A method and system for monitoring the safety of an existing building structure

Abstract

The application relates to a kind of existing building structure safety monitoring method and system, system is distributedly laid multiple monitoring instruments in building key component and important position, monitoring instrument includes: the real-time acquisition module of multi-source monitoring data, the high-precision data synchronization time module between monitoring instrument, and the edge computing module of data preprocessing and fusion processing.Therein, data fusion is based on the multi-source data after synchronization, the neural network of data-driven is used as model error compensator, and is combined with Kalman filter, realizes multi-source data fusion and estimates the physical quantity not monitored;Monitoring instrument is connected cloud platform through wireless communication, and cloud receives multi-point monitoring data and edge computing result, carries out existing building structure safety monitoring and evaluation.Compared with prior art, the application has the advantages of being able to realize microsecond-level data synchronization accuracy and efficiently estimating difficult-to-directly-observe physical parameters, and improves the precision and reliability of existing building structure safety monitoring.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of structural safety monitoring, and particularly relates to a method and system for safety monitoring of existing building structures such as high-rise buildings and large-span structures. BACKGROUND

[0002] With the acceleration of urbanization, especially the safety monitoring demand of existing structures such as high-rise buildings and large-span structures is increasing. Traditional monitoring methods often use single sensors or lack multi-device synchronous acquisition capability, which is difficult to meet the demand of multi-physical quantity synchronous monitoring of structures. With the increase of building height, the traditional structure safety monitoring equipment gradually exposes the deficiency, and the existing technology generally has the following problems:

[0003] 1. Heavy monitoring equipment, difficult to deploy: traditional monitoring equipment is often bulky, difficult to install, and the wiring between devices is complex, increasing the installation difficulty and maintenance cost.

[0004] 2. Asynchronous data acquisition: the existing equipment has precision problems in the synchronization between multiple devices, especially when multiple devices span a large range, it is difficult to ensure high-precision data synchronous acquisition.

[0005] 3. Single monitoring method: the existing equipment usually relies on a single sensor type, which cannot comprehensively monitor various physical quantities and states of building structures, resulting in a blind area of monitoring.

[0006] 4. Slow response: most existing systems rely on cloud data processing, resulting in large data upload delay, affecting real-time monitoring and timely evaluation and early warning effect.

[0007] The Chinese patent CN113126143B discloses a portable micro-vibration and strong-vibration real-time monitoring system. The system is composed of a monitoring end and a communication base station, and uses ultra-low frequency vibration acceleration sensors and high-precision three-axis intensity sensors to collect signals from the initial stage to the whole process of destruction of disaster events, thereby realizing long-distance and wide-range vibration monitoring in fields such as slope, mine fracture zone, and geological disasters. However, the monitoring instrument proposed in this patent still has some shortcomings. First, the monitored physical quantity is relatively single, only focusing on the collection of acceleration signals, and lacking the ability to perceive multiple parameters such as structure inclination, which limits the comprehensiveness of monitoring. Second, the system architecture is relatively complex, and the synchronization capability is not ideal. The monitoring end and the communication base station need to be deployed separately, and a communication base station and a support system need to be specially constructed, which makes the system deployment cost high and the flexibility poor, especially in limited space environments such as buildings, it is difficult to quickly build a dense monitoring network. Moreover, although the system can collect the whole process of disaster signals, it does not consider how to use local computing resources at the monitoring end to preliminarily process data, and all raw data must be transmitted back to the base station or data center, which increases the communication bandwidth and the pressure of cloud processing, making it difficult to meet the monitoring scenarios with high real-time requirements.

[0008] In order to overcome the above problems, there is an urgent need for a monitoring method and system with higher integration, synchronization accuracy and real-time response capability to adapt to the safety monitoring of high-rise buildings and large-span structures. SUMMARY

[0009] The purpose of the present application is to provide a method and system for monitoring the safety of existing building structures, which can be quickly deployed in complex environments and has higher integration, synchronization accuracy and real-time response capability, especially in the safety monitoring of high-rise buildings and large-span structures, and has stronger adaptability.

[0010] The purpose of the present application can be achieved by the following technical solutions:

[0011] A method for monitoring the safety of existing building structures, comprising the following steps:

[0012] Real-time acquisition of multi-source data collected by multiple monitoring instruments distributed in key components and important positions of the building, and data preprocessing and data synchronization;

[0013] Based on the synchronized multi-source data, a data-driven neural network is used as a model error compensator combined with an unscented Kalman filter for multi-source data fusion to estimate unmonitored physical quantities, and the collected data is combined for safety monitoring of existing building structures.

[0014] The method for data synchronization is:

[0015] Initial data synchronization is performed using GPS or NTP / PTP protocols;

[0016] The initially synchronized data is constructed into an input feature vector with spatial topological characteristics. This vector is then input into a graph convolutional neural network with gated attention for topological data processing. Nodes in the graph convolutional neural network update their own features by aggregating information from their neighbors. The aggregation weights are determined by the gated attention mechanism. The output of the gated attention mechanism is input into a long short-term memory module to capture the temporal patterns of the features. The output of the long short-term memory module is then input into an uncertainty-driven adaptive fusion module, which outputs the optimal clock offset estimate. High-precision data synchronization is then performed based on this optimal clock offset estimate.

[0017] The input feature vector is represented as follows:

[0018] ,

[0019] in, Indicates the first Each monitoring node The input feature vector corresponding to the data collected at each moment; This is the current monitoring node. exist The clock offset calculated using the PTP protocol at all times; Is the current monitoring node in Network latency measured at any given time; It is with nodes The set of all neighboring nodes that communicate directly. It is a node Receive from node Signal strength indicator It is a node To the node The number of jumps.

