historical reliability score of each sensor node, which is dynamically updated based on the frequency with which the sensor has been marked as drifting in the past and its accuracy in historical arbitration, and is incorporated as a weighting factor into the decision-making process.

[0037] ​The final arbitration logic is as follows: if the combined similarity between the primary calibration data and the consensus estimate is the highest in both Euclidean distance and cosine similarity dimensions, and is higher than their respective similarities with the original readings, then the data point is determined to be a true drift, and the primary calibration data is adopted as the final output; if the combined similarity between the original reading and the consensus estimate is the highest, then the reading is determined to be more likely to reflect the true structural state change, and the original reading will be retained, and the structural anomaly warning process will be triggered to remind operators to pay attention to the potential risks of the measurement point.

[0038] All arbitration results, including the final adopted data values, arbitration paths, and calculated confidence scores, are recorded in detail in the data quality confidence report, providing a transparent and traceable basis for decision-making in subsequent data use and model updates.

[0039] The online incremental learning process of the model self-update module is as follows: At fixed time windows, data samples with confidence levels higher than a threshold are selected from the secondary refined data to form a micro-batch training set. Using these high-confidence samples as supervision signals, the gradient descent method is used to fine-tune the weight parameters of the physical information neural network. The adjustment range is controlled by the learning rate and the confidence weighting factor. Meanwhile, the key relation parameters in the initial baseline mapping relation library are updated by moving average based on the latest data distribution, enabling the digital twin model to track the long-term performance evolution of the building structure. The model version management submodule retains historical model parameters. If the validation performance degrades over several consecutive periods, it will automatically roll back to the previous stable version.

[0040] It should be further explained that the model self-update module automatically selects data samples with scores higher than a set threshold from the secondary fine-tuning data stream output after multi-layer verification, using a fixed time window period. These high-confidence samples are then combined with their corresponding environmental parameters to form a micro-batch training dataset. Subsequently, the model parameter optimization process is initiated. Using these selected high-quality data as supervision signals, the stochastic gradient descent algorithm with momentum is used to fine-tune the weight parameters of the physical information neural network in a constrained manner. The learning rate is dynamically weighted according to the confidence score. The higher the confidence score, the greater the contribution weight of the sample to the gradient update during backpropagation. This ensures that the model optimization process fully utilizes reliable new data while effectively suppressing the interference of potential noise. While updating the neural network parameters, the key relational parameters in the initial benchmark mapping relational library that characterize the inherent physical correlation between sensors, including but not limited to the correlation coefficient matrix, covariance structure and transfer function coefficients, are also progressively updated using the exponential weighted moving average method. This enables the digital twin model to smoothly track the long-term performance evolution of building structures caused by material aging, accumulation of micro-damage, etc., and avoid systematic deviations between the model and the physical entity. In addition, the model version management submodule continuously monitors the model's performance metrics during the verification phase and fully saves snapshots of model parameters for each update cycle. If the system detects that the model's correction error for new data continues to increase and exceeds the preset tolerance within multiple consecutive time windows, it will automatically trigger the rollback mechanism to restore the network weights and relation parameters to the previous stable version and issue a maintenance alert at the same time. This ensures the system's continuous learning ability while effectively maintaining its long-term robustness and reliability.

[0041] Various types of sensors, including strain sensors, acceleration sensors, temperature sensors, and humidity sensors, are fixed to the beams, columns, nodes, and foundations of the building structure in a spatially distributed manner.

[0042] It should be further explained that the system uses various types of sensors, including strain sensors for monitoring local deformation characteristics of the structure, acceleration sensors for capturing dynamic response characteristics of the structure, and temperature and humidity sensors for directly measuring environmental factors. These sensors are fixed to key mechanical parts of the building structure according to the principle of spatial distribution. Among them, strain sensors are mainly pasted or welded to the mid-span and support areas of beam members bearing the main loads, and the root and middle areas of column members, for accurately measuring the strain distribution of the structure under load. Accelerometers are installed at the center of floor slabs, mid-span of large-span beams, and the top floor of buildings using magnetic bases or bolts to monitor the vibration response of the structure under dynamic excitations such as wind loads and seismic forces. Temperature and humidity sensors are arranged in pairs on the inner and outer surfaces of the structure, at different orientations, and at typical floor heights, directly exposed to the ambient air to accurately sense changes in temperature and humidity gradients. The placement of all sensors is determined through structural mechanics analysis to ensure that their spatial distribution can fully cover representative areas of the structure, forming a three-dimensional monitoring network capable of sensing the static response, dynamic characteristics, and environmental effects of the structure. This provides a sufficient and reliable raw data foundation for subsequent data fusion analysis and drift correction.

