Coal Mine Goaf Settlement Monitoring and Early Warning System Based on BeiDou Positioning
By constructing a two-dimensional spatiotemporal prior grid and a long short-term memory network for multipath errors, and combining them with an adaptive Kalman filter, the problem of overlapping multipath errors and weak deformation signals in the monitoring of surface subsidence in coal mine goaf areas was solved, achieving high-precision subsidence monitoring and early warning.
Patent Information
- Application Number
- CN202611119758.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-27
- Publication Date
- 2026-11-10
AI Technical Summary
In existing technologies for monitoring surface subsidence in coal mine goaf areas, the frequency bands of multipath errors and weak deformation signals overlap, making it difficult to effectively distinguish them through frequency domain feature separation. This results in weak deformation signals being over-smoothed or multipath errors remaining, making it impossible to obtain true subsidence deformation information.
A coal mine goaf settlement monitoring and early warning system based on BeiDou positioning was constructed. By extracting the carrier-to-noise ratio and elevation angle of satellite signals, a two-dimensional spatiotemporal prior grid of multipath error was established. The multipath error estimate was output using a long short-term memory network. Combined with an adaptive Kalman filter for data processing, the multipath error was removed and the BeiDou three-dimensional coordinates were extracted.
It effectively separated the multipath error and weak deformation signal of frequency band overlap, obtained the real BeiDou three-dimensional coordinate sequence, improved the continuity and reliability of settlement monitoring data, and realized high-precision surface settlement monitoring and early warning in mining subsidence areas.
Smart Images

Figure CN123065415A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of surveying and mapping technology, and discloses a coal mine goaf subsidence monitoring and early warning system based on Beidou positioning. Background Technology
[0002] Monitoring surface subsidence in coal mine goaf areas is a crucial aspect of ensuring safe production in mining areas. The BeiDou Navigation Satellite System (BDS) has become the mainstream method for goaf subsidence monitoring due to its all-weather operational capabilities. Conventional BeiDou subsidence monitoring techniques typically acquire dual-frequency carrier phase observations and pseudorange observations using a receiver, constructing observation equations that include subsidence deformation, ionospheric delay, tropospheric delay, and multipath error terms. In the data processing stage, a dual-frequency ionospheric-free combined formula is used to eliminate the influence of the first-order ionospheric delay term. The tropospheric delay term is corrected using empirical models or meteorological data. Subsequently, the observation equations are solved using the least squares method or Kalman filtering to obtain the three-dimensional coordinate time series of the monitoring points. The surface subsidence status of the goaf is assessed based on the changes in the three-dimensional coordinate time series. This conventional processing flow can obtain positioning results that meet accuracy requirements in open and unobstructed monitoring environments, providing fundamental data support for goaf subsidence assessment.
[0003] The complex surface environment of coal mine goaf areas, with mountains, buildings, or large machinery often obstructing the monitoring station, causes reflection and refraction of BeiDou signals, resulting in multipath errors. Existing technologies for handling multipath errors primarily focus on the data post-processing stage, typically employing empirical mode decomposition (EMD) or bandpass filtering to perform frequency domain analysis on the carrier phase residual sequence. The specific implementation process involves: after correcting ionospheric and tropospheric errors, calculating the carrier phase residual sequence; using EMD to decompose the residual sequence into multiple intrinsic mode functions (EMFs); eliminating high-frequency mode components to remove multipath errors; or designing a bandpass filter based on the frequency band range of the multipath error, filtering out signals within a specific frequency band as multipath errors, and retaining low-frequency components as subsidence deformation signals. This type of frequency domain feature separation method can reduce the impact of multipath errors in some stable scenarios.
[0004] The complex topography of mined-out areas changes dynamically with subsidence, resulting in multipath errors exhibiting strong nonlinearity and time- and spatially varying characteristics. Weak subsidence deformation signals and multipath errors severely overlap in frequency bands. Existing technologies employ empirical mode decomposition or bandpass filtering to process carrier phase residuals. When separating based on frequency domain features, it is difficult to distinguish between weak deformation signals and multipath errors with overlapping frequency bands. This can easily lead to over-smoothing of weak deformation signals or residual time-varying multipath errors, making it impossible to extract true subsidence deformation information. Since relying solely on frequency domain features cannot characterize the physical mapping relationship between multipath errors and signal quality parameters such as satellite elevation angle and carrier-to-noise ratio, and the filtering process lacks constraints on the spatial distribution characteristics of errors, existing frequency domain filtering-based multipath error elimination methods suffer from the core technical problem of failing to separate multipath errors with overlapping frequency bands from weak deformation signals. Summary of the Invention
[0005] The purpose of this invention is to provide a coal mine goaf subsidence monitoring and early warning system based on BeiDou positioning, which can solve the problems mentioned in the background art.
[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A coal mine goaf subsidence monitoring and early warning system based on BeiDou positioning includes: an observation data receiving end, which acquires BeiDou dual-frequency carrier phase observations and pseudorange observations, and constructs observation equations including subsidence deformation terms, ionospheric delay terms, tropospheric delay terms, and multipath error terms; a priori grid construction end, which, based on ionospheric and tropospheric error corrections, extracts the carrier-to-noise ratio and elevation angle of each satellite signal, and constructs a two-dimensional spatiotemporal priori grid for multipath error; and an error stripping and coordinate extraction end, which inputs the two-dimensional spatiotemporal priori grid and the carrier phase residual sequence into a long short-term memory network, outputs an estimated value of the multipath error and subtracts it from the carrier phase residual, and extracts the BeiDou three-dimensional coordinate sequence; wherein, the loss function of the long short-term memory network introduces a deformation continuity constraint regularization term.
[0007] Preferably, after the observation data receiving end acquires the BeiDou dual-frequency carrier phase observation value and pseudorange observation value, it eliminates the influence of the first-order term of the ionospheric delay term based on the ionospheric-free combination formula, and constructs wide-lane combined observation values using the pseudorange observation value and the carrier phase observation value with the first-order ionospheric influence eliminated; it solves the wide-lane ambiguity floating-point solution based on the wide-lane combined observation value, calculates the narrow-lane ambiguity floating-point solution by combining the ionospheric-free combined observation value, fixes the wide-lane ambiguity and narrow-lane ambiguity using the LAMBDA search method, and substitutes the fixed ambiguity back into the observation equation to obtain the carrier phase residual sequence.
[0008] Preferably, after extracting the carrier-to-noise ratio and elevation angle of each satellite signal, the prior grid construction end establishes a two-dimensional planar grid with elevation angle as the horizontal axis and carrier-to-noise ratio as the vertical axis based on the physical mapping relationship that the multipath effect is negatively correlated with the satellite elevation angle and positively correlated with the carrier-to-noise ratio fluctuation. The carrier phase residual variance values of each satellite signal falling into the corresponding grid cell of the two-dimensional planar grid in multiple historical epochs are statistically analyzed. The carrier phase residual variance values are used as the multipath error intensity characterization values of the grid cells. The two-dimensional spatiotemporal prior grid of the multipath error is constructed from the multipath error intensity characterization values.
