A slope displacement estimation method, device and equipment based on data fusion and a storage medium

Abstract

The application discloses a kind of based on data fusion's slope displacement estimation method, comprising: based on multiple sensors data acquisition and pre-processing, the data time and space alignment of acquisition;Based on the data after alignment, discrete observation point extraction and reference sampling section generation are carried out, and high-quality data section is screened by sliding window method;Based on neural differential equation, dynamic modeling is carried out to the data after screening, and nonlinear modeling is carried out to the deformation process of slope by neural network;Based on the prediction-update process of Kalman filter enhanced model, state transition matrix and observation matrix are dynamically adjusted;Based on multiple sensor data, fusion estimation and adaptive error correction are carried out;Through the combination of distributed computing and edge computing, real-time calculation of slope displacement estimation is carried out.The application also discloses a kind of based on data fusion's slope displacement estimation device, equipment and storage medium.The application can realize high-precision, high-reliability slope displacement monitoring by integrating data fusion enhancement technology.

Description

Technical Field

[0001] This application relates to the field of slope safety monitoring technology, and more specifically, to a slope displacement estimation method, device, equipment, and storage medium based on data fusion. Background Technology

[0002] With the rapid development of infrastructure construction, safety monitoring of large-scale engineering geological bodies such as railway and highway slopes and dams has become an important means to ensure engineering safety and the safety of people's lives and property. As a key component of the proactive early warning system, geological slope stability monitoring systems effectively identify potential landslides, collapses, and other disaster risks by collecting and analyzing slope deformation information in real time, significantly improving the ability to prevent and control engineering disasters. Among existing slope monitoring methods, the Global Navigation Satellite System (GNSS) is widely used in the field of geological disaster deformation monitoring due to its high-precision absolute position information. However, GNSS has limitations in signal integrity and continuity, especially in environments with signal obstruction such as tunnels, dense forests, and steep slopes, where it is prone to positioning anomalies and multipath errors. Furthermore, GNSS systems typically output data on a timescale of seconds or longer, resulting in insufficient ability to capture high-frequency, sudden deformations. Accelerometers, as the core component of the inertial measurement unit (IMU), can acquire high-frequency dynamic response information by measuring the micro-vibrations and acceleration changes of slope structures, compensating for the shortcomings of GNSS systems in terms of real-time performance.

[0003] To address the characteristics of multi-source heterogeneous sensors, current research often employs fusion algorithms to achieve complementary system advantages and enhance the stability and robustness of monitoring systems. Among these, Kalman filtering and its extended algorithms, as well as unscented Kalman filtering, are widely used in multi-sensor data fusion. Through a recursive optimization mechanism combining state prediction and observation updates, these filtering algorithms can effectively reduce observation noise and system errors, improving estimation accuracy. However, these methods are typically based on linear state transition models or approximate complex dynamic systems through Taylor expansion linearization. Their ability to model complex nonlinear spatiotemporal characteristics is limited, making it difficult to accurately characterize the creep deformation laws and sudden instability characteristics of slope systems. Furthermore, existing fusion methods still have significant shortcomings in handling the dynamic changes of complex geological systems, controlling long-term cumulative errors, and suppressing non-Gaussian noise. Especially in highly dynamic environments or when satellite signals are limited, the accuracy of IMU data fusion rapidly declines, making it impossible to achieve high-precision displacement monitoring over long periods, resulting in insufficient overall system robustness and reduced early warning performance. Summary of the Invention

[0004] In view of at least one defect or improvement need of the prior art, this application provides a slope displacement estimation method, apparatus, device and storage medium based on data fusion, which can solve at least one of the problems existing in the background art.

[0005] To achieve the above objectives, according to the first aspect of this application, a slope displacement estimation method based on data fusion is provided, the method comprising the following steps:

[0006] Data acquisition and preprocessing are performed using multiple sensors, aligning the acquired data in time and space.

[0007] Based on the aligned data, discrete observation points are extracted and benchmark sampling segments are generated, and high-quality data segments are selected using the sliding window method.

[0008] Dynamic modeling of the selected data is performed based on neural differential equations, and nonlinear modeling of the slope deformation process is performed using neural networks.

[0009] The prediction-update process based on the Kalman filter enhancement model dynamically adjusts the state transition matrix and the observation matrix.

[0010] Fusion estimation and adaptive error correction based on multi-sensor data;

[0011] By combining distributed computing with edge computing, slope displacement estimation can be calculated in real time.

[0012] Furthermore, the aforementioned slope displacement estimation method based on data fusion, specifically including the multi-sensor data acquisition and preprocessing, and the temporal and spatial alignment of the acquired data, includes:

[0013] The multi-sensor system includes a high-performance GNSS receiver and an IMU with integrated high-precision triaxial accelerometers;

[0014] Hardware synchronization between accelerometer sampling and GNSS time is achieved through the PPS signal provided by the GNSS receiver;

[0015] The accelerometer data is preprocessed with temperature drift compensation and zero bias compensation.

[0016] Perform quality assessment and outlier removal on GNSS-acquired data;

[0017] The coordinate data acquired by GNSS after removing outliers is converted into a geocentric coordinate system, and then further converted into a navigation coordinate system, which is the ENU coordinate system, with its origin set as the reference point of the monitoring area.

[0018] Furthermore, the aforementioned slope displacement estimation method based on data fusion, which involves extracting discrete observation points and generating benchmark sampling segments based on the aligned data, and filtering high-quality data segments using a sliding window method, specifically includes:

[0019] Construct a unified time series data structure and interpolate the data acquired by GNSS to align with the IMU sampling frequency;

[0020] The time series data are grouped using the sliding window method. Within each sliding window, data indicators are analyzed, and data windows that meet the conditions of coverage, temporal uniformity, and anomaly ratio are selected. The data are then merged based on the overlap between windows to generate a baseline sampling segment.

[0021] Furthermore, the aforementioned slope displacement estimation method based on data fusion, which involves dynamically modeling the selected data using neural differential equations and nonlinearly modeling the slope deformation process using neural networks, specifically includes:

[0022] The neural differential equation learns the derivative function of the system through a deep residual network. Its input vector includes recursive boundary state values ​​and historical observation data, and its output vector is the state increment.

[0023] The neural network uses a deep residual network as its basic architecture, including multiple residual units, each of which introduces a residual connection.

[0024] The training data for the neural network comes from historical slope deformation observation data, including fused GNSS and IMU observation results and baseline recursive boundary state sequences.

[0025] The training objective of the model is to minimize the error between the predicted state and the actual observed state.

