Rock deformation prediction method based on machine learning and array displacement meter data

Abstract

The present application relates to the technical field of machine learning, in particular to a surrounding rock deformation prediction method based on machine learning and array displacement meter data. The method obtains displacement time series formed by array displacement meters at different depth measuring points, carries out time alignment and incremental processing on the displacement time series, and generates measuring point state sequences containing cumulative displacement, displacement increment, deformation rate, deformation acceleration and displacement gradient of adjacent measuring points; constructs a space-time deformation graph according to the depth sequence relationship of the measuring points and the lag correlation relationship between the measuring points, inputs the space-time deformation graph into a machine learning model, and generates a surrounding rock state hidden vector through graph feature coding and time sequence feature coding. The present application constrains the prediction process by time evolution relationship, depth proximity relationship and lag propagation relationship, reduces the trend disconnection and gradient distortion of the prediction results of multiple measuring points in the depth direction, and improves the consistency of the surrounding rock deformation prediction results and the actual deformation transmission process.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and more specifically to a method for predicting surrounding rock deformation based on machine learning and array-type displacement gauge data. Background Technology

[0002] Surrounding rock deformation monitoring typically relies on array-type displacement gauges to acquire displacement change data at different depths. Multiple gauges are arranged along the borehole depth, with each gauge outputting displacement values ​​at a fixed sampling period. The monitoring system then generates a displacement time series based on this time sequence. Conventional processing methods often involve time alignment, outlier removal, smoothing filtering, and cumulative displacement calculation of the raw displacement data. The data is then displayed and judged based on single-gauge curves, differences between adjacent gauges, or cumulative deformation within a specified depth range. While this method can reflect the magnitude of deformation at different depths within the surrounding rock, its primary processing object remains a single gauge or a combination of local gauges. The depth sequence relationship between gauges usually exists only as a data identifier and does not participate in the structural construction of the prediction model.

[0003] In surrounding rock deformation prediction, existing technologies often employ regression models, time series models, or machine learning models to extrapolate historical displacement curves. Specifically, the displacement sequence formed at a given measurement point over consecutive sampling times is typically used as input, and the displacement values ​​at one or more future times are used as output. The prediction model is then trained using historical samples. Some schemes introduce derived features such as displacement increment, deformation rate, and deformation acceleration to enhance the model's ability to express deformation trends. The modeling focus of these schemes is curve fitting in the time dimension; the model can learn the historical pattern of a single measurement point changing from slow to rapid changes. However, they typically lack inherent constraints regarding deformation transfer, response lag, and abrupt changes in local gradients between measurement points at different depths within the same array.

[0004] Some surrounding rock monitoring and prediction schemes combine data from multiple depth measuring points into the same machine learning model. Common methods include concatenating the displacement sequences of each measuring point in depth order to form a feature vector, or using the cumulative displacement, displacement increment, and deformation rate of multiple measuring points as tabular features input into the model. Compared to single-point prediction, this type of scheme can utilize more monitoring data, but the correlation between different measuring points is mainly fitted by the model itself in high-dimensional features, lacking a graph structure representation that matches the depth spacing between measuring points, the displacement gradient of adjacent measuring points, and the hysteresis correlation between measuring points. When multiple measuring point data are input side by side, the physical proximity relationship between deep and shallow measuring points and the direction of deformation propagation are easily weakened, and the prediction results may show that the curves of each measuring point are reasonable, but the overall depth direction is inconsistent.

[0005] The main technical problem with existing technologies is that multi-depth displacement data collected by array-type displacement gauges simultaneously contains temporal evolution information and spatial coupling information. However, conventional prediction methods are mostly based on extrapolation of single-point sequences or splicing of features from multiple measurement points. They fail to construct a unified spatiotemporal deformation relationship that can be learned by machine learning models, considering the depth sequence of measurement points, the displacement gradient of adjacent measurement points, and the lag correlation between measurement points. As a result, when local deformation of the surrounding rock extends to adjacent depths, the model struggles to maintain the continuity and consistency of the predicted displacement in the depth direction. When local measurement points show accelerated deformation first, or when there is a response lag between deep and shallow measurement points, the prediction results are prone to problems such as disconnection of trends between adjacent measurement points, distortion of gradient changes, and insufficient representation of future deformation transmission paths. Summary of the Invention

[0006] The purpose of this invention is to provide a method for predicting surrounding rock deformation based on machine learning and array-type displacement gauge data, which can effectively solve the problems in the background art mentioned above.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for predicting surrounding rock deformation based on machine learning and array displacement gauge data includes: acquiring displacement time series formed by array displacement gauges at different depths, performing time alignment and incremental processing on the displacement time series, and generating a measurement point state sequence containing cumulative displacement, displacement increment, deformation rate, deformation acceleration, and displacement gradient of adjacent measurement points. A spatiotemporal deformation map is constructed based on the depth sequence relationship and the hysteresis correlation relationship between the measurement points at different depths. The hysteresis correlation relationship is determined by calculating the cross-correlation peak value of the displacement increment sequence under multiple hysteresis steps to determine the response hysteresis, propagation direction, and dynamic edge weight. The spatiotemporal deformation map includes the basic edge determined by the depth sequence relationship and the dynamic edge determined by the dynamic edge weight. The spatiotemporal deformation map is updated as the sampling window slides, and the neighborhood consistency of the dynamic edge is checked during the update. The spatiotemporal deformation diagram is input into a machine learning model, and the latent vector of the surrounding rock state is generated by graph feature encoding and temporal feature encoding. Based on the implicit vector of the surrounding rock state, the predicted displacement values ​​of each measuring point and the predicted gradient values ​​between measuring points at future times are output; and when the spatiotemporal deformation map is input into the machine learning model for training, displacement prediction loss and continuity constraint loss are used, and the continuity constraint loss is constructed based on the relationship between the predicted displacement difference and the evolution of historical displacement gradient.

[0008] Preferably, the construction of the spatiotemporal deformation graph includes: using the state sequence of the measurement point corresponding to each depth measurement point at the same sampling time as the graph node attribute; generating basic edges based on the physical proximity relationship between adjacent depth measurement points; generating dynamic edge weights based on the hysteresis cross-correlation peak values ​​between displacement increment sequences of different depth measurement points; and combining the basic edges and the dynamic edge weights into a graph adjacency matrix, wherein the graph adjacency matrix is ​​updated as the sampling window slides.

[0009] Preferably, the graph feature encoding includes: performing neighborhood aggregation on the node attributes of each depth measurement point in the spatiotemporal deformation graph, wherein the displacement gradient of adjacent measurement points, the hysteresis correlation edge weight, and the depth spacing encoding are used as the aggregation weight input during neighborhood aggregation to obtain the spatial coupling feature corresponding to each depth measurement point, and mapping the spatial coupling feature to a spatial feature sequence corresponding to the displacement time series according to the depth order of the measurement points.

[0010] Preferably, the temporal feature encoding includes: gating recursive encoding of the state sequence of each depth measurement point within a continuous sampling window to generate the temporal hidden state of the measurement point; A depth position attention weight is applied to the time latent state of the measuring points at different depths, and the depth position attention weight is fused with the spatial coupling feature obtained by the graph feature encoding to generate the latent vector of the surrounding rock state.