[0020] The gated attention mechanism dynamically controls the degree of information fusion through a gating vector:

[0021] ,

[0022] ,

[0023] ,

[0024] in, It is a node Aggregate neighbor nodes Attention weights at the time It is a leaky linear rectifier function. It is a trainable attention vector. It is a trainable weight matrix. It is a node The characteristics of the monitoring instrument are indicated. It is a node neighboring nodes The characteristics of the monitoring instrument are indicated. It is a learnable gating vector used to control whether to fuse information from neighboring nodes or retain its own information. It is a trainable gated weight matrix. This represents vector concatenation. It represents the Hadamah accumulation. Nodes that output the gating attention mechanism Its characteristics.

[0025] The Long Short-Term Memory module performs the following operations on its input:

[0026] ,

[0027] ,

[0028] in, Nodes that output the gating attention mechanism exist Characteristics of time, Indicates in Time of the first The hidden state of each node after processing by the Long Short-Term Memory module. Indicates in Time of the first The cell state of each node after processing by the Long Short-Term Memory module. It is a trainable weight matrix used in the output layer of the Long Short-Term Memory module. This refers to the historical offset sequence of the Long Short-Term Memory (LSTM) module. This indicates that the Long Short-Term Memory (LSTM) module is processing the data.

[0029] The adaptive fusion module performs the following operations to output the optimal clock offset estimate:

[0030] ,

[0031] ,

[0032] ,

[0033] ,

[0034] in, Variance of the PTP protocol is the mean square error of the long short-term memory module prediction result, is a variance function, and MSE is a mean square error function; refers to a historical offset sequence of the PTP protocol, refers to a historical offset sequence of the long short-term memory module, is the size of the sliding window, is a fusion weight, is the final estimation value of the clock offset of the th node at the th moment.

[0035] The preprocessing specifically comprises: adopting a method based on Fourier transform or wavelet transform to perform noise reduction and feature extraction on the collected data.

[0036] The multi-source data fusion comprises the following steps:

[0037] A state space model is constructed, and a system state vector is defined. The state vector includes an unknown physical quantity to be solved at the th moment and a low-frequency bias of a sensor at the th moment.

[0038] The evolution of the state over time is described by a state equation, and a light first neural network is introduced to learn the error after updating the state vector, which is defined as:

[0039] ,

[0040] wherein, is a state transition matrix, is process noise, is an observation vector, which is collected from the sensor;

[0041] The process of mapping the state to the observation value is described by an observation equation, and a light second neural network is introduced to learn the error after updating the observation vector:

[0042] ,

[0043] ,

[0044] wherein, is observation noise, is the height of the installation position of the monitor, is a displacement state, is an acceleration state, and Low-frequency deviations between the accelerometer and tilt sensor at all times;

[0045] Data fusion is performed using unscented Kalman filtering. First, an unscented transformation is performed, and a set of Sigma points is selected to capture the mean and covariance of the states.

[0046] ,

[0047] in, It is a set matrix of Sigma points. Is The optimal state estimate is obtained by updating the unscented Kalman filter at each time step. Is Covariance at time, It is a scaling parameter;

[0048] The prediction step of the unscented Kalman filter is shown below:

[0049] ,

[0050] ,

[0051] ,

[0052] in, Is The Sigma point in the state space at time step. Is The Sigma point in the state space at time step has been propagated to time, It is an unscented Kalman filter based on Information about time The optimal estimate of the system state at time t. It is the dimension of the state vector. It is the first The mean weight of each Sigma point Is The first time in the state space Sigma points and have been spread to time, It is the first Covariance weights of Sigma points It is the process noise covariance matrix;

[0053] The update steps for the unscented Kalman filter are shown below:

[0054] ,

[0055] in, is the predicted observation value corresponding to the Sigma point;

[0056] The mean, covariance and cross-covariance of the predicted value are calculated:

[0057] ,

[0058] ,

[0059] ,

[0060] where, is the predicted mean of the observation value, is the predicted observation value corresponding to the th Sigma point, is the covariance matrix of the predicted observation value, is the covariance matrix of the observation noise, is the cross-covariance matrix of the state and the observation value;

[0061] The Kalman gain is calculated and the state is updated:

[0062] ,

[0063] ,

[0064] ,

[0065] where, is the Kalman gain value, is the posterior estimation mean of the state vector at time , is the posterior covariance matrix of the state estimation at time is the covariance matrix of the predicted observation value.

[0066] The loss function of the two lightweight neural networks and introduced is defined as:

[0067] ,

[0068] where, is the observation model mapping to the observation space, is the next state predicted by the state transition model, and are the parameters of the neural networks and , , , This is a regularization hyperparameter of the loss function, used to balance the weights of each error term in the loss function;

[0069] Neural networks obtain state vectors by minimizing a loss function. , The first element in That is The unknown values ​​are estimated at any given time through data fusion.

[0070] A safety monitoring system for existing building structures, used to implement the method, the system comprising:

[0071] Multiple monitoring devices are distributed in key building components and important locations. Each monitoring device includes a data acquisition module, an edge computing module, and a data synchronization module. The data acquisition module is used to acquire multi-source data in real time. The edge computing module is used to perform data preprocessing and data fusion processing. The data fusion processing is based on synchronized multi-source data, using a data-driven neural network as a model error compensator, combined with an unscented Kalman filter to perform multi-source data fusion and estimate unmonitored physical quantities. The data synchronization module is used to synchronize the data acquired by the acquisition modules of different monitoring devices.