[0009] Preferably, before inputting the two-dimensional spatiotemporal prior grid and the carrier phase residual sequence into the long short-term memory network, the error stripping and coordinate extraction end performs an inter-epoch high-order difference operation on the carrier phase residual sequence to remove gross errors, and performs min-max normalization processing on the differenced carrier phase residual sequence and the grid values of the two-dimensional spatiotemporal prior grid respectively; the normalized carrier phase residual sequence and the normalized grid values are concatenated along the feature dimension to form a joint feature vector, which is then input into the input layer of the long short-term memory network.
[0010] Preferably, the loss function of the Long Short-Term Memory (LSTM) network incorporates a deformation continuity constraint regularization term as follows: based on the physical constraint that the surface subsidence in the goaf conforms to first-order or second-order difference smoothing between adjacent epochs, a regularization penalty term is constructed that includes the subsidence difference value between adjacent epochs corresponding to the estimated value of the multipath error output by the network; the regularization penalty term is weighted and summed with the mean square error loss between the multipath error value predicted by the LTM network and the actual multipath error label to form the total loss function of the LTM network, constraining the LTM network to maintain the temporal smoothness of subsidence deformation when fitting multipath errors.
[0011] Preferably, when using the LAMBDA search method to fix the wide-lane ambiguity and narrow-lane ambiguity, the RATIO test value after ambiguity fixation is calculated; if the RATIO test value is less than a preset fixed threshold, satellite observations in the current epoch with satellite elevation angles lower than the preset elevation angle threshold and carrier-to-noise ratio lower than the preset signal-to-noise ratio threshold are removed, and the dimensionality-reduced ambiguity search space of the remaining satellites is obtained; the LAMBDA search is re-executed in the dimensionality-reduced ambiguity search space and the updated RATIO test value is calculated, until the updated RATIO test value is greater than or equal to the preset fixed threshold, and ambiguity fixation is completed.
[0012] Preferably, after constructing a two-dimensional spatiotemporal prior grid of the multipath error from the multipath error intensity characterization value, based on the periodic repetition characteristics of the BeiDou satellite orbit, historical two-dimensional spatiotemporal prior grids corresponding to historical epochs with the same satellite configuration as the current epoch are extracted; the two-dimensional spatiotemporal prior grid constructed for the current epoch is fused with the historical two-dimensional spatiotemporal prior grid using an exponential smoothing weighting method; the two-dimensional spatiotemporal prior grid for the current epoch is corrected using the spatial distribution information of historical multipath errors, and the final dynamic spatiotemporal prior grid used for inputting into the Long Short-Term Memory network is obtained.
[0013] Preferably, the error stripping and coordinate extraction end introduces a multi-head self-attention mechanism after the hidden layer output of the Long Short-Term Memory network; the temporal hidden feature vector extracted by the Long Short-Term Memory network is mapped to a query matrix, a key matrix, and a value matrix; the attention weight distribution is obtained based on the dot product operation of the query matrix and the key matrix; the value matrix is weighted and summed using the attention weight distribution to strengthen the feature expression corresponding to the strong multipath error region in the joint feature vector; and the output vector of the multi-head self-attention mechanism is mapped to the estimated value of the multipath error.
[0014] Preferably, after the error stripping and coordinate extraction end extracts the BeiDou three-dimensional coordinate sequence, the BeiDou three-dimensional coordinate sequence is input into an adaptive Kalman filter; the innovation covariance is calculated in real time based on the innovation vector of the adaptive Kalman filter, and the observation noise covariance matrix and the process noise covariance matrix are dynamically estimated based on the innovation covariance within the sliding window; when the Mahalanobis distance of the innovation vector exceeds the chi-square distribution test threshold, an adaptive attenuation factor is constructed to reduce the observation weight of the current epoch, and the BeiDou three-dimensional coordinate sequence is smoothed and gross error suppressed.
[0015] Preferably, it also includes a tiered early warning terminal, which calculates the three-dimensional displacement rate and cumulative displacement of the surface of the goaf based on the extracted BeiDou three-dimensional coordinate sequence; constructs a multi-level settlement early warning threshold system according to the geological and mining conditions and overlying structure characteristics of the goaf, and compares the three-dimensional displacement rate and cumulative displacement with the corresponding level of early warning threshold; when any indicator of the three-dimensional displacement rate or cumulative displacement reaches the corresponding early warning threshold, the corresponding level of early warning signal is triggered, and the early warning signal is bound to the BeiDou coordinates of the corresponding monitoring point and pushed to the remote monitoring platform.
[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention extracts the carrier-to-noise ratio (CNR) and elevation angle of each satellite signal to construct a two-dimensional spatiotemporal prior grid for multipath errors. The physical mapping relationship between multipath errors and satellite elevation angle and CNR is transformed into prior grid features. This two-dimensional spatiotemporal prior grid and the carrier phase residual sequence are input into a long short-term memory (LSTM) network, outputting an estimated value of the multipath error, which is subtracted from the carrier phase residual. This changes the traditional method of relying on frequency domain features to separate errors, transforming the multipath error stripping process into a data-driven estimation process guided by physical features. This overcomes the signal confusion caused by frequency band overlap and avoids... The problem of weak deformation signals being over-smoothed or time-varying multipath errors being retained was addressed. The loss function of the Long Short-Term Memory Network (LSTM) introduced a deformation continuity constraint regularization term. Based on the physical law that surface subsidence in the goaf conforms to differential smoothing between adjacent epochs, a regularization penalty term containing the subsidence difference values of adjacent epochs was constructed and weighted and summed with the mean square error loss. This constrained the network to maintain the temporal smoothness of subsidence deformation when fitting multipath errors, preventing the network from fitting weak nonlinear subsidence deformation as multipath errors. This achieved the separation of weak deformation signals and multipath errors under frequency band overlap conditions and obtained the true BeiDou three-dimensional coordinate sequence.