[0011] Preferably, the generation of the hysteresis cross-correlation peak specifically includes: calculating the correlation values ​​of the displacement increment sequences of any two depth measurement points under multiple hysteresis steps within a preset historical sampling window, selecting the hysteresis step where the absolute value of the correlation value satisfies the local peak condition as the response hysteresis between measurement points, and encoding the response hysteresis, the corresponding correlation value, and the depth direction of the measurement point together as the edge attribute of the dynamic edge weight.

[0012] Preferably, the sliding update of the graph adjacency matrix specifically includes: comparing the differences of dynamic edge weights formed by two consecutive sampling windows, and marking edges whose differences exceed the stable range of edge weights as candidate edges for deformation propagation; Perform neighborhood consistency verification on the depth measurement points associated with the candidate edges of the deformation propagation, and write the edges that pass the verification into the graph adjacency matrix of the current sampling window.

[0013] Preferably, the neighborhood aggregation includes: dividing each depth measurement point into a shallow neighborhood and a deep neighborhood according to the depth direction; calculating displacement gradient aggregation terms and hysteresis correlation aggregation terms for the shallow neighborhood and the deep neighborhood respectively; and then inputting the two types of aggregation terms into a nonlinear mapping layer with shared weights to obtain the spatial coupling feature that distinguishes the difference between the deep and shallow responses.

[0014] Preferably, the determination of the depth position attention weights includes: inputting the temporal hidden state of each depth measurement point, the depth position encoding, the displacement gradient of adjacent measurement points, and the corresponding dynamic edge weights into the attention scoring function to obtain the contribution score of each depth measurement point in the current sampling window, and normalizing the contribution score to form the depth position attention weights used to fuse the spatial coupling features.

[0015] Preferably, the training of the machine learning model specifically includes: establishing a displacement prediction loss using the measured future displacement values ​​of each depth measurement point as supervision labels; constructing a continuity constraint loss using the predicted displacement difference between adjacent depth measurement points and the evolution relationship of historical displacement gradients; constructing a multi-step consistency constraint loss using the deformation rate change between multiple consecutive prediction steps; and using the displacement prediction loss, the continuity constraint loss, and the multi-step consistency constraint loss together for parameter updates.

[0016] Preferably, the parameter update includes: generating a side weight stability coefficient based on the change magnitude of the dynamic side weight in consecutive training batches, adding the side weight stability coefficient to the weight calculation of the continuity constraint loss, and performing a joint reverse update on the graph feature encoding parameters and temporal feature encoding parameters of the associated depth measurement points when the change direction of the predicted displacement difference is inconsistent with that of the displacement gradient of the adjacent measurement points.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. The proposed surrounding rock deformation prediction method transforms different depth measuring points of an array-type displacement gauge into nodes in a spatiotemporal deformation graph. It converts depth sequence relationships, displacement gradients of adjacent measuring points, and hysteresis correlations between measuring points into edges and edge weights in the graph, transforming multi-measuring point displacement data from parallel features into machine learning inputs with deep correlations. Graph feature encoding aggregates the spatial coupling relationships of adjacent measuring points, while temporal feature encoding models the displacement increment, deformation rate, and deformation acceleration within a continuous sampling window. The fusion results in a latent vector representing the surrounding rock state. Therefore, future displacement prediction no longer relies solely on the historical curve of a single measuring point but is simultaneously constrained by the gradients of adjacent measuring points, depth spacing, and hysteresis response relationships during the prediction process. This ensures that the predicted displacement and gradient prediction values ​​of each depth measuring point maintain a continuous relationship within the same array, reducing the possibility of correct local curve fitting but distorted depth direction relationships.

[0018] 2. The proposed surrounding rock deformation prediction method also updates the graph adjacency matrix through a sliding sampling window, allowing the dynamic edge weights between measuring points to adjust according to the hysteresis correlation changes in the displacement increment sequence, adapting to the changes in the surrounding rock deformation propagation relationship at different stages. During neighborhood aggregation, shallow neighborhood, deep neighborhood, displacement gradient aggregation term, and hysteresis correlation aggregation term are respectively involved in the formation of spatial coupling features. The depth position attention weight is then combined with the temporal latent state of the measuring point, depth position encoding, and dynamic edge weights to complete feature fusion, enabling the contribution of measuring points at different depths to the prediction results to change with the current deformation state. During training, displacement prediction loss, continuity constraint loss, and multi-step consistency constraint loss are introduced, so that the model parameter update is simultaneously constrained by the future measured displacement values, the difference relationship between adjacent measuring points, and the rate change between multiple prediction step sizes, thereby enhancing the consistency of the prediction sequence in the time and depth directions. Attached Figure Description

[0019] Figure 1 This is an overall flowchart of the surrounding rock deformation prediction method based on machine learning and array-type displacement gauge data of the present invention; Figure 2 A flowchart for constructing the spatiotemporal deformation diagram of this invention; Figure 3 This is a flowchart of the graph feature encoding, temporal feature encoding, and surrounding rock state latent vector generation process of the present invention; Figure 4 This is a flowchart of the machine learning model training and parameter update process of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please refer to Figure 1This embodiment provides a method for predicting surrounding rock deformation based on machine learning and array-type displacement gauge data, including: acquiring displacement time series formed by array-type displacement gauges at different depths; performing time alignment and incremental processing on the displacement time series to generate a measurement point state sequence containing cumulative displacement, displacement increment, deformation rate, deformation acceleration, and displacement gradient of adjacent measurement points; constructing a spatiotemporal deformation map based on the depth sequence relationship of the measurement points at different depths and the hysteresis correlation relationship between the measurement points; wherein, the hysteresis correlation relationship is determined by calculating the cross-correlation peak value of the displacement increment sequence at multiple hysteresis steps to determine the response hysteresis, propagation direction, and dynamics. The spatiotemporal deformation graph includes basic edges determined by depth order relations and dynamic edges determined by the dynamic edge weights. The spatiotemporal deformation graph is updated as the sampling window slides, and neighborhood consistency checks are performed on the dynamic edges during the update. The spatiotemporal deformation graph is input into a machine learning model, and a latent vector of the surrounding rock state is generated by graph feature encoding and temporal feature encoding. Based on the latent vector of the surrounding rock state, the predicted displacement values ​​of each measuring point and the predicted gradient values ​​between measuring points are output at future times. Furthermore, during model training, displacement prediction loss and continuity constraint loss are used. The continuity constraint loss is constructed based on the evolution relationship between the predicted displacement difference and the historical displacement gradient. In practical implementation, the depth measuring points of the array displacement gauge form a set of measuring points according to the depth order within the same monitoring section. Each measuring point corresponds to a displacement time series. After receiving the displacement values ​​generated by each measuring point at consecutive sampling times, the system aligns missing, repeated, and asynchronous times using a unified time axis as a reference. The time-aligned data is not directly input into the model. Instead, the displacement increment is formed by the displacement difference between adjacent sampling times, the deformation rate is formed by the change of the displacement increment relative to the sampling interval, the deformation acceleration is formed by the change of adjacent deformation rates, and the displacement of adjacent depth measuring points within the same sampling time or the same sampling window is used as the basis for further analysis. The displacement differential is used to form the displacement gradient between adjacent measuring points, thereby converting the original displacement value into a measuring point state sequence that simultaneously includes cumulative deformation, short-term changes, and depth differences. The spatiotemporal deformation map regards each depth measuring point as a node of the bearing state sequence, and regards the proximity relationship formed by the depth sequence of measuring points and the lag correlation relationship of the displacement increment sequence between measuring points as edge relationships. In the graph feature encoding stage, the machine learning model aggregates the deformation information of adjacent measuring points along the edge relationships, and in the temporal feature encoding stage, it extracts the deformation evolution information of each measuring point along the time direction. Then, the two types of encoding results are mapped into the hidden vector of the surrounding rock state, and the predicted displacement value and the predicted gradient value between measuring points at future time are output.