[0072] In addition, a cloud platform is used to combine estimated unmonitored physical quantities with collected data for existing building structural safety monitoring and to provide visualization.

[0073] The cloud platform and the monitoring instrument transmit data wirelessly.

[0074] Compared with the prior art, the present invention has the following beneficial effects:

[0075] (1) This invention adopts a multi-level synchronization method (GPS timing + hardware triggering + PTP / NTP protocol), and on this basis, an adaptive synchronization compensation algorithm based on topology perception and gating attention mechanism is used. This algorithm constructs the data acquisition synchronization problem as a graph neural network learning task, dynamically fuses multi-source information by utilizing the topological relationship between nodes and the gating attention mechanism, and can adaptively weight the protocol synchronization result and prediction result through the uncertainty of the data. Thus, it accurately compensates for the complex drift of the node crystal oscillator and the dynamic network delay, and stably improves the synchronization accuracy of data acquisition between multiple devices to the microsecond level. It solves the data time difference problem caused by transmission delay in traditional distributed monitoring systems, and provides a reliable data foundation for the overall modal analysis and damage location of the structure.

[0076] (2) The application introduces a hybrid Kalman filter multi-source data fusion framework based on physical information neural network, which not only realizes the synchronous collection of existing physical quantities, but also estimates the physical parameters that are difficult to directly observe with high precision by online learning and compensating for physical model errors. This breaks through the limitation of traditional monitoring equipment that can only obtain a single monitoring quantity and is easily affected by model errors, and realizes the breakthrough from simple monitoring to deep perception, providing a deeper data dimension for accurate diagnosis of the safety state of the structure.

[0077] (3) In order to improve the system response speed, the edge computing module and the cloud analysis are cooperated in the monitoring instrument, which can complete data preprocessing and intelligent calculation locally at the monitoring end, the system response is fast, and the safety state of the building structure can be real-time warned and evaluated.

[0078] (4) The application realizes lightweight wireless deployment, reduces implementation cost and interference. The device is small in size and light in weight, and the wireless communication mode eliminates the complex wiring engineering, the deployment efficiency is greatly improved compared with traditional equipment, and there is almost no interference to the use function of the structure, which is suitable for structure monitoring in complex and sensitive environment.

[0079] (5) The application can adapt to various complex application scenarios. The integrated monitoring device is not only suitable for existing structures such as high-rise buildings and large-span structures, but also can be widely used in structure safety monitoring in various complex environments.

[0080] The application realizes the deep integration of innovative algorithm kernel (multi-source data fusion, synchronous accuracy compensation) and efficient integrated hardware system, and builds a new generation of lightweight distributed intelligent monitoring system. It breaks through the bottleneck of high deployment cost, large response delay and data isolation of traditional system, optimizes the complete process from field perception to data transmission cooperation to cloud intelligent decision-making, and can provide strong technical support for safety evaluation and maintenance of building structure. BRIEF DESCRIPTION OF DRAWINGS

[0081] Figure 1 It is a system structure schematic diagram of the application;

[0082] Figure 2 It is a schematic diagram of the composition of the acquisition module;

[0083] Figure 3 It is a schematic diagram of the implementation process of the method of the application. DETAILED DESCRIPTION

[0084] The application will be described in detail below in combination with the drawings and specific embodiments. The embodiments are implemented on the premise of the technical scheme of the application, and detailed implementation methods and specific operation processes are given, but the protection scope of the application is not limited to the following embodiments.

[0085] Example 1

[0086] The embodiment first provides a kind of existing building structure safety monitoring system, it is suitable for long-term safety monitoring of existing building such as high-rise building, large-span structure etc..Traditional monitoring equipment often faces the problems such as bulky, wiring complex, poor data synchronization accuracy, slow response etc..In order to solve these problems, an innovative solution is proposed: monitoring instrument adopts compact and light design, small in size, light in weight, can be quickly deployed and does not interfere with structure.Various sensors are built-in monitoring instrument, can synchronously collect multiple physical quantities including vibration, inclination etc..Its data acquisition system combines multiple synchronization technologies such as global positioning system (GPS), high-precision time synchronization protocol (PTP) and network time protocol (NTP), and also adopts adaptive synchronization accuracy compensation algorithm, to ensure that multiple devices achieve microsecond-level high-precision data synchronization.Monitoring instrument also has built-in edge computing module, can carry out real-time intelligent calculation at data acquisition end.Through introducing multi-source data fusion algorithm, equipment can calculate unmonitored physical quantity according to the data collected by sensor.Processed data is uploaded to cloud platform in real time through 4G or Ethernet, for storage, depth analysis and visual display, user can remotely master the safety condition of structure at any time, and carries out performance evaluation.Compared with prior art, the present application combines light weight design, multi-parameter monitoring, high synchronization accuracy and edge intelligent processing, provides an efficient, accurate and real-time intelligent monitoring solution, significantly improves the reliability and efficiency of structure safety monitoring in complex environment.

[0087] As shown in Figure 1 The system comprises:

[0088] A plurality of monitoring instruments are distributed in key components and important positions of building, and the monitoring instrument comprises an acquisition module, an edge computing module, a data synchronization module and a data fusion module, wherein the acquisition module is used for real-time acquisition of multi-source data, and the edge computing module is used for data preprocessing and data fusion processing, wherein the data fusion processing is based on synchronized multi-source data, uses data-driven neural network as model error compensator, combines with unscented Kalman filter, carries out multi-source data fusion, estimates unmonitored physical quantity, and the data synchronization module is used for data synchronization of data collected by acquisition module of different monitoring instruments.