[0026] Furthermore, in the aforementioned slope displacement estimation method based on data fusion, the prediction-update process based on the Kalman filter enhancement model dynamically adjusts the state transition matrix and the observation matrix, specifically including:

[0027] The state transition matrix is ​​dynamically corrected, and the Jacobian matrix is ​​calculated based on the prediction results of the differential equation.

[0028] The Kalman gain matrix is ​​optimized, and the observation noise covariance matrix is ​​dynamically adjusted based on the observation residuals, which reflect the deviation between the Neural ODE predictions and the GNSS observations.

[0029] The observation mapping method is dynamically selected based on the validity of sensor data, and the observation matrix is ​​adaptively adjusted.

[0030] Furthermore, the aforementioned slope displacement estimation method based on data fusion, specifically includes the following steps: Fusion estimation and adaptive error correction based on multi-sensor data.

[0031] The state is jointly estimated based on Neural ODE prediction and IMU and GNSS observation data, and the multi-source observation joint estimation of the state vector is achieved through a weighted mapping function.

[0032] Drift is controlled by IMU dual-integral position estimation, and zero bias is dynamically and recursively corrected;

[0033] Calculate the residuals between GNSS observations and Neural ODE predicted states, dynamically adjust the observation weights, and adjust the observation noise covariance matrix based on the residual mean square error.

[0034] Furthermore, the aforementioned slope displacement estimation method based on data fusion, which combines distributed computing and edge computing to calculate slope displacement estimates in real time, specifically includes:

[0035] The data acquisition tasks from multiple sensors are distributed to multiple computing nodes to achieve distributed data processing;

[0036] The collected data is pre-processed at the edge computing nodes, including data preprocessing, feature extraction, and preliminary analysis.

[0037] The data processed by the edge computing nodes is transmitted to the central server for further fusion processing and analysis to provide real-time early warning of slope displacement.

[0038] According to a second aspect of this application, a slope displacement estimation device based on data fusion is also provided, comprising:

[0039] The data acquisition module is used for data acquisition and preprocessing based on multiple sensors, aligning the acquired data in time and space.

[0040] The data filtering module is used to extract discrete observation points and generate benchmark sampling segments based on the aligned data, and to filter high-quality data segments using the sliding window method.

[0041] The dynamic modeling module is used to dynamically model the filtered data based on neural differential equations and to perform nonlinear modeling of the slope deformation process using neural networks.

[0042] The dynamic adjustment module is used to dynamically adjust the state transition matrix and observation matrix in the prediction-update process based on the Kalman filter enhancement model.

[0043] The fusion estimation and error correction module is used for fusion estimation and adaptive error correction based on multi-sensor data.

[0044] The displacement estimation module is used to calculate slope displacement estimates in real time by combining distributed computing and edge computing.

[0045] According to a third aspect of this application, a slope displacement estimation device based on data fusion is also provided, which includes at least one processing unit and at least one storage unit, wherein the storage unit stores a computer program that, when executed by the processing unit, causes the processing unit to perform the steps of any of the methods described above.

[0046] According to a fourth aspect of this application, a storage medium is also provided, which stores a computer program executable by a data fusion-based slope displacement estimation device, which, when run on the data fusion-based slope displacement estimation device, causes the data fusion-based slope displacement estimation device to perform the steps of any of the methods described above.

[0047] In summary, compared with the prior art, the above-described technical solutions conceived in this application can achieve the following beneficial effects:

[0048] This application provides a slope displacement estimation method based on data fusion. It acquires data from multiple sensors and performs spatiotemporal alignment to ensure data synergy; it utilizes a sliding window method to filter high-quality data segments, improving data quality; it performs nonlinear modeling of the slope deformation process based on neural differential equations to capture complex deformation characteristics; it enhances the prediction-update process of the model through Kalman filtering, dynamically adjusting the state transition matrix and observation matrix to effectively suppress error accumulation; and it fuses multi-sensor data for estimation and adaptive error correction, further improving monitoring accuracy. Finally, it combines distributed computing and edge computing to achieve real-time displacement estimation, meeting the real-time requirements of slope safety monitoring. By integrating multi-sensor data acquisition, neural differential equation modeling, and Kalman filtering enhancement technologies, high-precision and high-reliability slope displacement monitoring can be achieved. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 This is a flowchart illustrating a slope displacement estimation method based on data fusion, provided as an embodiment of this application. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. Furthermore, the technical features involved in the various embodiments described below can be combined with each other as long as they do not conflict with each other.

[0052] The terms "first," "second," "third," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.

[0053] Figure 1 A flowchart illustrating a slope displacement estimation method based on data fusion, as provided in this application embodiment, is shown below. Figure 1 As shown in the figure, an embodiment of this application provides a slope displacement estimation method based on data fusion, which includes the following steps:

[0054] 101. Data acquisition and preprocessing based on multiple sensors, aligning the acquired data in time and space;

[0055] 102. Based on the aligned data, extract discrete observation points and generate benchmark sampling segments, and filter high-quality data segments using the sliding window method;

[0056] 103. Dynamically model the screened data based on neural differential equations, and use neural networks to perform nonlinear modeling of the slope deformation process;

[0057] 104. Prediction-update process based on Kalman filter enhancement model, dynamically adjusting state transition matrix and observation matrix;

[0058] 105. Fusion estimation and adaptive error correction based on multi-sensor data;

[0059] 106. By combining distributed computing and edge computing, slope displacement estimation is calculated in real time.

[0060] Specifically, in this embodiment, a high-performance GNSS receiver and an IMU integrating a high-precision triaxial accelerometer are used as the main sensors. These sensors are synchronously deployed at key monitoring points on the slope to monitor slope displacement changes in real time. The GNSS receiver provides three-dimensional position information of points on the slope surface, mainly used to capture low-frequency displacement; while the accelerometer captures instantaneous acceleration response through high-frequency sampling, and short-term displacement fluctuations can be reconstructed through integration. By fusing the complementary spatiotemporal data of both, displacement monitoring across time scales is achieved.

[0061] To ensure temporal and spatial alignment of data from different sensors, the sampling time of the GNSS and IMU (Inertial Measurement Unit) is synchronized using a PPS (Pulse Per Second) signal provided by the GNSS receiver. The PPS signal is emitted once per second, accurate to the microsecond level, ensuring consistent time references between the two sensors. The geographic coordinates (latitude, longitude, elevation) of the GNSS are converted to a unified geocentric-fixed coordinate system (ECEF), and then further converted to the local East-North-Up (ENU) coordinate system. The IMU data is converted to the same ENU coordinate system using attitude angles (roll, pitch, heading). Preprocessing includes converting the vehicle coordinate system acceleration to the navigation system acceleration using a rotation matrix, subtracting the gravity term to obtain net acceleration, and then using an integration method to convert the net acceleration into displacement information, i.e., the dynamic position information of the slope monitoring point in the navigation coordinate system.