[0017] 2. This invention, based on the periodic repetition characteristics of BeiDou satellite orbits, extracts historical two-dimensional spatiotemporal prior grids corresponding to historical epochs with the same satellite configuration as the current epoch. It then uses an exponential smoothing weighting method to fuse the two-dimensional spatiotemporal prior grid constructed for the current epoch with the historical two-dimensional spatiotemporal prior grids. Furthermore, it uses historical multipath error spatial distribution information to correct the current epoch's grid, obtaining a dynamic spatiotemporal prior grid. This reduces the random fluctuations of the current epoch's prior grid and enhances the stability of the input features. When the fixed ambiguity test value is less than a preset fixed threshold, satellites with elevation angles below a preset elevation angle threshold and carrier-to-noise ratios below a preset signal-to-noise ratio are removed from the current epoch. Satellite observations with a noise ratio threshold are used to obtain a dimension-reduced ambiguity search space, within which the search is re-executed, eliminating interference from low-quality satellite observations and improving the success rate of ambiguity fixation. The BeiDou 3D coordinate sequence is input into an adaptive Kalman filter, and the innovation covariance is calculated in real time based on the innovation vector. The observation noise covariance matrix and the process noise covariance matrix are dynamically estimated. When the Mahalanobis distance of the innovation vector exceeds the chi-square distribution test threshold, an adaptive attenuation factor is constructed to reduce the observation weight of the current epoch. This smooths and suppresses gross errors in the BeiDou 3D coordinate sequence, reducing the interference of abnormal observations on the coordinate sequence and improving the continuity and reliability of settlement monitoring data. Attached Figure Description
[0018] Figure 1 This is a flowchart illustrating the overall workflow of the coal mine goaf subsidence monitoring and early warning system based on BeiDou positioning according to the present invention. Figure 2This is a flowchart of the process for fixing the ambiguity of BeiDou dual-frequency observations and solving the carrier phase residual in this invention. Figure 3 This is a flowchart of the dynamic construction process of the two-dimensional spatiotemporal prior grid for multipath error in this invention; Figure 4 This is a flowchart of the joint feature preprocessing and LSTM multipath error intelligent estimation process of the present invention. Figure 5 This is a flowchart of the adaptive Kalman filter coordinate sequence smoothing and gross error suppression of the present invention; Figure 6 This is a flowchart of the goaf subsidence classification, early warning determination, and remote push notification process of the present invention. Detailed Implementation
[0019] refer to Figure 1 In one embodiment, a BeiDou-based coal mine goaf subsidence monitoring and early warning system is deployed at multiple monitoring stations on the surface of the coal mine goaf. Each monitoring station is equipped with a BeiDou dual-frequency receiver, and the receiver antenna is mounted on a forced centering base at the monitoring point, with the antenna phase center coinciding with the center of the monitoring point marker. The system achieves high-precision monitoring of goaf surface subsidence through the coordinated operation of the observation data receiving end, the prior grid construction end, and the error stripping and coordinate extraction end.
[0020] The observation data receiver acquires dual-frequency carrier phase observations and pseudorange observations broadcast by the BeiDou Navigation Satellite System in real time. The sampling interval of the observation data is set according to the monitoring accuracy requirements. The observation data receiver parses the acquired raw observations to extract the carrier phase observations, pseudorange observations, satellite ephemeris data, satellite clock bias parameters, and receiver clock bias parameters for each visible satellite at each epoch. Based on the extracted observations and parameters, the observation data receiver constructs observation equations that include subsidence deformation terms, ionospheric delay terms, tropospheric delay terms, and multipath error terms.
[0021] For BeiDou dual-frequency observations, the mathematical expression for the carrier phase observation equation is:
[0022] in, Indicates receiver In the calendar For satellite The Frequency point carrier phase observations, in cycles; Indicates receiver With satellite In the calendar The geometric distance, in meters; The speed of light in a vacuum is expressed in meters per second. This indicates the receiver clock error, in seconds. Indicates satellite In the calendar The clock difference, in seconds; Indicates satellite The wavelength of the carrier signal (in meters), corresponding to the wavelengths of the B1I and B2I frequencies respectively in the dual-frequency case; Indicates receiver For satellite The Frequency point carrier phase integer ambiguity; Indicates satellite Signal in epoch Ionospheric delay, in meters; Indicates satellite Signal in epoch The tropospheric delay, in meters; Indicates satellite Signal in epoch The multipath error is expressed in meters. Indicates satellite Carrier phase observations at epochs Random noise, in meters.
[0023] The mathematical expression for the pseudorange observation equation is:
[0024] in, Indicates receiver In the calendar For satellite The pseudorange observations are in meters; Indicates receiver With satellite In the calendar geometric distance; Indicates satellite pseudorange observations at epochs The random noise is expressed in meters.
[0025] To facilitate subsequent error correction and parameter solving, the observation data receiver transforms the aforementioned observation equations into a linearized form with the receiver's three-dimensional coordinates, clock error, ambiguity, ionospheric delay, tropospheric delay, and multipath error as unknown parameters. The linearization process uses the approximate coordinates of the monitoring points as a reference, and adjusts the geometric distance... Perform a Taylor expansion at the approximate coordinates and take the first-order term. The approximate coordinates of the monitoring point are: ,satellite In the calendar The coordinates are Then the geometric distance The approximate value is:
[0026] Expanding the geometric distance at approximate coordinates using Taylor series and taking the first-order term, we obtain:
[0027] in, , , These represent the differences between the actual and approximate coordinates of the monitoring point in three directions, respectively; the formula for calculating the partial derivative is:
[0028] Substituting the above expansion into the carrier phase observation equation, we obtain the linearized observation equation:
[0029] in, Represents the residual of the carrier phase observation; The constant term represents the carrier phase observation value.
[0030] After the observation data receiver completes the construction of the observation equations, preliminary corrections are made to the ionospheric delay and tropospheric delay terms. For the ionospheric delay term, an ionospheric-free combination method of dual-frequency observations is used to eliminate the influence of the first-order ionospheric delay. For the tropospheric delay term, the Saastamoinen model is used to calculate the dry tropospheric delay component, and a random walk model is used to estimate the wet tropospheric delay component. The tropospheric delay is divided into two parts: dry delay and wet delay. The dry delay is caused by dry gas components in the atmosphere, and the wet delay is caused by water vapor in the atmosphere. The formula for calculating the dry tropospheric delay using the Saastamoinen model is as follows:
[0031] in, This indicates the tropospheric dry delay, in meters. This indicates the ground air pressure at the observation station, expressed in hectopascals (hPa). Indicates the geographical latitude of the observation station, in radians; This indicates the altitude of the observation station, expressed in kilometers. This represents the zenith distance of a satellite, expressed in radians. The relationship between zenith distance and elevation angle is as follows: .
[0032] The tropospheric wet delay is estimated using a random walk model, and its state equation is:
[0033] in, Represents the epoch Tropospheric wet delay; The process noise is represented by a Gaussian distribution with a mean of zero. In the observation equation, the tropospheric wet delay is taken as the parameter to be estimated and solved together with the receiver's three-dimensional coordinates, clock error, and ambiguity. After ionospheric and tropospheric error corrections, the main remaining error sources in the observation equation are multipath error and observation noise. The parameters to be estimated mainly include the receiver's three-dimensional coordinates, clock error, and carrier phase integer ambiguity.
[0034] The prior mesh construction end obtains the signal-to-noise ratio (SNR) and elevation angle information of each visible satellite at each epoch from the observation data receiver. The satellite elevation angle is calculated from satellite ephemeris data, specifically the satellite's elevation angle in the observation station's horizontal coordinate system; the SNR is the ratio of the signal carrier power to the noise power density output by the receiver, measured in dB-Hz. Based on the physical mapping relationship between multipath effects and satellite elevation angle (negatively correlated with SNR fluctuations), the prior mesh construction end establishes a two-dimensional planar mesh with elevation angle as the horizontal axis and SNR as the vertical axis.