[0089] And cloud platform is used for existing building structure safety monitoring in combination with estimated unmonitored physical quantity and collected data, and provides visual display.

[0090] Wherein, data transmission is carried out between cloud platform and monitoring instrument by wireless communication mode.

[0091] In particular, the acquisition module adopts a standardized, modular sensor interface and data acquisition architecture. In a typical embodiment, as shown in FIG. 3, the acquisition module integrates a high-precision acceleration sensor, a high-resolution tilt sensor, and a data acquisition module, etc. Meanwhile, the architecture reserves an expansion interface and protocol, and can be compatible with and integrate other types of sensors (such as strain gauges, temperature sensors, etc.), to achieve comprehensive and synchronous monitoring of multiple physical quantities of the building structure. Figure 2

[0092] The specific processing procedures of the edge computing module and the data synchronization module are described in the method embodiments below.

[0093] In a preferred embodiment, the lightweight distributed synchronous intelligent monitoring instrument, in addition to the standard acceleration and tilt sensors, can also flexibly select and add other types of sensor modules according to specific monitoring requirements, such as shown in FIG. 4. The monitoring instrument includes: Figure 1

[0094] (1) The acceleration sensor is used to monitor the vibration characteristics of the building structure in various directions. It is fixed on the sensor mounting plate through bolts to ensure that it can stably monitor the vibration of the structure. The acceleration sensor can select a high-precision low-frequency vibration pickup or a three-axis MEMS accelerometer. The selection of these sensors can be based on the monitoring requirements, the type and vibration characteristics of the building structure, and the monitoring scene, etc. Each acceleration sensor connects its standard analog output port to the corresponding input channel of the data acquisition module through a shielded cable. The data acquisition module converts the analog signal output by the acceleration sensor into a digital signal through the built-in ADC, for subsequent processing.

[0095] The signal output by the acceleration sensor can be represented as:

[0096]

[0097] wherein, is the amplitude of the vibration, is the frequency of the vibration, is the time, is the phase angle. The acceleration sensor will output the corresponding signal according to the vibration frequency response of the building, to ensure accurate monitoring of the dynamic response of the building structure.

[0098] (2) The tilt sensor is used to monitor the inclination characteristics of the key components of the building structure, especially the small changes of the lateral resisting members. The tilt sensor is fixed on the mounting plate by bolts and kept horizontal to ensure that it can stably monitor the inclination characteristics of the structure. The tilt sensor needs to have high resolution, and its technical parameters are determined according to the monitoring accuracy requirements. The tilt sensor transmits the voltage signals and ​​Respectively connected to the analog input end of the data acquisition module corresponding to the sensor. The data acquisition module samples and converts the analog signal of the tilt sensor through the internal ADC, digitizes the signal and inputs it into the processing module for further processing and output.

[0099] The relationship between the output signal of the tilt sensor and the tilt characteristics of the building structure can be expressed as follows:

[0100]

[0101] Where, and are the voltage outputs of the sensor in the Y and X axis directions respectively. The tilt angle output by the tilt sensor reflects the change in the tilt characteristics of the building, suitable for monitoring small structural tilts.

[0102] (3) The data acquisition module has multi-channel, 24-bit data acquisition capability and edge computing function, responsible for collecting sensor data and performing preliminary processing. The analog input end of the data acquisition module is connected to the output end of the sensor, and the data signal is transmitted to the ADC module of the data acquisition module through the transmission line for digital processing.

[0103] (4) In order to ensure high-precision distributed synchronous data acquisition in the case of multiple monitors working together, the device is equipped with a GPS antenna for receiving satellite signals. The GPS antenna is installed on the top of the monitor shell, ensuring that the antenna can receive satellite signals and avoiding the influence of signal reception due to the shielding of the device shell or other objects. The GPS antenna is connected to the built-in GPS receiving module through a radio frequency coaxial cable. The cable passes through the special hole of the device shell and is connected to the GPS interface in the data acquisition module.

[0104] (5) The main power supply of the monitor is 220V AC, which provides all the power required for the daily operation of the device. In order to ensure that the monitor can continue to work when the main power is cut off, a backup power supply is built into the device. This backup power supply is a set of high-performance lithium batteries that can provide stable power supply when the main power is disconnected. The specific design is as follows:

[0105] When the main power is cut off, the backup power supply will automatically access the power supply of the monitor. The backup power supply outputs a stable voltage through a DC-DC converter, ensuring that all critical components (such as sensors, data acquisition modules, communication modules, etc.) can continue to operate. The capacity of the backup power supply is designed to ensure that the device can continue to work for more than 12 hours without the main power supply.

[0106] The standby power switching mechanism adopts an automatic switching mode. When the 220V main power is detected to be disconnected, the standby power is automatically switched to ensure the continuous operation of the monitoring instrument without manual intervention. This design effectively avoids problems such as loss of monitoring data or system downtime caused by power interruption.

[0107] (6) All sensors, data acquisition modules, and standby power supplies are integrated into a waterproof housing with an IP67 protection rating, ensuring that the device can operate stably in harsh environments. The housing needs to have high rigidity to ensure that the device can collect the most realistic structural responses. The overall design of the device is compact, and the housing is equipped with hanging ear designs to facilitate installation.