[0062] After completing the temporal and spatial alignment of the data, the next step is to extract discrete observation points and generate a baseline sampling segment. A sliding window method is used to select high-quality data segments. The aligned GNSS and IMU data are merged into a unified time-series dataset according to timestamps. Since the sampling frequency of GNSS is relatively low (e.g., 10Hz) while that of IMU is relatively high (e.g., 100Hz), interpolation is performed on the GNSS data to make its sampling frequency consistent with that of the IMU. A sliding window with a length of 500ms and a sliding step size of 300ms can be defined. Within each window, the data coverage, time interval uniformity, and outlier ratio are calculated. Only when the coverage is higher than 90%, the standard deviation of the time interval is less than 2ms, and the outlier ratio is less than 5%, the data within that window is selected as the baseline sampling segment.

[0063] Neural ODEs are used to dynamically model the filtered baseline sampling data. Neural ODEs learn the state changes of the system through deep neural networks (DNNs), employing a deep residual network (ResNet) as the basic architecture of the DNN, containing multiple residual blocks. Each residual block contains two convolutional layers and one skip connection to facilitate gradient flow and improve training efficiency. Historical slope deformation data is used to train the Neural ODE model. Input data includes IMU acceleration and angular velocity, as well as GNSS location information. The model's objective is to minimize the mean squared error (MSE) between the predicted state and the actual observed state.

[0064] A trained Neural ODE model is combined with a Kalman filter to improve the accuracy of state estimation. The Neural ODE model is used to predict the next state of the slope, including its location, velocity, and attitude. The state prediction output of the Neural ODE is used to provide the dynamic transition matrix and residuals, driving the Kalman filter optimization. The Kalman gain is dynamically adjusted based on the difference between the predicted state and the observed data. This helps to balance the weights of model predictions and actual observations. The slope state estimate is updated by combining the Kalman gain and the observed data.

[0065] Further, multi-sensor data is used for fusion estimation and adaptive error correction. High-frequency data from the IMU and high-precision position data from GNSS are fused, and a more accurate displacement estimate is obtained through a weighted model of the mean square error of the multi-source observation residuals. This process is based on the results of the Neural ODE model and Kalman filtering, dynamically adjusting and correcting the zero bias and cumulative errors in the IMU data, as well as the multipath errors in the GNSS data.

[0066] Finally, by combining distributed computing and edge computing, and based on the updated state and covariance, the final fusion estimation is completed using IMU / GNSS data, achieving real-time slope displacement estimation. Local computing resources (such as high-performance embedded computers) are deployed to process sensor data in real time, avoiding data transmission delays. For large-scale slope monitoring, a distributed computing architecture is adopted, distributing computing tasks across multiple nodes to improve processing efficiency. Once a displacement exceeding a preset threshold is detected, the system immediately issues an alarm to facilitate timely intervention.

[0067] This application provides a slope displacement estimation method based on data fusion. It collects data from multiple sensors and performs spatiotemporal alignment to ensure data synergy; it uses a sliding window method to filter high-quality data segments, improving data quality; it performs nonlinear modeling of the slope deformation process based on neural differential equations to capture complex deformation characteristics; it enhances the prediction-update process of the model through Kalman filtering, dynamically adjusting the state transition matrix and observation matrix to effectively suppress error accumulation; and it fuses multi-sensor data for estimation and adaptive error correction, further improving monitoring accuracy. Finally, it combines distributed computing and edge computing to achieve real-time displacement estimation, meeting the real-time requirements of slope safety monitoring. By integrating multi-sensor data acquisition, neural differential equation modeling, and Kalman filtering enhancement technologies, it is possible to achieve high-precision and high-reliability slope displacement monitoring.

[0068] Optionally, the slope displacement estimation method based on data fusion provided in this application embodiment, wherein the data acquisition and preprocessing based on multiple sensors, aligning the acquired data in time and space, specifically includes:

[0069] The multi-sensor system includes a high-performance GNSS receiver and an IMU with integrated high-precision triaxial accelerometers;

[0070] Hardware synchronization between accelerometer sampling and GNSS time is achieved through the PPS signal provided by the GNSS receiver;

[0071] The accelerometer data is preprocessed with temperature drift compensation and zero bias compensation.

[0072] Perform quality assessment and outlier removal on GNSS-acquired data;

[0073] The coordinate data acquired by GNSS after removing outliers is converted into a geocentric coordinate system, and then further converted into a navigation coordinate system, which is the ENU coordinate system, with its origin set as the reference point of the monitoring area.

[0074] Specifically, the first step is to complete the multi-sensor data acquisition of the GNSS and IMU systems.

[0075] In the system design, GNSS data output is geographic coordinates (latitude, longitude, elevation), and IMU data output is triaxial acceleration (α). x a y a z ) and triaxial angular velocity (ω) x ω y ω zTo ensure time consistency of the output data from the two sensor systems, a unified time reference is achieved through a PPS (Pulse Per Second) signal provided by the GNSS receiver. The GNSS receiver emits a PPS signal once per second, which is connected to the external trigger port of the IMU's data acquisition system, enabling hardware synchronization between IMU sampling and GNSS time. The IMU acquisition system drives its internal clock based on the PPS signal, assigning a timestamp consistent with the GNSS timestamp to each set of IMU data during data acquisition. The time synchronization accuracy is better than 1 microsecond, ensuring consistency in sampling time between the two types of sensors, thereby achieving spatiotemporal alignment of the data.

[0076] In the IMU data preprocessing stage, temperature drift compensation is performed first. The temperature value T at the sampling time is measured in real time using an internal temperature sensor. The system obtains a temperature compensation model through experimental calibration. ,in, The compensated acceleration value, The raw acceleration value output by the sensor. This is the temperature drift coefficient (determined by calibration experiments). The temperature at the current sampling time. The calibration reference temperature is used. This formula is applied to acceleration data along three axes, calculating and correcting them separately. This method effectively reduces the impact of temperature changes on IMU measurement accuracy.

[0077] Zero-bias compensation employs a combination of static calibration and dynamic correction. During system initialization, the IMU sensor is stationary, and static data is collected. The zero-bias value is estimated by calculating the average of the triaxial acceleration data over a long time window.

[0078]

[0079] Where N is the number of sampling points in the static state. This represents the original acceleration value at the i-th sampling time. During dynamic monitoring, the zero-bias drift is dynamically corrected based on temperature changes.

[0080]

[0081] in, This is the drift change coefficient. The final correction value for acceleration is: .