[0043] The data acquisition module employs multi-channel synchronous sampling technology to ensure timestamp alignment of data from different sensors, and the sampling frequency is adaptively adjusted according to the dynamic response characteristics of the structure.

[0044] It should be further explained that the data acquisition module adopts a multi-channel synchronous sampling technology based on precision clock synchronization. By integrating a high-stability temperature-compensated crystal oscillator as the system clock source in the main control unit, and using the IEEE 1588 precision clock protocol or equivalent hardware trigger signal, the sensor acquisition channels distributed in different physical locations are ensured to achieve microsecond-level time synchronization, thereby ensuring that all sensor data have strictly aligned timestamps, laying the foundation for timing consistency for subsequent multi-sensor data fusion analysis. Meanwhile, the module features adaptive sampling frequency adjustment. Its built-in intelligent sampling control unit continuously monitors the dynamic response characteristics of the structure. By calculating the energy spectral density of the dominant frequency component of the accelerometer signal in real time, when the structure is identified as being in a stable state and the response frequency components are mainly concentrated in the low-frequency band, the sampling frequency is automatically adjusted to the lowest effective frequency that satisfies the Nyquist sampling theorem to reduce data volume and power consumption. When the structure is detected to have a significant dynamic response due to excitations such as wind load, traffic vibration, or earthquake, the sampling frequency is immediately dynamically increased to more than five times the dominant frequency based on the detected dominant frequency components, ensuring that the transient response characteristics of the structure can be fully captured without aliasing distortion. This sampling strategy optimization based on the actual dynamic characteristics of the structure not only ensures the complete acquisition of key data but also improves the overall energy efficiency of the system, providing high-quality and effective data input for the back-end processing system.

[0045] The results output module provides a graphical interface that displays the secondary calibration data, data quality confidence report, and structural health status assessment results in real time. When a verified structural anomaly is detected, a warning message is issued through an audible and visual alarm and a remote communication interface.

[0046] It should be further explained that the results output module provides a graphical user interface with multi-layered visualization areas. The main display area dynamically displays the high-confidence secondary calibration data after secondary verification by overlaying trend curves. At the same time, it displays the level information of the data quality confidence report in the form of background coloring or transparency changes. In the structural health status assessment area, the system uses a traffic light-style color coding system to intuitively display the safety status of each key part of the building structure based on the fusion analysis results of the calibrated multi-sensor data. When the drift arbitrator confirms that a certain sensor data or a group of sensor data reflects a real structural anomaly and exceeds the safety threshold, the system immediately activates a multi-level alarm mechanism. The local device emits alternating audible and visual alarm signals at a specific frequency. At the same time, it sends structured alarm information containing the location, type, severity and timestamp of the anomaly to the preset management terminal through the integrated wired or wireless remote communication interface. This module also provides data traceability, allowing users to click on any sensor data point to view its complete processing history, including raw readings, calibration data at all levels, consensus estimates, and arbitration decision paths, thereby providing comprehensive, transparent, and intuitive information support for structural safety assessment and decision-making.

[0047] In the arbitration process, the drift arbitrator also incorporates the sensor's historical reliability score as a weighting factor. The historical reliability score is dynamically updated based on the frequency at which the sensor has been marked as drifting in the past and the accuracy of the arbitration results.

[0048] To clarify the quantitative standards and update mechanism for historical reliability scoring, and to ensure the objectivity and operability of the weighting factors, the specific definitions are as follows: The core metrics are defined as follows: Drift frequency F d The ratio of the number of times a sensor is marked as a "suspected drift point" within a statistical period to the total number of sampling periods is given by the following formula: , where N drift N represents the number of suspected drift markers within the statistical period. total This represents the total number of sampling periods within the statistical period, which is set to 7 days by default, consistent with the model's self-update time window; F d ∈[0,1],F d The smaller the value, the lower the probability of sensor drift.