[0108] The monitoring instrument in the system adopts a compact and lightweight design: the monitoring instrument is small in size and light in weight, making it easy to install and deploy without interfering with the normal use of the building structure. The device is equipped with hanging ear round holes, which are installed through expansion bolts, making the installation process quick and simple, and suitable for structural monitoring in complex environments.

[0109] Moreover, the system has a built-in 4G communication module that can upload the collected data to the cloud platform in real time through the 4G network. The cloud platform stores, analyzes, and visualizes the data in real time. The monitoring data is displayed through a graphical interface, making it easy for operators to view the safety status of the building structure in real time.

[0110] In a preferred embodiment, the system can also monitor the acceleration and inclination data of the structure in real time. When the data exceeds the preset threshold, the early warning mechanism can be automatically triggered to send an early warning signal in a timely manner. Through real-time analysis by the cloud platform, real-time evaluation of the structural safety can be provided, and a corresponding structural state evaluation report can be generated.

[0111] Embodiment 2

[0112] This embodiment is based on the system of embodiment 1 and provides a method for monitoring the safety of existing building structures, as shown in Figure 3 The method comprises the following steps:

[0113] S1, determine the required monitoring physical quantities according to the monitoring target, and select the sensors of the monitoring instrument. In a basic configuration, the device hardware mainly consists of acceleration sensors, inclination sensors, data acquisition modules, etc. At the same time, the architecture reserves extension interfaces and protocols, which can be compatible with and integrated with other types of sensors (such as strain gauges, temperature sensors, etc.).

[0114] S2, integrate all hardware, communication modules, built-in power supplies, etc. into a waterproof box with high rigidity and high protection performance, complete the connection and assembly of various components and modules in the box, and ensure that the monitoring instrument can work stably.

[0115] S3, After the completion of the integration and assembly of the monitoring instrument components, it is necessary to find a suitable location in the building structure for installation. During the on-site installation process, it is necessary to ensure a solid and rigid connection between the equipment and the ground to avoid errors in monitoring data due to vibration or tilting.

[0116] S3 specifically includes the following steps:

[0117] S31, Arrange multiple devices at key points in the building structure, such as beams, columns, and support points, to ensure that the devices can effectively monitor the dynamic response of the building structure.

[0118] S32, When fixing the device, ensure that the bottom of the device is in direct contact with the concrete layer to obtain the most accurate monitoring data. If there is a decorative layer such as floor tiles that interfere with the direct contact of the device with the concrete layer, consider using a high-vibration coupling agent between the monitoring instrument and the floor tiles, or use a cutting machine to cut the floor tiles or concrete, expose the concrete structure layer, and use an angle grinder and sandpaper to smooth it.

[0119] S33, Use an impact drill hammer to drill holes at the device hanging ear position, with a hole depth of about 40mm and a hole diameter of 8mm, and use M6x40 expansion bolts for fixation. The expansion bolts are made of 304 stainless steel to ensure that the device is firm and durable.

[0120] S34, After the completion of the device installation, check the stability of the device. If there are gaps around the device, fill them with sand or quick-drying cement to reduce the impact of external interference on the monitoring results.

[0121] S4, After the completion of the installation of the monitoring instrument on the structure to be measured, enter the data collection phase. The monitoring instrument collects real-time vibration and inclination data of the building structure through acceleration sensors and inclination sensors. The data collection process is as follows:

[0122] S41, Sensor operation: The acceleration sensor and the inclination sensor begin to monitor the dynamic response of the building structure in real time. The acceleration sensor monitors the vibration level of the building structure and outputs an analog signal; the inclination sensor monitors the inclination characteristics of the building structure and outputs a voltage signal proportional to the inclination angle.

[0123] S42, Data collection and conversion: The data collection module samples the collected analog signals through the digital converter (ADC) and converts the analog signals to digital signals. The data of each sensor is independently collected and digitized.

[0124] S43, Data storage and preprocessing: The data collection module stores the collected real-time monitoring data in the local cache and prepares to upload to the cloud platform. The built-in edge computing module in the device can perform preliminary processing on the collected data, such as noise reduction and feature extraction.

[0125] S5, in order to ensure that multiple monitoring instruments can work together, the system adopts a distributed architecture, which can realize distributed synchronization between multiple devices, so as to realize high-precision data synchronization acquisition, and upload the data to the cloud platform in real time through the 4G communication module.

[0126] S5 specifically includes the following steps:

[0127] S51, each device transmits the collected data to the cloud platform through the built-in 4G communication module through wireless network. The cloud platform stores, analyzes and displays the safety state of the building structure through the visual interface.

[0128] S52, in order to ensure the distributed synchronization accuracy between multiple devices, GPS or NTP / PTP protocol is used for preliminary data synchronization.

[0129] In this embodiment, GPS synchronization mode is preferred to ensure that the timestamp accuracy of them reaches microsecond level and the data acquisition time consistency of all devices is ensured. If there is no GPS antenna or the device cannot receive stable GPS signal, PTP time synchronization can be performed by connecting time server.

[0130] For the data synchronization accuracy requirement between multiple devices, the following synchronization error formula can be used:

[0131]

[0132] Wherein, is the synchronization error between the whole device, is the number of devices to be synchronized, is the time synchronization error of the i device.

