A surface rock movement disaster early warning method based on multi-modal monitoring data

Abstract

The application discloses a kind of surface rock movement disaster early warning methods based on multi-modal monitoring data, comprising: obtaining multi-modal monitoring data and preprocessing, mapping to rock movement influence unit;Dual-phase crack image registration, improve PSPNet to extract abnormal area, form image modal abnormal evolution field;Extract multi-modal abnormal information, structure projection generates abnormal transmission relationship, constructs structure abnormal field;Based on structure abnormal field, extract connected relationship and propagation path, construct rock movement precursor topological skeleton;Identify critical bridging unit, judge continuous through path, determine fission channel trigger state;Construct phase change state variable with hysteresis memory, determine rock movement evolution stage, output early warning result.The application extracts crack abnormal area by improving PSPNet network and carries out structured topological analysis in combination with multi-modal monitoring data, realizes the early identification and stable early warning of surface rock movement disaster from discrete anomaly to through instability process.

Description

Technical Field

[0001] This invention relates to the field of geological disaster monitoring and early warning technology, and in particular to a method for early warning of surface rock movement disasters based on multimodal monitoring data. Background Technology

[0002] Surface rock movement hazards are widespread in mining areas, mountainous construction areas, and regions with complex geological conditions, mainly manifesting as surface subsidence, crack propagation, and local instability. To achieve effective early warning for such hazards, existing technologies typically employ multi-source monitoring to acquire information such as surface displacement, deep displacement, remote sensing deformation, and environmental triggering data, combined with image monitoring technology to obtain crack development data. In terms of data processing, threshold judgment, time series analysis, or simple data fusion methods are often used to analyze various monitoring data to achieve early warning and judgment of rock movement hazards.

[0003] However, existing technologies still have shortcomings in utilizing multimodal data. Most methods analyze monitoring data from different sources independently or perform simple overlay processing, lacking in-depth exploration of the spatial correlations and structural evolution characteristics between multimodal data. In particular, they struggle to accurately depict the spatial connectivity between crack propagation, areal deformation, and displacement anomalies. While image data can be used to identify crack regions, it typically only operates at the single-temporal segmentation level, failing to effectively combine multi-temporal information to reflect the evolutionary process of anomalous regions. Consequently, the identification of disaster precursors remains at the level of local anomalies.

[0004] Existing early warning methods are mostly based on instantaneous monitoring data or short-term trends, lacking a continuous description of the abnormal evolution process. They are difficult to identify the key stage from discrete anomalies to overall connectivity instability, and are prone to problems such as delayed early warning, false alarms, or missed alarms.

[0005] Therefore, how to provide a method for early warning of surface rock movement disasters based on multimodal monitoring data is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a method for early warning of surface rock movement disasters based on multimodal monitoring data. This invention introduces an improved PSPNet network to segment abnormal regions in crack images and constructs a structural anomaly field by combining multimodal data. It further reconstructs the topological skeleton of rock movement precursors, identifies critical bridging units, and constructs rock movement phase transition state variables with hysteresis memory. This enables early identification and dynamic early warning of the evolution of surface rock movement disasters from discrete anomalies to through instability. It has the advantages of high early warning accuracy, low false alarm rate, and strong result stability.

[0007] A method for early warning of surface rock movement disasters based on multimodal monitoring data according to an embodiment of the present invention includes:

[0008] Acquire multimodal monitoring data within the monitoring area, preprocess the multimodal monitoring data, and map various types of monitoring data to pre-divided rock movement influence units;

[0009] After registering the crack image data of the current time with the previous time, the data is input into the improved PSPNet network to extract the anomalous regions of each time phase, generate the temporal evolution results of the anomalous regions, and map them to the corresponding rock movement influence units to form an image modal anomalous evolution field.

[0010] Extract the unit anomaly intensity, anomaly direction, and anomaly duration corresponding to each modal data. Combine the image modal anomaly evolution field with the spatial adjacency relationship, geological continuity relationship, and consistency of anomaly propagation direction between rock movement-affected units to perform structural projection, obtain the anomaly transmission relationship, and generate the structural anomaly field.

[0011] Based on the structural anomaly field, the abnormal connectivity, abnormal propagation path, abnormal bifurcation structure and non-connected sections between each rock movement influence unit are extracted to construct the topological skeleton of rock movement precursors.

[0012] In the topological framework of rock movement precursors, rock movement influence units located between different anomalous branches and satisfying the conditions of consistent propagation direction, continuous geological structure, and enhanced local anomalousness are identified as critical bridging units. Based on the critical bridging units, it is determined whether a continuous through path is formed, thereby determining the triggering state of the fission channel.