[0082] In terms of GNSS data preprocessing, multipath effect detection is performed first. By observing the carrier-to-noise ratio (C / N0) of each satellite signal, if C / N0 is less than 35 dBHz, the data is considered to potentially have multipath effects and the data epoch is discarded. Secondly, the Position Dilution of Precision (PDOP) index is used to screen the data. A PDOP threshold of 2.0 is set; GNSS observations exceeding this threshold are discarded.

[0083] Anomaly detection uses Mahalanobis distance as the discrimination method. For the current GNSS observation position vector, the Mahalanobis distance between it and the mean and covariance matrix of the samples within the historical window is calculated: ,like Exceeding the confidence interval threshold (e.g.) If the value of the observation is within the 95% confidence level, then the observation point is considered abnormal and is removed.

[0084] The latitude, longitude, and altitude coordinates obtained from GNSS measurements need to be converted to a unified geocentric-fixed coordinate system (ECEF). The conversion formula is as follows: , , ,in, Latitude Longitude For the height of the ellipsoid, , where a is the semi-major axis of the WGS84 ellipsoid, e 2 This is the square of the first eccentricity. Inertial Measurement Units (IMUs) and Global Navigation Satellite Systems (GNSS) often operate in different coordinate systems. IMUs typically output data in the vehicle coordinate system, while GNSS positions are mostly measured in a geocentric coordinate system. To achieve consistent fusion of multi-sensor data, coordinate system transformation and initialization of the navigation coordinate system must be performed.

[0085] First, establish a navigation coordinate system, typically using the ENU (East-North-Up) coordinate system. Its origin is set as the reference point for the monitoring area, with the x-axis pointing due east, the y-axis due north, and the z-axis pointing towards the zenith. Convert GNSS latitude, longitude, and altitude to the ENU coordinate system:

[0086]

[0087]

[0088]

[0089]

[0090]

[0091] in, , , The latitude, longitude, and elevation of the reference point are given by M and N, respectively, which can be calculated using the WGS84 ellipsoid model.

[0092] IMU data conversion from the carrier coordinate system to the navigation coordinate system requires attitude angle information (i.e., roll angle). Pitch angle Heading angle The three-axis acceleration and angular velocity require coordinate transformation using a rotation matrix. The rotation matrix is ​​determined by the attitude angles and is expressed as follows:

[0093]

[0094] in

[0095]

[0096]

[0097]

[0098] The acceleration of the carrier coordinate system is achieved through this rotation matrix. Converted to navigation system acceleration : Subsequently, the acceleration in the navigation coordinate system needs to be subtracted from the gravity term to obtain the net acceleration. The gravity vector in the navigation coordinate system is represented as...

[0099]

[0100] Where g is the local gravitational acceleration, typically taken as 9.80665 m / s². The net acceleration is calculated as follows: After converting the acceleration data, velocity and position are obtained using integration methods. The trapezoidal integral method is employed, and the first step is the integration from acceleration to velocity:

[0101]

[0102] The second step is the integration from velocity to displacement.

[0103]

[0104] in, The IMU sampling period is 10ms (100Hz). The dynamic position information of the slope monitoring point in the navigation coordinate system can be obtained through the above integration process.

[0105] Optionally, the slope displacement estimation method based on data fusion provided in this application embodiment, wherein the step of extracting discrete observation points and generating benchmark sampling segments based on the aligned data, and filtering high-quality data segments using a sliding window method, specifically includes:

[0106] Construct a unified time series data structure and interpolate the data acquired by GNSS to align with the IMU sampling frequency;

[0107] The time series data are grouped using the sliding window method. Within each sliding window, data indicators are analyzed, and data windows that meet the conditions of coverage, temporal uniformity, and anomaly ratio are selected. The data are then merged based on the overlap between windows to generate a baseline sampling segment.

[0108] Specifically, after completing multi-sensor data acquisition and spatiotemporal benchmark unification, in order to achieve high accuracy in subsequent recursive calculations and state estimation, it is necessary to extract key discrete observation points based on synchronized and preprocessed GNSS and IMU sampled data and generate highly complete data benchmark sampling segments.

[0109] First, the multi-sensor data output in step 101 is sorted by timestamp to construct a unified time-series data structure. Since the GNSS data sampling frequency is 10Hz while the IMU data sampling frequency is 100Hz, there is a difference in sampling time resolution between the two. Therefore, under a unified time reference, an interpolation algorithm is used to upsample the GNSS data to the IMU sampling frequency, ensuring strict alignment of the two data sources in the time dimension.

[0110] GNSS data interpolation employs either linear interpolation or cubic spline interpolation. Taking linear interpolation as an example, for any two adjacent GNSS sampling times... and Given that their corresponding coordinate values ​​are respectively and Then at any time t between these two times, the interpolation point The calculation formula is

[0111] .

[0112] IMU data itself has a high sampling rate and can be directly organized in chronological order. Ultimately, all sensor data is organized into a time-series set of observation points with a uniform time step, denoted as...

[0113]

[0114] in, For time GNSS observation point coordinates These are the triaxial acceleration observations of the IMU at that moment. This is the IMU temperature reading at that moment.

[0115] After constructing the time series data structure, a sliding window method was used to group the time series data for further analysis and selection of high-quality data segments. Each window is 500ms long, and the window sliding step is 60% of the window length, meaning the window overlap is 40%. Assuming a sampling frequency of 100Hz, a 500ms time interval corresponds to 50 IMU observation points. Therefore, the window size W can be expressed as:

[0116]

[0117] When the window is slid, the sliding step size s is 30, which means that the window moves backward 300ms each time.

[0118] Within each sliding window, multiple indicators such as data coverage, time interval distribution, and anomaly ratio need to be analyzed and quantified to ensure the integrity and high reliability of the generated benchmark sampling segment.

[0119] Data coverage primarily targets IMU observation data and GNSS interpolation data, checking for missing data within the window. Assuming the sliding window should theoretically contain M observation points, the actual number of effective observation points is... Then the coverage

[0120]

[0121] Windows with a coverage rate greater than 90% are considered valid windows, while windows with a coverage rate lower than this threshold will be directly rejected.

[0122] Although the system performs time synchronization, under extreme conditions (such as data anomalies or signal loss), the time intervals between observation points may still be uneven. To analyze the uniformity of the time intervals, the time difference between adjacent observation points within the statistical window is calculated, and the mean is determined. and standard deviation Ideally... = 10ms$, The closer the value is to 0, the more uniform the time interval distribution. A standard deviation threshold of 2ms is set; windows exceeding this threshold are considered time anomaly windows and are not included in the baseline segment generation.