[0133] S53, in order to further optimize the data synchronization acquisition process between multiple devices, on the basis of GPS or NTP / PTP protocol synchronization, deep learning algorithm is introduced to predict the clock drift and network delay of each monitor and to dynamically compensate, so as to realize high-precision data acquisition synchronization. The algorithm steps are as follows:

[0134] The initially synchronized data is constructed into an input feature vector with spatial topological characteristics. This vector is then input into a graph convolutional neural network with gated attention for topological data processing. Nodes in the graph convolutional neural network update their own features by aggregating information from their neighbors. The aggregation weights are determined by the gated attention mechanism. The output of the gated attention mechanism is input into a long short-term memory module to capture the temporal patterns of the features. The output of the long short-term memory module is then input into an uncertainty-driven adaptive fusion module, which outputs the optimal clock offset estimate. Based on this optimal clock offset estimate, high-precision data synchronization is performed, thereby overcoming the instability of the hardware clock and achieving ultra-high precision synchronization.

[0135] Specifically, the first step is to construct an input feature vector with spatial topological characteristics. At any given moment, for the monitoring network, the first The monitoring devices are arranged in a certain number of positions, and an extended feature vector is constructed as the model input:

[0136] ,

[0137] in, Indicates the first Each monitoring node The input feature vector corresponding to the data collected at each moment; It is the current monitoring node exist The clock offset calculated using the PTP protocol at all times; Is the current monitoring node in Network latency measured at any given time; It is with nodes The set of all neighboring nodes that communicate directly. It is a node Receive from node Signal strength indicator It is a node To the node The number of jumps.

[0138] By incorporating the spatial topology information of the monitoring network as input features, the model can perceive the spatial relationships between various monitoring devices, rather than relying on isolated single-point data. The topological data is then processed using a graph convolutional neural network with gated attention. The graph convolution operation is as follows:

[0139] ,

[0140] in, It is a node In the Feature representation in layered networks Represents a node the degree of the node, is a trainable weight matrix, is a Sigmoid activation function.

[0141] The node updates its own features by aggregating the information of neighbor nodes, and the aggregated weights are determined by the gated attention mechanism. The gated attention mechanism dynamically controls the degree of information fusion through the gating vector, which can better improve the adaptability and robustness of the model to complex and dynamic networks. Its core formula is as follows:

[0142] ,

[0143] ,

[0144] ,

[0145] wherein, is the feature representation of the node aggregating neighbor nodes , is a leaky linear rectifier function, is a trainable attention vector, is a trainable weight matrix, is the feature representation of the node monitor at the node , is the feature representation of the node monitor at the neighbor node of the node is a learnable gating vector for controlling whether to fuse the information of neighbor nodes or to retain its own information, is a trainable gating weight matrix, denotes vector concatenation, denotes Hadamard product, is the feature of the node output by the gated attention mechanism.

[0146] The output of the gated attention mechanism is then input into a long short-term memory (LSTM) module to capture the temporal patterns of the features:

[0147] ,

[0148] ,

[0149] wherein, is the feature of the node output by the gated attention mechanism at time , denotes the th feature of the node at timeThe hidden state of the i-th node after processing by the long short-term memory module, represents the time at which the final estimate of the clock offset of the i-th node at the time The unit state of the i-th node after processing by the long short-term memory module, is a trainable weight matrix for the output layer of the long short-term memory module, refers to the historical offset sequence of the PTP protocol, represents the processing of the long short-term memory module.

[0150] Finally, the optimal clock offset estimation value is output by an adaptive fusion module driven by uncertainty:

[0151] ,

[0152] ,

[0153] ,

[0154] ,

[0155] wherein, the variance of the PTP protocol, is the mean square error of the prediction result of the long short-term memory module, is a variance function, and MSE is a mean square error function; refers to the historical offset sequence of the PTP protocol, refers to the historical offset sequence of the long short-term memory module, is the size of the sliding window, is the fusion weight, is the time at which the final estimate of the clock offset of the i-th node at the time

[0156] Here, an adaptive fusion method based on online uncertainty estimation is adopted, and the fusion weight is not fixed, but is dynamically adjusted according to the accuracy of the PTP protocol and neural network prediction. Finally, the timestamps of the nodes are dynamically adjusted according to , so as to realize higher-precision data acquisition synchronization.

[0157] S6, the edge computing module built-in the monitor can preprocess the collected data, including noise reduction, data compression and feature extraction, to reduce the computing burden of the cloud platform.

[0158] The edge computing module denoises the collected vibration and inclination data by Fourier transform and wavelet transform, and extracts the most important monitoring features. This process can significantly improve data transmission efficiency and reduce data redundancy.​

[0159] Fourier transform is used to convert time-domain signals into frequency-domain signals, which can help analyze the spectral distribution of signals. By performing Fourier transform on acceleration and inclination signals, their frequency components can be identified, and then noise can be removed or features can be extracted. The Fourier transform of acceleration and inclination signals is processed as follows:

[0160] ,

[0161] where, is the signal representation in the frequency domain, is the sampled signal in the time domain, is the frequency, is the sampling length of the signal, is the imaginary unit.

[0162] Wavelet transform is used to analyze different scale components of signals, which can effectively denoise and extract features of non-stationary signals. Unlike Fourier transform, wavelet transform can provide local characteristics of signals in both time and frequency domains, and the formula is as follows:

[0163] ,

[0164] where, is the discrete wavelet coefficient, is the discrete wavelet basis function, usually the scaling and translation of the mother wavelet function, and are the discrete indices of scale and position, is the number of sampling points of the signal.

[0165] S7, a multi-source data fusion algorithm is introduced in the edge computing module built-in the monitor, which realizes the estimation of unmonitored physical quantities using collected multi-source data. In this way, not only can the advantages of the monitor in collecting multi-physical quantity data be realized, but also the cost of obtaining unmonitored physical quantities through algorithms can be saved. Specifically, based on the synchronized multi-source data, a data-driven neural network is used as a model error compensator combined with an unscented Kalman filter to perform multi-source data fusion and estimate unmonitored physical quantities. Combined with the collected data, existing building structure safety monitoring can be performed.