[0022] In this embodiment, time alignment uses the set of sampling times as a unified index. If multiple readings exist at a certain index position for the same measurement point, the reading with the closest time distance and normal data status is used as the corresponding value of that index. If a measurement point lacks readings within a short time interval, a placeholder value is formed by the displacement change trend of adjacent times and the synchronous change trend of adjacent measurement points. The placeholder value participates in the sequence structure construction but can be given low weight in the training loss. The cumulative displacement uses the difference relative to the initial stable segment, the displacement increment uses the difference between the current displacement and the displacement of the previous alignment time, the deformation rate uses the ratio of the displacement increment to the sampling interval, and the deformation acceleration uses the adjacent... The ratio of the difference in deformation rate to the sampling interval, and the ratio of the displacement gradient of adjacent measuring points to the displacement difference of adjacent depth measuring points and the depth spacing encoding of two measuring points. The depth spacing encoding is not limited by hardware structural parameters, but is used as a scale factor formed on the data side by the known arrangement relationship of measuring points. The measuring point state sequence is arranged into a state matrix according to the depth direction of the measuring points. The rows of the state matrix correspond to the depth measuring points, the columns correspond to the sampling time within the time window, and the channels correspond to the cumulative displacement, displacement increment, deformation rate, deformation acceleration, and displacement gradient of adjacent measuring points. Subsequent graph encoding and temporal encoding extract features from the same state matrix to avoid establishing multiple sets of inconsistent data inputs separately.

[0023]

[0024]

[0025]

[0026]

[0027] ; in, This represents the state vector of the i-th depth measuring point at the t-th sampling time. Indicates cumulative displacement. Indicates the displacement increment. Indicates the deformation rate. Indicates deformation acceleration, This represents the displacement gradient between adjacent measuring points. This represents the time interval between adjacent sampling times. This represents the depth spacing encoding between the i-th depth measurement point and the (i+1)-th depth measurement point. and Increase in the same direction and When the value changes from near 0 to positive, the calculation result represents the state where the shallow measuring point has a stronger displacement change relative to the deep measuring point. When the value changes from negative to positive, the calculation result indicates that the deformation rate of the corresponding measuring point is accelerating. The above value is directly entered into the subsequent graph node attributes instead of being simplified to a single displacement value.

[0028] Further, refer to Figure 2 The construction of the spatiotemporal deformation graph includes: using the state sequence of each depth measuring point at the same sampling time as the graph node attribute; generating basic edges based on the physical proximity between adjacent depth measuring points; generating dynamic edge weights based on the hysteresis cross-correlation peak values ​​between displacement increment sequences of different depth measuring points; and combining the basic edges and the dynamic edge weights into a graph adjacency matrix. The graph adjacency matrix is ​​updated as the sampling window slides. In specific implementation, the basic edges are determined by the depth order of the measuring points, and undirected or directed basic edges are established between any adjacent depth measuring points. The basic edges are used to ensure the continuity of the surrounding rock along the depth direction. The deformation relationship is incorporated into the model. The dynamic edge weights are calculated by the hysteresis cross-correlation between the displacement increment sequences. They are used to express the degree of correlation between the displacement change of a certain measuring point and the displacement change of another measuring point after several sampling steps. When the basic edge and the dynamic edge weights are synthesized, the basic edge provides fixed topological support, and the dynamic edge weights provide the strength relationship that changes with the deformation stage of the surrounding rock. The synthesized graph adjacency matrix is ​​updated once in each sliding sampling window. The edge weights of the previous window will not directly cover the current window, but will be fused with the hysteresis correlation results calculated in the current window to preserve the continuous change trajectory of the deformation propagation relationship.

[0029] In a preferred embodiment, the establishment of basic edges does not require all measuring points to be connected pairwise. Instead, adjacent points in depth order are used as primary connections, and non-adjacent measuring points whose hysteresis correlation strength meets the set conditions are used as secondary connections. The primary connections are used to describe the continuous deformation constraints of adjacent measuring points, and the secondary connections are used to describe the synchronous or hysteresis response across depth regions. The node attributes adopt the measuring point state sequence within the current sampling window rather than the reading at a single moment. The edge attributes include at least the depth direction of the measuring points at both ends of the edge, the displacement increment hysteresis correlation value, the response hysteresis, and the gradient change direction. Each element of the graph adjacency matrix is ​​related to whether there is a basic connection between the two measuring points and the size of the dynamic edge weight. If the two measuring points are adjacent depth measuring points, the corresponding element retains the basic adjacency component. If the two measuring points have a stable hysteresis correlation peak within the historical sampling window, the corresponding element is superimposed with the dynamic adjacency component. If the two measuring points have no adjacency relationship and the hysteresis correlation peak does not meet the local peak condition, the corresponding element is set as a weak connection or not connected. Thus, the graph structure contains both fixed connections determined by the arrangement of the array displacement gauges and data connections derived from the surrounding rock deformation data itself.

[0030]

[0031] ; in, This represents the graph adjacency weight between the i-th depth measurement point and the j-th depth measurement point within the t-th sampling window. This indicates the values ​​of the base edges, determined by depth order. This indicates the hysteresis correlation strength of the displacement increment sequence. This represents the fusion coefficient between the weights of the basic edges and the dynamic edges. Indicates the candidate lag step size. Denotes the set of candidate lag steps. Indicates the current sampling window. and These represent the average displacement increments of the i-th and j-th measuring points within the window, respectively. and At this time, the adjacency weight includes both adjacent depth constraints and data correlation constraints. but At this time, the adjacency weight expresses the hysteresis response relationship derived from the data between non-adjacent measurement points. When the value is close to 0, non-adjacent measurement points will not form a strong connection in the current window.

[0032] Further, refer to Figure 3 The graph feature encoding includes: performing neighborhood aggregation on the node attributes of each depth measurement point in the spatiotemporal deformation graph. During neighborhood aggregation, the displacement gradient of adjacent measurement points, the hysteresis correlation weight, and the depth spacing encoding are used as the aggregation weight input to obtain the spatial coupling features corresponding to each depth measurement point. The spatial coupling features are then mapped to a spatial feature sequence corresponding to the displacement time series according to the depth order of the measurement points. In specific implementation, the graph feature encoding does not replace neighborhood aggregation with ordinary feature concatenation. Instead, it weights and merges the states of adjacent nodes of each node. The weights are formed by graph adjacency weights, displacement gradient directions, and depth spacing encoding. The displacement gradient is used to distinguish whether there is local deformation concentration between adjacent measurement points. The hysteresis correlation weight is used to distinguish whether there is delayed response between adjacent or non-adjacent measurement points. The depth spacing encoding is used to avoid imbalance in the expression of relationship strength caused by depth distance differences. The spatial coupling features obtained by neighborhood aggregation are rearranged according to the depth order of the measurement points so that subsequent time series feature encoding can receive a spatial feature sequence consistent with the depth order of the original displacement time series. The graph encoding result maintains a corresponding relationship with the original time series.

