Nonlinear mapping pattern recognition method for extracting multi-dimensional deformation law of geologic body

Abstract

The present application relates to the technical field of data recognition and data representation, and discloses a nonlinear mapping pattern recognition method for extracting multi-dimensional deformation law of geological bodies, which comprises the following steps: dividing observation points into spatially adjacent characteristic bodies, subtracting displacement average to generate local spatial decoupling deformation components; stacking and constructing a space-time characteristic tensor, and generating a space-time evolution trajectory tensor through time delay reconstruction; outputting a multi-dimensional deformation intrinsic trajectory matrix through Gaussian kernel mapping and local preservation projection dimension reduction; constructing a topological distance matrix according to the Euclidean distance of points in a monitoring window, and solving a local topological self-recurrent Boolean matrix through conversion; calculating a topological self-recurrent rate index, determining a transition to an irreversible sliding instability mode state and outputting an alarm signal when the index is lower than a mutation damping threshold and the manifold curvature time limit integral value of the multi-dimensional deformation intrinsic trajectory matrix diverges; and the present application eliminates common mode displacement disturbance noise and excludes environmental alternating interference.

Description

Technical Field

[0001] This invention belongs to the field of data recognition and data representation technology, and in particular relates to a nonlinear mapping pattern recognition method for extracting the multidimensional deformation laws of geological bodies. Background Technology

[0002] Currently, spatiotemporal remote sensing methods such as satellite navigation positioning and synthetic aperture radar measurement are commonly used to obtain the spatial displacement sequence of the observed object's surface. The displacement data of each observation point are then arranged into a time series matrix. By solving the gradient changes of the matrix in the spatiotemporal dimension, the deformation trend is characterized, thereby achieving safety status identification and classification representation.

[0003] When conducting anomaly pattern recognition under conditions such as open-pit mine slopes, the geological body, as an open system, is inevitably driven by physical environmental fields such as rainfall infiltration, water level fluctuations, and temperature changes, resulting in displacement of the geological body surface. This causes large-scale trajectory drift of the displacement components in the low-dimensional feature space. However, existing methods rigidly use a static reference subspace as the judgment criterion, making the system prone to misclassifying this normal trajectory drift as abrupt abnormal displacement changes when facing the alternating fluctuations of the above environment, thus generating high-frequency false alarms. For example, Chinese invention patent application CN121482997A discloses a landslide displacement monitoring method based on the Internet of Things, which... The method uses signal strength and latency metrics between network nodes to optimize transmission links and screen stable data sequences. This approach determines data validity based on the communication characteristics of network transmission, assuming that stable transmission quality equates to a valid data sequence. However, when a geological body enters a nonlinear deformation stage, its deformation signals are often accompanied by complex environmental disturbances. Even if the communication link maintains a high signal-to-noise ratio, the topological features inside the geological body may have undergone irreversible slip evolution. This comparative technique ignores the nonlinear dynamic evolution law inside the geological body and the non-correlation with external physical quantities, resulting in the inability to identify potential signs of instability under complex and non-ideal working conditions, and has limitations such as loss of true data and misjudgment of risks.

[0004] Therefore, the technical problem to be solved by this invention is how to establish a feature self-recursive representation mechanism that does not rely on peripheral sensor compensation and retains the inflection point of abrupt change, adaptively offset long-period dynamic drift, and realize the self-repair of deformation features and classification of abrupt change state under the interference of interrupted flow. Summary of the Invention

[0005] To address the problems in the background art, the technical solution of the present invention is as follows: A nonlinear mapping pattern recognition method for extracting multidimensional deformation laws of geological bodies, comprising the following steps: Step S1: Obtain the displacement data stream of the observation point, divide the observation point into spatial nearest neighbor feature bodies according to the spatial relative distance, subtract the mean displacement of the measurement point within the spatial nearest neighbor feature bodies to reduce common mode noise, and generate local spatial decoupled deformation components. Step S2: Within the data monitoring window, the local spatial decoupled deformation components are stacked along the time axis to construct a spatiotemporal feature tensor. After time delay reconstruction, the topological extension of the phase space is completed, and the spatiotemporal evolution trajectory tensor is generated. Step S3: Project the spatiotemporal evolution trajectory tensor onto the regenerated kernel space using Gaussian kernel mapping to generate a kernel feature matrix. Reduce the dimensionality of the kernel feature matrix based on the local preservation projection criterion to output a multidimensional deformation eigentrajectory matrix. Step S4: Extract a local sliding monitoring window from the multidimensional deformation intrinsic trajectory matrix, calculate the Euclidean distance between each vector point pair within the local sliding monitoring window to construct a symmetric topological distance matrix, and use 1.5 times the median of the Euclidean distance between point pairs in the historical health stage as an adaptive tolerance threshold to binarize the symmetric topological distance matrix and solve for the local topological autorecursive Boolean matrix. Step S5: Calculate the proportion of non-zero elements in the upper triangular region of the local topological self-recursive Boolean matrix to obtain the topological self-recursion rate index. When the topological self-recursion rate index is continuously lower than the abrupt damping threshold of 0.15 and the time-limited integral value of the manifold curvature of the multidimensional deformation intrinsic trajectory matrix continues to diverge and exceeds the set threshold, the state transition of the geological body to the irreversible slip instability mode is determined, and an instability alarm signal is output.

[0006] Preferably, step S1 specifically includes the following sub-steps: Step S11, synchronously collect observation point displacement data streams from multiple independent sources using a multi-channel data interface; Step S12, perform group clustering on the multi-source observation points according to the geometric distance of the physical space dimension, and divide spatially adjacent sampling nodes with common-mode correlation into the same spatial nearest neighbor feature body; Step S13, calculate the spatial mean of the data of each sampling node within the spatial nearest neighbor feature body as the mean of the measurement point displacement, and eliminate the common-mode displacement disturbance noise caused by the long-period environmental physical field on spatially adjacent measurement points by subtracting the mean of the measurement point displacement from the observation point displacement data stream, thereby generating a local spatial decoupled deformation component.

[0007] Preferably, step S1 further includes a data self-repair step triggered when data loss is detected: calculating the first-order differential gradient difference between the immediately preceding and following intact time-series data of the missing interval; determining the reciprocal of the squared term of the first-order differential gradient difference plus 1 as the dynamic penalty weight factor, and introducing the dynamic penalty weight factor as a regularization constraint term into the reconstruction loss function with the goal of minimizing the nuclear norm; releasing the smoothing constraint to fill in random packet loss in the steady phase; and relaxing the smoothness constraint by increasing the dynamic penalty weight factor in the accelerated sliding critical phase, thereby maintaining the local slope change inflection point characteristics in data reconstruction.

[0008] Preferably, step S3 specifically includes the following sub-steps: Step S31, using a Gaussian kernel function to perform nonlinear similarity operations on the spatiotemporal evolution trajectory tensor to construct a global kernel matrix; Step S32, based on the adjacency degree of adjacent state samples in the original feature space, setting a local topology-preserving adjacency weight matrix in the regenerated kernel space, wherein for non-adjacent samples, the corresponding weight value is set to 0; Step S33, by solving the generalized eigenvalue decomposition equations about the global kernel matrix and the local topology-preserving adjacency weight matrix, selecting the eigenvectors corresponding to the smallest non-zero eigenvalue of the first term to form a low-dimensional projection matrix, reducing the high-dimensional feature linear projection to a low-dimensional intrinsic manifold space, and generating a multidimensional deformable intrinsic trajectory matrix.