[0035] The horizontal axis elevation angle range of the two-dimensional planar grid is set to 0° to 90°, and the vertical axis carrier-to-noise ratio range is set to 20dB-Hz to 50dB-Hz. The size of the grid cells is set according to the complexity of the monitoring environment, and the grid step size in the elevation angle direction is consistent with the grid step size in the carrier-to-noise ratio direction. For each visible satellite at each epoch, the prior grid construction terminal determines its corresponding grid cell in the two-dimensional planar grid based on its elevation angle and carrier-to-noise ratio.
[0036] The prior grid construction stage statistically analyzes the variance of carrier phase residuals for each satellite signal falling into the corresponding grid cell across multiple historical epochs. The carrier phase residual is the difference between the observed value and the theoretical value calculated based on approximate parameters. For each grid cell, the variance of all historical carrier phase residuals falling into that cell is calculated, and this variance is used as the multipath error intensity characterization value for that grid cell. The larger the multipath error intensity characterization value, the more severe the multipath effect on the satellite signal corresponding to that grid cell. The two-dimensional matrix composed of the multipath error intensity characterization values of all grid cells constitutes the two-dimensional spatiotemporal prior grid of multipath error.
[0037] The error stripping and coordinate extraction end obtains the carrier phase residual sequence from the observation data receiving end and the two-dimensional spatiotemporal prior grid from the prior grid construction end. The error stripping and coordinate extraction end inputs the two-dimensional spatiotemporal prior grid and the carrier phase residual sequence into the long short-term memory network, outputs the estimated value of the multipath error and subtracts it from the carrier phase residual, and extracts the BeiDou three-dimensional coordinate sequence.
[0038] Long Short-Term Memory (LSTM) networks employ a multi-layered stacked structure, including an input layer, hidden layers, and an output layer. The input layer receives the joint feature vector, the hidden layers consist of multiple LSM units used to extract temporal features, and the output layer outputs an estimate of the multipath error. The internal structure of a LSM unit includes a forget gate, an input gate, a cell state, and an output gate. The forget gate determines which information to discard from the cell state; its expression is:
[0039] in, Represents the epoch The output of the forget gate; This represents the sigmoid activation function; The weight matrix representing the forget gate; Represents the epoch The hidden layer output; Represents the epoch The joint feature vector input to the Long Short-Term Memory network (containing normalized carrier phase residual sequence values and two-dimensional spatiotemporal prior grid values). This represents the bias vector of the forget gate.
[0040] The input gate determines which new information is stored in the cell state, and includes two parts: input gate activation and candidate cell state.
[0041]
[0042] in, Represents the epoch Input gate output; This represents the weight matrix of the input gate; This represents the bias vector of the input gate; Represents the epoch Candidate cell state; A weight matrix representing the state of candidate cells; A bias vector representing the state of a candidate cell; This represents the hyperbolic tangent activation function.
[0043] The formula for updating cell state is:
[0044] in, Represents the epoch cellular state; Represents the epoch The cellular state.
[0045] The output gate is used to determine the output value of the hidden layer, and its expression is:
[0046]
[0047] in, Represents the epoch The output gate outputs; This represents the weight matrix of the output gate; This represents the bias vector of the output gate; Represents the epoch The hidden layer output.
[0048] The loss function of the Long Short-Term Memory (LSTM) network incorporates a deformation continuity constraint regularization term to ensure that the network's output multipath error estimate conforms to the physical laws of surface subsidence in the goaf area. The overall loss function expression for the LSM network is:
[0049] in, Represents the total loss function; This represents the mean squared error loss, used to measure the difference between the multipath error value predicted by the network and the true multipath error label; Represents the weight coefficient of the regularization term; This represents the regularization term for deformation continuity constraints.
[0050] Mean square error loss The expression is:
[0051] in, This represents the total number of epochs in the training samples; Represents the epoch The true multipath error label; Indicates the network in epochs Output multipath error estimate.
[0052] Deformation continuity constraint regularization term Based on the physical constraint that the surface subsidence of the goaf conforms to first-order difference smoothing between adjacent epochs, its expression is:
[0053] in, This represents the epoch calculated based on the multipath error estimate from the network output. With the era The settlement difference between them; Represents the epoch With the era The actual settlement difference between them.
[0054] During network training, the total loss function is minimized using the backpropagation algorithm to update the network's weight parameters. After training, the Long Short-Term Memory (LSTM) network can accurately estimate the magnitude of the multipath error based on the input carrier phase residual sequence and the two-dimensional spatiotemporal prior grid.
[0055] The error stripping and coordinate extraction end subtracts the multipath error estimate output by the Long Short-Term Memory network from the corresponding carrier phase residual, obtaining a residual sequence containing only subsidence deformation signal and residual random noise. Based on the corrected residual sequence, the error stripping and coordinate extraction end resolves the observation equation to obtain the BeiDou three-dimensional coordinate sequence of the monitoring points. The BeiDou three-dimensional coordinate sequence contains the longitude, latitude, and elevation information of each monitoring point in the Earth coordinate system, where the elevation information is used to characterize the subsidence deformation of the surface in the goaf.
[0056] To clearly explain the physical meaning and value range of each parameter in the above observation equation, this embodiment provides the definition table of BeiDou dual-frequency observation parameters shown in Table 1.
[0057] Table 1. Definition of BeiDou Dual-Frequency Observation Parameters
[0058] This embodiment constructs a two-dimensional spatiotemporal prior grid containing the physical mapping relationship of multipath errors, transforming the spatial distribution characteristics of multipath errors into quantifiable prior information. This prior information, along with the carrier phase residual sequence, is input into a long short-term memory network for multipath error estimation. The deformation continuity constraint regularization term introduced in the loss function effectively prevents the network from misjudging weak settlement deformation signals as multipath errors, achieving separation of multipath errors and settlement deformation signals under overlapping frequency band conditions.
[0059] refer to Figure 2 In one optional embodiment, after the observation data receiver acquires the BeiDou dual-frequency carrier phase observations and pseudorange observations, it eliminates the influence of the first-order term of the ionospheric delay term based on the ionospheric-free combination formula, and constructs wide-lane combined observations using the pseudorange observations and the carrier phase observations with the first-order ionospheric influence eliminated. The wide-lane ambiguity floating-point solution is solved based on the wide-lane combined observations, and the narrow-lane ambiguity floating-point solution is calculated by combining the ionospheric-free combined observations. The LAMBDA search method is used to fix the wide-lane and narrow-lane ambiguities, and the fixed ambiguities are substituted back into the observation equations to obtain the carrier phase residual sequence.
[0060] The expression for the carrier phase observation value of the ionospheric combination is:
[0061] in, Represents the epoch For satellite The ionospherically-free combined carrier phase observations, in cycles; The carrier frequency of the BeiDou B1I frequency point ( =1561.098MHz); The carrier frequency of the BeiDou B2I frequency point ( =1207.140MHz). This indicates that receiver r is at epoch s for satellite The carrier phase observations at the B1I frequency point, in cycles; This indicates that receiver r is at epoch s for satellite The carrier phase observations at the B2I frequency point are in cycles.