[0033] In this embodiment, the input of neighborhood aggregation consists of the node's own state, the states of neighboring nodes, and edge attributes. The node's own state is used to preserve the local deformation trend of the corresponding measurement point. The states of neighboring nodes are used to introduce deformation transmission information from the surrounding depth region. The edge attributes are used to adjust the contribution of neighboring nodes to the current node. In specific processing, a neighborhood set is established for each depth measurement point. The neighborhood set includes adjacent measurement points connected by basic edges and associated measurement points whose dynamic edge weights satisfy the preservation conditions. The model performs a linear mapping on the state vector of each node in the neighborhood set, and then multiplies it by the aggregation coefficient calculated by graph adjacency weight, gradient direction consistency, and depth spacing decay. All neighborhood mapping results are summed and synthesized with the node's own mapping results to form the spatial coupling feature of the current measurement point. If the displacement increment of a shallow measurement point changes before that of a deep measurement point and the lagging related edge weight is higher, then the shallow measurement point feature will have a larger input weight in the spatial coupling feature of the deep measurement point. If the displacement gradient direction of adjacent measurement points is opposite to the historical gradient direction, then the aggregation coefficient will be suppressed and the disturbance of abnormal gradients to spatial features will be reduced.

[0034]

[0035] ; in, This represents the spatial coupling characteristics of the i-th depth measurement point at the t-th sampling time. Represents a nonlinear mapping function. This represents the node's own state mapping matrix. This represents the neighborhood node state mapping matrix. This represents the neighborhood set of the i-th measurement point within the current window. This represents the weight when the j-th measurement point is aggregated to the i-th measurement point. This represents the gradient consistency adjustment coefficient. This indicates the consistency of the displacement gradient direction between the i-th and j-th measuring points. When the displacement gradient directions of the two measuring points are the same and... When the neighboring input is retained as a strong contributor, and the displacement gradient directions of the two measuring points are opposite, the neighboring input is retained as a strong contributor. At that time, the neighboring input is suppressed. This represents the depth spacing encoding between two measurement points. The larger the depth spacing, the weaker the neighborhood contribution under the same correlation strength.

[0036] Furthermore, the temporal feature encoding includes: performing gated recursive encoding on the state sequence of each depth measuring point within a continuous sampling window to generate a latent temporal state of the measuring point; applying depth position attention weights to the latent temporal states of measuring points at different depths, and fusing the depth position attention weights with the spatial coupling features obtained from the graph feature encoding to generate the latent state vector of the surrounding rock. In specific implementation, the gated recursive encoding reads the state sequence of the same measuring point along the time direction, and at each sampling moment, it inputs the cumulative displacement, displacement increment, deformation rate, deformation acceleration, and spatial coupling features. The recursive unit further... The new gate and candidate states retain historical deformation trends. The forget gate reduces the impact of earlier stable segments on the current acceleration segment. The time hidden state of the measurement point corresponds to the time evolution representation of each depth measurement point at the end of the current window. The depth position attention weight is calculated based on the time hidden state of the measurement point, the depth position encoding, the displacement gradient of adjacent measurement points, and the dynamic edge weight. The calculation result is used to adjust the contribution of each depth measurement point in the surrounding rock state hidden vector. The surrounding rock state hidden vector is formed by the weighted combination of the time hidden state of each measurement point and the spatial coupling features. The subsequent prediction layer outputs the future displacement prediction value of multiple measurement points and the gradient prediction value between measurement points based on the surrounding rock state hidden vector.

[0037] In a preferred embodiment, the temporal feature encoding uses a gated recursive unit with shared parameters to process measurement points at different depths, maintaining the same mapping scale for different measurement points in temporal evolution modeling. Depth position encoding is superimposed at the input to distinguish between shallow and deep measurement points. The input of the gated recursive unit is not limited to the state of the measurement point itself, but also includes the spatial coupling features generated by graph feature encoding, enabling the temporal recursion process to simultaneously read local and neighborhood changes. The depth position attention weight is calculated once within the current sampling window. After the weight calculation is completed, the temporal latent states of all measurement points are normalized and weighted, and the weighted temporal latent states are concatenated or added to the spatial coupling features of the corresponding depth measurement points to generate a surrounding rock state latent vector. The prediction layer performs multi-output mapping on the surrounding rock state latent vector, where the future displacement prediction value is arranged according to the depth measurement point, and the gradient prediction value between measurement points is arranged according to adjacent measurement point pairs. The two outputs are generated on the same latent vector, avoiding the use of independent feature sources for displacement prediction and gradient prediction.

[0038] Furthermore, the generation of the hysteresis cross-correlation peak includes: calculating the correlation values ​​of the displacement increment sequences of any two depth measuring points under multiple hysteresis steps within a preset historical sampling window; selecting the hysteresis step where the absolute value of the correlation value satisfies the local peak condition as the inter-measuring point response hysteresis; and encoding the response hysteresis, the corresponding correlation value, and the depth direction of the measuring point together as the edge attribute of the dynamic edge weight. In specific implementation, the preset historical sampling window slides forward over time, and the displacement increment sequences of any two measuring points are staggered and aligned according to the candidate hysteresis step, with each hysteresis step corresponding to a correlation value. If the absolute value of the correlation value of a certain lag step is greater than the absolute value of the correlation value of the adjacent candidate lag step, and the corresponding correlation direction is consistent with the deformation transmission direction in the current depth direction, then the corresponding lag step is used as the response lag. If multiple local peaks appear at the same time, the local peaks that are consistent with the basic edge direction and appear more frequently in the continuous sampling window are selected first. The response lag is used to describe the order of occurrence of changes between two measurement points, the corresponding correlation value is used to describe the strength relationship between the changes of the two measurement points, and the depth direction of the measurement point is used to determine the direction encoding of the edge attribute. The three types of information together form the edge attribute of the dynamic edge weight.

[0039] In this embodiment, the local peak condition can be determined by comparing adjacent lag step lengths and the consistency of continuous windows. Isolated related peaks appearing in a single window are not directly written into the dynamic graph structure, but are compared with candidate peaks of the same measurement point pair in the previous and next windows. If the lag step length corresponding to the candidate peak changes too much and the related direction switches repeatedly, the corresponding edge attribute is marked as unstable and the edge weight is reduced. If the lag step length corresponding to the candidate peak remains adjacent or the same value in the continuous window, the corresponding edge attribute is marked as stable lag state and retained. When the lag step length is positive, it means that the displacement increment of the first measurement point changes before the second measurement point. When the lag step length is negative, it means that the displacement increment of the second measurement point changes before the first measurement point. When the lag step length is 0, it means that the displacement increment of the two measurement points changes synchronously. The direction encoding in the edge attribute combines the above-mentioned sequential relationship with the depth order relationship, so that the model can distinguish different data relationships such as shallow to deep transmission, deep to shallow response, and local synchronous deformation.