[0009] Preferably, in step S5, the state of the closed loop attractor during normal dynamic fluctuations is characterized by the topological self-recursion rate index: when the multidimensional deformation intrinsic trajectory matrix is ​​shifted due to environmental alternation, the deformation dynamic evolution trajectory within the local sliding monitoring window is confined within the self-recursive closed loop attractor in the low-dimensional intrinsic manifold space; in this state, the symmetric topological distance matrix within the local sliding monitoring window exhibits a self-recursive distribution, and the topological self-recursion rate index is stably maintained in the range of not less than 0.90, so as to eliminate interference signals generated by environmental alternation.

[0010] Preferably, step S5 specifically includes the following sub-steps: Step S51, fitting the multidimensional deformation intrinsic trajectory matrix into a continuous dynamic trajectory curve in the low-dimensional intrinsic manifold space, and solving the differential geometric local manifold curvature in real time along the tangent direction of the dynamic trajectory curve; Step S52, performing a sliding time-limited numerical integration on the local manifold curvature within a set sliding time limit to obtain the manifold curvature time-limited integral value; Step S53, when the topological self-recursion rate index is continuously lower than 0.15, and the manifold curvature time-limited integral value breaks the steady-state fluctuation threshold and continues to diverge beyond the set threshold, confirming the state transition to the irreversible slip instability mode.

[0011] Preferably, before outputting the instability alarm signal, the method further includes a step of quantifying and compensating for the acquisition time phase lag deviation: calculating the sliding cross-correlation function between each sampling channel through cross-correlation analysis in the sliding time domain, extracting the time offset corresponding to the maximum value as the relative time phase lag between each channel; aligning the time phase components of the state transition determination point based on the relative time phase lag, eliminating the system identification delay caused by the asynchronous transmission delay of multiple channels, and completing the quantization compensation for the state transition time.

[0012] Preferably, the output instability alarm signal specifically involves the underlying mapping steps that drive the alarm hardware to generate discrete response actions: converting the state transition time after quantization compensation into a hard real-time interrupt data packet; directly writing the hard real-time interrupt data packet into the external interrupt trigger register mapping area of ​​the alarm output controller through the bus interface, forcibly switching to the highest interrupt priority to drive the alarm hardware to generate discrete response actions.

[0013] Preferably, the method is applied to the monitoring and early warning of geological landslides on the slopes of large open-pit mines or the banks of reservoirs of major water conservancy projects. When high-frequency environmental disturbances and sensor temperature drift occur in the acquired multi-channel displacement sensor signals, the data processing unit reduces the common-mode displacement disturbance noise on adjacent measuring points by decoupling the deformation components in the local space. Under the premise of using the topological self-recursive rate index to offset the long-period environmental drift, the system uses the joint verification judgment that the topological self-recursive rate index is continuously lower than 0.15 and the time limit integral value of the manifold curvature exceeds the divergence threshold. The state transition time after quantization compensation is used as the start time of the forward sliding segment and input to the alarm output controller to trigger the response action of the millisecond-level external alarm hardware.

[0014] Compared with existing technologies, the nonlinear mapping pattern recognition method for extracting multidimensional deformation laws of geological bodies in this invention has the following advantages: 1. In the nonlinear mapping pattern recognition of multidimensional deformation law extraction of geological bodies, the processor adopts spatial nearest neighbor difference transformation to subtract the spatial mean of the corresponding nearest neighbor monitoring unit from the cumulative displacement component collected by each spatial observation point, generating local spatial decoupled displacement value, and reconstructing the third-order geometric deformation spatiotemporal tensor, effectively offsetting the common-mode mechanical jitter noise generated by external strong winds, rain, snow or navigation satellite transient disturbances on the nearest measuring point group; on this basis, when reconstructing lost data, the processor introduces the differential gradient of the intact time sequence before and after as a regularization penalty factor for reconstruction loss, while maintaining the consistency of the dimension of the feature matrix, and forcibly retaining the nonlinear slope characteristics of the local abrupt inflection point in the data reconstruction.

[0015] 2. The processor performs a dynamic phase space topological extension on the time dimension of the deformation feature tensor through time-domain time-delay reconstruction, transforming the static displacement deformation features into trajectory manifold features containing dynamic change trajectories, and calculating the sample nearest neighbor topological weight matrix in the high-dimensional kernel space. Based on this, the processor solves the generalized eigenvalue decomposition based on the local topology-preserving projection criterion, adaptively projecting the features in the high-dimensional mapping space to reduce the dimensionality to the low-dimensional intrinsic manifold space, generating a multi-dimensional deformation intrinsic trajectory matrix, and reconstructing motion trajectory patterns that do not drift with the measurement scale in the low-dimensional intrinsic manifold, avoiding interference of feature nonlinear interleaving on classification performance.

[0016] 3. The processor uses the low-dimensional deformation intrinsic trajectory matrix to calculate the symmetric topological distance matrix within the sliding monitoring sub-window, and determines the topological adaptive tolerance threshold based on global historical trajectory characteristics. The judgment result generates a local topological autorecursive Boolean matrix. This feature autorecursive relationship represents the periodic environmental deformation as an autorecursive closed-loop attractor in the low-dimensional intrinsic feature space. This ensures that when trajectory drift is caused by environmental alternation, the topological autorecursion rate index within the local sliding window remains stable at a high level. This solves the defect of traditional data comparison methods that misjudge normal trajectory drift caused by environmental alternation as abnormal instability, and achieves effective hedging against normal dynamic fluctuations. Attached Figure Description

[0017] Figure 1 This is a flowchart of the deformation data reconstruction and instability determination method of the present invention; Figure 2 This is a hardware interaction diagram for slope displacement acquisition and alarm in this invention. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0019] A nonlinear mapping pattern recognition method for extracting multidimensional deformation patterns of geological bodies includes the following steps: Step S1: Obtain the displacement data stream of the observation point, divide the observation point into spatial nearest neighbor feature bodies according to the spatial relative distance, subtract the mean displacement of the measurement point within the spatial nearest neighbor feature bodies to reduce common mode noise, and generate local spatial decoupled deformation components. Step S2: Within the data monitoring window, the local spatial decoupled deformation components are stacked along the time axis to construct a spatiotemporal feature tensor. After time delay reconstruction, the topological extension of the phase space is completed, and the spatiotemporal evolution trajectory tensor is generated. Step S3: Project the spatiotemporal evolution trajectory tensor onto the regenerated kernel space using Gaussian kernel mapping to generate a kernel feature matrix. Reduce the dimensionality of the kernel feature matrix based on the local preservation projection criterion to output a multidimensional deformation eigentrajectory matrix. Step S4: Extract a local sliding monitoring window from the multidimensional deformation intrinsic trajectory matrix, calculate the Euclidean distance between each vector point pair within the local sliding monitoring window to construct a symmetric topological distance matrix, and use 1.5 times the median of the Euclidean distance between point pairs in the historical health stage as an adaptive tolerance threshold to binarize the symmetric topological distance matrix and solve for the local topological autorecursive Boolean matrix. Step S5: Calculate the proportion of non-zero elements in the upper triangular region of the local topological self-recursive Boolean matrix to obtain the topological self-recursion rate index. When the topological self-recursion rate index is continuously lower than the abrupt damping threshold of 0.15 and the time-limited integral value of the manifold curvature of the multidimensional deformation intrinsic trajectory matrix continues to diverge and exceeds the set threshold, the state transition of the geological body to the irreversible slip instability mode is determined, and an instability alarm signal is output.