[0049] Arbitration accuracy A a The ratio of the number of times a sensor's arbitration result is subsequently verified as "correct" within a statistical period (e.g., manual review, structural state backtracking) to the total number of arbitrations is given by the following formula: , where N correct For the correct number of arbitrations, N arbit Total number of arbitrations; A a ∈[0,1],A a The larger the value, the higher the reliability of the sensor data.

[0050] The reliability scoring formula includes: historical reliability scores R are calculated using a weighted summation, with weighting coefficients determined using the Delphi method. The Delphi method involves evaluating the scores by five experts in the field of building structure monitoring, and taking the average. The drift frequency weight ω is used as an example. d =0.3, Arbitration accuracy weight ω a =0.7, the formula is: R=ω d ·(1-F d )+ω a ·A aR∈[0,1], R≥0.8 indicates a high reliability sensor, and R≤0.5 indicates a low reliability sensor. Low reliability sensors will trigger additional calibration checks.

[0051] The rating update rules include the following: Update cycle: Synchronized with the time window of the model self-update module, i.e., updated once every 7 days; Moving average processing: To avoid the impact of a single abnormal arbitration result on the score, an exponential moving average is used for updating, with the formula: R new =α·R curr +(1-α)·R last Where α=0.6 is the current period's scoring weight, R curr R is the score calculated for the current period. last This is a historical score from the previous period; Abnormal reset mechanism: If a sensor is replaced due to hardware failure, its historical reliability score is reset to the initial value of 0.8 and the statistics start again to avoid old failure data affecting the reliability assessment of the new sensor; hardware failures include disconnection and zero drift exceeding the threshold.

[0052] It should be further explained that the decision-making mechanism of the drift arbitrator enhances the scientific rigor and accuracy of the arbitration decision by introducing and dynamically maintaining a sensor historical reliability scoring system. This historical reliability score is calculated based on two core dimensions: first, the frequency with which the sensor is marked as a suspected drift point during long-term operation, i.e., the ratio of the cumulative number of markings to the total operating time; and second, the accuracy of the sensor's judgment in historical arbitration, i.e., the proportion of cases where the arbitrator ultimately determines it to be a true drift out of all marked cases.

[0053] The system quantifies and evaluates these two dimensions of each sensor at a fixed cycle, and calculates a current reliability score between zero and one using a weighted summation formula, where the weighting coefficients can be adjusted according to specific application scenarios. When the drift arbitrator performs three-party data comparison decisions, this score is transformed into a weighting factor that directly participates in the calculation of Euclidean distance and cosine similarity. Specifically, when calculating distance and similarity, the influence of data from sensors with high reliability scores in the decision-making process is amplified, while the reference value of data from sensors with low scores is appropriately suppressed.

[0054] This dynamically weighted decision-making mechanism enables the system to learn and optimize its judgment preferences based on the long-term performance of sensors. For example, for a sensor that has been repeatedly proven to have drift problems, even if its current reading differs from the consensus estimate, the system will be more inclined to adopt the corrected data; conversely, for a sensor with a good historical record, its original reading will receive a higher trust weight. The implementation of this technical solution improves the system's intelligence in distinguishing between real structural changes and sensor drift in complex environments, giving the entire monitoring system the ability to self-optimize based on historical experience.

[0055] The model self-update module also includes a model health monitoring submodule, which periodically injects test data with known drift patterns into the simulation environment to verify the model correction performance. If the correction error continues to exceed the tolerance range, the model reconstruction process is triggered.