[0040] Furthermore, the sliding update of the graph adjacency matrix includes: comparing the differences in dynamic edge weights formed by two consecutive sampling windows, marking edges whose differences exceed the stable range of edge weights as deformation propagation candidate edges; performing neighborhood consistency verification on the depth measurement points associated with the deformation propagation candidate edges, and writing the edges that pass the verification into the graph adjacency matrix of the current sampling window. In specific implementation, the dynamic edge weights of two consecutive sampling windows are compared using the same measurement point pair as the index. If the dynamic edge weight in the current window changes significantly compared to the previous window, it indicates that the displacement increment correlation between the corresponding measurement point pair may have changed in stages. The system does not immediately write it into the graph adjacency matrix as a true deformation propagation relationship, but marks the corresponding edge as a deformation propagation candidate edge, and then checks whether the adjacent measurement points at both ends of the candidate edge have the same or similar displacement increment direction, gradient change direction, and response hysteresis direction. If the neighborhood data of the measurement points at both ends of the candidate edge supports the change of the corresponding edge, the candidate edge is written into the graph adjacency matrix of the current sampling window. If the neighborhood data does not support it, the candidate edge only participates in the current window calculation as a low-weight edge or is filtered out.

[0041] ; in, This represents the change in the edge between the i-th and j-th measurement points within the current sampling window. This indicates the hysteresis correlation strength of the previous sampling window. and These represent the response lag between the current sampling window and the previous sampling window, respectively. This represents the lag adjustment coefficient, when... It changed from 0.3 to 0.75 and When the values ​​remain similar, the change in edge intensity is mainly contributed by the change in correlation intensity and expresses the enhanced relationship between the measurement points. Small changes but When the value changes from positive to negative, the change in edge quantity represents a shift in the order of responses. Candidate edges need to undergo neighborhood consistency verification before entering the graph adjacency matrix.

[0042] Preferably, the neighborhood consistency check can be completed by the shallow neighboring points, deep neighboring points, and historical edge attributes of the two ends of the candidate edge. Specifically, the two ends of the candidate edge are expanded into local neighborhood sets, and the proportion of points with consistent displacement increment direction, the proportion of points with consistent displacement gradient direction, and the proportion of edges with consistent response hysteresis direction are calculated. If the three proportions support the change direction of the candidate edge, the candidate edge is identified as conforming to the deformation propagation relationship of the current sampling window. If the candidate edge appears only on a single point pair and there is no corresponding change in other points in the neighborhood, the candidate edge is retained as a temporary edge and its weight in the graph feature encoding is reduced. Through this processing method, the update of the dynamic graph structure is based on the common data relationship between the current point pair and the surrounding point group, rather than being determined by the instantaneous change of a single correlation value.

[0043] Further, the neighborhood aggregation includes: dividing each depth measurement point into a shallow neighborhood and a deep neighborhood according to the depth direction; calculating displacement gradient aggregation terms and hysteresis correlation aggregation terms for the shallow neighborhood and the deep neighborhood respectively; and inputting the two types of aggregation terms into a nonlinear mapping layer with shared weights to obtain the spatial coupling feature that distinguishes the difference between shallow and deep responses. In specific implementation, with the current measurement point as the center, related measurement points with a depth less than the current measurement point constitute the shallow neighborhood, and related measurement points with a depth greater than the current measurement point constitute the deep neighborhood. The shallow and deep neighborhoods are not mixed in the calculation, but are aggregated separately according to the displacement gradient and hysteresis correlation edge weights. The displacement gradient aggregation term is formed by encoding the displacement gradient direction and gradient magnitude between the neighboring measurement point and the current measurement point, and the hysteresis correlation aggregation term is formed by the response hysteresis and correlation strength between the neighboring measurement point and the current measurement point. The aggregation terms in both directions are input into the same nonlinear mapping layer, so that the model distinguishes between shallow and deep responses while maintaining the same mapping rules. The output spatial coupling feature includes the deformation relationship of the current measurement point relative to the shallow and deep regions.

[0044] In a preferred embodiment, the shallow neighborhood aggregation term and the deep neighborhood aggregation term retain the direction sign respectively. When the displacement increment of the shallow measuring point changes first and the deep measuring point changes subsequently, the shallow hysteresis correlation aggregation term has a leading direction code. When the deep measuring point changes first and the shallow measuring point changes subsequently, the deep hysteresis correlation aggregation term has a reverse direction code. The displacement gradient aggregation term uses the sign and absolute value of the displacement difference between adjacent measuring points to participate in the mapping, avoiding the loss of the deformation direction between the shallow and deep parts by only expressing the deformation concentration with the gradient magnitude. The nonlinear mapping layer uses shared weights to process the shallow aggregation term, the deep aggregation term and the current measuring point's own state. After output, the depth position code is superimposed. The spatial coupling feature thus retains the relative position of the current measuring point in the array depth direction. When the subsequent temporal encoding reads this feature, it can distinguish the different association meanings of the same displacement increment in the shallow and deep measuring points.

[0045] Further, the determination of the depth position attention weights includes: inputting the temporal latent state of each depth measuring point, the depth position encoding, the displacement gradient of adjacent measuring points, and the corresponding dynamic edge weights into the attention scoring function to obtain the contribution score of each depth measuring point within the current sampling window, and normalizing the contribution scores to form the depth position attention weights used to fuse the spatial coupling features. In specific implementation, the temporal latent state of the measuring point provides the deformation evolution representation of the current measuring point in the time dimension, the depth position encoding provides the relative depth of the current measuring point in the array, the displacement gradient provides the degree of difference between the current measuring point and adjacent measuring points, the dynamic edge weights provide the hysteresis correlation strength between the current measuring point and other measuring points, and the attention scoring function maps the above inputs into scalar scores. The measuring point with a higher score occupies a larger weight in the surrounding rock state latent vector. The normalization process ensures that the weights of each measuring point within the same sampling window are at the same scale. The spatial coupling features are weighted and fused according to the depth position attention weights to form a unified state representation.

[0046] In this embodiment, the attention scoring function can adopt a combination of linear mapping and nonlinear activation. The input vector is formed by concatenating the temporal hidden state of the measuring point, the depth position encoding, the displacement gradient statistics of adjacent measuring points, and the dynamic edge weight statistics. The dynamic edge weight statistics include the average dynamic edge weight between the current measuring point and the shallow neighborhood, the average dynamic edge weight with the deep neighborhood, and the maximum hysteresis correlation edge weight. The displacement gradient statistics include the gradient between the current measuring point and the adjacent shallow measuring points, the gradient with the adjacent deep measuring points, and the gradient change within the current window. The depth position encoding uses a monotonic position vector to express the order of the measuring points. The normalization process uses exponential normalization or homogeneous normalization to convert the contribution scores of all measuring points into a weight set. The weight set is multiplied by the spatial coupling features generated by the graph feature encoding point by point and then summed. It is then fused with the temporal hidden state generated by the gated recursive unit to obtain the surrounding rock state hidden vector used for displacement prediction and gradient prediction.