[0020] Preferably, step S1 specifically includes the following sub-steps: Step S11, synchronously collect observation point displacement data streams from multiple independent sources using a multi-channel data interface; Step S12, perform group clustering on the multi-source observation points according to the geometric distance of the physical space dimension, and divide spatially adjacent sampling nodes with common-mode correlation into the same spatial nearest neighbor feature body; Step S13, calculate the spatial mean of the data of each sampling node within the spatial nearest neighbor feature body as the mean of the measurement point displacement, and eliminate the common-mode displacement disturbance noise caused by the long-period environmental physical field on spatially adjacent measurement points by subtracting the mean of the measurement point displacement from the observation point displacement data stream, thereby generating a local spatial decoupled deformation component.

[0021] Preferably, step S1 further includes a data self-repair step triggered when data loss is detected: calculating the first-order differential gradient difference between the immediately preceding and following intact time-series data of the missing interval; determining the reciprocal of the squared term of the first-order differential gradient difference plus 1 as the dynamic penalty weight factor, and introducing the dynamic penalty weight factor as a regularization constraint term into the reconstruction loss function with the goal of minimizing the nuclear norm; releasing the smoothing constraint to fill in random packet loss in the steady phase; and relaxing the smoothness constraint by increasing the dynamic penalty weight factor in the accelerated sliding critical phase, thereby maintaining the local slope change inflection point characteristics in data reconstruction.

[0022] Preferably, step S3 specifically includes the following sub-steps: Step S31, using a Gaussian kernel function to perform nonlinear similarity operations on the spatiotemporal evolution trajectory tensor to construct a global kernel matrix; Step S32, based on the adjacency degree of adjacent state samples in the original feature space, setting a local topology-preserving adjacency weight matrix in the regenerated kernel space, wherein for non-adjacent samples, the corresponding weight value is set to 0; Step S33, by solving the generalized eigenvalue decomposition equations about the global kernel matrix and the local topology-preserving adjacency weight matrix, selecting the eigenvectors corresponding to the smallest non-zero eigenvalue of the first term to form a low-dimensional projection matrix, reducing the high-dimensional feature linear projection to a low-dimensional intrinsic manifold space, and generating a multidimensional deformable intrinsic trajectory matrix.

[0023] Preferably, in step S5, the state of the closed loop attractor during normal dynamic fluctuations is characterized by the topological self-recursion rate index: when the multidimensional deformation intrinsic trajectory matrix is ​​shifted due to environmental alternation, the deformation dynamic evolution trajectory within the local sliding monitoring window is confined within the self-recursive closed loop attractor in the low-dimensional intrinsic manifold space; in this state, the symmetric topological distance matrix within the local sliding monitoring window exhibits a self-recursive distribution, and the topological self-recursion rate index is stably maintained in the range of not less than 0.90, so as to eliminate interference signals generated by environmental alternation.

[0024] Preferably, step S5 specifically includes the following sub-steps: Step S51, fitting the multidimensional deformation intrinsic trajectory matrix into a continuous dynamic trajectory curve in the low-dimensional intrinsic manifold space, and solving the differential geometric local manifold curvature in real time along the tangent direction of the dynamic trajectory curve; Step S52, performing a sliding time-limited numerical integration on the local manifold curvature within a set sliding time limit to obtain the manifold curvature time-limited integral value; Step S53, when the topological self-recursion rate index is continuously lower than 0.15, and the manifold curvature time-limited integral value breaks the steady-state fluctuation threshold and continues to diverge beyond the set threshold, confirming the state transition to the irreversible slip instability mode.

[0025] Preferably, before outputting the instability alarm signal, the method further includes a step of quantifying and compensating for the acquisition time phase lag deviation: calculating the sliding cross-correlation function between each sampling channel through cross-correlation analysis in the sliding time domain, extracting the time offset corresponding to the maximum value as the relative time phase lag between each channel; aligning the time phase components of the state transition determination point based on the relative time phase lag, eliminating the system identification delay caused by the asynchronous transmission delay of multiple channels, and completing the quantization compensation for the state transition time.

[0026] Preferably, the output instability alarm signal specifically involves the underlying mapping steps that drive the alarm hardware to generate discrete response actions: converting the state transition time after quantization compensation into a hard real-time interrupt data packet; directly writing the hard real-time interrupt data packet into the external interrupt trigger register mapping area of ​​the alarm output controller through the bus interface, forcibly elevating it to the highest interrupt priority to drive the alarm hardware to generate discrete response actions.

[0027] Preferably, the method is applied to the monitoring and early warning of geological landslides on the slopes of large open-pit mines or the banks of reservoirs of major water conservancy projects. When high-frequency environmental disturbances and sensor temperature drift occur in the acquired multi-channel displacement sensor signals, the data processing unit reduces the common-mode displacement disturbance noise on adjacent measuring points by decoupling the deformation components in the local space. Under the premise of using the topological self-recursion rate index to offset the long-period environmental drift, the topological self-recursion rate index is continuously lower than 0.15, and the bidirectional causality verification is determined by combining the manifold curvature time limit integral value exceeding the divergence threshold. The state transition time after quantization compensation is used as the start time of the forward sliding segment and input to the alarm output controller to trigger the response action of the millisecond-level external alarm hardware.

[0028] Example 1: The current data processing unit synchronously collects data distributed on the surface of the monitoring area through a multi-channel data interface. A three-dimensional cumulative displacement data stream of physical space observation points, wherein... The total number of physical space observation points is 12 in this embodiment. The processor calculates the geometric distance between each pair of observation points based on their spatial coordinates and determines the spatial distance threshold according to the spacing of the observation points. Spatially adjacent sampling nodes with a distance not greater than the threshold are assigned to the same candidate group. Subsequently, the correlation between the displacement sequences of each sampling node in the candidate group is calculated. Sampling nodes with the same displacement change direction and the correlation reaching the preset condition are identified as having common mode correlation and assigned to the same spatial nearest neighbor feature body. The processor calculates the spatial mean of the data of each sampling node within the spatial nearest neighbor feature body and uses this spatial mean as the mean of the measurement point displacement. Then, the mean of the measurement point displacement is subtracted from the displacement data stream of each observation point to eliminate the common mode displacement disturbance noise caused by the long-period environmental physical field on the spatially adjacent measurement points and generate local spatial decoupled deformation components.

[0029] When rainfall causes short-term signal obstruction or random packet loss at some physical space observation points, the processor detects the missing data intervals within the current data monitoring window, initiates the data self-repair process, and calculates the first-order differential gradients of the immediately preceding and following intact time-series data to obtain the first-order differential gradient of the preceding interval. and subsequent first-order differential gradient The dynamic penalty weight factor is calculated according to the following formula. : The processor introduces a dynamic penalty weighting factor as a regularization constraint into the reconstruction loss function, which aims to minimize the kernel norm. This factor adjusts the relative interaction between the kernel norm term and the reconstruction residual term. During the stationary deformation phase, the processor maintains the smoothness constraint of the kernel norm term to fill in data gaps caused by random packet loss. During the accelerated sliding critical phase, when the intact time-series data before and after the missing interval maintain the same slope change trend, the difference between the first-order differential gradient of the preceding interval and the first-order differential gradient of the following interval decreases, and the dynamic penalty weighting factor increases. Accordingly, the processor reduces the relative weight of the kernel norm term in the reconstruction loss function, relaxes the smoothness constraint, and makes the reconstructed data closer to the original observation data containing slope changes, preserving the characteristics of local slope abrupt change inflection points.