[0062] The pseudorange observation expression for the ionospheric combination is:
[0063] in, This represents pseudorange observations for the ionospheric-free combination, in meters. The pseudorange observation value at the B1I frequency point is represented in meters. This represents the pseudorange observation value at the B2I frequency point, in meters.
[0064] Ionospheric-free combined observations can eliminate the influence of the first-order term of ionospheric delay, and the observation equations no longer contain the ionospheric delay parameter, simplifying the subsequent parameter solution process.
[0065] The wide-lane combined observation value is obtained by subtracting the carrier phase observation values at frequency B1I and frequency B2I, and its expression is:
[0066] in, This represents the carrier phase observation value of the wide-lane combination, in cycles. The wavelength of the wide-lane combination is:
[0067] in, This represents the wavelength of the wide-lane combination, measured in meters. Because the wavelength of the wide-lane combination is relatively long, its ambiguity is easier to fix.
[0068] The observation data receiver constructs an observation equation using pseudorange observations and wide-lane combined carrier phase observations, and solves for the floating-point solution of the wide-lane ambiguity using the least squares method. During the solution process, pseudorange observations are used as prior information to constrain the range of wide-lane ambiguity values.
[0069] After obtaining the floating-point solution for the wide-lane ambiguity, the observation data receiver calculates the floating-point solution for the narrow-lane ambiguity by combining the observations with the ionosphere-free combined observations. The expression for the narrow-lane combined observations is:
[0070] in, This represents the carrier phase observation value for the narrow-lane combination, in cycles. The wavelength of the narrow-lane combination is:
[0071] in, The wavelength of the narrow alley combination is represented in meters.
[0072] The observation data receiver employs the LAMBDA search method to perform integer searches on the floating-point solutions of wide-lane and narrow-lane ambiguities, obtaining fixed integer solutions for the ambiguities. The LAMBDA search method reduces search complexity by performing Cholesky decomposition on the covariance matrix of the floating-point solutions of the ambiguities, transforming the ambiguity space into an orthogonal integer space. The objective function of the LAMBDA search method is... The Mahalanobis distance between the integer and floating-point solutions of ambiguity is expressed as follows:
[0073] in, Represents the candidate integer ambiguity vector; Represents the floating-point solution vector of ambiguity; This represents the covariance matrix of the floating-point solution to the ambiguity. The smaller the objective function value, the smaller the difference between the candidate integer ambiguity vector and the floating-point solution vector, and the higher the probability that it is the true ambiguity vector.
[0074] During the search process, the covariance matrix is first analyzed. Perform Cholesky decomposition to obtain the lower triangular matrix. , making By transformation Transform the objective function into:
[0075] The transformed objective function is a vector. The search space is transformed into a hypersphere centered at the origin by squared modulus, which greatly reduces the search complexity. During the search process, the objective function value of each candidate integer ambiguity combination is calculated, and the combination with the smallest objective function value is the optimal fixed ambiguity solution.
[0076] When using the LAMBDA search method to fix the wide-lane and narrow-lane ambiguities, the RATIO test value after ambiguity fixing is calculated. The RATIO test value is the ratio of the objective function value of the suboptimal ambiguity combination to the objective function value of the optimal ambiguity combination, used to measure the reliability of ambiguity fixing. If the RATIO test value is less than a preset fixing threshold, the current ambiguity fixing result is considered unreliable, and ambiguity fixing needs to be performed again.
[0077] When the RATIO test value is less than a preset fixed threshold, the observation data receiver removes satellite observations in the current epoch that have a satellite elevation angle lower than the preset elevation angle threshold and a carrier-to-noise ratio lower than the preset signal-to-noise ratio threshold. Satellite signals with low elevation angles and low carrier-to-noise ratios are more severely affected by multipath effects and atmospheric refraction, resulting in lower quality observations and a higher likelihood of ambiguity fixing failure. After removing low-quality satellite observations, the remaining satellite observations constitute a dimension-reduced ambiguity search space.
[0078] The observation data receiver re-executes the LAMBDA search in the dimensionality-reduced ambiguity search space to calculate the updated RATIO test value. If the updated RATIO test value is still less than the preset fixed threshold, the satellite observations with the lowest elevation angle and lowest carrier-to-noise ratio among the remaining satellites are removed, the ambiguity search space is narrowed again, and the search is re-executed. This process is repeated until the updated RATIO test value is greater than or equal to the preset fixed threshold, at which point ambiguity fixing is complete.
[0079] After substituting the fixed wide-lane ambiguity and narrow-lane ambiguity back into the corresponding observation equation, the formula for calculating the carrier phase residual is:
[0080] in, Represents the epoch For satellite The carrier phase residual; This represents the fixed carrier phase integer ambiguity. The residual includes multipath error, residual ionospheric delay, residual tropospheric delay, and subsidence deformation signal, and is the basic data for subsequent multipath error estimation.
[0081] To clearly explain the meaning and judgment criteria of each inspection parameter in the ambiguity fixing process, this embodiment provides the ambiguity fixing quality inspection parameter table shown in Table 2.
[0082] Table 2. Quality Inspection Parameters for Ambiguity Fixation
[0083] This embodiment improves the efficiency of ambiguity fixing by constructing a combination of wide-lane and narrow-lane observations, solving for ambiguity floating-point solutions in stages, and using the LAMBDA search method for integer fixing. When the RATIO test value does not meet the requirements, the ambiguity search space is narrowed by eliminating low-quality satellite observations, thus eliminating interference from low-quality observations and improving the success rate of ambiguity fixing.
[0084] refer to Figure 3In one optional embodiment, after extracting the carrier-to-noise ratio (CNR) and elevation angle of each satellite signal at the prior grid construction end, a two-dimensional planar grid is established based on the physical mapping relationship that the multipath effect is negatively correlated with the satellite elevation angle and positively correlated with the CNR fluctuation. The grid has the elevation angle as the horizontal axis and the CNR as the vertical axis. The carrier phase residual variance values of each satellite signal falling into the corresponding grid cells of the two-dimensional planar grid in multiple historical epochs are statistically analyzed. The carrier phase residual variance values are used as the multipath error intensity characterization values of the grid cells. A two-dimensional spatiotemporal prior grid of multipath error is constructed from the multipath error intensity characterization values.
[0085] The elevation angle horizontal axis of the two-dimensional planar grid is divided into multiple grid cells at equal intervals from 0° to 90°, and the carrier-to-noise ratio vertical axis is divided into the same number of grid cells at equal intervals from 20dB-Hz to 50dB-Hz. For each visible satellite at each epoch, the prior grid construction end determines its corresponding row and column index in the two-dimensional planar grid based on its elevation angle and carrier-to-noise ratio values, thus locating the specific grid cell.
[0086] The prior mesh construction module maintains a historical residual database to store carrier phase residual data and corresponding satellite elevation angles and carrier-to-noise ratios for all epochs within a specified time period. The historical residual database is updated using a sliding window mechanism with a window length of 7 days. At the end of each day, all carrier phase residual data and corresponding satellite elevation angles and carrier-to-noise ratios for that day are added to the database, while historical data outside the window is deleted. This sliding window update mechanism ensures that the data in the historical residual database reflects recent multipath error distribution characteristics, improving the timeliness of the two-dimensional spatiotemporal prior mesh.