[0047]

[0048] ; in, This represents the attention weight of the depth position of the i-th depth measurement point within the t-th sampling window. This indicates that the measurement point time is hidden. Indicates depth location encoding. This represents the statistical value of the displacement gradient between adjacent measuring points. Let represent the dynamic edge weight statistics associated with the i-th measurement point, P, Q, M, and N represent the mapping matrices for different input terms, and u represent the score vector. This represents a nonlinear function, where n represents the number of depth measurement points. This represents the implicit vector of the surrounding rock state, when a certain measuring point... The rate of expression deformation continues to increase and At higher levels, the corresponding score numerator increases and causes When a measurement point has only local fluctuations but weak adjacent gradients and dynamic edge weights, the corresponding weight decreases after normalization.

[0049] Further, refer to Figure 4 The training of the machine learning model includes: establishing a displacement prediction loss using the measured future displacement values ​​of each depth measurement point as supervision labels; establishing a continuity constraint loss using the predicted displacement difference between adjacent depth measurement points and the evolution relationship of historical displacement gradients; and establishing a multi-step consistency constraint loss using the deformation rate change between consecutive prediction steps. The displacement prediction loss, the continuity constraint loss, and the multi-step consistency constraint loss are used together for parameter updates. In specific implementation, the displacement prediction loss is used to constrain the deviation between the predicted displacement of each measurement point and the corresponding measured displacement; the continuity constraint loss is used to constrain the predicted displacement difference between adjacent depth measurement points from deviating from the displacement gradient evolution direction within the historical window; and the multi-step consistency constraint loss is used to constrain the deformation rate change between consecutive prediction steps to maintain consistency with the acceleration trend expressed by the time series encoding. The training samples consist of multiple sliding sampling windows. Each sample inputs the state sequence and spatiotemporal deformation map within the current window and outputs the multi-measurement point displacement values ​​and gradient values ​​at multiple future times. The model parameters are updated by backpropagation using the three types of losses, so that the graph feature encoding parameters, time series feature encoding parameters, and output mapping parameters are adjusted under the same training objective.

[0050] In a preferred embodiment, the continuity constraint loss does not use a fixed gradient threshold. Instead, a reference gradient term is constructed based on the mean direction and trend of the displacement gradients of adjacent measurement points within the historical sampling window. When the predicted displacement difference and the reference gradient term are in the same direction and the magnitude of change is within the historical evolution range, the continuity constraint loss is small. When the predicted displacement difference and the reference gradient term are in opposite directions or there is an abrupt reversal between adjacent measurement points, the continuity constraint loss increases. The multi-step consistency constraint loss is based on the displacement difference between adjacent step sizes within the predicted sequence to construct the prediction rate. Then, the consistency between the direction of change of the prediction rate and the direction of deformation acceleration at the end of the historical window is compared. If the end of the historical window expresses an accelerating change while the predicted sequence suddenly turns into a decelerating change, the loss term increases. If the end of the historical window expresses a stable change while the predicted sequence produces an unprovoked jump, the loss term also increases. This allows the training process to simultaneously cover future displacement values, depth direction difference relationships, and time direction rate relationships.

[0051]

[0052]

[0053] ; in, Indicates joint training losses, Indicates the displacement prediction loss. This represents the loss due to continuity constraints. This represents the loss due to multi-step consistency constraints. This represents the weight of the continuity constraint loss. This represents the weight of the multi-step consistency constraint loss. This represents the predicted displacement of the i-th measuring point at the r-th prediction step in the future. This represents the corresponding measured displacement, and m represents the number of prediction step sizes. This represents the reference displacement gradient of adjacent measuring points obtained from the historical window, when... and When the directions are consistent, The value is relatively low, and when the two directions are opposite, Increase and drive the model to adjust the prediction difference between adjacent measurement points.

[0054] Further, the parameter update includes: generating a side weight stability coefficient based on the change magnitude of the dynamic side weights in consecutive training batches; adding the side weight stability coefficient to the weight calculation of the continuity constraint loss; and performing joint reverse update on the graph feature encoding parameters and temporal feature encoding parameters of the associated depth measurement points when the change direction of the predicted displacement difference is inconsistent with that of the displacement gradient of the adjacent measurement points. In specific implementation, the side weight stability coefficient is calculated from the dynamic side weight fluctuation of the same measurement point pair in consecutive training batches or consecutive sampling windows. When the dynamic side weight fluctuation is small and the direction is stable, the side weight stability coefficient is high, and the continuity constraint loss weight of the corresponding measurement point pair increases. When the dynamic side weight fluctuation is large or the response lag direction changes frequently, the side weight stability coefficient decreases, and the continuity constraint loss weight of the corresponding measurement point pair decreases. When the change direction of the predicted displacement difference is inconsistent with that of the historical adjacent measurement point displacement gradient, the training process not only updates the output layer parameters, but also propagates backward along the graph aggregation path and time recursion path related to the measurement point pair. The graph feature encoding parameters and temporal feature encoding parameters of the associated depth measurement points are jointly adjusted.

[0055] In this embodiment, the joint reverse update uses associated depth measurement points as the local update range. Associated depth measurement points include two adjacent measurement points with inconsistent predicted displacement difference directions, adjacent measurement points connected to the two measurement points by a basic edge, and associated measurement points with stable dynamic edge weights connected to the two measurement points. The update sources of graph feature encoding parameters include neighborhood aggregation weights, edge attribute mapping matrices, and shallow and deep aggregation mapping matrices. The update sources of temporal feature encoding parameters include state mapping parameters of gated recursive units, depth position attention scoring parameters, and surrounding rock state latent vector fusion parameters. If the dynamic edge weight stability coefficient of a certain measurement point pair is high and the predicted difference direction is repeatedly opposite to the reference gradient, then the continuity constraint forms a strong gradient input for the measurement point pair. The model will adjust the contribution of the corresponding edge in the neighborhood aggregation and the weight of the corresponding measurement point in the temporal encoding. If the dynamic edge weight stability coefficient is low, the gradient input of the continuity constraint for the measurement point pair is reduced, and the model relies more on displacement prediction loss and multi-step consistency constraints to complete the update.

[0056] In a specific implementation scenario, historical monitoring data is organized into continuous sampling windows. Each window contains a multi-depth displacement sequence with several consecutive sampling times. Several sampling times after the end of the window serve as supervision labels. During the training phase, the state matrix, graph adjacency matrix, edge attribute matrix, and depth position encoding within the window are input into the machine learning model. The graph feature encoding layer outputs the spatial coupling features of each measuring point at each sampling time. The temporal feature encoding layer reads the spatial coupling features and the measuring point state vector in chronological order and outputs the temporal hidden state of the measuring point. The depth position attention layer merges the temporal hidden states of all measuring points into the surrounding rock state hidden vector. The output layer generates a multi-step predicted displacement matrix and a gradient prediction matrix. The inference phase uses the same data processing and graph construction method, but no longer receives future measured labels. The prediction results are output in the order of depth measuring points, and the gradient prediction values ​​are output in the order of adjacent measuring point pairs. The output data maintains a fixed mapping relationship with the input array displacement gauge measuring points.