[0030] Within the data monitoring window, the processor stacks the local spatially decoupled deformation components along the time axis to construct a spatiotemporal feature tensor. Spatiotemporal feature tensor The size is The three displacement components correspond to three spatial directions, respectively. The total number of sampling time points within the data monitoring window is 48 in this embodiment. The processor sets the time delay to 2 hours and, according to the preset reconstruction order, extracts spatiotemporal feature tensor slices distributed at 2-hour intervals from each valid sampling time. The slices are arranged along the newly added time delay direction. The preset reconstruction order remains unchanged within the same data monitoring window. Each group of reconstructed data contains the same number of time delay slices. The reconstruction results are arranged according to the order of the valid sampling times to complete the topological extension of the phase space and generate the spatiotemporal evolution trajectory tensor. Spacetime evolution trajectory tensor Each valid state sample contains spatial displacement features corresponding to the sampling time and its historical time delay position, which are used to represent the dynamic trajectory of displacement data evolving over time.

[0031] The processor will convert the spacetime evolution trajectory tensor The time delay feature corresponding to each valid sampling time point is expanded into a state sample vector. A Gaussian kernel function is used to calculate the nonlinear similarity between any two state sample vectors, and the calculation results are arranged according to the sample index to generate a kernel feature matrix. In this embodiment, the kernel feature matrix It also serves as a global kernel matrix, used to represent the similarity relationship of each state sample after being projected onto the regenerated kernel Hilbert space.

[0032] In constructing a local topology, the adjacency weight matrix is ​​maintained. At that time, the processor calculates the Euclidean distance between each state sample vector in the original feature space, and selects a preset number of neighboring state samples for each state sample in ascending order of distance; for neighboring state samples, the kernel feature matrix is... The Gaussian kernel similarity is used as the adjacency weight. For non-adjacent samples, the corresponding weight value is set to 0, and the adjacency weight matrix is ​​preserved by the local topology. The sum of the weights of each row forms a degree matrix, and then the local topology-preserving adjacency weight matrix is ​​subtracted from the degree matrix. The Laplacian matrix is ​​obtained, and the processor uses the global kernel matrix based on the local preservation projection criterion. Left and right multiplication by the Laplace matrix constitutes one side of the generalized eigenvalue decomposition equation, with the global kernel matrix as the basis. The left and right multiplication degree matrices form the other side, and the corresponding generalized eigenvalues ​​and eigenvectors are solved.

[0033] The processor selects the first feature value in ascending order. The eigenvectors corresponding to the smallest non-zero eigenvalues ​​form a low-dimensional projection matrix, where, To determine the low-dimensional spatial dimension, which is 3 in this embodiment, the kernel feature matrix is ​​used by the low-dimensional projection matrix. Linear projection is performed on each state sample to reduce the dimensionality of the high-dimensional features to a smaller value. In a low-dimensional eigenmanifold space, the projection vectors are arranged in the order of effective sampling times to generate a multidimensional deformation eigentrajectory matrix. Multidimensional deformation eigentrajectory matrix Each column in the data corresponds to a valid sampling time. Each element represents the coordinates of the intrinsic trajectory in the low-dimensional intrinsic manifold space at that sampling moment.

[0034] Processor in multidimensional deformation eigentrajectory matrix The length of the cut-off section is The local sliding monitoring window, in which, The number of sampling points contained in the local sliding monitoring window is 48 in this embodiment. The processor calculates any two sampling times within the local sliding monitoring window. and The Euclidean distance between the eigentrajectories of the vectors is denoted as . Then, arrange all Euclidean distances according to the sampling time index to construct a symmetric topological distance matrix. .

[0035] The processor calculates the median of the Euclidean distance between pairs of intrinsic trajectory vectors during each historical health phase, and sets 1.5 times this median as the topology adaptive tolerance threshold. Subsequently, a topology-adaptive tolerance threshold was used. Symmetric topological distance matrix The elements in Perform binarization transformation and solve for the local topological autorecursive Boolean matrix. ,in, Local topological self-recursive Boolean matrix The logical decision element in; when hour, Take 1, when hour, Take 0.

[0036] Processor statistical local topology autorecursive Boolean matrix The topological self-recursion rate index is obtained by calculating the number of non-zero elements in the upper triangle (excluding diagonal elements) and the proportion of this number to the total number of elements in the upper triangle. When the monitored area is only subject to periodic environmental physical field disturbances, and the deformation evolution trajectory remains within a self-recursive closed-loop attractor, the symmetric topological distance matrix... It exhibits a self-recursive distribution, and the topological self-recursion rate index The value remains stable within a range not lower than 0.90, used to eliminate interference from alternating temperature and rainfall. When shear slip occurs within the soil and rock medium and evolves towards instability, the intrinsic trajectory gradually deviates from the self-recursive closed-loop attractor and shifts unidirectionally outward from the attractor. This is the topological self-recursion rate index. It then decreased and fell below the mutation damping threshold of 0.15.

[0037] In calculating the topological autorecursion rate index Simultaneously, the processor processes the multidimensional deformation eigentrajectories according to the sampling time sequence. The intrinsic trajectory vectors are fitted to a continuous dynamic trajectory curve, and the curvature of the local manifold in differential geometry is calculated along the tangent direction of the dynamic trajectory curve. In this embodiment, the time span between the start and end sampling times of the current local sliding monitoring window is used as the set sliding time limit. Within this set sliding time limit, the processor calculates the differential geometric local manifold curvature of two adjacent sampling times according to the time interval between adjacent sampling times. Take the average value, multiply the average value by the corresponding time interval, and then sum them up segment by segment to obtain the time-limited integral value of the manifold curvature.

[0038] The threshold for setting the time-limited integral value of manifold curvature is calibrated based on the manifold curvature characteristics of the geological body during its historical stable deformation stages. During system initialization or initial calibration, the data processing unit continuously collects the time-limited integral value of the manifold curvature when the geological body is in a state of no slippage and creep. It calculates the statistical mean and standard deviation of this integral value, and uses the sum of the statistical mean and three times the standard deviation as the threshold for setting the time-limited integral value of manifold curvature. For the medium-sized rock slope used in this embodiment, this threshold is calibrated within the range of 1.2 to 1.8. When the topological self-recursion rate index... When the value remains below 0.15 and the time-limited integral value of the manifold curvature continues to diverge and exceeds the set threshold of the time-limited integral value of the manifold curvature, the processor determines that the geological body has transitioned to an irreversible slip instability mode.

[0039] Before outputting the instability alarm signal, the processor quantifies and compensates for the phase lag deviation between each sampling channel. Within the current local sliding monitoring window, the processor selects the sampling channel with complete data as the reference channel and performs mean-removal processing on the displacement sequences of the reference channel and the other sampling channels respectively. Subsequently, the processor changes the relative time offset between the two displacement sequences sequentially according to an integer number of sampling periods, and calculates the product and summation of the corresponding sample values ​​of the two sequences at each time offset position to form a sliding cross-correlation function. For each sampling channel to be aligned, the processor extracts the time offset corresponding to the maximum value of the sliding cross-correlation function and uses it as the relative phase lag of the sampling channel relative to the reference channel.

[0040] The processor moves the time index of the corresponding sampling channel according to the direction and magnitude of the relative phase lag, so that the data components corresponding to the same deformation process in each sampling channel are aligned to the same sampling phase. The sampling period is used as the quantization unit of the time offset. After the phase components of each channel are aligned, the processor corrects the state transition determination point according to the relative phase lag, and uses the corrected sampling time as the quantized compensated state transition time to eliminate the system identification delay caused by the asynchronous transmission delay of multiple channels.

