System and method for automatically monitoring structural deformation based on GNSS (Global Navigation Satellite System) three-dimensional displacement curve

Abstract

The invention discloses a structure deformation automatic monitoring system and method based on a GNSS three-dimensional displacement curve. The method comprises the following steps: acquiring three-dimensional coordinate initial data of key points from a GNSS receiver, performing quality screening and simple smoothing, unifying timestamps, and outputting an initial position sequence; on the basis of the initial position sequence, extracting relative motion and frequency spectrum features after linear interpolation, and splicing the relative motion and frequency spectrum features to obtain a monitoring point feature vector; establishing a joint model of a structure deformation state and a vibration state by combining the feature vectors of the monitoring points, and outputting an accurate deformation track and an accurate vibration state by the joint model after filtering; based on the accurate deformation track and the accurate vibration state, calculating a vibration health index to evaluate the health degree of the vibration level; based on the accurate deformation track and the accurate vibration state, geometric calculation is carried out on the key point coordinates, and an alarm state identifier is output; and inputting a graph neural network based on the accurate deformation track and the accurate vibration state to realize classification of the hinge region failure mode.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of hinge area monitoring of hydropower stations, and more particularly, to a structure deformation automatic monitoring system and method based on GNSS three-dimensional displacement curves. BACKGROUND

[0002] Structure deformation monitoring is a core link in the operation safety management of large-scale hydraulic structures. Traditional monitoring methods rely on manual regular inspection, optical total station measurement, or the laying of sensor arrays such as strain gauges and displacement meters. These methods can reflect the static displacement of the structure to some extent, but they are difficult to achieve long-period continuous observation, and the data acquisition is limited in complex environments such as water flow impact, electromagnetic interference or night operation, and cannot provide real-time, global and fine deformation information.

[0003] In recent years, the global navigation satellite system (GNSS) has been gradually introduced into the safety monitoring of hydraulic structures due to its all-weather, long-time and high-precision displacement monitoring capability. Existing GNSS-based researches mainly focus on the static solution of key point displacement, and analyze the overall settlement, horizontal displacement and rotation of the structure by comparing time series. This kind of method usually relies on single time series data, combined with filtering algorithm or simple spectrum analysis to judge abnormalities, but it has shortcomings in identifying local hinge area deformation patterns, rapid vibration characteristics and state decoupling under multi-source interference conditions.

[0004] In actual operation, the hinge area is often a stress concentration and failure sensitive part of the structure. When the water flow, opening and closing conditions or external load change suddenly, the hinge area will show complex deformation-vibration coupling characteristics. If the existing GNSS-based monitoring method only relies on displacement sequence for static analysis, it is often difficult to fully reflect the dynamic response characteristics of the structure under multiple working conditions, especially in identifying early abnormalities and distinguishing interference factors from real deformations. If these subtle changes in the hinge area cannot be captured in time and accurately, local damage may accumulate gradually, leading to operational obstacles, and even affecting the safety and stability of the overall structure in severe cases. SUMMARY

[0005] The purpose of the present application is to provide a structure deformation automatic monitoring system and method based on GNSS three-dimensional displacement curves, which can overcome the shortcomings of existing monitoring methods in real-time, dynamic response and early abnormality identification, and accurately monitor and intelligently diagnose the structure deformation and vibration state of key parts under complex environmental conditions.

[0006] In order to achieve the above purpose, the present application provides the following technical scheme:

[0007] A structure deformation automatic monitoring method based on GNSS three-dimensional displacement curves, comprising:

[0008] Step S1, obtaining initial data of three-dimensional coordinates of key points from a GNSS receiver, the initial data of three-dimensional coordinates of key points being subjected to quality screening and simple smoothing and unified timestamp, and outputting a preliminary position sequence;

[0009] Specifically, the quality screening and simple smoothing and unified timestamp include:

[0010] S11, collecting initial data of three-dimensional coordinates of key points from a GNSS receiver, and performing quality screening to obtain effective observations;

[0011] S12, performing median filtering and superimposing moving average on the effective observations according to time to obtain a smoothed position sequence;

[0012] S13, linearly resampling the smoothed position sequence according to a unified timestamp and filling in missing measurements to output a preliminary position sequence;

[0013] Specifically, the quality screening includes:

[0014] The initial data of three-dimensional coordinates of key points is combined with a carrier-to-noise ratio, a geometric dilution of precision, an autonomous integrity monitoring residual, and a cycle slip marker to generate a quality index sequence, the quality index sequence is formed into a quality mask according to a threshold set, and the initial data stream is subjected to the quality mask to obtain initial observations. On this basis, jump observations are removed by using coordinate difference, non-steady-state observations are removed by using steady-state consistency markers, and complete samples are selected by combining time consistency masks, and finally effective observations are obtained.

[0015] Step S2, based on the preliminary position sequence, linear interpolation is performed to extract relative motion and spectral features and splice them to obtain a monitoring point feature vector;

[0016] Specifically, the step of obtaining the monitoring point feature vector includes:

[0017] S21, linear interpolation is performed based on the preliminary position sequence to form a three-dimensional displacement sequence with equidistant timestamps;

[0018] S22, key point distance changes and leaf normal angle changes are calculated from the three-dimensional displacement sequence to obtain key point relative motion parameters, and an analysis time window for spectral analysis is determined according to the key point relative motion parameters;

[0019] S23, the three-dimensional displacement sequence is extracted in combination with the analysis time window to obtain a dominant frequency, a super-low frequency displacement component is obtained by approximating low-pass filtering on a low-frequency passband below the dominant frequency with a moving average, and a main frequency shift rate is obtained by the change of the dominant frequency relative to a reference;

[0020] S24, the monitoring point feature vector is constructed in combination with the super-low frequency displacement component, the key point relative motion parameters, and the main frequency shift rate.

[0021] Step S3, a joint model of structural deformation state and vibration state is established in combination with the monitoring point feature vector, and the joint model outputs the accurate deformation trajectory and the accurate vibration state after filtering; specifically, the outputting of the accurate deformation trajectory and the vibration state comprises:

[0022] S31, a joint state space model is constructed in combination with the seven feature parameters of the monitoring point feature vector;

[0023] S32, the process noise covariance and the observation noise covariance of the joint state space model are dynamically adjusted in combination with the real-time vibration energy intensity and the real-time carrier-to-noise ratio, to obtain the dynamically scaled process noise covariance and the corrected observation noise covariance, respectively;

[0024] S33, the dynamically scaled process noise covariance and the corrected observation noise covariance are used to perform prediction and update iteration through joint Kalman filtering, to output the optimally estimated deformation state and vibration state;

[0025] Specifically, the process noise covariance of the joint state space model is adjusted as follows:

[0026] The dynamic scaling rule of the process noise covariance of the joint state space model is as follows:

[0027] (a) When the real-time vibration energy intensity exceeds the vibration threshold multiple of the reference value, the process noise covariance is enlarged by a corresponding multiple, to reflect the uncertainty brought by the abnormal increase of vibration;

[0028] (b) When the real-time carrier-to-noise ratio is lower than the set lower limit of the carrier-to-noise ratio, the process noise covariance is enlarged by a corresponding multiple, to reflect the uncertainty brought by the increase of electromagnetic interference;

[0029] (c) When both of the above two situations occur, that is, the vibration energy intensity abnormally increases and the carrier-to-noise ratio significantly decreases, the process noise covariance is enlarged to a higher multiple, to reflect the increase of uncertainty caused by coupled interference.