[0013] Based on the structural changes of the topological framework of rock movement precursors, the activity level of critical bridging units, and the triggering state of fission channels, a rock movement phase transition state variable with hysteresis memory is constructed, and the current rock movement evolution stage is determined. According to the rock movement evolution stage, the corresponding surface rock movement disaster early warning results are output.

[0014] Optionally, the multimodal monitoring data includes surface displacement data, deep displacement data, remote sensing deformation data, remote sensing image data, and environmental triggering data.

[0015] Optionally, the preprocessing of multimodal monitoring data, mapping various types of monitoring data to pre-divided rock movement influence units, includes:

[0016] The multimodal monitoring data is processed for time alignment and spatial coordinate unification. Abnormal data is denoised and smoothed. Rock movement influence units are divided according to the spatial range of the monitoring area. Multimodal monitoring data are mapped to the corresponding rock movement influence units according to the spatial location of each monitoring data.

[0017] Optionally, the step of generating the temporal evolution results of the anomalous region and mapping them to the corresponding rock movement influence units to form an image modal anomaly evolution field includes:

[0018] Acquire the crack image data at the current moment and the crack image data at the previous moment. Perform denoising, grayscale normalization and spatial registration on the crack image data at the two moments in sequence to obtain the registered image at the current moment and the registered image at the previous moment. Establish the correspondence between the image pixel region and the rock movement influence unit according to the spatial range of the rock movement influence unit.

[0019] An improved PSPNet network is constructed, comprising a dual-temporal input layer, a high-resolution coding backbone layer, a differential guided fusion layer, a boundary refinement layer, a pyramid pooling layer, and a segmentation output layer, wherein:

[0020] The high-resolution coding backbone layer adopts the HRNet-W18 architecture. The first input branch receives the registered image at the current time step, and the second input branch receives the registered image at the previous time step. The two input branches are respectively processed by the first stage convolutional layer and the second stage high-resolution parallel convolutional layer to extract shallow texture features.

[0021] After the high-resolution parallel convolutional layer in the third stage, a difference-guided fusion layer is set up to concatenate the current time-time features, the previous time-time features and their element-wise difference features through channel splicing and then compressing them through a one-to-one convolutional layer to output the fused features.

[0022] A boundary refinement layer is set after the high-resolution parallel convolutional layer in the fourth stage. The boundary refinement layer is composed of a 3x3 convolutional layer, a dilated convolutional layer, and a 1x1 convolutional layer connected in sequence.

[0023] The pyramid pooling layer is connected after the boundary refinement layer, and uses four different pooling scales to perform contextual aggregation on the fused features;

[0024] The segmentation output layer is connected to the pyramid pooling layer and outputs the anomaly region segmentation map at the current time and the anomaly region segmentation map at the previous time, respectively.

[0025] A training sample set is constructed, which includes crack image samples at the current time, crack image samples at the previous time, and corresponding abnormal region annotation maps. The training sample set is input into the improved PSPNet network, and the improved PSPNet network is trained using a joint loss function that includes pixel-level segmentation constraints, boundary constraints, and dual temporal consistency constraints, to obtain the trained improved PSPNet network.

[0026] The current and previous registered images are input into the trained improved PSPNet network, which outputs the current and previous abnormal region segmentation maps. The network extracts the boundary expansion, area increase / decrease, connectivity changes, and adjacent region expansion of the abnormal regions to form the temporal evolution results of the abnormal regions.

[0027] The temporal evolution results of the anomalous region are mapped to the corresponding rock movement influence unit according to the spatial location of the anomalous region. The coverage range, expansion direction, connectivity changes and continuous activity of the anomalous region in each rock movement influence unit are statistically analyzed to generate the image modal anomalous evolution value corresponding to each rock movement influence unit, thereby forming the image modal anomalous evolution field.

[0028] Optionally, obtaining the anomaly propagation relationship and generating the structural anomaly field includes:

[0029] Unit-level anomaly extraction was performed on surface displacement data, deep displacement data, remote sensing deformation data, and environmental triggering data to obtain the anomaly intensity, anomaly direction, and anomaly duration of each rock movement influence unit, and the image modal anomaly evolution value of each rock movement influence unit was obtained.

[0030] The abnormal intensity, abnormal direction, abnormal duration, and image modal abnormal evolution value of each rock movement-affected unit are uniformly characterized to form the current abnormal state information of each rock movement-affected unit.