[0041] The processor converts the quantized and compensated state transition time into a hard real-time interrupt data packet, and writes the hard real-time interrupt data packet directly into the external interrupt trigger register mapping area of ​​the alarm output controller through the bus interface, forcing the track to the highest interrupt priority, and driving the alarm hardware to generate discrete response actions.

[0042] The data processing unit adopts a dual-time task execution architecture. Phase space topology extension, kernel mapping dimensionality reduction, and state transition determination run in the background non-real-time thread of the computing unit and are updated and evaluated according to a fixed period. When the background non-real-time thread confirms that the geological body has transitioned to an irreversible slip instability mode and completes the quantization compensation of the state transition time, it converts the determination result into a high-priority hard real-time interrupt request. This interrupt request is mapped to a physical interrupt pin through the hardware abstraction layer. The physical interrupt pin is connected to the external interrupt input register of the alarm output controller through the bus interface. After the interrupt control register detects the edge trigger signal, it interrupts the processor's normal task queue within microseconds, executes the alarm driver program with the highest interrupt priority, and outputs a high-level pulse to the external alarm hardware to form an instability alarm signal.

[0043] Through the above processing, the data processing unit sequentially completes the generation of local spatial decoupling deformation components, construction of spatiotemporal feature tensors, generation of spatiotemporal evolution trajectory tensors, dimensionality reduction of kernel feature matrices, solution of local topological self-recursive Boolean matrices, and state transition determination. The periodic deformation fluctuations caused by temperature and rainfall remain as self-recursive closed-loop attractors in the low-dimensional intrinsic manifold space, while the intrinsic trajectory evolving towards the irreversible slip instability mode is manifested as a unidirectional shift away from the attractor. Thus, without adding external environmental sensor compensation, the processor offsets long-period environmental drift and retains the local slope change inflection point characteristics before accelerated instability, converting the deformation state into an instability alarm signal.

[0044] Example 2: The current data processing unit synchronously collects three-dimensional displacement data of 12 physical spatial observation points on the surface of the geological body at continuous sampling times through a general data interface, generating a multi-dimensional cumulative displacement data stream. The data used comes from a publicly available landslide displacement monitoring dataset. The processor selects the total number of sampling time points within the data monitoring window based on the deformation signal mutation rate, signal frequency domain bandwidth, and matrix multiplication and addition operation load. and the length of the local sliding monitoring window When the bandwidth of the high-frequency components of the deformation signal increases, the processor size increases. To preserve high-frequency deformation components; as the number of sampling points increases, the kernel feature matrix... As the scale grows quadratically, the computational and storage requirements for eigenvalue decomposition increase accordingly. The processor calculates the first-order difference variance of the displacement data stream within the previous data monitoring window online. When this variance exceeds a preset limit, it is determined that the high-frequency glitches are enhanced, and... Set to 48, Set to 48.

[0045] In the experiment, Gaussian white noise with a signal-to-noise ratio of 20dB and power frequency interference harmonics with a frequency of 50Hz were superimposed on the multidimensional cumulative displacement data stream. Data interruption was set from the 10th sampling time to the 15th sampling time, so that the interruption and packet loss rate of the original input was 15%. Comparative Examples 1 to 3, Examples 1 to 3, and two groups of local sliding monitoring window length exceeding the range were all based on the above 12 physical space observation points, 20dB noise injection signal-to-noise ratio, and 15% interruption and packet loss rate as input conditions.

[0046] For the sample group of this invention, the processor first performs grouping of observation points according to physical spatial geometric distance, subtracts the mean displacement of the measurement points within each spatially nearest feature body to generate local spatial decoupled deformation components, and calculates the first-order differential gradients of the immediately preceding and following intact time series data for the missing data interval. and The dynamic penalty weight factor is determined by adding 1 to the reciprocal of the squared term of the first-order differential gradient difference. The kernel norm is introduced as a regularization constraint into the reconstruction loss function aimed at minimizing the kernel norm. In the stationary deformation stage, the processor completes random packet loss through the smoothing constraint of the kernel norm term. In the accelerated sliding critical stage, when the intact data on both sides of the missing interval maintain the same slope change trend, the dynamic penalty weight factor increases. The processor reduces the relative weight of the kernel norm term according to the established mapping relationship, relaxes the smoothness constraint, and retains the local slope change inflection point characteristics. The control group does not perform spatial decoupling and dynamic penalty weight constraint. The interruption interval and high-frequency noise form non-physical jumps on the low-dimensional trajectory curve, affecting the stability of subsequent eigenvalue decomposition.

[0047] After completing the data self-repair, the processor stacks the local spatial decoupled deformation components along the time axis to construct a spatiotemporal feature tensor, and then reconstructs it with time delay to generate a spatiotemporal evolution trajectory tensor. Then, using Gaussian kernel mapping, Projecting onto the regenerating kernel Hilbert space and solving the generalized eigenvalue decomposition equation based on the local preservation projection criterion, the high-dimensional features are reduced to a three-dimensional low-dimensional eigenmanifold space, outputting a multidimensional deformable eigentrajectory matrix. ,exist Under the condition of 48, the processor from A local sliding monitoring window is extracted, the Euclidean distance between each pair of intrinsic trajectory vector points within the window is calculated, a symmetric topological distance matrix is ​​constructed, and a local topological autorecursive Boolean matrix is ​​obtained after binarization transformation.

[0048] The topology adaptive tolerance threshold was determined based on sensitivity test results of historical multi-source landslide monitoring datasets under different noise levels and packet loss rates. When this threshold is lower than 1.2 times the median Euclidean distance between point pairs in the historical healthy stage, high-frequency random noise will cause more spurious zero values ​​in the local topological autorecursive Boolean matrix. The topology autorecursion rate index... Non-physical fluctuations occur during the healthy phase; when the threshold is higher than 2.0 times the median, some unidirectional escape trajectories during the unstable phase are still judged as self-recursive, increasing false negatives. Based on this, the processor uses 1.5 times the median of the Euclidean distance between historical healthy phase point pairs as the topology adaptive tolerance threshold.

[0049] Regarding the abrupt damping threshold, statistical results from multiple sets of instability samples indicate that after the geological body enters the accelerated sliding stage, the self-recursive closed-loop attractor gradually breaks down, and the intrinsic trajectory shifts unidirectionally outward from the attractor. The sharp drop inflection point is between 0.10 and 0.18. The processor sets 0.15 as the abrupt change damping threshold and... The condition of continuously falling below the threshold and the time-limited integral value of manifold curvature continuously diverging and exceeding the set threshold is used as a two-way causal verification condition to confirm the state transition of the geological body to an irreversible slip instability mode.

[0050] Under the same input interference conditions and Under the condition of 48, the results of control group 1, control group 2 and control group 3 were obtained The values ​​are 0.54, 0.41, and 0.38, respectively, corresponding to pattern recognition accuracies of 68.3%, 61.2%, and 58.6%, respectively. The state transition identification delays are 42ms, 45ms, and 48ms, respectively. The above three sets of delay values ​​are lower than those of the sample group of this invention, but their pattern recognition accuracy decreases simultaneously, indicating that the shorter output time does not correspond to stable unstable pattern recognition results.