[0030] Specifically, the observation noise covariance of the joint state space model is adjusted as follows:

[0031] The dynamic scaling rule of the observation noise covariance of the joint state space model is as follows:

[0032] When the real-time carrier-to-noise ratio is less than the lower limit threshold of the carrier-to-noise ratio, the observation noise covariance is enlarged to the reference value × a1;

[0033] When the real-time vibration energy intensity is greater than the vibration energy reference value × k, the observation noise covariance is enlarged to the reference value × a2;

[0034] When the carrier-to-noise ratio is lower than the lower threshold value and the vibration energy intensity is greater than the vibration energy reference value x k at the same time, the observation noise covariance is amplified to the reference x a3;

[0035] Step S4, based on the accurate deformation trajectory and the accurate vibration state, calculate the vibration health index to evaluate the health degree of the vibration level; based on the accurate deformation trajectory and the accurate vibration state, perform geometric calculation on the key point coordinates to output the alarm state identifier; based on the accurate deformation trajectory and the accurate vibration state, input the graph neural network to realize classification of the hinge area failure mode.

[0036] Specifically, the evaluation of the health degree of the vibration level, the output of the alarm state identifier and the classification of the hinge area failure mode include:

[0037] S41, in combination with the vibration amplitude of the accurate vibration state, calculate the vibration health index for evaluating the health degree of the current vibration level relative to the historical extreme value.

[0038] S42, based on the accurate deformation trajectory, perform geometric calculation on the key point coordinates to obtain the distance offset and the angle offset, and combine the historical statistical reference to form a residual parameter, and output the alarm state identifier according to the hierarchical alarm rule;

[0039] S43, fuse the accurate deformation trajectory and the vibration state, input the graph neural network to realize classification of the hinge area failure mode.

[0040] Specifically, the graph structure of the graph neural network of step S43 is defined as:

[0041] The node set includes the left hinge shaft point A, the right hinge shaft point B and the leaf connecting point C.

[0042] The edge set includes the physical connection edge A-B, the B-C and the implicit mechanical edge A-C, forms a fully connected topology, inputs the graph neural network and outputs the 8-class failure mode classification result.

[0043] A structure deformation automatic monitoring system based on GNSS three-dimensional displacement curve comprises:

[0044] A preliminary position sequence acquisition module: obtains key point three-dimensional coordinate initial data from a GNSS receiver, the key point three-dimensional coordinate initial data is subjected to quality screening and simple smoothing and unified timestamp, and a preliminary position sequence is output;

[0045] A monitoring point feature vector extraction module: used for extracting relative motion and frequency spectrum features after linear interpolation based on the preliminary position sequence, and splicing to obtain a monitoring point feature vector;

[0046] Joint state estimation module: used to establish a joint model of structural deformation state and vibration state based on the monitoring point feature vector, to estimate the two types of states synchronously through filtering method, and to output accurate deformation trajectory and accurate vibration state;

[0047] Comprehensive analysis and alarm module: used to calculate the vibration health index to evaluate the health degree of vibration level based on the accurate deformation trajectory and the accurate vibration state, to perform geometric calculation on key point coordinates based on the accurate deformation trajectory and the accurate vibration state, and to output alarm state identification, and to input the graph neural network based on the accurate deformation trajectory and the accurate vibration state to realize classification of failure modes in the hinge area.

[0048] Compared with the prior art, the beneficial effects of the present application are:

[0049] The present application can identify abnormal displacement trends in the early stage by collecting GNSS three-dimensional coordinate data in real time at key points, and combining quality screening and time unified processing; further, the present application can realize sensitive capture of dynamic response under complex working conditions by using interpolation and frequency domain analysis to extract ultra-low frequency displacement, relative motion characteristics and frequency offset parameters, and constructing a monitoring point feature vector; then, the present application can effectively separate the real deformation of the structure and the vibration interference by means of joint state space modeling and dynamic filtering algorithm, and improve the accuracy and stability of state estimation; finally, the present application can automatically output graded alarm information and identify failure modes based on geometric parameter calculation, health index evaluation and intelligent classification model. Through the above multi-level processing, the present application can significantly improve the real-time performance and refinement degree of structural deformation monitoring, effectively reduce the dependence on manual inspection and false alarm rate, and ensure the safe operation of key hydraulic structures, which has wide engineering application value and promotion prospect. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 The present application is based on the GNSS three-dimensional displacement curve of the structural deformation automatic monitoring method flowchart;

[0051] Figure 2 The present application is based on the description of the structure of the hinge area topology schematic diagram;

[0052] Figure 3 The present application is based on the GNSS three-dimensional displacement curve of the structural deformation automatic monitoring system schematic diagram. DETAILED DESCRIPTION

[0053] The technical solutions in the embodiments of the present application will be described in detail below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0054] Embodiment 1

[0055] Referring to Figure 1 As shown in the figure, the embodiment provides a structure deformation automatic monitoring method based on GNSS three-dimensional displacement curve, comprising:

[0056] Step S1, obtaining key point three-dimensional coordinate initial data from a GNSS receiver, the key point three-dimensional coordinate initial data is subjected to quality screening, simple smoothing and unified timestamp, and a preliminary position sequence is outputted;

[0057] Further, step S1 comprises:

[0058] S11, collecting key point three-dimensional coordinate initial data stream from a GNSS receiver, and performing quality screening to obtain effective observation.

[0059] Specifically, the initial data stream is continuously collected at key points A, B and C by the GNSS receiver; and a quality index sequence is generated from the initial data stream in combination with the carrier-to-noise ratio sequence, the geometric precision factor sequence, the autonomous integrity monitoring residual sequence and the cycle slip mark sequence recorded by the receiver synchronously.

[0060] Among them, the carrier-to-noise ratio refers to the ratio of useful signal power to noise power, which is obtained by receiver statistics; the geometric precision factor refers to the amplification coefficient of satellite geometry to positioning error, which is obtained by ephemeris and observation configuration solution; the autonomous integrity monitoring residual refers to the fitting residual of measurement and consistency test model, which is calculated by consistency test; the cycle slip mark refers to the determination mark of carrier phase discontinuity, which is determined by the receiver kernel.