[0123] Based on the joint analysis of the above three indicators—coverage, temporal uniformity, and anomaly ratio—data windows meeting the following criteria were selected: and and , The abnormal proportion is determined by batch detection and filtering of sliding windows. Multiple window segments that meet the criteria are selected and merged to generate a baseline sampling segment based on the overlap between windows. If two windows have more than 40% overlap and the rate of change of the abnormal proportion between the two windows is less than 5%, the two windows can be merged into a longer sampling segment. After recursive merging operations, a set of baseline sampling segments is obtained, each containing complete IMU and GNSS observation data, timestamps, and quality assessment information.

[0124] The final generated baseline sampling segments need to be standardized. The coordinate values, velocity, acceleration, and other observations of each data segment are normalized to improve the numerical stability and computational efficiency of subsequent recursive calculations.

[0125] Optionally, the slope displacement estimation method based on data fusion provided in this application embodiment, wherein the dynamic modeling of the screened data based on neural differential equations and the nonlinear modeling of the slope deformation process using neural networks specifically include:

[0126] The neural differential equation learns the derivative function of the system through a deep residual network. Its input vector includes recursive boundary state values ​​and historical observation data, and its output vector is the state increment.

[0127] The neural network uses a deep residual network as its basic architecture, including multiple residual units, each of which introduces a residual connection.

[0128] The training data for the neural network comes from historical slope deformation observation data, including fused GNSS and IMU observation results and baseline recursive boundary state sequences.

[0129] The training objective of the model is to minimize the error between the predicted state and the actual observed state.

[0130] Specifically, the stability of slope geological bodies is controlled by a complex mechanical environment, and their deformation evolution process exhibits significant nonlinearity, time-varying characteristics, and uncertainty. Traditional dynamic models based on physical laws often require a large number of slope soil and rock parameters (such as pore water pressure, internal friction angle, and cohesion) as input, and assume that the system's dynamic behavior follows linear or quasi-linear control laws. However, the spatial distribution of the physical properties of soil and rock masses in actual engineering is complex and difficult to measure accurately; at the same time, their dynamic response process is also affected by multiple disturbances such as environmental vibration and surface water infiltration. These factors prevent traditional methods from fully characterizing the dynamic behavior of the system.

[0131] Therefore, this application introduces Neural ODE as the core of dynamic state modeling. Neural ODE learns the derivative function of the system through a deep neural network (DNN), enabling the dynamic process of state evolution over time to be modeled in continuous time form, thus avoiding sampling errors and information loss caused by traditional discrete-time methods.

[0132] The fundamental idea of ​​neural differential equations originates from ordinary differential equations (ODEs). In the standard state-space model, the evolution of a state over time can be represented as...

[0133]

[0134] in, Let f represent the state vector of the system at time t, and let f be the state derivative function. These are model parameters.

[0135] In the Neural ODE framework, the function f is no longer explicitly defined by manual design or analytical expression, but is parametrically modeled through neural networks, denoted as f.

[0136]

[0137] The input to a neural network (NN) is the current state. Given time t, the output is the first derivative of the state, reflecting the system's changing trend. The ODE solver can integrate the state from the initial time to the target time, predicting the time evolution path of the system state.

[0138] To predict slope displacement, the input and output of the Neural ODE model need to be rationally designed based on sensor-collected data and recursive boundary conditions. Input vector It includes recursive boundary state values ​​and historical observation data (IMU and GNSS fusion data), and its structure is defined as follows:

[0139]

[0140] in, , , These represent the three-dimensional position (East, North, Up), three-dimensional velocity, and boundary attitude angle of the recursive boundary, respectively. , , These represent the acceleration, temperature, and GNSS coordinate observations obtained from the IMU, respectively. Output vector. The state increment is the predicted change in the current state, including the rate of change of position, velocity, and attitude angle.

[0141] .

[0142] In this application, a deep residual network (ResNet) is used as the basic architecture for the neural network. ResNet alleviates the gradient vanishing problem in deep neural networks by introducing residual connections, thus enhancing the model's ability to fit dynamic time processes. The network model includes l residual blocks, each with the following structure:

[0143]

[0144] in, The input features of the l-th layer, and These are the weight matrix and the bias vector, respectively. This is the activation function.

[0145] This model uses the ReLU (Rectified Linear Unit) activation function, whose mathematical expression is:

[0146]

[0147] The output layer employs linear activation, directly outputting the state increment. Training data is derived from historical slope deformation observation data, combined with fused GNSS and IMU observations, and a baseline recursive boundary state sequence. The training set is divided into time segments, each containing an input observation sequence and a target predicted state increment sequence. The model training objective is to minimize the error between the predicted state and the actual observed state; the loss function is the weighted mean squared error (WMSE).

[0148]

[0149] in, , and These are the weighting coefficients for position, velocity, and attitude prediction errors, respectively, which are adjusted according to the actual monitoring accuracy requirements and typically meet the following criteria.

[0150]

[0151] The optimizer uses the Adam algorithm with an initial learning rate of 0.001 and adaptively adjusts gradient updates.

[0152] The trained neural ODE model is embedded into the ODE solver for state prediction. The ODE solver employs the classic fourth-order Runge-Kutta integral method with a time step of [missing information]. Depending on the IMU's sampling frequency (typically 100Hz, or 10ms), the four intermediate estimation formulas for the Runge-Kutta method are:

[0153]

[0154] The formula for updating the state prediction value is as follows:

[0155]

[0156] By integrating over continuous time, the dynamic evolution prediction results of the slope system state over time are obtained.

[0157] Due to the presence of zero-bias drift and noise accumulation in IMU data, and the influence of multipath effects and occlusion on GNSS data, both observation data exhibit time-varying errors. The neural ODE model, through historical data learning and nonlinear modeling, possesses a certain degree of error suppression capability, particularly in characterizing the dynamic propagation of errors. The error state model employs a mechanism combining state increment prediction and observation error feedback to predict the state. and observation status The difference is

[0158]

[0159] The error dynamic feedback control formula is:

[0160]

[0161] in, The error feedback gain coefficient is dynamically adjusted using the Kalman gain matrix.

[0162] Optionally, the slope displacement estimation method based on data fusion provided in this application, wherein the prediction-update process based on the Kalman filter enhancement model dynamically adjusts the state transition matrix and the observation matrix, specifically includes:

[0163] The state transition matrix is ​​dynamically corrected, and the Jacobian matrix is ​​calculated based on the prediction results of the differential equation.

[0164] The Kalman gain matrix is ​​optimized, and the observation noise covariance matrix is ​​dynamically adjusted based on the observation residuals, which reflect the deviation between the Neural ODE predictions and the GNSS observations.

[0165] The observation mapping method is dynamically selected based on the validity of sensor data, and the observation matrix is ​​adaptively adjusted.