[0056] It should be further explained that a model health monitoring submodule has been added to the model self-update module. This submodule continuously evaluates and ensures the correction performance of the digital twin model by constructing a closed-loop test environment. Specifically, this submodule periodically injects a series of predefined test data sequences with known drift modes and amplitudes into the currently serving digital twin model in the background simulation environment. These test data simulate various typical fault scenarios such as common sensor zero-point drift, sensitivity drift, and nonlinear drift. The model health monitoring submodule then records and analyzes the correction output of the digital twin model for these injected drifts. By calculating the error index between the corrected data and the known drift-free baseline, it quantitatively evaluates the model's performance in the current state. If the correction error of the model for one or more typical drift modes continues to exceed the system's preset tolerance range in multiple consecutive test cycles, the model health is determined to be insufficient and cannot meet the monitoring accuracy requirements, and the model reconstruction process is automatically triggered. This reconstruction process is not a simple parameter reset, but rather a partial backtracking to the initial training phase of the virtual calibration twin building module. The system will call the latest batch of high-confidence field data stored in the system, and combine it with the expanded injected test dataset to retrain and optimize the structure and parameters of the physical information neural network until the new model shows stable and qualified performance in the test environment. Only then can it be redeployed to the online monitoring system, thereby ensuring the long-term reliability and accuracy of the core algorithm of the entire monitoring system without human intervention.

[0057] This system effectively solves the problem of data drift caused by environmental fluctuations by constructing a virtual calibration twin that interacts in real time with the physical sensor network and establishing a multi-layered data verification mechanism. During online monitoring, the system can automatically identify and correct abnormal deviations in sensor data. Through primary screening and secondary consensus verification, it ensures the accuracy and reliability of the output data, thereby improving the credibility of the assessment of the building's structural health status.

[0058] This system possesses self-learning and continuous optimization capabilities. Through a model self-updating mechanism, the digital twin model can adapt to the long-term performance evolution of building structures. This feature not only reduces reliance on external manual calibration and maintenance costs, but also enables timely identification of real structural state changes and sensor malfunctions, providing stable and reliable technical support for safety early warning and long-term performance evaluation of building structures.

[0059] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0060] 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 building structure health monitoring system based on multi-sensor fusion, comprising a sensor network consisting of various types of sensors deployed at key parts of the building structure, a data acquisition module, a data processing center, and a result output module, wherein the data acquisition module is used to collect raw monitoring data from the sensor network, the data processing center performs fusion analysis on the collected data, and the result output module is used for display and alarm functions, characterized in that, The data processing center further includes a virtual calibration twin construction module, a multi-layer data verification module, and a model self-updating module; During the system initialization phase, the virtual calibration twin construction module establishes an initial benchmark mapping relationship library describing the inherent physical correlation between the readings of each sensor based on the multi-sensor benchmark data of the building structure under no-load and environmentally stable conditions. This mapping relationship library is obtained through machine learning model training and is used to simulate the sensor data interaction mode of the building structure under healthy conditions. The multi-layer data verification module includes a primary drift screening unit and a secondary consensus verification unit. The primary drift screening unit receives real-time sensor data and environmental parameters, calculates the theoretically expected correlations by embedding a physical information neural network with simplified mechanical equations of the building structure, generates primary correction data, and marks suspected drift points. The secondary consensus verification unit calls the initial benchmark mapping relationship library and combines it with the current reliable sensor data. It uses a graph signal processing algorithm to reconstruct the consensus estimate of the suspected measurement points. It compares the original readings, primary correction data, and consensus estimate through an arbitration mechanism to output high-confidence secondary fine-calibration data. The model self-updating module feeds back the secondary fine-calibration data and the corresponding data quality confidence report to the physical information neural network to realize online incremental learning and adaptive optimization of the model.

2. The building structure health monitoring system based on multi-sensor fusion according to claim 1, characterized in that: The virtual calibration twin construction module performs the following steps when establishing the initial benchmark mapping relationship library: Under the baseline conditions of no external load on the building structure and stable ambient temperature and humidity, raw data from all sensors are collected simultaneously for multiple consecutive cycles. Extract the time-domain and frequency-domain features of each sensor data, and calculate the correlation coefficient, covariance matrix, and transfer function between different sensor data. The above features and relational parameters are input into a deep spatiotemporal graph neural network for training. The network takes environmental temperature and humidity data as input and the physical constraint relationship between sensor readings as the hidden layer supervision signal, and learns a digital twin model that can predict the expected readings of each sensor under given environmental conditions. The trained digital twin model and its parameters are stored as an initial benchmark mapping relationship library, which is used for real-time data comparison and drift detection during the online monitoring phase.