[0061] The mean and standard deviation of the quality index sequence in the starting steady window are estimated and a threshold set is generated. Specifically, the threshold set is determined by the rules as follows:

[0062] Rule 1: carrier-to-noise ratio lower limit = carrier-to-noise ratio mean - a x carrier-to-noise ratio standard deviation;

[0063] Rule 2: geometric precision factor upper limit = geometric precision factor mean + a x geometric precision factor standard deviation;

[0064] Rule 3: residual upper limit = residual mean + a x residual standard deviation, wherein a is a dimensionless coefficient between 2 and 3.

[0065] The three-threshold gating method is used to calculate the quality mask by combining the threshold set with the quality index sequence. Specifically, the quality mask is defined as: when the carrier-to-noise ratio is greater than or equal to the set lower limit, the geometric precision factor is less than or equal to the set upper limit, the autonomous integrity monitoring residual is less than or equal to the set residual upper limit, and the cycle slip flag is equal to zero, it is determined that all conditions are met, at this time the quality mask takes the value of 1, the quality mask takes 1, otherwise 0. The initial data stream is acted on by the quality mask to obtain the initial observation. The coordinate difference sequence of adjacent epochs is calculated from the preliminary screening observation to generate a jump threshold. Specifically, the coordinate difference is Δposition = ||position(t)-position(t-1)||, when Δposition is greater than the jump threshold (taking three times the standard deviation), the epoch is removed, and the jump-removed observation is output. The speed near-zero test and variance lower limit test are calculated in the short window to generate a steady-state consistency flag, and the samples that do not meet the steady-state consistency are removed to obtain the steady-state observation. The missing epoch is marked from the steady-state observation and the original timestamp is retained to generate a time consistency mask, according to which the complete samples on the same time axis are screened, and finally the effective observation is obtained.

[0066] S12, the effective observation is subjected to median filtering and superimposed moving average according to time to obtain a smoothed position sequence.

[0067] Specifically, a fixed-length sliding window is established in time sequence based on the effective observation, and the median value and standard deviation of the three components of the three-dimensional coordinates are calculated in each window to determine the abnormal points in the window; when the deviation of a certain time point from the median value exceeds three times the standard deviation, the coordinate of the time point is replaced with the window median value to form an anti-jump sequence that removes isolated jump points. Specifically, the three times standard deviation criterion is used to quickly identify sudden deviation points, and the median replacement is used to restore reasonable values without lowering the overall trend; after the determination and replacement are completed, the anti-jump sequence is output.

[0068] After obtaining the anti-jump sequence, a symmetric adjacent range is taken for each time point along the time axis, and an equal-weight moving average is performed on the coordinate values in the range, and the average result is taken as the smoothed coordinate of the time point, and the continuous smoothed position sequence can be obtained by advancing point by point.

[0069] Specifically, the equal-weight moving average refers to taking the current time point and several points before and after the current point to calculate the arithmetic mean, thereby suppressing high-frequency jitter and retaining the true deformation trend of slow change; the original timestamp and order are strictly preserved during the calculation process to ensure seamless connection in subsequent steps. After completing the moving average, the smoothed position sequence is output

[0070] S13, the smoothed position sequence is linearly resampled according to the unified timestamp and the missing data is filled to output the preliminary position sequence.

[0071] The time mapping table is generated by aligning the smoothed position sequence with the target time axis point by point, wherein the target time axis refers to a uniform timestamp set arranged in a fixed step length within a current monitoring period, and the time mapping table refers to a one-to-one correspondence between the original timestamp and the uniform timestamp.

[0072] The linear resampling sequence is obtained by linear interpolation between the time and coordinates of the two adjacent observed positions according to the time mapping table. Specifically, the linear resampling refers to taking a linear transition in the three-dimensional coordinates between two known time points according to the time ratio. When there is only one side observation at the beginning or end of the sequence, the coordinates of the last smoothed position sequence are kept unchanged and written into the linear resampling sequence.

[0073] After obtaining the linear resampling sequence, the missing data marker table is generated according to the missing condition of each uniform timestamp identified by the time mapping table. The missing data marker table and the linear resampling sequence are combined to form the completed sequence. For single-point or continuous missing data with a small number, the corresponding linear interpolation result is directly written. For a short gap with continuous missing data exceeding a preset number, the linear extension is preferentially filled according to the change trend of the adjacent positions, and if the extension direction is obviously inconsistent with the change trend of the adjacent positions, the segmented strategy of keeping the last valid value is adopted, and the backfill result is uniformly written into the completed sequence.

[0074] The three-point mean processing is performed on the completed sequence along the target time axis to obtain a preliminary position sequence that is continuous in time and complete in three points. Specifically, the three-point mean processing refers to taking the arithmetic mean of the coordinates of the current time point and its adjacent time points on both sides to weaken the small discontinuity at the junction of interpolation and backfilling.

[0075] Step S2, based on the preliminary position sequence, linear interpolation is performed to extract relative motion and spectral features and splice to obtain a monitoring point feature vector.

[0076] Further, step S2 includes:

[0077] S21, linear interpolation is performed based on the preliminary position sequence to form a three-dimensional displacement sequence with equal interval timestamps.

[0078] A uniform equal interval timestamp set is constructed based on the preliminary position sequence, and linear interpolation is performed between adjacent time points according to the time ratio to generate a three-dimensional displacement sequence corresponding to the equal interval timestamps. The equal interval timestamp set refers to a time point sequence arranged in a fixed step length, which is used to unify the time bases of points A, B and C. The three-dimensional displacement sequence refers to a structured time-coordinate matrix formed by compiling the three-dimensional coordinate values of points A, B and C at each equal interval timestamp.

[0079] Specifically, the original time mark is extracted from the preliminary position sequence, and a set of equidistant time stamps is generated with a time starting point, a time ending point, and a fixed step. Then, the original time interval in which each equidistant time stamp falls is paired with its two end coordinates, and a straight line transition is made on the coordinate component according to the time proportion to obtain the interpolated coordinates of the time point. Sequentially advancing can form a three-dimensional displacement sequence covering the whole segment. To ensure the availability of the boundary, the time point located at the beginning and end of the sequence and having only one side observation is written into the three-dimensional displacement sequence by using the nearest available coordinate unchanged; to suppress accidental interpolation distortion, a consistency comparison is made between the preliminary position sequence and the three-dimensional displacement sequence, the continuity of the difference between adjacent points before and after interpolation is counted, and if a single point is found to increase suddenly and does not meet the adjacent continuity rule, the interpolated coordinates of the time point are rolled back to the arithmetic mean of the adjacent two points and updated to the three-dimensional displacement sequence. Thus, the three-dimensional displacement sequence is equidistant in time and three-point complete in space, meeting the unified data format requirements of subsequent relative motion and spectral analysis.