[0051] Sample groups 1, 2, and 3 of this invention, after undergoing spatial decoupling, data self-repair, time-delay reconstruction, Gaussian kernel mapping, and local preserving projection processing, are obtained in the healthy evolution stage. The values ​​are 0.94, 0.92, and 0.91 respectively. When the intrinsic trajectory position drifts due to diurnal temperature differences and long-period environmental physical fields, the trajectory within the local sliding monitoring window remains within the self-recursive closed-loop attractor. It remains in the range not lower than 0.90; when the local mass evolves towards an irreversible slip instability mode, the intrinsic trajectory detaches from the attractor and is... Together with the time-limited integral value of manifold curvature, the state transition determination was completed. The pattern recognition accuracy of the three sets of samples was 96.7%, 95.8% and 94.4% respectively, and the state transition recognition time delay was 162ms, 165ms and 168ms respectively.

[0052] To verify the impact of the local sliding monitoring window length on the processing results, Group 1, which exceeded the range, will... Reduced to 24, with the remaining input conditions remaining unchanged, the group obtained The accuracy rate was 71.5%, with a state transition identification delay of 85ms. However, due to the short time series contained within the window, the time delay reconstruction failed to fully cover the closed trajectory formed by long-period environmental fluctuations. Furthermore, the local topological self-recursive Boolean matrix could not stably represent the self-recursive closed loop attractor. Both the accuracy of pattern recognition and the accuracy of pattern recognition are lower than The present invention sample is 48.

[0053] Group 2 (out of range) Expanding to 96, we get The accuracy rate was 0.88, the pattern recognition accuracy was 83.1%, and the state transition identification latency was 412ms. With an increased window length, the processor needed to perform point-to-point distance calculations and matrix operations on more intrinsic trajectory vectors, resulting in a corresponding increase in the feature matrix size and thus a higher state transition identification latency. Although... It is higher than Group 1 (out-of-range), but its pattern recognition accuracy is still lower than Group 1 (out-of-range). The three groups of the present invention are 48.

[0054] In the above experiments, the local spatial decoupling deformation component is used to reduce common-mode displacement disturbance noise at spatially adjacent measurement points; the dynamic penalty weighting factor is used to retain local slope abrupt change inflection points during data completion; and the topological self-recursion rate index is used to eliminate trajectory position changes caused by long-period environmental drift. When the value remains below 0.15 and the integral value of the manifold curvature over time continues to diverge and exceeds the set threshold, the processor confirms the state transition of the irreversible slip instability mode, outputs an instability alarm signal, and generates corresponding data classification and identification instructions and alarm control packages.

[0055] Example 3: The current data processing unit receives data distributed across the surface of the monitoring area on the slope of a large open-pit mine. The three-dimensional cumulative displacement data were synchronously uploaded from several spatial observation points, among which... If we set the value to 12, before constructing the spatiotemporal feature tensor, the processor first checks whether there are continuous missing data or random packet loss within the current data monitoring window, and performs data self-repair on the detected missing data intervals.

[0056] The processor calculates the first-order differential gradients of the immediately preceding and following intact time-series data within the missing interval. and The dynamic penalty weight factor is obtained by adding 1 to the square of the difference between the two terms. Then, it is introduced as a regularization constraint into the reconstruction loss function, which aims to minimize the nuclear norm. Represents the reconstructed deformation observation slice, with This represents the original deformation data. The low-rank structure of the reconstructed data is constrained by the nuclear norm term, and the deviation between the reconstructed data and the original deformation data is constrained by the Frobenius norm residual term. Used to adjust the relative effects of two constraints.

[0057] During the stable deformation phase, the processor maintains the smoothness constraint of the kernel norm terms and fills in the data gaps caused by random packet loss. During the accelerated sliding critical phase, when the intact time-series data before and after the missing interval maintain the same slope trend, the dynamic penalty weight factor increases. The processor reduces the relative weight of the kernel norm terms in the reconstruction loss function according to a predetermined mapping relationship, relaxing the smoothness constraint. Approaching areas with slope changes It preserves the characteristics of local slope abrupt change in inflection points.

[0058] When solving the reconstruction loss function, the processor adopts the alternating direction multiplier algorithm, introduces auxiliary matrix variables, and decomposes the original optimization problem into a kernel norm solution subproblem and a reconstruction residual solution subproblem. The two subproblems are updated alternately by the singular value shrinkage operator and the least squares mapping, respectively. When the relative change of the reconstruction matrix obtained by two adjacent iterations is less than 0.01%, or the number of iterations reaches 500, the iteration stops and the completed deformation observation data is output.

[0059] After completing the data self-repair, the processor stacks the completed data stream along the time axis to construct a spatiotemporal feature tensor. The size of this tensor is ,in, The total number of sampling time points within the data monitoring window is 48 in this embodiment, and the processor uses a 2-hour time delay. The time dimension is used for time delay reconstruction. The spatial displacement features of the current sampling moment and the corresponding historical time delay position are arranged in chronological order to complete the topological extension of the phase space and generate the spatiotemporal evolution trajectory tensor. .

[0060] The processor uses a Gaussian kernel function to calculate the spatiotemporal evolution trajectory tensor. The nonlinear similarity between the trajectory vectors of each sample is calculated, and a kernel feature matrix is ​​constructed according to the sample index. Matrix elements From the sample trajectory vector and Gaussian kernel similarity determination, Gaussian bandwidth Take 1.2 times the median distance of samples within the current data monitoring window. Then, construct a local topology-preserving adjacency weight matrix based on the adjacency relationships of each state sample in the original feature space. For non-adjacent samples, the corresponding weight value is set to 0.

[0061] The processor uses the local preservation projection criterion to process the kernel feature matrix. Adjacency weight matrix with local topology Generalized eigenvalue decomposition is performed. The eigenvectors corresponding to the three smallest non-zero eigenvalues ​​are selected in ascending order to form a low-dimensional projection matrix. This low-dimensional projection matrix is ​​then used to linearly project the high-dimensional kernel features, reducing the dimensionality of each state sample to a three-dimensional low-dimensional eigenmanifold space. The projection results are then arranged according to the order of sampling times to obtain the multidimensional deformation eigentrajectory matrix. .

[0062] Processor in multidimensional deformation eigentrajectory matrix The above captures a local sliding monitoring window with a time span of 48 hours. The number of sampling points included within the window is... This means that for any two sampling times within the window... and Calculate the Euclidean distance between the corresponding intrinsic trajectory vectors. Then, all distances are arranged according to the sampling time index to construct a symmetric topological distance matrix. The processor extracts the median of the Euclidean distance between pairs of intrinsic trajectory vector points during historical health phases, and uses 1.5 times this median as the topology adaptive tolerance threshold. And then Binarize each element in the matrix to generate a locally topological autorecursive Boolean matrix. ,when Not greater than At that time, logical decision elements Take 1; when Greater than hour, Take 0.

[0063] Processor statistical local topology autorecursive Boolean matrix The proportion of non-zero elements in the upper triangular region excluding the diagonal elements yields the topological self-recursion rate index. Simultaneously, the multidimensional deformation eigentrajectory matrix The intrinsic trajectory vectors are fitted to a continuous dynamic trajectory curve, and the curvature of the local manifold in differential geometry is calculated along the tangent direction of the curve. Numerical integration is performed within a set sliding time limit to obtain the time-limited integral value of the manifold curvature.