[0166] Specifically, the state estimation equation for the discrete-time Kalman filter is as follows:

[0167] State transition equation (Prediction)

[0168]

[0169] Predicting covariance matrix

[0170]

[0171] Observational update (Correction)

[0172]

[0173] Status update

[0174]

[0175] Covariance Update

[0176]

[0177] in, For state vectors, For the observation vector, Here is the state transition matrix. For the observation matrix, and These are the covariance matrices of process noise and observation noise, respectively. This is the Kalman gain matrix.

[0178] In traditional Kalman filtering algorithms, the state transition matrix is ​​typically constructed based on static physical models or empirical data, and its form often fails to capture the nonlinear evolution process in complex systems. This application uses state prediction results from Neural ODE. The state transition matrix is ​​dynamically modified to more accurately reflect the time-varying and nonlinear evolution trends of state variables such as slope displacement, velocity, and attitude. Based on the state prediction results of Neural ODE, an enhanced state transition matrix is ​​defined. This matrix is ​​obtained by calculating the Jacobian matrix of the state prediction result to the state input.

[0179]

[0180] The Jacobian matrix reflects the sensitivity of the system's dynamic evolution to changes in state input, thus enabling dynamic adjustment of the linear approximation capability of the state prediction model. For the process noise covariance matrix... Traditional methods rely on fixed empirical values, while this application incorporates dynamic adjustment of the residual error from the Neural ODE model. The residual error is...

[0181]

[0182] Calculate the dynamic process noise covariance matrix based on the residual error.

[0183]

[0184] in, This is a weighting factor (with a value range of 0.01-0.1, used to balance the weight of residuals on process noise). This is the covariance matrix of the basic process noise. After the above processing, the state prediction equation is enhanced to...

[0185]

[0186] During the observation update process, the Kalman gain matrix controls the response strength of the state estimate to the observation data. In traditional gain matrix calculations, the weight ratio between state prediction and observation data is fixed, making it difficult to adapt to dynamic changes in the quality of the observation data. This application predicts the state using Neural ODE. Combined with actual GNSS observation data Construct dynamic observation residual weights.

[0187] First, calculate the observation residuals.

[0188]

[0189] The residual reflects the deviation between the Neural ODE predictions and GNSS observations. The observation noise covariance matrix is ​​then dynamically adjusted based on the residual magnitude.

[0190]

[0191] in, This is an adjustment factor (usually set to 0.05-0.2). The fundamental noise covariance matrix of GNSS observation data is typically calculated based on signal-to-noise ratio metrics such as C / N0 and PDOP of the GNSS data. The optimized Kalman gain matrix is...

[0192]

[0193] The state update equation is rewritten as follows

[0194]

[0195] Covariance updated to

[0196] .

[0197] Traditional Kalman filter observation matrix The linear relationship between the mapping state vector and the observation vector is established. However, in complex slope environments, GNSS observation data often only provides observational information for position state variables, while the IMU indirectly affects velocity and attitude state estimation. This application dynamically adjusts the observation matrix based on the validity of sensor data. When the sensor observation quality is good, the element values ​​adopt full-dimensional observation mapping, while when the observation quality is low (such as insufficient GNSS satellites or significant IMU drift), only high-confidence data is used to update some state variables. For example, when the number of available GNSS satellites is less than 4, only Neural ODE is used to predict the state to update the velocity and attitude states, while the position state is directly transmitted by the predicted state. At this time, the observation matrix becomes a sparse matrix.

[0198] Optionally, the slope displacement estimation method based on data fusion provided in this application embodiment, wherein the fusion estimation and adaptive error correction based on multi-sensor data specifically includes:

[0199] The state is jointly estimated based on Neural ODE prediction and IMU and GNSS observation data, and the multi-source observation joint estimation of the state vector is achieved through a weighted mapping function.

[0200] Drift is controlled by IMU dual-integral position estimation, and zero bias is dynamically and recursively corrected;

[0201] Calculate the residuals between GNSS observations and Neural ODE predicted states, dynamically adjust the observation weights, and adjust the observation noise covariance matrix based on the residual mean square error.

[0202] Specifically, in slope monitoring systems, sensor observation data often exhibit varying degrees of observation errors and data inconsistencies due to environmental interference, hardware limitations, and system noise. In particular, IMU sensors, due to zero-bias drift and accumulated errors, show an exponential increase in displacement estimation errors during long-term integration; while GNSS often exhibits discontinuities and abnormal fluctuations due to multipath effects and obstructed environments. This application introduces a fusion estimation and adaptive error correction mechanism to fully utilize the advantages of different sensor data, dynamically identify and suppress data error sources, and achieve stable and reliable state quantity estimation.

[0203] Based on the theory of integrated navigation of inertial navigation and satellite positioning systems, multi-source sensor data fusion aims to improve the robustness and accuracy of the system by jointly estimating state variables and complementing the advantages and disadvantages of sensor observations. This application uses the Kalman filter algorithm framework, introduces Neural ODE to predict state information, supplements it with IMU and GNSS observation data, and uses a state mapping transformation model to complete the joint estimation of multi-sensor data.

[0204] Let the system state vector be

[0205]

[0206] in, It is a three-dimensional position vector. It is a three-dimensional velocity vector. Here is the attitude angle vector. State prediction is provided by NeuralODE: The goal of the fusion estimation is to achieve joint estimation of the state vector from multiple sources using a weighted mapping function $f(\cdot)$. The specific mapping relationship is as follows:

[0207]

[0208] The IMU inertial navigation system first eliminates the influence of gravity by observing three-axis acceleration data from the accelerometer, obtaining pure dynamic acceleration. The key to eliminating the influence of gravity lies in coordinate system transformation; the IMU's onboard data must be converted to the navigation coordinate system. Attitude angles ( , , The constructed Directional Cosine Matrix (DCM) is as follows:

[0209]

[0210] The dynamic acceleration vector is

[0211]

[0212] Where g is the local gravity vector.

[0213] The velocity integral is calculated as follows

[0214]

[0215] The position integral is calculated as follows

[0216]

[0217] Drift control is achieved through zero bias correction and temperature drift model, which has been completed in step 101 and needs to be dynamically and recursively corrected here.

[0218] GNSS observations, calculated using RTK differential methods, have absolute reference value, but are subject to jumps and fluctuations due to multipath effects and occlusion. To ensure the stationarity of the fused state estimate, it is necessary to calculate the residuals between the Neural ODE predicted state and the GNSS observed state, and dynamically adjust the observation weights.