[0080] S22, calculate the key point distance change and the door leaf normal angle change from the three-dimensional displacement sequence, obtain the key point relative motion parameters, and determine the analysis time window of the spectral analysis according to the key point relative motion parameters;

[0081] The AB distance change of points A and B and the door leaf normal angle change defined by points A, B, and C are calculated from the three-dimensional displacement sequence to generate the key point pair relative motion parameters, and the analysis time window is determined under the stability and continuity constraints of relative motion. The AB distance change refers to the deviation of the spatial distance sequence composed of the coordinates of points A and B from the reference length, and the reference length can be the statistical mean of the static horizontal stable stage; the door leaf normal angle change refers to the deviation of the normal of the door leaf plane composed of points A, B, and C from the reference normal, and the reference normal can be the statistical direction of the static horizontal stable stage. Specifically, the spatial distance between points A and B is calculated from the three-dimensional displacement sequence at each time point to form an AB length sequence, and the AB distance change is obtained by subtracting the reference length; simultaneously, the normal direction of the door leaf plane composed of points A, B, and C is constructed at each time point, and then compared with the reference normal to obtain the door leaf normal angle change. Aligning the AB distance change and the door leaf normal angle change on the same time axis and concatenating them in a predetermined order generates the key point pair relative motion parameters containing "length deviation" and "normal deviation".

[0082] To ensure that the subsequent spectrum processing is in a representative dynamic section, the relative motion parameters of the key points are used to determine the analysis time window. Specifically, the fluctuation amplitude, change rate and mutation count in the short window are calculated based on the AB distance change and the normal angle change of the leaf, and a stability score sequence is generated to measure the degree of smoothness. Then, the stability score sequence is evaluated on the whole segment by sliding, and the continuous interval with better score and not less than the minimum analysis time is selected to obtain the candidate analysis time window set. When there are multiple intervals in the candidate set, the "relative motion coupling consistency" rule is introduced for sorting: in the same time window, the interval with higher consistency of change direction and smaller phase lag between AB distance change and normal angle change of the leaf is given higher priority. If there are still parallel intervals, the interval with lower fluctuation amplitude and lower mutation count is selected as the final analysis time window. When there is a significant working condition switching (such as opening state change or external force disturbance) during monitoring, the whole segment is divided into multiple segments according to the switching point, and the above scoring and sorting rules are repeated in each segment to form a multi-window analysis time window result. Through the above processing, the relative motion parameters of the key point pairs describe the geometric relative changes of the structure, and the analysis time window locks the representative time domain interval, which together provides a unified, reliable and traceable data basis for subsequent frequency spectrum extraction and feature splicing.

[0083] S23, combining the analysis time window, the three-dimensional displacement sequence is subjected to fast Fourier transform to extract the dominant frequency, and the super-low frequency displacement component is obtained by sliding average approximation low-pass filtering in the low-frequency passband below the dominant frequency. The real-time vibration energy intensity is represented by the amplitude adjacent to the dominant frequency, and the dominant frequency offset rate is obtained by the change of the dominant frequency relative to the reference.

[0084] Specifically, the three-dimensional displacement sequence is subjected to fast Fourier transform according to the coordinate components in combination with the analysis time window, and the spectral peak with the maximum amplitude is found on the amplitude spectrum of each component and comprehensively determined to obtain the dominant frequency corresponding to the time window. The dominant frequency represents the vibration characteristic frequency with the most concentrated energy in the segment, serving as a frequency domain reference.

[0085] Sliding average filtering is used to approximate low-pass filtering in the low-frequency range below the dominant frequency, and only the slowly changing components are retained to obtain the super-low frequency displacement X component, the super-low frequency displacement Y component and the super-low frequency displacement Z component. Specifically, the sliding average filtering refers to taking a number of adjacent points before and after the current time point on the time axis and performing arithmetic average with the current point, and replacing the current coordinate value with the average value. The window is shortened at the beginning and end according to the number of available adjacent points to maintain the time sequence unchanged.

[0086] The dominant frequency of this paragraph is compared with the reference dominant frequency obtained in the static water closing stage to calculate the relative change ratio of the two, and the dominant frequency offset rate is obtained. When the dominant frequency is higher than the reference value, it is recorded as positive offset, and when it is lower than the reference value, it is recorded as negative offset, which is used to measure whether the vibration frequency has a significant drift.

[0087] After the above processing, the three-dimensional displacement sequence obtains the dominant frequency reference in the frequency domain, and the three-axis ultra-low frequency displacement component matched with it is obtained in the time domain, and the dominant frequency offset rate for frequency stability measurement is formed.

[0088] S24, the ultra-low frequency displacement component, the relative motion parameter of the key point pair, the vibration energy intensity and the dominant frequency offset rate are spliced in order, and the monitoring point feature vector is output.

[0089] The ultra-low frequency displacement component is split according to the three-dimensional coordinate direction to form the ultra-low frequency displacement X component, the ultra-low frequency displacement Y component and the ultra-low frequency displacement Z component, which are used to describe the slow structural deformation characteristics of the monitoring point in the low frequency range and can sensitively reflect the cumulative effects of settlement or slip.

[0090] The hinge axis length parameter and the hinge angle parameter are extracted from the relative motion parameter of the key point pair: the hinge axis length parameter refers to the spatial distance change between the key point A and the key point B, which is used to describe whether the hinge axis is abnormal in extension; the hinge angle parameter refers to the included angle change between the straight line AB and the straight line AC, which is used to reflect whether the door leaf plane is twisted or deflected.

[0091] On this basis, combined with the projection calculation result in the relative motion parameter of the key point pair, the projection distance parameter is obtained; the projection distance parameter refers to the projection change of the key point in the selected sensitive direction (for example, along the hinge axis direction or along the normal direction of the door leaf), which is used to capture local small geometric offset, and the direction selection is based on the structural sensitive direction determined in S22.

[0092] The above geometric and frequency domain elements are combined with the dominant frequency offset rate, and assembled into the monitoring point feature vector in a predetermined order, which is explicitly written as:

[0093] The monitoring point feature vector = [ultra-low frequency displacement X component, ultra-low frequency displacement Y component, ultra-low frequency displacement Z component, hinge axis length parameter, hinge angle parameter, projection distance parameter, dominant frequency offset rate].

[0094] Step S3, combined with the monitoring point feature vector, a joint model of structural deformation state and vibration state is established, and the accurate deformation trajectory and vibration state are output after filtering by the joint model

[0095] Further, step S3 includes:

[0096] S31, combined with the seven feature parameters of the monitoring point feature vector, a joint state space model is constructed.