3. The building structure health monitoring system based on multi-sensor fusion according to claim 1, characterized in that: The physical information neural network in the primary drift screening unit adopts a deep learning architecture constrained by partial differential equations. The forward propagation process of this network integrates the constitutive equation of the building structure and the spatial topological relationship of the sensor layout. The network input layer receives real-time sensor data vectors and environmental temperature and humidity vectors. The hidden layer learns the nonlinear mapping between environmental parameters and sensor readings through residual connections. The output layer generates the theoretical expected values ​​of each sensor. The difference analysis submodule calculates the residual sequence between the original readings and the theoretical expected values, and performs time-series smoothness and correlation analysis on the residuals. If the sliding variance of the residuals exceeds the preset threshold or shows a strong correlation with environmental changes, the data point is marked as a suspected drift point. The compensation generation submodule generates dynamic compensation coefficients for marked suspected drift points based on the magnitude and direction of the residuals, performs linear or nonlinear transformations on the original data, and generates first-level correction data.

4. A building structure health monitoring system based on multi-sensor fusion according to claim 1, characterized in that: The execution process of the secondary consensus verification unit includes: The graph signal processing algorithm uses the sensor network topology as a graph structure and reliable sensor data as node signals. It reconstructs the consensus estimate of the suspected measurement point in the current state through graph Laplace regularization. The drift arbitrator uses a majority decision mechanism to compare the Euclidean distance and cosine similarity of the included angle between the original readings of the suspected point, the first-level correction data, and the consensus estimate. If the similarity between the first-level correction data and the consensus estimate is higher than the similarity between the two and the original reading, then the point is determined to be a true drift, and the first-level correction data is adopted as the output. If the original reading has the highest similarity to the consensus estimate, then the point is determined to be likely to reflect the actual structural change, the original reading is retained, and a structural anomaly warning is triggered. The arbitration result, along with the confidence score, is recorded in the data quality confidence report.

5. A building structure health monitoring system based on multi-sensor fusion according to claim 1, characterized in that: The online incremental learning process of the model self-updating module is as follows: At fixed time windows, data samples with confidence levels higher than a threshold are selected from the secondary refined data to form a micro-batch training set. Using these high-confidence samples as supervision signals, the gradient descent method is used to fine-tune the weight parameters of the physical information neural network. The adjustment range is controlled by the learning rate and the confidence weighting factor. Meanwhile, the key relation parameters in the initial baseline mapping relation library are updated by moving average based on the latest data distribution, enabling the digital twin model to track the long-term performance evolution of the building structure. The model version management submodule retains historical model parameters. If the validation performance degrades over several consecutive periods, it will automatically roll back to the previous stable version.

6. A building structure health monitoring system based on multi-sensor fusion according to claim 1, characterized in that: The various types of sensors include strain sensors, acceleration sensors, temperature sensors, and humidity sensors, which are fixed to the beams, columns, nodes, and foundations of the building structure in a spatially distributed manner.

7. A building structure health monitoring system based on multi-sensor fusion according to claim 1, characterized in that: The data acquisition module employs multi-channel synchronous sampling technology to ensure that the timestamps of data from different sensors are aligned, and the sampling frequency is adaptively adjusted according to the dynamic response characteristics of the structure.

8. A building structure health monitoring system based on multi-sensor fusion according to claim 1, characterized in that: The result output module provides a graphical interface that displays the secondary calibration data, data quality confidence report, and structural health status assessment results in real time. When a verified structural abnormality is detected, a warning message is issued through an audible and visual alarm and a remote communication interface.

9. A building structure health monitoring system based on multi-sensor fusion according to claim 4, characterized in that: In the arbitration process, the drift arbitrator also incorporates the sensor's historical reliability score as a weighting factor. The historical reliability score is dynamically updated based on the frequency at which the sensor has been marked as drifting in the past and the accuracy of the arbitration results.

10. A building structure health monitoring system based on multi-sensor fusion according to claim 5, characterized in that: The model self-update module also includes a model health monitoring submodule, which periodically injects test data of known drift patterns into the simulation environment to verify the model correction performance. If the correction error continues to exceed the tolerance range, the model reconstruction process is triggered.

Images