[0064] When the topology recursion rate metric When the sampling period is below the abrupt damping threshold of 0.15 for three consecutive sampling periods, and the time-limited integral value of the manifold curvature continues to diverge and exceeds the set threshold, the processor determines that the geological body has transitioned to an irreversible slip instability mode. Subsequently, the relative temporal lag between each sampling channel is calculated through sliding time-domain cross-correlation analysis, and the temporal components of the state transition determination point are aligned according to the relative temporal lag to obtain the quantized compensated state transition time.

[0065] The processor converts the quantized and compensated state transition moments into hard real-time interrupt data packets. It then writes these data packets into the external interrupt trigger register mapping area of ​​the alarm output controller via the bus interface, triggering the alarm driver program with the highest interrupt priority, which in turn drives the field alarm hardware to generate discrete response actions.

[0066] Example 4: This example combines Figures 1 to 2 A description of a nonlinear mapping pattern recognition method for extracting multidimensional deformation patterns of geological bodies, such as... Figure 1 As shown, the method includes several sequential steps, where step S1 is to subtract the displacement mean to generate a local spatial decoupling component, step S2 is to stack the decoupling components to generate a spatiotemporal evolution trajectory tensor, step S3 is to reduce the dimension through kernel mapping to output the deformation intrinsic trajectory matrix, step S4 is to calculate the Euclidean distance to solve the topological self-recursive Boolean matrix, and step S5 is to determine instability and output an alarm based on the index and curvature divergence.

[0067] like Figure 2 As shown, physical space observation points are set up on the slope of a large open-pit mine. These observation points are connected to multi-channel displacement sensors. The multi-channel displacement sensors transmit the generated three-dimensional cumulative displacement data stream to the multi-channel data interface inside the data processing unit. After the multi-channel data interface completes synchronous acquisition, it transfers the data to the computing unit. The computing unit writes the data down to the reference parameter configuration area of ​​the memory. At the same time, the computing unit generates a hard real-time interrupt data packet and sends it directly to the external interrupt trigger register mapping area in the alarm output controller through the bus interface. The external interrupt trigger register mapping area outputs discrete response actions to the external alarm hardware of the terminal based on the highest interrupt priority.

[0068] Example 5: Before the monitoring area is fully configured, the data processing unit performs initial state calibration. The processor reads the three-dimensional displacement static reference signals of 12 physical spatial observation points deployed in the monitoring area through a multi-channel data interface, and continuously records the observation data under static conditions for 24 hours. For the displacement components in the three directions of each observation point, the arithmetic mean and standard deviation are calculated respectively to obtain the mean drift vector. and standard deviation vector .

[0069] The processor constructs three-standard-deviation confidence intervals for each displacement component: The processor will shift the mean vector Standard deviation vector The baseline parameter configuration area, along with its three-standard-deviation confidence interval, is written to the memory. During subsequent geological deformation pattern recognition, the processor reads... The input data is processed in real time by subtracting the corresponding component of the mean drift vector from the input displacement component in the corresponding direction of each observation point. The three-standard-deviation confidence interval is used to check whether the input data after offset processing is still within the background fluctuation range under static conditions. After this processing, the pseudo displacement components caused by sensor hardware temperature drift and stress release of the mounting base are deducted from the input data. The input data under different geological structure scenarios adopt the same static benchmark.

[0070] During operation, the processor uses a 168-hour sliding update window to continuously monitor the fluctuation characteristics of the multidimensional cumulative displacement data stream. The processor first calculates the first-order difference sequence of the displacement signal within the current sliding update window, then divides the data into several continuous intervals according to the difference amplitude, counts the frequency of samples appearing in each amplitude interval, and converts the frequency into the corresponding probability. Subsequently, it accumulates each probability and its base-2 logarithm product, taking the opposite value to obtain the displacement signal fluctuation entropy. .

[0071] When the displacement signal fluctuation entropy When the performance degradation threshold of 1.2 bits is continuously exceeded, the processor determines that the environmental disturbance characteristics of the current monitoring area have deviated from the original monitoring benchmark. This change may be caused by seasonal changes in surface vegetation cover, groundwater circulation, or other long-term environmental conditions. The processor then initiates model reconstruction, clears the old version monitoring benchmark data in the cache that exceeds the 168-hour sliding update window, and refits the self-recursive closed loop attractor with the real-time observation data collected in the current window.

[0072] Updating the topology adaptive tolerance threshold At that time, the processor uses the Euclidean distance between each pair of intrinsic trajectory vector points within the current sliding update window as input and adjusts the algorithm using the least squares method. This allows the local topological autorecursive Boolean matrix obtained by binarizing according to the threshold to... Matching the autorecursive distribution of the current health stage, the updated topology adaptive tolerance threshold is written back to the baseline parameter configuration area as the decision boundary for subsequent local topology autorecursive Boolean matrices. Thus, when the soil and rock stress state and environmental disturbance characteristics in the monitoring area undergo long-term changes, the processor uses recent observation data to update the monitoring baseline, reducing the decrease in pattern recognition accuracy and false alarms caused by model aging.

[0073] Example 6: Before deploying the monitoring terminal to a steep slope, the data processing unit first performs offline benchmark parameter calibration to correct the influence of sensor temperature drift and structural installation deviations on the original three-dimensional displacement signal. The operator places the monitoring terminal on a horizontal benchmark platform and continuously collects 24 hours of static working condition data, and then processes the data... The arithmetic mean and standard deviation of the displacement components in each direction at each observation point are calculated to obtain the mean drift vector. with standard deviation vector .

[0074] The processor constructs an initial error compensation matrix based on the mean shift vector. : The processor will initialize the error compensation matrix. The data is written to the register mapping area of ​​the monitoring terminal, and the processor reads it when acquiring the displacement data stream of the observation point. and the original sensor displacement input amount Point-by-point compensation was performed to obtain the processed deformation observation slices. : because equal The opposite of the above operation is equivalent to subtracting the mean drift vector point by point from the original sensor displacement input. After offline calibration, the output of each sensing unit is unified to the same physical reference. The systematic deviation caused by individual hardware differences is deducted from the input data. The compensated deformation observation slice is used for subsequent spatiotemporal feature tensor construction and nonlinear mapping processing.

[0075] After the monitoring terminal is put into operation, the processor performs online reference drift calibration, and the processor sets a sliding update window with a length of 168 hours. And calculate in real time the sampling times within the time window. Mean of displacement eigenvector ,when It maintains the same direction of change during consecutive sampling times, and its offset relative to the current reference offset exceeds the drift tolerance threshold. At that time, the processor initiates online benchmark drift calibration.

[0076] When updating the baseline offset, the processor calculates the mean of the current displacement feature vector. offset from the old version Perform low-pass exponential smoothing to obtain the updated baseline offset. : ,in, The forgetting factor is set to 0.05 in this embodiment, and the processor uses... The original baseline value is replaced and used as a constant in the subsequent deformation feature reconstruction. As a result, the slow baseline offset caused by long-period creep and environmental changes is gradually deducted from the displacement features, while the local changes caused by critical accelerated slip are preserved.

[0077] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit of this application and the scope of protection of this invention, and all of these forms are within the protection scope of this application.