[0097] wherein the joint state space model refers to a state vector containing both structural deformation state and vibration state, and the two types of states are estimated synchronously through joint Kalman filtering method, so as to realize the separated modeling of the real deformation and vibration of the structure. Specifically, the following steps are included:

[0098] S311, in the joint state space model, the state vector is defined to consist of deformation state and vibration state:

[0099] The deformation state refers to the real three-dimensional coordinates of the key point A, the key point B and the key point C, with the unit of millimeter, which is used to describe the deformation trend of the structure at low frequency and slow scale;

[0100] The vibration state refers to the vibration amplitude (millimeter) and the corresponding vibration main frequency (hertz) of the key point in X, Y and Z directions, wherein the vibration amplitude represents the instantaneous displacement fluctuation range caused by the water flow impact, and the vibration main frequency represents the frequency band where the energy is concentrated.

[0101] By taking the vibration amplitude and the vibration main frequency as independent state components and jointly estimating them with the deformation state, the physical decoupling of noise can be realized, so as to ensure that the real deformation trajectory of the structure can still be accurately identified when the vibration interference is strong.

[0102] S312, a state transition model is constructed in combination with the main frequency shift rate in the feature vector of the monitoring point. The state transition model includes a deformation state evolution model and a vibration state evolution model.

[0103] Specifically, the deformation state evolution model is:

[0104] The deformation state at the current time is obtained by superimposing a small perturbation noise on the basis of the deformation state at the previous time;

[0105] The covariance of the small perturbation noise is:

[0106] Deformation state covariance = default covariance + 0.5×absolute value of main frequency shift rate.

[0107] Wherein, the default covariance refers to the baseline uncertainty set by the system in the absence of frequency drift, and the main frequency shift rate refers to the relative change ratio of the dominant frequency at the current time window relative to the dominant frequency of the static water baseline, which is used to reflect the influence of vibration drift on the deformation uncertainty.

[0108] The vibration state evolution model is:

[0109] The vibration state at the current time is obtained based on the vibration state at the previous time in combination with a first-order autoregressive model;

[0110] The first-order autoregressive model is defined as:

[0111] Current vibration state = Autoregressive coefficient x previous time vibration state + Disturbance noise.

[0112] Wherein, the covariance matrix of the disturbance noise adopts a diagonal structure: the first three diagonal items respectively take a fixed proportional coefficient of the square of the vibration amplitude in each direction, for quantifying the randomness of amplitude fluctuation; the last three diagonal items take a constant value equal to the variance of the reference frequency, for maintaining the stability of the vibration main frequency estimation.

[0113] Through the above modeling method, the state transition model can explicitly introduce the main frequency deviation rate into the uncertainty adjustment of the deformation state, and at the same time, describe the time evolution of the vibration state through the autoregressive method, so as to maintain the stability of state prediction and the ability of physical decoupling under strong vibration interference.

[0114] S313, combined with the observation coordinate data of GNSS, an observation model is established. Wherein, the observation model decomposes the original GNSS observation value into three parts:

[0115] The observation model decomposes the original GNSS observation value into three parts: deformation state component, vibration state component and observation noise component.

[0116] Specifically, the deformation state component is the true three-dimensional coordinates of the key points A, B and C, which is used to represent the true geometric deformation of the structure under low frequency and slow scale.

[0117] The vibration state component is the vibration amplitude and the vibration main frequency, which is synthesized into a simple harmonic vibration waveform in X, Y and Z three directions according to the corresponding parameters in the feature vector of the monitoring point, and is defined as:

[0118] Vibration displacement X = vibration amplitude X x sin (2π x vibration main frequency x time) x X direction unit vector

[0119] Vibration displacement Y = vibration amplitude Y x sin (2π x vibration main frequency x time) x Y direction unit vector

[0120] Vibration displacement Z = vibration amplitude Z x sin (2π x vibration main frequency x time) x Z direction unit vector.

[0121] Wherein, the vibration amplitude X, the vibration amplitude Y and the vibration amplitude Z are the vibration amplitudes in three directions, the vibration main frequency is the frequency where the energy is concentrated, and the X / Y / Z direction unit vector is the three-dimensional coordinate basis vector.

[0122] The observation noise component is the measurement random error, which is represented by the covariance matrix "observation noise covariance";

[0123] The observation noise covariance is dynamically generated from the real-time carrier-to-noise ratio (C / N0) and multipath effect estimation of the GNSS receiver: when the C / N0 decreases, the observation noise covariance is increased to reflect the risk of electromagnetic interference; when the multipath effect is significant, the variance component in the corresponding direction is corrected to ensure the precision.

[0124] In this way, the GNSS observation value is explicitly decomposed into: real deformation component + vibration simulation component + noise error component.

[0125] This observation model ensures that the vibration component is no longer mixed into the observation noise, but exists as an independent term, so that the separation estimation of the real deformation and vibration is realized in the Kalman filtering process, and the accurate deformation trajectory and vibration state of the structure are still stably output under the working conditions of electromagnetic interference or multipath.

[0126] S32, in combination with the real-time vibration energy intensity and the real-time carrier-to-noise ratio (C / N0) of the GNSS receiver, dynamically adjusts the process noise covariance and the observation noise covariance of the joint state space model, respectively obtaining the dynamically scaled process noise covariance and the corrected observation noise covariance.

[0127] Further, step S32 includes:

[0128] S321, dynamically scaling the process noise covariance based on the real-time vibration energy intensity and the C / N0 state.

[0129] The process noise covariance refers to the uncertainty parameter introduced in the state transition process, and its size determines the sensitivity of state prediction to sudden interference.

[0130] The dynamic scaling rule is defined as follows:

[0131] (a) When the real-time vibration energy intensity exceeds the vibration energy reference value × a1, the process noise covariance is enlarged to the reference value × a2; wherein a1 is the vibration energy threshold multiple, a2 is the amplification coefficient in the high vibration case, and the vibration energy reference value is obtained by statistics during the static water closed door phase;

[0132] (b) When the real-time C / N0 is lower than the lower limit threshold of C / N0, the process noise covariance is enlarged to the reference value × a3; wherein a3 is the amplification coefficient in the low signal-to-noise ratio (SNR) case, and the lower limit threshold of C / N0 is the minimum acceptable SNR of the GNSS receiver in the stable working condition;

[0133] (c) When the C / N0 is lower than the lower limit threshold and the vibration energy intensity is greater than the vibration energy reference value × a1 at the same time, the process noise covariance is enlarged to the reference value × a4; wherein a4 is the amplification coefficient in the double interference case.

[0134] The values of parameters a2, a3, and a4 are obtained through laboratory control tests.

[0135] Determine a2 by measuring the amplification ratio of the state prediction residual variance under the condition of applying different amplitude vibrations step by step without electromagnetic interference.

[0136] Determine a3 by measuring the amplification ratio of the state prediction residual variance under the condition of reducing the carrier-to-noise ratio step by step without electromagnetic interference.