[0166] The gating attention mechanism dynamically controls the degree of information fusion through the gating vector:

[0167] First, a state space model is constructed, and the system state vector is defined as :

[0168] ,

[0169] where, , For the displacement and velocity (unknown physical quantities) at time t, For the low frequency bias of the acceleration sensor and the tilt sensor at time t, which is used to suppress the integral drift.

[0170] The state evolution over time is described by the state equation, and a lightweight first neural network is introduced to learn the error of the state vector update, defined as:

[0171] ,

[0172] where, is the state transition matrix, is the process noise, is the observation vector, which comes from the data collected by the sensor.

[0173] Then the observation model is constructed, and the observation vector comes from the data collected by the sensor:

[0174] ,

[0175] where, is the measurement value of the acceleration sensor and the tilt sensor at time t. The state mapping to the observation value is described by the observation equation, and a lightweight second neural network

[0176] is introduced to learn the error of the observation vector update:

[0177] ,

[0178] ,

[0179] where, is the observation noise, is the height of the installation position of the monitor, is the displacement state, is the acceleration state, are the low frequency biases of the acceleration sensor and the tilt sensor at time t, respectively. Unscented Kalman filtering is used for data fusion. First, unscented transformation is performed, and a set of Sigma points is selected to capture the mean and covariance of the state:

[0180]

[0181] ,

[0182] where, is the set matrix of Sigma points,​​ Is The optimal state estimate is obtained by updating the unscented Kalman filter at each time step. Is Covariance at time, It is a scaling parameter;

[0183] The prediction step of the unscented Kalman filter is shown below:

[0184] ,

[0185] ,

[0186] ,

[0187] in, Is The Sigma point in the state space at time step. Is The Sigma point in the state space at time step has been propagated to time, It is an unscented Kalman filter based on Information about time The optimal estimate of the system state at time t. It is the dimension of the state vector. It is the first The mean weight of each Sigma point Is The first time in the state space Sigma points and have been spread to time, It is the first Covariance weights of Sigma points This is the process noise covariance matrix; it is important to note that the state equation here includes... The output has been compensated.

[0188] The update step of the unscented Kalman filter propagates the Sigma points through the observation equation, and similarly, this step is also incorporated here. Error compensation is performed as follows:

[0189] ,

[0190] in, This refers to the predicted observation value corresponding to the Sigma point;

[0191] Calculate the mean, covariance, and cross-covariance of the predicted values:

[0192] ,

[0193] ,

[0194] ,

[0195] where, is the predicted mean of the observation, is the predicted observation value corresponding to the th Sigma point, is the covariance matrix of the predicted observation value, is the covariance matrix of the observation noise, is the cross-covariance matrix of the state and the observation.

[0196] The Kalman gain is calculated and the state is updated:

[0197] ,

[0198] ,

[0199] ,

[0200] where, is the Kalman gain value, is the posterior estimation mean of the state vector at time , is the posterior covariance matrix of the state estimation at time is the covariance matrix of the predicted observation value.

[0201] The loss function of the two lightweight neural networks and is defined as:

[0202] ,

[0203] where, is the observation model mapping to the observation space, is the next state predicted by the state transition model, and are the parameters of the neural networks and , , , is a regularization hyperparameter of the loss function, used to balance the weights of the error terms in the loss function;

[0204] The neural networks obtain the state vector , by minimizing the loss function, and the first element in The unknown quantity estimation value after data fusion at the moment.

[0205] S8, data transmission: the processed data is uploaded to the cloud platform in real time through the 4G communication module. The cloud platform stores and further analyzes the data to generate a building safety state evaluation report.

[0206] S9, remote analysis and decision: the cloud platform can deeply analyze the collected data, automatically detect abnormalities and generate early warnings. When the device detects an abnormality, the cloud platform will issue an alarm signal according to the preset threshold and provide repair suggestions for the user.

[0207] The above describes the preferred embodiments of the application in detail. It should be understood that those skilled in the art can make many modifications and changes without creative labor according to the concept of the application. Therefore, any technical solution obtained by logical analysis, reasoning or limited experiment by those skilled in the art on the basis of the prior art according to the concept of the application shall be within the protection scope determined by the claims.

Claims

1. A method of safety monitoring of an existing building structure, characterized by, The method comprises the following steps: Real-time acquisition of multi-source data collected by multiple monitors distributed in key components and important positions of the building, data preprocessing and data synchronization; Based on the synchronized multi-source data, a data-driven neural network is used as a model error compensator combined with an unscented Kalman filter for multi-source data fusion to estimate unmonitored physical quantities, and combined with the collected data for existing building structure safety monitoring; The data synchronization method is as follows: Preliminary data synchronization is performed using GPS or NTP / PTP protocol; The preliminarily synchronized data is constructed into an input feature vector with spatial topological characteristics, which is input into a graph convolutional neural network with gate attention for topological data processing, the nodes in the graph convolutional neural network update their features by aggregating the information of neighboring nodes, and the aggregation weights are determined by the gate attention mechanism, the output of the gate attention mechanism is input into a long short-term memory module for capturing the temporal pattern of the features, and the output of the long short-term memory module is input into an adaptive fusion module driven by uncertainty to output an optimal clock offset estimation value, and high-precision data synchronization is performed based on the optimal clock offset estimation value.