[0087] For each grid cell, the prior grid builder retrieves all carrier phase residual data falling within that cell from the historical residual database and calculates the variance of these residual data. The formula for calculating the variance is:
[0088] in, Indicates the index number of the grid cell in a two-dimensional planar grid ( =1,2,…, =1,2,…, , (Total number of grid cells); Indicates the first The variance of carrier phase residuals in each grid cell; Indicates falling into the first The number of historical residual data for each grid cell; Indicates falling into the first The first grid cell of the first grid cell Historical residual data ( =1,2,…, ); Indicates falling into the first The average of all historical residual data for each grid cell.
[0089] The calculated residual variance value of each grid cell is filled into a two-dimensional matrix to form an initial two-dimensional spatiotemporal prior grid. The initial two-dimensional spatiotemporal prior grid reflects the average intensity distribution of multipath errors under different elevation angles and carrier-to-noise ratios.
[0090] After constructing a two-dimensional spatiotemporal prior grid of multipath error using the multipath error intensity characterization values, the prior grid construction end extracts the historical two-dimensional spatiotemporal prior grids corresponding to historical epochs with the same satellite configuration as the current epoch, based on the periodic repetition characteristics of the BeiDou satellite orbit. The orbital period of the BeiDou satellite is approximately 12 hours, so the satellite's spatial configuration will repeat approximately every 12 hours.
[0091] The prior mesh construction stage calculates the satellite configuration similarity between the current epoch and historical epochs. Satellite configuration similarity is measured by the root mean square error (RMSE) of the elevation and azimuth angles of all commonly visible satellites in both epochs. The calculation of satellite configuration similarity first extracts the elevation and azimuth angles of all commonly visible satellites in the current and historical epochs. For each commonly visible satellite, the difference in elevation and azimuth angles between the two epochs is calculated. The formula for calculating the RMS error of satellite configuration similarity is:
[0092] in, Indicates the number of satellites that are commonly visible; Indicates the satellite in the current epoch The altitude angle; Satellites in historical eras The altitude angle; Indicates the satellite in the current epoch The azimuth angle; Satellites in historical eras The azimuth angle. When the RMSE is less than the preset similarity threshold, the satellite configurations of the two epochs are considered to be the same.
[0093] After extracting the historical epoch with the same satellite configuration as the current epoch, the prior grid construction end obtains the corresponding historical two-dimensional spatiotemporal prior grid. The two-dimensional spatiotemporal prior grid constructed for the current epoch is then fused with the historical two-dimensional spatiotemporal prior grid using an exponential smoothing weighting method. The formula for exponential smoothing weighting is:
[0094] in, Indicates the fusion of the first The multipath error intensity characterization value of each grid cell; This represents the 2nd-dimensional spatiotemporal prior grid constructed in the current epoch. The characterization value of each grid cell; Represents the first in a two-dimensional spatiotemporal prior grid of history The characterization value of each grid cell; This represents the smoothing coefficient, with a value ranging from 0 to 1.
[0095] Smoothing coefficient The value of is set based on the reliability of historical data; the more recent the historical data, the higher its reliability, and the smaller the corresponding smoothing coefficient. Through exponential smoothing weighted fusion, the spatial distribution information of historical multipath errors is used to correct the two-dimensional spatiotemporal prior grid of the current epoch, reducing the random fluctuations of the current epoch's prior grid and enhancing the stability of the input features. The fused dynamic spatiotemporal prior grid serves as the prior feature of the final input Long Short-Term Memory network.
[0096] To clearly illustrate the structure and parameter meaning of the two-dimensional spatiotemporal prior grid, this embodiment provides a two-dimensional spatiotemporal prior grid cell parameter table as shown in Table 3. Table 3 lists the parameter values of some typical grid cells.
[0097] Table 3. Parameters of Two-Dimensional Spatiotemporal Prior Grid Elements
[0098] This embodiment constructs a two-dimensional spatiotemporal prior grid for multipath errors by statistically analyzing the variance of historical residuals, quantifying the physical mapping relationship between multipath errors and satellite elevation angle and carrier-to-noise ratio into grid features. Based on the periodic repetition characteristics of BeiDou satellite orbits, it integrates prior grids with the same historical satellite configurations and uses historical multipath error distribution information to correct the current grid, thereby improving the accuracy and stability of the prior grid.
[0099] refer to Figure 4 In one optional embodiment, before inputting the two-dimensional spatiotemporal prior grid and the carrier phase residual sequence into the Long Short-Term Memory (LSTM) network at the error stripping and coordinate extraction end, an inter-epoch high-order differencing operation is performed on the carrier phase residual sequence to remove gross errors. Then, min-max normalization is performed on both the differencing carrier phase residual sequence and the grid values of the two-dimensional spatiotemporal prior grid. The normalized carrier phase residual sequence and the normalized grid values are concatenated along the feature dimension to form a joint feature vector, which is then input into the input layer of the LTM network.
[0100] Interepoch higher-order difference operations employ second-order differencing to eliminate linear trend terms and constant terms in the carrier phase residual sequence, highlighting the time-varying characteristics of multipath errors. The formula for calculating second-order differencing is:
[0101] in, Represents the epoch The second-order difference residual; Represents the epoch The original carrier phase residual; Represents the epoch The original carrier phase residual; Represents the epoch The original carrier phase residual.
[0102] Gross errors are detected in the residual sequence after second-order differencing. The gross error detection method is... Criterion. For the residual sequence after second-order differencing, calculate its standard deviation. If the second-order difference residual of a certain epoch satisfies If the observations at that epoch are considered to have gross errors, then linear interpolation is used to correct them. The interpolation formula is:
[0103] in, Represents the epoch The observation time; Represents the epoch The observation time; Represents the epoch The observation time.
[0104] After gross error removal, the error stripping and coordinate extraction stages perform min-max normalization on the differential carrier phase residual sequence and the grid values of the two-dimensional spatiotemporal prior grid, mapping the data to the [0,1] interval to eliminate the influence of different feature dimensions. The min-max normalization formula is:
[0105] in, Represents the normalized eigenvalues; Represents the original eigenvalues; This represents the minimum value in the characteristic sequence; This represents the maximum value in the characteristic sequence.
[0106] The normalized carrier phase residual sequence and the grid values of the normalized two-dimensional spatiotemporal prior grid are concatenated along the feature dimension to form a sequence with dimension . The joint feature vector, where Indicates the length of the input sequence. This represents the total number of grid cells in the two-dimensional spatiotemporal prior grid. The joint feature vector is input to the input layer of the Long Short-Term Memory network.