[0137] Determine a4 by measuring the compound amplification effect of the residual variance under the condition of superimposing electromagnetic interference and vibration.

[0138] The value of parameter a1 is determined according to the statistical results of GNSS vibration energy intensity distribution for at least 24 hours in the static water closed door stage, and the 95% quantile is taken as the threshold multiple for distinguishing normal fluctuations and abnormal enhancement.

[0139] S322, correct the observation noise covariance based on real-time vibration energy intensity and real-time carrier-to-noise ratio

[0140] The observation noise covariance refers to the uncertainty of GNSS observation value error, and its value determines the weight of the observation value in the filtering calculation.

[0141] The correction rule is defined as follows:

[0142] (a) When the real-time carrier-to-noise ratio is less than the lower threshold of the carrier-to-noise ratio, the observation noise covariance is amplified to the reference value × b1; wherein b1 is the amplification coefficient under the condition of insufficient carrier-to-noise ratio;

[0143] (b) When the real-time vibration energy intensity is greater than the vibration energy reference value × b2, the observation noise covariance is amplified to the reference value × b3; wherein b2 is the vibration energy threshold multiple, and b3 is the amplification coefficient under the condition of high vibration;

[0144] (c) When the carrier-to-noise ratio is lower than the lower threshold and the vibration energy intensity is greater than the vibration energy reference value × b2 at the same time, the observation noise covariance is amplified to the reference value × b4; wherein b4 is the amplification coefficient under the condition of double interference.

[0145] Wherein, the values of parameters b1, b3 and b4 are obtained through laboratory control test:

[0146] Determine b1 by measuring the amplification ratio of the observation residual variance under the condition of superimposing different intensity electromagnetic noise step by step under the condition of constant carrier-to-noise ratio.

[0147] Determine b3 by measuring the amplification ratio of the observation residual variance under the condition of applying different amplitude mechanical vibration step by step without electromagnetic interference.

[0148] When the electromagnetic noise and mechanical vibration are superimposed simultaneously, the composite amplification effect of the measurement residual variance is measured, and b4 is determined accordingly.

[0149] The value of the parameter b2 is determined according to the GNSS vibration energy intensity distribution statistical result of the static water closed door phase for at least 24 hours, and the 95% quantile thereof is taken as the threshold multiple.

[0150] The final output is the corrected observation noise covariance, which physically reflects the influence of external interference on GNSS observation. By reducing the weight of abnormal observation, it can prevent the abnormal value from misleading the joint Kalman filter, thereby maintaining the stable description of the true deformation and vibration state.

[0151] S33, using the dynamically scaled process noise covariance and the corrected observation noise covariance, performing prediction and update iteration through joint Kalman filtering to output the optimal estimated deformation state and vibration state.

[0152] Further, step S33 comprises:

[0153] S331, performing a prediction step, specifically: for the deformation state, since the structural deformation process is slow and continuous, the optimal estimated value of the previous moment can be directly inherited to form the prediction value of the current moment. For the vibration state, the autoregressive model defined in step S312 is called to recursively calculate the prediction value of the current moment based on the vibration state of the previous moment and the autoregressive coefficient.

[0154] At the same time of generating the prediction value, the dynamically scaled process noise covariance is superimposed to the prediction result to form the prior covariance matrix. The prior covariance matrix refers to the description of the uncertainty range in the state prediction stage, which is used to provide the error boundary for the subsequent update link.

[0155] S332, performing an update step, specifically: calculating the Kalman gain according to the observation matrix, and calling the corrected observation noise covariance to finally determine the weight of the observation value in the update. Two types of abnormality judgment are introduced in the state correction process:

[0156] When the real-time carrier-to-noise ratio drops by more than 30%, the observation noise covariance R is further amplified to reduce the weight of the current observation value, preventing abnormal observation from interfering with state estimation;

[0157] When the vibration state prediction value deviates from the ultra-low frequency displacement feature in the monitoring point feature vector by more than 50%, the ultra-low frequency displacement feature in the monitoring point feature vector is taken as the reference to reset the observation value, thereby ensuring that the update result is consistent with the physical constraint.

[0158] Finally, the residual between the predicted value and the observed value is fused to obtain the accurate deformation trajectory and the accurate vibration state at the current time.

[0159] In step S4, the vibration health index is calculated based on the accurate deformation trajectory and the accurate vibration state to evaluate the health degree of the vibration level; the key point coordinates are geometrically calculated based on the accurate deformation trajectory and the accurate vibration state to output the alarm state identifier; and the graph neural network is input based on the accurate deformation trajectory and the accurate vibration state to realize classification of the failure mode of the hinge area.

[0160] Further, step S4 includes:

[0161] In S41, the vibration health index is calculated in combination with the vibration amplitude of the accurate vibration state, and is used to evaluate the health degree of the current vibration level relative to the historical extreme value.

[0162] The vibration health index is the ratio of the current vibration amplitude to the historical maximum vibration amplitude. Specifically, the evaluation criteria are divided into three levels:

[0163] When the vibration health index is less than the threshold H1, it is determined to be in a normal state, indicating that the water flow is running smoothly and the structure has not appeared significant vibration.

[0164] When the vibration health index is greater than the threshold H2, it is determined to be in an alarm state, indicating that the structure may have a potential resonance risk.

[0165] When the vibration health index is between H1 and H2, it is determined to be in a continuous monitoring state, indicating that the vibration level has increased but is still within a controllable range, and observation needs to be maintained.

[0166] H1 is the threshold value of the normal state, which can be obtained by statistically analyzing the stable range of vibration amplitudes from historical operation data; H2 is the threshold value of the alarm state, which can be determined by laboratory resonance test or finite element simulation to predict the potential risk level; in addition, to avoid excessive dependence on a single source, the final values of H1 and H2 can be corrected by regression combined with long-term monitoring results on site to ensure the adaptability of the threshold values to different working conditions.

[0167] For example, if H1=0.3 and H2=1.2 are set, when the vibration health index is less than 0.3, it is determined to be normal; when it is greater than 1.2, it is determined to be in an alarm state; and when it is between 0.3 and 1.2, it is in a continuous monitoring state. The vibration health index output finally serves as a direct quantitative indicator of the safety of the structure in operation, providing a reference basis for subsequent abnormal diagnosis and failure mode recognition.

[0168] In S42, the key point coordinates are geometrically calculated based on the accurate deformation trajectory to obtain the distance offset and the angle offset, and the residual parameters are formed in combination with the historical statistical reference to output the alarm state identifier according to the graded alarm rules.

[0169] Specifically, based on the accurate deformation trajectory, the real-time coordinates of the key points A, B and C are obtained. The real-time hinge length is calculated through the coordinates of points A and B, and the distance offset is obtained based on the theoretical hinge length of 2.000 meters. The distance offset refers to the difference between the real-time hinge length and the theoretical hinge length, which is used to quantify the risk of radial slip.