[0057] In another preferred embodiment, the method can be applied in incremental training or periodic retraining. After new array displacement gauge monitoring data enters the system, the state matrix is ​​first updated according to the existing measurement point depth order, and then the dynamic edge weights and edge weight stability coefficients are calculated according to the new sliding sampling window. If the graph adjacency matrix of the current window changes little compared to the historical stable window, the model parameters maintain the original training results and the prediction is performed directly. If the edge change of multiple measurement point pairs in the current window passes the neighborhood consistency check, the corresponding window is added to the retraining sample set. The continuity constraint weights in the joint loss are redistributed according to the edge weight stability coefficients. The parameter update still propagates along the graph feature encoding and temporal feature encoding. The incremental training process does not change the measurement point arrangement relationship and output dimension, so that the model can absorb new deformation propagation data while maintaining the original depth mapping relationship.

[0058] In this embodiment, the gradient prediction values ​​between measurement points output by the model can be directly calculated from the predicted displacement matrix, or generated by the gradient branch set in the output layer. If the direct calculation method is used, the future gradient prediction value is obtained by dividing the difference of the future displacement prediction values ​​of adjacent measurement points by the corresponding depth spacing encoding. If the gradient branch method is used, the gradient branch and the displacement branch share the surrounding rock state latent vector and are subject to continuity constraint loss during the training phase. Consistency constraints can also be set between the gradient branch output value and the predicted displacement difference to ensure that the displacement prediction results and gradient prediction results within the same model do not contradict each other. During inference output, the future displacement prediction value is used to express the deformation trend of each depth measurement point, and the gradient prediction value is used to express the difference change trend between adjacent depth measurement points. Both types of outputs come from the same spatiotemporal deformation map and the same surrounding rock state latent vector.

[0059] In terms of data storage and computation organization, the method represents the data of each sampling window as three types of tensors: the node state tensor is used to store the state sequence of each depth measurement point within the window; the edge attribute tensor is used to store the basic edge, dynamic edge weight, response hysteresis, correlation value, and depth direction encoding; and the label tensor is used to store the measured displacement at multiple future sampling times and the reference gradient calculated from the measured displacement. During training, samples are read according to the window time order and adjacent windows are allowed to share some historical sampling data. The graph adjacency matrix is ​​recalculated as the window slides. The graph feature encoding and the temporal feature encoding share the same window input, avoiding repeated interpolation or alignment between the graph model and the temporal model. When the model is saved, the depth order of the measurement points, the channel order of the state vector, and the order of the output measurement points are recorded so that the results can be generated according to the same mapping relationship in the subsequent inference stage.

[0060] In terms of abnormal data processing, the method does not rely on additional hardware structures, nor does it replace the prediction model with a fixed alarm threshold. If a single measuring point exhibits isolated reading fluctuations in the current window, the reading source marker is retained during the time alignment stage, and the displacement increment and deformation rate are still calculated in the state sequence. During the graph construction stage, the corresponding relationship between fluctuations is determined by hysteresis cross-correlation and neighborhood consistency verification. If the fluctuation does not form a stable dynamic edge weight, the aggregation contribution of the corresponding measuring point pair in the graph feature encoding is reduced. If the fluctuation forms a displacement gradient change and response hysteresis relationship with the same direction in multiple adjacent measuring points, the dynamic graph structure retains the corresponding edge attributes, and the subsequent time-series encoding and training loss can read this change. Therefore, the model adopts different data processing paths for the propagation of local deformation of the surrounding rock and single-point reading fluctuations.

[0061] During the prediction process after model deployment, the system receives the displacement data of the array displacement gauges within the latest sliding sampling window, generates a sequence of measurement point states according to a predetermined time index, establishes basic edges according to the depth order of the measurement points, calculates the dynamic edge weights and response hysteresis of the current window according to the displacement increment sequence, writes the basic edges, dynamic edge weights, and edge attributes into the graph adjacency matrix, and then inputs the node states and the graph adjacency matrix into the trained machine learning model. Graph feature encoding generates spatial coupling features, temporal feature encoding generates the temporal hidden states of the measurement points, and the depth position attention layer generates the hidden vector of the surrounding rock state. The output layer outputs the predicted displacement values ​​of each measurement point and the predicted gradient values ​​of adjacent measurement points at multiple future sampling times at once. All outputs are arranged according to the original depth measurement point order of the array displacement gauges, which facilitates coaxial comparison with subsequent monitoring data.

[0062] In a preferred extended embodiment, the method can also perform intra-model consistency verification on the prediction results. The verification objects include the future displacement prediction matrix, the gradient prediction matrix, and the multi-step deformation rate matrix. If the adjacent measurement point difference calculated by the displacement prediction matrix is ​​inconsistent with the direction of the gradient prediction matrix, the corresponding measurement point pair is marked as having inconsistent prediction relationship, and the continuity constraint participation of the corresponding sample is increased during the retraining stage. If the rate change between continuous prediction steps conflicts with the acceleration direction expressed by the time hidden state, the multi-step consistency constraint participation of the corresponding measurement point is increased. The consistency verification does not change the input structure of the trained model, nor does it introduce new sensing objects. Instead, it forms a closed-loop data constraint between the existing displacement prediction value, gradient prediction value, graph adjacency matrix, and measurement point state sequence.

[0063] In this embodiment, the machine learning model can be implemented using a combination of graph convolutional coding layers, gated recursive coding layers, depth position attention layers, and multi-output prediction layers. Alternatively, it can be implemented using graph message passing coding layers and temporal attention coding layers with the same input-output relationship. The replacement of the specific model structure does not change the core data processing path. The core data processing path always includes the state representation of the multi-depth displacement time series of array displacement gauges, the construction of spatiotemporal deformation graphs based on depth order and hysteresis correlation, the generation of latent vectors of surrounding rock state based on graph feature coding and temporal feature coding, training and updating based on joint loss and edge weight stability coefficients, and prediction of future displacement and gradient based on the same latent vector. Any implementation that uses the above data relationships and training constraints to complete the surrounding rock deformation prediction can correspond to the technical route disclosed in this embodiment.

[0064] Specifically, when the surrounding rock is in a stable deformation stage, the displacement increment of each measuring point is small, the dynamic edge weight is mainly maintained by the foundation edge, the graph adjacency matrix tends to be stable, and the spatial coupling characteristics reflect the continuous relationship between adjacent depth measuring points more. The temporal feature encoding output shows a slow change in the temporal hidden state of the measuring points. When a local measuring point begins to show accelerated deformation, the displacement increment, deformation rate, and deformation acceleration of that measuring point change in the state sequence. If adjacent measuring points subsequently show related changes, the hysteresis cross-correlation peak will correspond to the measuring point pair and be written into the dynamic edge weight. The graph feature encoding introduces this propagation relationship in the neighborhood aggregation, and the temporal feature encoding retains the change trend in the subsequent window. The future displacement and gradient prediction values ​​in the prediction output are thus constrained by both the changes of the local measuring point itself and the response relationship of adjacent measuring points.

[0065] During data training, sample partitioning is based on temporal continuity. Earlier historical windows are used for initial model parameter training, while subsequent windows are used to verify the difference between the predicted output and the measured labels. Training samples must not be randomly shuffled to the point of disrupting the temporal order. Sliding windows are allowed to overlap to maintain the continuity of deformation evolution. Labels are generated using the future sampling time after the end of the window as the prediction target. The reference gradient is calculated from the future measured displacement difference. In each round of model training, the graph adjacency matrix corresponding to the current window is reread, instead of using a fixed adjacency matrix across all samples. After the parameters are updated, the historical change records of displacement prediction loss, continuity constraint loss, and multi-step consistency constraint loss are retained to determine whether the training only fits the displacement values ​​while ignoring the depth relationship.