2. The method of claim 1, wherein, The input feature vector is represented as: ， in, Indicates the first Each monitoring node The input feature vector corresponding to the data collected at each moment; It is the current monitoring node exist The clock offset calculated using the PTP protocol at all times; Is the current monitoring node in Network latency measured at any given time; It is with nodes The set of all neighboring nodes that communicate directly. It is a node Receive from node Signal strength indicator It is a node To the node The number of jumps.

3. The method of claim 1, wherein the method further comprises: The gate attention mechanism dynamically controls the fusion degree of information through a gate vector: ， ， ， wherein, is a node aggregating neighboring nodes attention weights at time is a leaky linear rectification function, is a trainable attention vector, is a trainable weight matrix, is a node characteristic representation of a monitor at is a node neighboring node characteristic representation of a monitor at is a learnable gating vector to control whether to fuse the information of neighboring nodes or to keep its own information, is a trainable gating weight matrix, denotes vector concatenation, denotes Hadamard product, characteristic of a node output by the gated attention mechanism.

4. The method of claim 1, wherein, The long short-term memory module performs the following operations on its input: ， ， wherein, node output by the gating attention mechanism at a time instant, denotes the hidden state of the th node after processing by the long short-term memory module at a time instant, denotes the cell state of the th node after processing by the long short-term memory module at a time instant, is a trainable weight matrix for the output layer of the long short-term memory module, is a sequence of historical offset quantities for the long short-term memory module, denotes processing by the long short-term memory module.

5. The method of claim 1, wherein the method further comprises: The adaptive fusion module performs the following operations to output the optimal clock offset estimation value: ， ， ， ， wherein, the variance of the PTP protocol, is the mean square error of the long short-term memory module prediction result, is the variance function, and MSE is the mean square error function; denotes a historical offset sequence of the PTP protocol, denotes a historical offset sequence of the long short-term memory module, is the size of the sliding window, is the fusion weight, is the final estimation value of the clock offset of the node at the moment.

6. The method of claim 1, wherein, The preprocessing specifically includes denoising and feature extraction of the collected data by a Fourier transform or wavelet transform based method.

7. The method of claim 1, wherein the method further comprises: The multi-source data fusion comprises the following steps: Constructing a state space model, defining a system state vector , the state vector comprises unknown physical quantities to be solved at the moment and low frequency bias of the sensor at the moment The evolution of the state over time is described by a state equation, and a lightweight first neural network is introduced. The error after learning the state vector update is defined as: ， wherein, is the state transition matrix, is the process noise, is the observation vector, from the data collected by the sensors; The state mapping to observations is described by an observation equation, and a light-weight second neural network is introduced to learn the error of the observation vector update: ​ ， ， wherein, is an observation noise, is a height at which the monitor is installed, is a displacement state, is an acceleration state, are low frequency biases of the acceleration sensor and the tilt sensor, respectively, are low frequency biases of the acceleration sensor and the tilt sensor, respectively, Data fusion is performed using unscented Kalman filtering, which first performs unscented transformation by selecting a set of Sigma points to capture the mean and covariance of the state: ， in, It is a set matrix of Sigma points. Is The optimal state estimate is obtained by updating the unscented Kalman filter at each time step. Is Covariance at time, It is a scaling parameter; The prediction step of the unscented Kalman filter is as follows: ， ， ， wherein, is a Sigma point in the state space at time , is a Sigma point in the state space at time and has been propagated to time , is the optimal estimate of the system state at time based on information up to time , is the dimension of the state vector, is the mean weight of the th Sigma point, is the th Sigma point in the state space at time and has been propagated to time , is the covariance weight of the th Sigma point, is the process noise covariance matrix; The update step of the unscented Kalman filter is as follows: ， wherein is the predicted observation corresponding to the Sigma point; The mean, covariance and cross-covariance of the predicted value are calculated: ， ， ， wherein, is the predicted mean of the observation, is the predicted observation corresponding to the th Sigma point, is the covariance matrix of the predicted observation, is the covariance matrix of the observation noise, is the cross-covariance matrix of the state and the observation. The Kalman gain is calculated and the state is updated: ， ， ， wherein is a Kalman gain value, is is a posteriori estimation mean value of the state vector at the time instant, is a posteriori covariance matrix of the state estimation at the time instant, is a covariance matrix of the predicted observation.

8. A method of monitoring the safety of an existing building structure according to claim 7, wherein The two lightweight neural networks introduced and The loss function is defined as: ， wherein, is an observation model mapping to an observation space, is a next state predicted by a state transition model, and are parameters of neural networks and respectively, , , is a regularization hyperparameter of the loss function for balancing the weights of error terms in the loss function; The neural network obtains the state vector by minimizing a loss function , The first element of is the estimate of the unknown quantity after data fusion at time .

9. A system for monitoring the safety of an existing building structure, characterized by The system for implementing the method of any one of claims 1-8 comprises: A plurality of monitors distributed in key components and important positions of the building, the monitors comprising a collection module, an edge computing module and a data synchronization module, wherein the collection module is used for real-time collection of multi-source data, the edge computing module is used for data preprocessing and data fusion processing, wherein the data fusion processing is based on synchronized multi-source data, a data-driven neural network is used as a model error compensator combined with an unscented Kalman filter for multi-source data fusion to estimate unmonitored physical quantities, and the data synchronization module is used for data synchronization of data collected by the collection modules of different monitors; and a cloud platform for existing building structure safety monitoring combined with the estimated unmonitored physical quantities and the collected data, and providing visual display; Wherein, the cloud platform and the monitors transmit data through wireless communication.

Images