[0107] Error stripping and coordinate extraction are performed by introducing a multi-head self-attention mechanism after the hidden layer output of the Long Short-Term Memory (LSTM) network. The hidden layer output of the LSM network is in dimension [missing information]. The temporal hidden feature vector, where This indicates the number of units in the hidden layer. The multi-head self-attention mechanism maps the temporal hidden feature vector into a query matrix, a key matrix, and a value matrix. It obtains the attention weight distribution based on the dot product operation of the query matrix and the key matrix, and uses the attention weight distribution to perform a weighted summation on the value matrix, thereby strengthening the feature representation in the joint feature vector corresponding to the strong multipath error region.
[0108] The calculation process of the multi-head self-attention mechanism is as follows: First, the temporal hidden feature vector output by the hidden layer... Through three different linear transformation layers, each is mapped to a query matrix. Key matrix AND-value matrix :
[0109] in, , , These represent the linear transformation weight matrices for the query, key, and value, respectively.
[0110] Then, the dot product of the query matrix and the key matrix is calculated, and divided by the square root of the key matrix dimension to obtain the attention score matrix:
[0111] Where Q represents the query matrix; Represents the key matrix; Represents a value matrix; Key matrix The dimension; The function is used to convert attention scores into a probability distribution.
[0112] The multi-head self-attention mechanism divides the query matrix, key matrix, and value matrix into three parts: There are 3 heads, each performing the attention calculation independently. Finally, the outputs of all heads are concatenated and passed through a linear transformation layer to obtain the final output vector.
[0113] in, Indicates the first The attention output of each head; This indicates a concatenation operation along the feature dimension; This indicates the output linear transformation weight matrix.
[0114] The output vector of the multi-head self-attention mechanism is mapped to an estimate of the multipath error through a fully connected layer. The fully connected layer uses a linear activation function to ensure that the numerical range of the output multipath error estimate is consistent with the true value.
[0115] refer to Figure 5 After the error stripping and coordinate extraction stages extract the BeiDou 3D coordinate sequence, the BeiDou 3D coordinate sequence is input into an adaptive Kalman filter for smoothing and gross error suppression. The state vector of the adaptive Kalman filter contains the 3D coordinates and 3D velocity of the monitoring points, and the state equation adopts a uniform motion model. The state vector of the adaptive Kalman filter is:
[0116] in, , , Representing epochs The longitude, latitude, and elevation of the monitoring point in the Earth coordinate system; , , Representing epochs The speed of the monitoring point in three directions.
[0117] State transition matrix for:
[0118] in, This indicates the time interval between adjacent epochs.
[0119] The observation equation is:
[0120] in, Represents the epoch The observation vector, i.e. the extracted BeiDou three-dimensional coordinates; Represents the observation matrix; This represents the observation noise vector, which follows a Gaussian distribution with a mean of zero. Observation matrix for:
[0121] The adaptive Kalman filter calculates the innovation covariance in real time based on the innovation vector and dynamically estimates the observation noise covariance matrix and the process noise covariance matrix based on the innovation covariance within a sliding window. The innovation vector is the difference between the observed and predicted values.
[0122] in, Represents the epoch The new information vector; Represents the epoch The predicted state value.
[0123] The estimated value of the new information covariance is:
[0124] in, Represents the epoch The estimated value of the new information covariance; This indicates the length of the sliding window.
[0125] Based on the information covariance estimate, the observation noise covariance matrix is dynamically updated. With process noise covariance matrix :
[0126]
[0127] in, Represents the epoch The state prediction covariance matrix; Represents the epoch The Kalman gain matrix.
[0128] When the Mahalanobis distance of the innovation vector exceeds the chi-square distribution test threshold, the observation value of the current epoch is considered to be anomaly, and an adaptive decay factor is constructed to reduce the observation weight of the current epoch. The formula for calculating the Mahalanobis distance is:
[0129] in, Represents the epoch Mahalanobis distance of the new information vector.
[0130] like ,in This indicates a degree of freedom of 3 and a significance level of 1. If the chi-square distribution critical value is found, then the adaptive decay factor is calculated. :
[0131] Correcting the Kalman gain matrix using an adaptive attenuation factor:
[0132] The state vector and state covariance matrix are updated based on the corrected Kalman gain matrix to complete the suppression of outlier observations.
[0133] refer to Figure 6The system also includes a tiered early warning terminal, which calculates the three-dimensional displacement rate and cumulative displacement of the surface of the goaf area based on the extracted BeiDou three-dimensional coordinate sequence. The three-dimensional displacement rate is the difference between the three-dimensional coordinates of adjacent epochs divided by the time interval, and the cumulative displacement is the difference between the three-dimensional coordinates of the current epoch and the three-dimensional coordinates of the initial epoch.
[0134] The tiered early warning system constructs a multi-level settlement early warning threshold system based on the geological and mining conditions and overlying rock structure characteristics of the goaf area. The early warning threshold system includes four levels: blue, yellow, orange, and red, each corresponding to different three-dimensional displacement rate and cumulative displacement thresholds. The tiered early warning system compares the calculated three-dimensional displacement rate and cumulative displacement with the corresponding level's early warning threshold. When either the three-dimensional displacement rate or the cumulative displacement reaches the corresponding early warning threshold, the corresponding level's early warning signal is triggered.
[0135] After a warning signal is triggered, the tiered warning terminal binds the warning signal to the BeiDou coordinates of the corresponding monitoring point, generating warning information that includes the warning level, monitoring point location, displacement data, and trigger time. This warning information is then pushed to the remote monitoring platform. Upon receiving the warning information, the remote monitoring platform issues audible and visual alarms and displays relevant data, providing a basis for decision-making regarding mine safety production.
[0136] To clearly illustrate the specific content of the goaf subsidence classification and early warning threshold system, this embodiment provides the goaf subsidence classification and early warning threshold table shown in Table 4.
[0137] Table 4. Early Warning Thresholds for Subsidence in Goaf Areas
[0138] This embodiment improves the quality of input features by performing high-order differencing and normalization on the carrier phase residual sequence. A multi-head self-attention mechanism is introduced into the Long Short-Term Memory network to strengthen the feature representation of regions with strong multipath error and improve the accuracy of multipath error estimation. An adaptive Kalman filter is used to smooth and suppress gross errors in the BeiDou 3D coordinate sequence, reducing the impact of abnormal observations. The tiered early warning system, based on a multi-level early warning threshold system, enables automatic early warning of subsidence status, ensuring safe production in the mining area.
Claims
1. A coal mine goaf subsidence monitoring and early warning system based on BeiDou positioning, characterized in that, include: At the observation data receiving end, the BeiDou dual-frequency carrier phase observation value and pseudorange observation value are obtained, and the observation equation including the subsidence deformation term, ionospheric delay term, tropospheric delay term and multipath error term is constructed. At the prior grid construction end, based on the ionospheric and tropospheric error correction, the carrier-to-noise ratio and elevation angle of each satellite signal are extracted to construct a two-dimensional spatiotemporal prior grid for multipath error; At the error stripping and coordinate extraction end, the two-dimensional spatiotemporal prior grid and the carrier phase residual sequence are input into the long short-term memory network, and the estimated value of the multipath error is output and subtracted from the carrier phase residual to extract the BeiDou three-dimensional coordinate sequence. The loss function of the Long Short-Term Memory network incorporates a deformation continuity constraint regularization term.