[0170] Based on the accurate deformation trajectory, the door leaf plane is constructed based on the coordinates of points A, B and C, the angle between the normal vector of the door leaf plane and the ideal vertical direction is calculated, and the angle offset is obtained. The angle offset Δθ refers to the difference between the normal vector of the door leaf plane and the 90-degree direction, which is used to quantify the risk of local distortion.

[0171] Under the condition of static water closed door, 24 hours of normal working condition data are continuously collected, and the historical mean and standard deviation of the distance offset and the historical mean and standard deviation of the angle offset are obtained as the alarm reference. During real-time monitoring, the residual parameters include the distance residual and the angle residual, and the specific calculation method is as follows:

[0172] Distance residual = real-time distance offset - distance offset historical mean;

[0173] Angle residual = real-time angle offset - angle offset historical mean.

[0174] Among them, the distance residual and the angle residual are directly used for anomaly detection, and are quantitative indicators for measuring the degree of real-time constraint violation.

[0175] Establish a hierarchical alarm rule:

[0176] First-level alarm (radial slip risk): triggered when the distance residual is greater than 3 times the distance offset historical standard deviation and lasts for 5 minutes;

[0177] Second-level alarm (distortion risk): triggered when the angle residual is greater than 4 times the angle offset historical standard deviation and lasts for 3 minutes;

[0178] Emergency alarm (compound failure): triggered when the first-level alarm and the second-level alarm occur at the same time.

[0179] The final output result is the alarm state identifier, which can provide a basis for rapid response for operation and maintenance personnel.

[0180] S43, the accurate deformation trajectory and the vibration state are fused, input into a graph neural network, and the failure mode of the hinge area is classified.

[0181] Further, step S43 includes:

[0182] S431, define the graph structure Please refer to Figure 2As shown, the node set contains three key monitoring positions, i.e., the left hinge point (A), the right hinge point (B), and the leaf connecting point (C). The edge set design includes the physical connection edges A-B (characterizing the hinge line deformation), B-C (characterizing the force transmission path of the support arm), and the implicit mechanical edge A-C (capturing the rigid conduction effect of the door leaf). The nodes and edges constitute a small graph describing the topology of the hinge area. Please refer to Figure 3 As shown, the key innovation of the full connection topology of A-B-C lies in the explicit introduction of the A-C "implicit mechanical edge". This physically reflects the strong correlation between the deformations of A point (left hinge) and C point (door leaf-support arm connecting point) caused by the rigidity of the door leaf (even without direct physical connection). This design enables the network to effectively capture cross-node coupling effects such as "left hinge jamming (A) leading to stress concentration on the right side of the door leaf (near C)".

[0183] S432, extract node features, each node outputs a 7-dimensional feature vector at time t: F1-F3: corresponding to the three-dimensional displacement of the deformation state; F4: corresponding to the energy of the ultra-low frequency displacement component, derived from step S22; F5: corresponding to the instantaneous distance change rate of this node relative to point A; F6: corresponding to the angle change rate of this node relative to point B; F7: corresponding to the main frequency of vibration.

[0184] Under a 50Hz sampling rate, a 100-frame sliding time window (corresponding to 2 seconds) is used, and the input tensor dimension is 3x100x7. The output time series feature tensor provides input for subsequent network learning.

[0185] S433, build and train the graph neural network and put it into use, which includes:

[0186] Spatial convolution: based on the adjacency matrix defined in S421, aggregate topologically related node features;

[0187] Temporal convolution: use one-dimensional convolution (convolution kernel size is 3) to extract time series features;

[0188] Comprehensive feature generation: after multiple layers of spatio-temporal convolution blocks, obtain the comprehensive feature vector of the hinge area within the time window through global pooling;

[0189] Output layer classification: output 8 categories of failure modes, including: normal, high-energy vibration, radial slip, axial slip, local twist, slip+twist, loose bolt, and bearing jamming.

[0190] The training data uses 5000 groups of finite element simulation samples and 200 groups of field injection fault samples, and the model parameters are optimized through distributed learning.

[0191] The final output result is the failure mode classification result, which serves as the basis for high-level safety assessment and decision support.

[0192] Example 2:

[0193] Referring to Figure 3 The embodiment shown provides a structure deformation automatic monitoring system based on GNSS three-dimensional displacement curve on the basis of Embodiment 1, and the system comprises the following functional modules:

[0194] A preliminary position sequence acquisition module: three-dimensional coordinate initial data of key points are acquired from a GNSS receiver, the three-dimensional coordinate initial data of key points are subjected to quality screening, simple smoothing and unified timestamp, and a preliminary position sequence is output.

[0195] A monitoring point feature vector extraction module: used for extracting relative motion and frequency spectrum features and splicing after linear interpolation on the basis of the preliminary position sequence, and a monitoring point feature vector is obtained.

[0196] A joint state estimation module: used for establishing a joint model of structure deformation state and vibration state in combination with the monitoring point feature vector. The two types of states are synchronously estimated by a filtering method, and an accurate deformation trajectory and an accurate vibration state are output.

[0197] A comprehensive analysis and alarm module: used for calculating a vibration health index to evaluate the health degree of vibration level based on the accurate deformation trajectory and the accurate vibration state; performing geometric calculation on key point coordinates based on the accurate deformation trajectory and the accurate vibration state, and outputting an alarm state identifier; and inputting a graph neural network based on the accurate deformation trajectory and the accurate vibration state to realize classification of hinge area failure modes.

[0198] The specific embodiments described above further illustrate the purposes, technical solutions and beneficial effects of the present application. It should be understood that the above description is only a specific embodiment of the present application and is not used to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A method for automatic monitoring of structural deformation based on GNSS three-dimensional displacement curve, characterized in that, Comprise: Step S1, obtain the initial data of the three-dimensional coordinates of the key points from the GNSS receiver, and perform quality screening, simple smoothing and unified timestamping on the initial data of the three-dimensional coordinates of the key points, and output a preliminary position sequence; Step S2, on the basis of the preliminary position sequence, extract relative motion and frequency spectrum characteristics after linear interpolation, and splice to obtain a feature vector of the monitoring point; Step S3, combine the feature vector of the monitoring point to establish a joint model of the structural deformation state and the vibration state, and output an accurate deformation trajectory and an accurate vibration state from the joint model after filtering; Step S4, based on the accurate deformation trajectory and the accurate vibration state, calculate the vibration health index to evaluate the health degree of the vibration level; Based on the accurate deformation trajectory and the accurate vibration state, perform geometric calculation on the key point coordinates to output an alarm state identifier; Based on the accurate deformation trajectory and the accurate vibration state, input a graph neural network to realize classification of the failure mode of the hinge area.