Claims

1. A nonlinear mapping pattern recognition method for extracting multidimensional deformation patterns of geological bodies, characterized in that, Includes the following steps: Step S1: Obtain the displacement data stream of the observation point, divide the observation point into spatial nearest neighbor feature bodies according to the spatial relative distance, subtract the mean displacement of the measurement point within the spatial nearest neighbor feature bodies to reduce common mode noise, and generate local spatial decoupled deformation components. Step S2: Within the data monitoring window, the local spatial decoupled deformation components are stacked along the time axis to construct a spatiotemporal feature tensor. After time delay reconstruction, the topological extension of the phase space is completed, and the spatiotemporal evolution trajectory tensor is generated. Step S3: Project the spatiotemporal evolution trajectory tensor onto the regenerated kernel space using Gaussian kernel mapping to generate a kernel feature matrix. Reduce the dimensionality of the kernel feature matrix based on the local preservation projection criterion to output a multidimensional deformation eigentrajectory matrix. Step S4: Extract a local sliding monitoring window from the multidimensional deformation intrinsic trajectory matrix, calculate the Euclidean distance between each vector point pair within the local sliding monitoring window to construct a symmetric topological distance matrix, and use 1.5 times the median of the Euclidean distance between point pairs in the historical health stage as an adaptive tolerance threshold to binarize the symmetric topological distance matrix and solve for the local topological autorecursive Boolean matrix. Step S5: Calculate the proportion of non-zero elements in the upper triangular region of the local topological self-recursive Boolean matrix to obtain the topological self-recursion rate index. When the topological self-recursion rate index is continuously lower than the abrupt damping threshold of 0.15 and the time-limited integral value of the manifold curvature of the multidimensional deformation intrinsic trajectory matrix continues to diverge and exceeds the set threshold, the state transition of the geological body to the irreversible slip instability mode is determined, and an instability alarm signal is output.

2. The nonlinear mapping pattern recognition method for extracting multidimensional deformation laws of geological bodies according to claim 1, characterized in that, Step S1 specifically includes the following sub-steps: Step S11, synchronously collect observation point displacement data streams from multiple independent sources using a multi-channel data interface; Step S12, perform group clustering on the multi-source observation points according to the geometric distance of the physical space dimension, and divide spatially adjacent sampling nodes with common mode association into the same spatial nearest neighbor feature body; Step S13, calculate the spatial mean of the data of each sampling node within the spatial nearest neighbor feature body as the mean of the measurement point displacement, and eliminate the common mode displacement disturbance noise caused by the long-period environmental physical field on spatially adjacent measurement points by subtracting the mean of the measurement point displacement from the observation point displacement data stream, thereby generating a local spatial decoupled deformation component.

3. The nonlinear mapping pattern recognition method for extracting multidimensional deformation laws of geological bodies according to claim 1, characterized in that, Step S1 also includes a data self-repair step triggered when missing data is detected: calculating the first-order differential gradient difference between the immediately preceding and following intact time-series data of the missing interval; determining the reciprocal of the squared term of the first-order differential gradient difference plus 1 as the dynamic penalty weight factor, and introducing the dynamic penalty weight factor as a regularization constraint term into the reconstruction loss function with the goal of minimizing the nuclear norm; releasing the smoothing constraint to fill in random packet loss in the stationary phase; and relaxing the smoothness constraint by increasing the dynamic penalty weight factor in the accelerated sliding critical phase, thus maintaining the local slope change inflection point characteristics in data reconstruction.

4. The nonlinear mapping pattern recognition method for extracting multidimensional deformation laws of geological bodies according to claim 1, characterized in that, Step S3 specifically includes the following sub-steps: Step S31, using a Gaussian kernel function to perform nonlinear similarity operations on the spatiotemporal evolution trajectory tensor to construct a global kernel matrix; Step S32, based on the adjacency degree of adjacent state samples in the original feature space, setting a local topology-preserving adjacency weight matrix in the regenerated kernel space, where for non-adjacent samples, the corresponding weight value is set to 0; Step S33, by solving the generalized eigenvalue decomposition equations about the global kernel matrix and the local topology-preserving adjacency weight matrix, selecting the eigenvectors corresponding to the smallest non-zero eigenvalues ​​of the preceding terms to form a low-dimensional projection matrix, reducing the high-dimensional feature linear projection to a low-dimensional intrinsic manifold space, and generating a multidimensional deformable intrinsic trajectory matrix.

5. The nonlinear mapping pattern recognition method for extracting multidimensional deformation laws of geological bodies according to claim 1, characterized in that, In step S5, the state of the closed loop attractor during normal dynamic fluctuations is characterized by the topological self-recursion rate index: when the multidimensional deformation intrinsic trajectory matrix is ​​shifted due to environmental alternation, the deformation dynamic evolution trajectory within the local sliding monitoring window is confined within the self-recursive closed loop attractor in the low-dimensional intrinsic manifold space; under this state, the symmetric topological distance matrix within the local sliding monitoring window exhibits a self-recursive distribution, and the topological self-recursion rate index is stably maintained in the range of not less than 0.90 to eliminate interference signals generated by environmental alternation.

6. The nonlinear mapping pattern recognition method for extracting multidimensional deformation laws of geological bodies according to claim 1, characterized in that, Step S5 specifically includes the following sub-steps: Step S51, fitting the multidimensional deformation intrinsic trajectory matrix into a continuous dynamic trajectory curve in the low-dimensional intrinsic manifold space, and solving the differential geometric local manifold curvature in real time along the tangent direction of the dynamic trajectory curve; Step S52, performing a sliding time-limited numerical integration on the local manifold curvature within a set sliding time limit to obtain the time-limited integral value of the manifold curvature; Step S53, when the topological self-recursion rate index is continuously lower than 0.15, and the time-limited integral value of the manifold curvature breaks the steady-state fluctuation threshold and continues to diverge beyond the set threshold, confirming the state transition to the irreversible slip instability mode.

7. The nonlinear mapping pattern recognition method for extracting multidimensional deformation laws of geological bodies according to claim 1, characterized in that, Before outputting the instability alarm signal, the process also includes a step of quantifying and compensating for the phase lag deviation of the acquisition: the sliding cross-correlation function between each sampling channel is calculated by cross-correlation analysis in the sliding time domain, and the time offset corresponding to the maximum value is extracted as the relative phase lag between each channel; the phase components of the state transition determination point are aligned based on the relative phase lag, the system identification delay caused by the asynchronous transmission delay of multiple channels is eliminated, and the quantification compensation for the state transition time is completed.

8. The nonlinear mapping pattern recognition method for extracting multidimensional deformation laws of geological bodies according to claim 1, characterized in that, The output instability alarm signal is specifically the underlying mapping steps that drive the alarm hardware to generate discrete response actions: the state transition time after quantization compensation is converted into a hard real-time interrupt data packet; the hard real-time interrupt data packet is directly written into the external interrupt trigger register mapping area of ​​the alarm output controller through the bus interface, and the track is forced to the highest interrupt priority to drive the alarm hardware to generate discrete response actions.

9. The nonlinear mapping pattern recognition method for extracting multidimensional deformation laws of geological bodies according to claim 1, characterized in that, The method is applied to the monitoring and early warning of geological landslides on the slopes of large open-pit mines or the banks of reservoirs of major water conservancy projects. When high-frequency environmental disturbances and sensor temperature drift occur in the acquired multi-channel displacement sensor signals, the data processing unit reduces the common-mode displacement disturbance noise on adjacent measuring points by decoupling the deformation components in the local space. Under the premise of using the topological self-recursion rate index to offset long-period environmental drift, the bidirectional causality check is performed by combining the topological self-recursion rate index being continuously lower than 0.15 with the manifold curvature time-limit integral value exceeding the divergence threshold. The state transition time after quantization compensation is used as the start time of the forward sliding segment and input to the alarm output controller to trigger the response action of the millisecond-level external alarm hardware.

Images