[0219] The residual is calculated as follows

[0220]

[0221] The residual mean square error (MSE) is

[0222]

[0223] The observation noise covariance matrix is ​​dynamically adjusted based on MSE. and affect the state observation matrix Weight allocation

[0224]

[0225] in, This is the gain coefficient, which controls the impact of MSE on observation noise estimation, and its value is usually in the range of 0.1 to 0.5.

[0226] After the joint estimation of multi-sensor data is completed, the system dynamically analyzes the stability of the fused state based on the changing trends of the observed state variables and the residual values, and the optimized state estimation results are used as the fused estimation output.

[0227] Optionally, the slope displacement estimation method based on data fusion provided in this application embodiment, which combines distributed computing and edge computing to calculate the slope displacement estimate in real time, specifically includes:

[0228] The data acquisition tasks from multiple sensors are distributed to multiple computing nodes to achieve distributed data processing;

[0229] The collected data is pre-processed at the edge computing nodes, including data preprocessing, feature extraction, and preliminary analysis.

[0230] The data processed by the edge computing nodes is transmitted to the central server for further fusion processing and analysis to provide real-time early warning of slope displacement.

[0231] Specifically, this application adopts a technical architecture that combines distributed computing and edge computing. By deploying edge computing terminals at field monitoring stations, the real-time acquisition and fusion computing tasks of multi-source sensor data such as GNSS and inertial measurement units (IMU) are pushed to the terminal side, realizing the localization of data processing. This significantly reduces the bandwidth pressure and latency risk of remote data transmission, ensuring that the monitoring system has real-time early warning capabilities.

[0232] In one specific embodiment, the processor of this application adopts an ARM architecture Cortex-A72 quad-core or an x86 architecture Intel Core i7 8th generation or higher model, with a main frequency of not less than 1.8GHz and a single-core floating-point operation capability greater than 10GFLOPS. The Cortex-A72 has excellent computing efficiency and low power consumption characteristics, making it suitable for continuous sampling and state estimation. The processor integrates a high-bandwidth memory controller, paired with LPDDR4 memory, with a memory capacity of not less than 8GB, ensuring efficient memory management and dynamic loading of the Neural ODE model during operation. The processor integrates a hardware floating-point unit (FPU) and an AI acceleration module (such as the NEONSIMD engine), providing hardware support for neural network forward inference and Kalman filter recursive optimization calculations. The Neural ODE model adopts a custom optimized TensorRT or ONNX Runtime deployment framework to achieve efficient inference in edge devices, with single-step state prediction latency controlled within 1ms. The system uses an industrial-grade motherboard, has strong anti-electromagnetic interference capabilities, supports wide temperature operation (-40°C to +85°C), and meets the long-term deployment requirements of complex outdoor environments on slopes. The equipment enclosure has an IP67 protection rating and is designed to be waterproof, dustproof, and lightning-proof.

[0233] The data acquisition system must support concurrent sampling from multiple channels and heterogeneous sensors, providing a stable high-speed data interface. The GNSS receiver uses the standard NMEA protocol output, with an RS232 or RS485 interface and a baud rate of 115200bps. The IMU accelerometer uses SPI or CAN bus protocols for sampling, with a sampling rate of up to 100Hz. The SPI interface communication rate is 10Mbps, and the CAN interface communication rate is no less than 500kbps. The edge computing terminal has no fewer than four RS485 interfaces and two CAN interfaces, supporting asynchronous acquisition and concurrent processing. The interface input voltage range is 5V to 12V, and a built-in EMC filter prevents surge interference and electrostatic damage. The PPS pulse signal output from the GNSS receiver is input to the external clock synchronization pin of the IMU system to achieve unified sampling clock and data timestamp alignment, ensuring spatiotemporal consistency of the sampled data. The interface protocol software is written in C language, and the underlying hardware access adopts bare-metal mode or RTOS (such as FreeRTOS) scheduling, possessing millisecond-level real-time performance. The data sampling buffer is managed using a circular queue. Data sampling points are enqueued and dequeued according to their timestamps to ensure high concurrency and non-blocking data sampling and processing.

[0234] The system operates on a Linux operating system, using a lightweight embedded Linux distribution (such as a customized version built from Buildroot or the Yocto Project). The kernel is trimmed to 4.x or 5.x LTS (Long Term Support) version, and unnecessary modules are removed to ensure a system startup time of less than 10 seconds and a system load of less than 0.5. The core algorithm is implemented in C++, with a Python interface to support model training and debugging. The Neural ODE-KF fusion algorithm module includes sub-modules for data synchronization management, IMU / GNSS data decoding and preprocessing, state estimation, Kalman recursive filtering, and data anomaly detection. The program adopts object-oriented design, with clear module division and standardized interfaces. The main program control flow is managed using a finite state machine (FSM) to ensure orderly state transitions and clear logic. The algorithm model is stored in a specific directory in the terminal file system. The Neural ODE model file format is ONNX or TensorFlow SavedModel, and it is loaded via a dynamic link library (Shared Object, .so file). The Kalman filter state variables are stored in a shared memory region, and multiple threads achieve synchronization through mutexes and condition variables.

[0235] The system supports remote SSH access and OTA (Over-the-Air) remote upgrades, allowing algorithm optimization and model updates to be performed remotely. System logs are managed using system-level syslog, with log files rotated daily, and critical status information can be uploaded to the central cloud server via the MQTT protocol.

[0236] To ensure reliable data storage and fast read / write speeds during long-term operation, the system employs non-volatile storage media. The main storage device uses eMMC 5.1 or industrial-grade SSDs with a capacity of at least 32GB, supporting fast random read / write with I / O latency below 1ms. The file system uses ext4, with journaling enabled to prevent file system corruption. Noatime is used to optimize write performance for critical system directories. Data sampling results, status estimation data, fused status variables, and alarm information are stored in different directories, with file names using timestamps for a clear structure and easy retrieval. Long-term archived data is compressed monthly using the Zstandard (.zst) compression algorithm, offering high compression ratios and fast decompression. Compressed files are automatically backed up weekly to an external industrial-grade TF card or USB storage device with hot-swappable interfaces, supporting plug-in / plug-out detection and automatic file system mounting.

[0237] Optionally, embodiments of this application also provide a slope displacement estimation device based on data fusion, comprising:

[0238] The data acquisition module is used for data acquisition and preprocessing based on multiple sensors, aligning the acquired data in time and space.

[0239] The data filtering module is used to extract discrete observation points and generate benchmark sampling segments based on the aligned data, and to filter high-quality data segments using the sliding window method.

[0240] The dynamic modeling module is used to dynamically model the filtered data based on neural differential equations and to perform nonlinear modeling of the slope deformation process using neural networks.

[0241] The dynamic adjustment module is used to dynamically adjust the state transition matrix and observation matrix in the prediction-update process based on the Kalman filter enhancement model.