2. The coal mine goaf subsidence monitoring and early warning system based on BeiDou positioning according to claim 1, characterized in that, After the observation data receiving end obtains the BeiDou dual-frequency carrier phase observation value and pseudorange observation value, it eliminates the influence of the first-order term of the ionospheric delay term based on the ionospheric-free combination formula, and constructs the wide-lane combined observation value using the pseudorange observation value and the carrier phase observation value that eliminates the first-order ionospheric influence. The wide-lane ambiguity floating-point solution is obtained based on the wide-lane combined observations, and the narrow-lane ambiguity floating-point solution is calculated by combining the ionosphere-free combined observations. The LAMBDA search method is used to fix the wide-lane ambiguity and the narrow-lane ambiguity, and the fixed ambiguity is substituted back into the observation equation to obtain the carrier phase residual sequence.
3. The coal mine goaf subsidence monitoring and early warning system based on BeiDou positioning according to claim 1, characterized in that, After extracting the carrier-to-noise ratio and elevation angle of each satellite signal, the prior mesh construction end establishes a two-dimensional planar mesh with elevation angle as the horizontal axis and carrier-to-noise ratio as the vertical axis based on the physical mapping relationship that the multipath effect is negatively correlated with the satellite elevation angle and positively correlated with the carrier-to-noise ratio fluctuation. The carrier phase residual variance values of each satellite signal falling into the corresponding grid cell of the two-dimensional planar grid in multiple historical epochs are statistically analyzed. The carrier phase residual variance values are used as the multipath error intensity characterization values of the grid cells. The two-dimensional spatiotemporal prior grid of the multipath error is constructed from the multipath error intensity characterization values.
4. The coal mine goaf subsidence monitoring and early warning system based on BeiDou positioning according to claim 1, characterized in that, Before the error stripping and coordinate extraction end inputs the two-dimensional spatiotemporal prior grid and the carrier phase residual sequence into the long short-term memory network, it performs an interepoch high-order difference operation on the carrier phase residual sequence to remove gross errors, and performs minimum-maximum normalization processing on the differenced carrier phase residual sequence and the grid value of the two-dimensional spatiotemporal prior grid respectively. The normalized carrier phase residual sequence and the normalized grid values are concatenated along the feature dimension to form a joint feature vector, which is then input into the input layer of the long short-term memory network.
5. The coal mine goaf subsidence monitoring and early warning system based on BeiDou positioning according to claim 1, characterized in that, The loss function of the Long Short-Term Memory Network introduces a deformation continuity constraint regularization term, which is specifically: based on the physical constraint that the surface subsidence of the goaf meets the first or second order difference smoothing between adjacent epochs, a regularization penalty term is constructed that includes the estimated value of the subsidence difference between adjacent epochs corresponding to the multipath error output by the network. The regularization penalty term is weighted and summed with the mean squared error loss between the multipath error value predicted by the Long Short-Term Memory Network and the actual multipath error label to form the total loss function of the Long Short-Term Memory Network, which constrains the Long Short-Term Memory Network to maintain the smoothness of the settlement deformation time series when fitting the multipath error.
6. The coal mine goaf subsidence monitoring and early warning system based on BeiDou positioning according to claim 2, characterized in that, When using the LAMBDA search method to fix the wide alley ambiguity and the narrow alley ambiguity, calculate the RATIO test value after the ambiguity is fixed; If the RATIO test value is less than a preset fixed threshold, satellite observations in the current epoch with satellite elevation angles lower than the preset elevation angle threshold and carrier-to-noise ratio lower than the preset signal-to-noise ratio threshold are removed, and the dimensionality reduction ambiguity search space of the remaining satellites is obtained. In the reduced ambiguity search space, the LAMBDA search is re-executed and the updated RATIO test value is calculated until the updated RATIO test value is greater than or equal to the preset fixed threshold, at which point the ambiguity is fixed.
7. The coal mine goaf subsidence monitoring and early warning system based on BeiDou positioning according to claim 3, characterized in that, After constructing a two-dimensional spatiotemporal prior grid of the multipath error from the multipath error intensity characterization value, based on the periodic repetition characteristics of the BeiDou satellite orbit, the historical two-dimensional spatiotemporal prior grid corresponding to the historical epochs with the same satellite configuration as the current epoch before the current epoch is extracted. The two-dimensional spatiotemporal prior grid constructed in the current epoch is fused with the historical two-dimensional spatiotemporal prior grid using the exponential smoothing weighting method. The two-dimensional spatiotemporal prior grid in the current epoch is then corrected using the spatial distribution information of historical multipath error, thus obtaining the final dynamic spatiotemporal prior grid used as input to the long short-term memory network.
8. The coal mine goaf subsidence monitoring and early warning system based on BeiDou positioning according to claim 4, characterized in that, The error stripping and coordinate extraction ends introduce a multi-head self-attention mechanism after the hidden layer output of the Long Short-Term Memory network; The temporal hidden feature vectors extracted by the Long Short-Term Memory Network are mapped to a query matrix, a key matrix, and a value matrix. The attention weight distribution is obtained based on the dot product operation of the query matrix and the key matrix. The value matrix is weighted and summed using the attention weight distribution to enhance the feature representation in the joint feature vector corresponding to the strong multipath error region, and the output vector of the multi-head self-attention mechanism is mapped to the estimated value of the multipath error.
9. The coal mine goaf subsidence monitoring and early warning system based on BeiDou positioning according to claim 1, characterized in that, After the error stripping and coordinate extraction end extracts the BeiDou three-dimensional coordinate sequence, the BeiDou three-dimensional coordinate sequence is input into an adaptive Kalman filter; The innovation covariance is calculated in real time based on the innovation vector of the adaptive Kalman filter, and the observation noise covariance matrix and the process noise covariance matrix are dynamically estimated based on the innovation covariance within the sliding window. When the Mahalanobis distance of the new information vector exceeds the chi-square distribution test threshold, an adaptive decay factor is constructed to reduce the observation weight of the current epoch, and the BeiDou three-dimensional coordinate sequence is smoothed and gross error suppressed.
10. The coal mine goaf subsidence monitoring and early warning system based on BeiDou positioning according to claim 1, characterized in that, It also includes a tiered early warning system that calculates the three-dimensional displacement rate and cumulative displacement of the surface of the goaf area based on the extracted BeiDou three-dimensional coordinate sequence. A multi-level settlement early warning threshold system is constructed based on the geological and mining conditions and overlying structure characteristics of the goaf area. The three-dimensional displacement rate and cumulative displacement are compared with the corresponding level of early warning threshold. When any of the three-dimensional displacement rate or cumulative displacement reaches the corresponding warning threshold, a warning signal of the corresponding level is triggered, and the warning signal is bound to the BeiDou coordinates of the corresponding monitoring point and then pushed to the remote monitoring platform.