2. The method according to claim 1, wherein, The quality screening, simple smoothing and unified timestamping comprise: S11, collect the initial data stream of the three-dimensional coordinates of the key points from the GNSS receiver, perform quality screening, and obtain valid observations; S12, perform median filtering on the valid observations according to time and superimpose moving average to obtain a smoothed position sequence; S13, linearly resample the smoothed position sequence according to a unified timestamp and fill in missing data to output a preliminary position sequence.

3. The method according to claim 2, wherein, The quality screening comprises: Combine the initial data stream of the three-dimensional coordinates of the key points with the carrier-to-noise ratio, the geometric dilution of precision, the autonomous integrity monitoring residual and the cycle slip marker to generate a quality index sequence, form a quality mask according to a threshold set, and use the quality mask on the initial data stream to obtain the preliminary observations; On this basis, use coordinate difference to remove jump observations in the preliminary observations, use steady consistency marker to remove non-steady observations in the preliminary observations, and combine time consistency mask to screen complete samples, and finally obtain valid observations.

4. The method according to claim 3, wherein, The step of obtaining the feature vector of the monitoring point comprises: S21, linearly interpolate based on the preliminary position sequence to form a three-dimensional displacement sequence with equidistant timestamps; S22, calculate the key point distance change and the leaf normal angle change from the three-dimensional displacement sequence to obtain the key point relative motion parameters, and determine the analysis time window for frequency spectrum analysis according to the key point relative motion parameters; S23, combine the analysis time window to extract the dominant frequency from the three-dimensional displacement sequence, approximate low-pass on the low-frequency passband below the dominant frequency with moving average to obtain the ultra-low frequency displacement component, obtain the main frequency shift rate from the change of the dominant frequency relative to the reference, and represent the real-time vibration energy intensity with the amplitude adjacent to the dominant frequency; S24, combine the ultra-low frequency displacement component, the key point relative motion parameters, the real-time vibration energy intensity and the main frequency shift rate to construct the feature vector of the monitoring point.

5. The method according to claim 4, wherein, The step of outputting the accurate deformation trajectory and the vibration state comprises: S31, combine the feature parameters of the feature vector of the monitoring point to construct a joint state space model; S32, combine the real-time vibration energy intensity and the real-time carrier-to-noise ratio to dynamically adjust the process noise covariance and the observation noise covariance of the joint state space model, respectively, to obtain the dynamically scaled process noise covariance and the corrected observation noise covariance. S33, using the dynamically scaled process noise covariance and the corrected observation noise covariance, performing prediction and update iterations by joint Kalman filtering to output the optimally estimated deformation state and vibration state.

6. The method according to claim 5, wherein, The process noise covariance of the joint state space model is adjusted as follows: The dynamic scaling rule of the process noise covariance of the joint state space model is as follows: When the real-time vibration energy intensity exceeds a1 times the vibration energy reference value, the process noise covariance is enlarged to a2 times the reference value; wherein a1 is the vibration energy threshold multiple, a2 is the amplification coefficient in the high vibration case, and the vibration energy reference value is obtained by static water closed door phase statistics; When the real-time carrier-to-noise ratio is lower than the lower threshold of the carrier-to-noise ratio, the process noise covariance is enlarged to a3 times the reference value; wherein a3 is the amplification coefficient in the low signal-to-noise ratio case, and the lower threshold of the carrier-to-noise ratio is the minimum acceptable signal-to-noise ratio of the GNSS receiver in the stable working condition; When the real-time carrier-to-noise ratio is lower than the lower threshold and the vibration energy intensity is greater than a1 times the vibration energy reference value, the process noise covariance is enlarged to a4 times the reference value; wherein a4 is the amplification coefficient in the double interference case.

7. The method according to claim 6, wherein, The observation noise covariance of the joint state space model is adjusted as follows: The dynamic scaling rule of the observation noise covariance of the joint state space model is as follows: When the real-time carrier-to-noise ratio is less than the lower threshold of the carrier-to-noise ratio, the observation noise covariance is enlarged to b1 times the reference value; wherein b1 is the amplification coefficient in the case of insufficient carrier-to-noise ratio; When the real-time vibration energy intensity is greater than b2 times the vibration energy reference value, the observation noise covariance is enlarged to b3 times the reference value; wherein b2 is the vibration energy threshold multiple, and b3 is the amplification coefficient in the high vibration case; When the real-time carrier-to-noise ratio is lower than the lower threshold and the vibration energy intensity is greater than b2 times the vibration energy reference value, the observation noise covariance is enlarged to b4 times the reference value; wherein b4 is the amplification coefficient in the double interference case.

8. The method according to claim 6, wherein, The evaluation of the health degree of the vibration level, the output of the alarm state identifier, and the classification of the hinge area failure mode include: S41, combining the vibration amplitude of the accurate vibration state, calculating the vibration health index for evaluating the health degree of the current vibration level relative to the historical extreme value; S42, based on the geometric calculation of the key point coordinates by the accurate deformation trajectory, obtaining the distance offset and the angle offset, and combining the historical statistical reference to form the residual parameter, and outputting the alarm state identifier according to the grading alarm rule; S43, fusing the accurate deformation trajectory and the vibration state, inputting the graph neural network to realize the classification of the hinge area failure mode.

9. The method according to claim 8, wherein, The graph structure of the graph neural network is defined as: The node set includes the left hinge shaft point A, the right hinge shaft point B, and the leaf connecting point C; The edge set includes the physical connection edge A-B, B-C and the implied mechanical edge A-C, forming a fully connected topology, inputting the graph neural network to output the classification results of 8 types of failure modes.

10. A system for automatic monitoring of structural deformation based on GNSS three-dimensional displacement curves to perform the steps of the method according to any one of claims 1-9, characterized in that, It includes: A preliminary position sequence acquisition module: obtaining key point three-dimensional coordinate initial data from a GNSS receiver, performing quality screening and simple smoothing on the key point three-dimensional coordinate initial data, and unifying the time stamp to output a preliminary position sequence; The monitoring point feature vector extraction module is configured to extract relative motion and frequency spectrum features and splice them to obtain a monitoring point feature vector on the basis of the preliminary position sequence after linear interpolation; The joint state estimation module is configured to establish a joint model of structural deformation state and vibration state in combination with the monitoring point feature vector, and to perform synchronous estimation on the two types of states through a filtering method to output an accurate deformation trajectory and an accurate vibration state; The comprehensive analysis and alarm module is configured to calculate a vibration health index to evaluate the health degree of the vibration level based on the accurate deformation trajectory and the accurate vibration state; Based on the accurate deformation trajectory and the accurate vibration state, geometric calculation is performed on key point coordinates to output an alarm state identifier; Based on the accurate deformation trajectory and the accurate vibration state, a graph neural network is input to realize classification of the failure mode of the hinge area.

Images