[0242] The fusion estimation and error correction module is used for fusion estimation and adaptive error correction based on multi-sensor data.

[0243] The displacement estimation module is used to calculate slope displacement estimates in real time by combining distributed computing and edge computing.

[0244] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method. The computer-readable storage medium may include, but is not limited to, any type of disk, including floppy disks, optical disks, DVDs, CD-ROMs, microdrives, as well as magneto-optical disks, ROMs, RAMs, EPROMs, EEPROMs, DRAMs, VRAMs, flash memory devices, magnetic cards or optical cards, nanosystems (including molecular memory ICs), or any type of medium or device suitable for storing instructions and / or data.

[0245] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0246] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0247] In the several embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some service interface; the indirect coupling or communication connection between devices or units may be electrical or other forms.

[0248] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0249] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0250] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0251] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, which may include: a flash drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0252] The foregoing description is merely an exemplary embodiment of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure. Those skilled in the art will readily conceive of embodiments of this disclosure upon considering the specification and practicing the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described herein. The specification and embodiments are to be considered exemplary only, and the scope and spirit of this disclosure are defined by the claims.

[0253] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0254] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A slope displacement estimation method based on data fusion, characterized in that, Includes the following steps: Data acquisition and preprocessing are performed using multiple sensors, aligning the acquired data in time and space. Based on the aligned data, discrete observation points are extracted and benchmark sampling segments are generated, and high-quality data segments are selected using the sliding window method. Dynamic modeling of the selected data is performed based on neural differential equations, and nonlinear modeling of the slope deformation process is performed using neural networks. The prediction-update process based on the Kalman filter enhancement model dynamically adjusts the state transition matrix and the observation matrix. Fusion estimation and adaptive error correction based on multi-sensor data; By combining distributed computing and edge computing, slope displacement estimation is calculated in real time; The process of extracting discrete observation points and generating benchmark sampling segments based on the aligned data, and filtering high-quality data segments using a sliding window method, specifically includes: Construct a unified time series data structure and interpolate the data acquired by GNSS to align with the IMU sampling frequency; The time series data is grouped using the sliding window method. Within each sliding window, data indicators are analyzed, and data windows that meet the conditions of coverage, temporal uniformity, and anomaly ratio are selected. Based on the overlap between windows, a baseline sampling segment is generated by merging them. The dynamic modeling of the screened data based on neural differential equations, and the nonlinear modeling of the slope deformation process using neural networks, specifically include: The neural differential equation learns the derivative function of the system through a deep residual network. Its input vector includes recursive boundary state values ​​and historical observation data, and its output vector is the state increment. The neural network uses a deep residual network as its basic architecture, including multiple residual units, each of which introduces a residual connection. The training data for the neural network comes from historical slope deformation observation data, including fused GNSS and IMU observation results and baseline recursive boundary state sequences. The goal of model training is to minimize the error between the predicted state and the actual observed state; The prediction-update process based on the Kalman filter enhancement model dynamically adjusts the state transition matrix and the observation matrix, specifically including: The state transition matrix is ​​dynamically corrected, and the Jacobian matrix is ​​calculated based on the prediction results of the differential equation. The Kalman gain matrix is ​​optimized, and the observation noise covariance matrix is ​​dynamically adjusted based on the observation residuals, which reflect the deviation between the Neural ODE predictions and the GNSS observations. The observation mapping method is dynamically selected based on the validity of sensor data, and the observation matrix is ​​adaptively adjusted.

2. The slope displacement estimation method based on data fusion as described in claim 1, characterized in that, The process of data acquisition and preprocessing based on multiple sensors, aligning the acquired data in time and space, specifically includes: The multi-sensor system includes a high-performance GNSS receiver and an IMU with integrated high-precision triaxial accelerometers; Hardware synchronization between accelerometer sampling and GNSS time is achieved through the PPS signal provided by the GNSS receiver; The accelerometer data is preprocessed with temperature drift compensation and zero bias compensation. Perform quality assessment and outlier removal on GNSS-acquired data; The coordinate data acquired by GNSS after removing outliers is converted into a geocentric coordinate system, and then further converted into a navigation coordinate system, which is the ENU coordinate system, with its origin set as the reference point of the monitoring area.

3. The slope displacement estimation method based on data fusion as described in claim 1, characterized in that, The fusion estimation and adaptive error correction based on multi-sensor data specifically includes: The state is jointly estimated based on Neural ODE prediction and IMU and GNSS observation data, and the multi-source observation joint estimation of the state vector is achieved through a weighted mapping function. Drift is controlled by IMU dual-integral position estimation, and zero bias is dynamically and recursively corrected; Calculate the residuals between GNSS observations and Neural ODE predicted states, dynamically adjust the observation weights, and adjust the observation noise covariance matrix based on the residual mean square error.

4. The slope displacement estimation method based on data fusion as described in claim 1, characterized in that, The method of combining distributed computing and edge computing to calculate slope displacement estimates in real time specifically includes: The data acquisition tasks from multiple sensors are distributed to multiple computing nodes to achieve distributed data processing; The collected data is pre-processed at the edge computing nodes, including data preprocessing, feature extraction, and preliminary analysis. The data processed by the edge computing nodes is transmitted to the central server for further fusion processing and analysis to provide real-time early warning of slope displacement.

5. A slope displacement estimation device based on data fusion, based on the slope displacement estimation method based on data fusion as provided in claim 1, characterized in that, include: The data acquisition module is used for data acquisition and preprocessing based on multiple sensors, aligning the acquired data in time and space. The data filtering module is used to extract discrete observation points and generate benchmark sampling segments based on the aligned data, and to filter high-quality data segments using the sliding window method. The dynamic modeling module is used to dynamically model the filtered data based on neural differential equations and to perform nonlinear modeling of the slope deformation process using neural networks. The dynamic adjustment module is used to dynamically adjust the state transition matrix and observation matrix in the prediction-update process based on the Kalman filter enhancement model. The fusion estimation and error correction module is used for fusion estimation and adaptive error correction based on multi-sensor data. The displacement estimation module is used to calculate slope displacement estimates in real time by combining distributed computing and edge computing.

6. A slope displacement estimation device based on data fusion, characterized in that, It includes at least one processing unit and at least one storage unit, wherein the storage unit stores a computer program that, when executed by the processing unit, causes the processing unit to perform the steps of the method according to any one of claims 1 to 4.

7. A storage medium, characterized in that, It stores a computer program executable by a data fusion-based slope displacement estimation device, which, when run on the data fusion-based slope displacement estimation device, causes the data fusion-based slope displacement estimation device to perform the steps of the method according to any one of claims 1 to 4.

Images