[0066] In a preferred embodiment, if some depth measurement points of the array displacement gauge have short-term missing measurements within a certain window, the node state of the missing measurement point can be generated into a temporary state vector by combining the state of the same measurement point before and after the previous time and the current state of adjacent depth measurement points. The temporary state vector carries a data source label and participates in graph construction. When calculating dynamic edge weights, the displacement increment corresponding to the missing measurement position is not used as the main basis for the correlation peak. When encoding graph features, the missing measurement point still retains the basic edge input. When encoding temporal features, the gated recursive unit retains the historical trend through the hidden state of the previous time. In the training loss, the missing measurement position is not used as the actual measurement label for displacement prediction. If the label is complete in the future, the missing measurement window can still be used as a training sample to participate in model learning, so that the model maintains the consistency of the input structure when the actual monitoring data of the array displacement gauge is incomplete.

Claims

1. A method for predicting surrounding rock deformation based on machine learning and array-type displacement gauge data, characterized in that, include: The displacement time series formed by array displacement gauges at different depths is obtained, and the displacement time series is time aligned and incrementally processed to generate a measurement point state sequence containing cumulative displacement, displacement increment, deformation rate, deformation acceleration and displacement gradient of adjacent measurement points. A spatiotemporal deformation map is constructed based on the depth sequence relationship and the hysteresis correlation relationship between the measurement points at different depths. The hysteresis correlation relationship is determined by calculating the cross-correlation peak value of the displacement increment sequence under multiple hysteresis steps to determine the response hysteresis, propagation direction, and dynamic edge weight. The spatiotemporal deformation map includes the basic edge determined by the depth sequence relationship and the dynamic edge determined by the dynamic edge weight. The spatiotemporal deformation map is updated as the sampling window slides, and the neighborhood consistency of the dynamic edge is checked during the update. The spatiotemporal deformation diagram is input into a machine learning model, and the latent vector of the surrounding rock state is generated by graph feature encoding and temporal feature encoding. Based on the implicit vector of the surrounding rock state, the predicted displacement values ​​of each measuring point and the predicted gradient values ​​between measuring points at future times are output; and when the spatiotemporal deformation map is input into the machine learning model for training, displacement prediction loss and continuity constraint loss are used, and the continuity constraint loss is constructed based on the relationship between the predicted displacement difference and the evolution of historical displacement gradient.

2. The method for predicting surrounding rock deformation based on machine learning and array-type displacement gauge data according to claim 1, characterized in that, The construction of the spatiotemporal deformation graph includes: using the state sequence of the measurement point corresponding to each depth measurement point at the same sampling time as the graph node attribute; generating basic edges based on the physical proximity relationship between adjacent depth measurement points; generating dynamic edge weights based on the hysteresis cross-correlation peak values ​​between displacement increment sequences of different depth measurement points; and combining the basic edges and the dynamic edge weights into a graph adjacency matrix, which is updated as the sampling window slides.

3. The method for predicting surrounding rock deformation based on machine learning and array-type displacement gauge data according to claim 1, characterized in that, The graph feature encoding includes: performing neighborhood aggregation on the node attributes of each depth measurement point in the spatiotemporal deformation graph. During neighborhood aggregation, the displacement gradient of adjacent measurement points, the hysteresis correlation edge weight, and the depth spacing encoding are used as the aggregation weight input to obtain the spatial coupling feature corresponding to each depth measurement point. The spatial coupling feature is then mapped to a spatial feature sequence corresponding to the displacement time series according to the depth order of the measurement points.

4. The method for predicting surrounding rock deformation based on machine learning and array-type displacement gauge data according to claim 1, characterized in that, The temporal feature encoding includes: gating recursive encoding of the state sequence of each depth measurement point within a continuous sampling window to generate the temporal hidden state of the measurement point; A depth position attention weight is applied to the time latent state of the measuring points at different depths, and the depth position attention weight is fused with the spatial coupling feature obtained by the graph feature encoding to generate the latent vector of the surrounding rock state.

5. The method for predicting surrounding rock deformation based on machine learning and array-type displacement gauge data according to claim 2, characterized in that, The generation of the hysteresis cross-correlation peak specifically includes: calculating the correlation values ​​of the displacement increment sequences of any two depth measurement points under multiple hysteresis steps within a preset historical sampling window, selecting the hysteresis step where the absolute value of the correlation value satisfies the local peak condition as the response hysteresis between measurement points, and encoding the response hysteresis, the corresponding correlation value, and the depth direction of the measurement point together as the edge attribute of the dynamic edge weight.

6. The method for predicting surrounding rock deformation based on machine learning and array-type displacement gauge data according to claim 2, characterized in that, The sliding update of the graph adjacency matrix specifically includes: comparing the differences of dynamic edge weights formed by two consecutive sampling windows, and marking edges whose differences exceed the stable range of edge weights as candidate edges for deformation propagation; Perform neighborhood consistency verification on the depth measurement points associated with the candidate edges of the deformation propagation, and write the edges that pass the verification into the graph adjacency matrix of the current sampling window.

7. The method for predicting surrounding rock deformation based on machine learning and array-type displacement gauge data according to claim 3, characterized in that, The neighborhood aggregation includes: dividing each depth measurement point into a shallow neighborhood and a deep neighborhood according to the depth direction; calculating displacement gradient aggregation terms and hysteresis correlation aggregation terms for the shallow neighborhood and the deep neighborhood respectively; and then inputting the two types of aggregation terms into a nonlinear mapping layer with shared weights to obtain the spatial coupling feature that distinguishes the difference between the deep and shallow responses.

8. The method for predicting surrounding rock deformation based on machine learning and array-type displacement gauge data according to claim 4, characterized in that, The determination of the depth position attention weights includes: inputting the temporal hidden state of each depth measurement point, the depth position encoding, the displacement gradient of adjacent measurement points, and the corresponding dynamic edge weights into the attention scoring function to obtain the contribution score of each depth measurement point in the current sampling window, and then normalizing the contribution score to form the depth position attention weights used to fuse the spatial coupling features.

9. The method for predicting surrounding rock deformation based on machine learning and array-type displacement gauge data according to claim 8, characterized in that, The training of the machine learning model specifically includes: establishing a displacement prediction loss using the measured future displacement values ​​of each depth measurement point as supervision labels; constructing a continuity constraint loss using the predicted displacement difference between adjacent depth measurement points and the evolution relationship of historical displacement gradients; constructing a multi-step consistency constraint loss using the deformation rate change between multiple consecutive prediction steps; and using the displacement prediction loss, the continuity constraint loss, and the multi-step consistency constraint loss together for parameter updates.

10. The method for predicting surrounding rock deformation based on machine learning and array-type displacement gauge data according to claim 9, characterized in that, The parameter update includes: generating a side weight stability coefficient based on the change magnitude of the dynamic side weight in consecutive training batches, adding the side weight stability coefficient to the weight calculation of the continuity constraint loss, and performing a joint reverse update on the graph feature encoding parameters and temporal feature encoding parameters of the associated depth measurement points when the change direction of the predicted displacement difference is inconsistent with that of the displacement gradient of the adjacent measurement points.

Images