[0031] Centered on each rock movement-affected unit, a candidate transmission unit set is established according to the spatially adjacent units. Based on the anomaly expansion direction in the current anomaly state information, the candidate transmission unit set is filtered by direction. Adjacent units located in the anomaly expansion forward region are retained as effective transmission units. The target units located in the anomaly edge expansion zone and the anomaly non-penetrating gap zone are identified by combining the image modal anomaly evolution value.

[0032] For each rock movement influence unit, the abnormal transmission value from the rock movement influence unit to each effective transmission unit is calculated. The abnormal transmission value is determined based on the abnormal intensity, abnormal duration, consistency between the abnormal expansion direction and the relative position of the target unit, the edge expansion activity corresponding to the image modality abnormal evolution value, and whether the target unit is located within the abnormal non-penetrating gap zone.

[0033] Anomaly transmission relationships are established based on the anomalous transmission values ​​between each rock movement influence unit and its effective transmission unit. The current anomalous state information of each rock movement influence unit is then fused with the corresponding anomalous transmission relationships to generate a structural anomalous field.

[0034] Optionally, the construction of the rock movement precursor topological framework includes:

[0035] Rock movement influence units that reach the preset conditions in the structural anomaly field are extracted as explicit anomaly units, and rock movement influence units located in the anomaly edge extension zone or the anomaly non-connected gap zone and have an abnormal continuous change state are extracted as potential extension units.

[0036] Based on the anomalous transmission relationship between obvious anomalous units, obvious anomalous connected regions are constructed. In each obvious anomalous connected region, the main propagation segment is extracted according to the anomalous transmission direction and the order of anomalous transmission strength. The main propagation segment is composed of continuous adjacent rock movement influence units with consistent anomalous transmission direction.

[0037] Based on the abnormal transmission relationship between potential extension units and adjacent explicit abnormal connected regions, potential extension paths located at the edge of explicit abnormal connected regions and distributed along the extension direction of the main propagation segment are extracted. The potential extension paths are determined as segments to be connected, and the connection positions between each segment to be connected and the corresponding explicit abnormal connected region are recorded.

[0038] Branch propagation paths in each visible anomaly connected region are identified. When the same rock movement influence unit is connected to two or more anomaly transmission paths at the same time, the rock movement influence unit is identified as a bifurcation unit. A hierarchical connection relationship is established by combining the main propagation segment, the segment to be connected and the bifurcation unit. The visible anomaly connected region and the main propagation segment constitute the visible skeleton, and the segment to be connected constitutes the hidden extension skeleton.

[0039] The explicit framework, implicit extension framework, bifurcation units, and unconnected sections are integrated to construct the topological framework of rock movement precursors.

[0040] Optionally, the step of determining whether a continuous through-path is formed based on critical bridging units, thereby determining the fission channel triggering state, includes:

[0041] Obtain the topological skeleton of rock movement precursors, extract the visible anomaly connected regions and their corresponding main propagation segments, and determine the unconnected segments and potential extension paths between adjacent visible anomaly connected regions.

[0042] In the unconnected sections and potential extension paths, rock movement influence units that simultaneously meet the following requirements—being consistent with the main propagation section of the adjacent visible anomaly connected area, having an anomaly transmission relationship with the adjacent anomaly unit, and having a continuous anomaly change state—are selected as candidate bridging units.

[0043] For each candidate bridging unit, the corresponding anomaly intensity value, anomaly duration value, and image modal anomaly evolution value are obtained. Combined with the spatial positional relationship and anomaly propagation relationship with the adjacent explicit anomaly connected regions, the bridging potential value of the candidate bridging unit is determined. The bridging potential value is jointly determined by the anomaly intensity, anomaly duration, the consistency of anomaly propagation direction, and the degree of edge expansion activity of the anomaly region.

[0044] Based on the bridging potential value and spatial distribution relationship of each candidate bridging unit, rock movement influence units that meet the preset conditions and form a continuous connection relationship in space are selected as critical bridging units, and the critical bridging units are associated with the corresponding obvious anomaly connected areas.

[0045] Based on the connection relationship between the critical bridging unit and the obvious abnormal connected region, it is determined whether there is a continuous through path consisting of the obvious abnormal connected region, the main propagation segment and the critical bridging unit. When a continuous through path exists, the fission channel is determined to be in the triggered state. When no continuous through path exists, the fission channel is determined to be in the untriggered state.

[0046] Optionally, the construction of rock movement phase transition state variables with hysteresis memory, the determination of the current rock movement evolution stage, and the output of corresponding surface rock movement disaster early warning results based on the rock movement evolution stage include:

[0047] Obtain the topological skeleton of rock movement precursors, as well as the triggering states of critical bridging units and fission channels. Extract the skeleton connectivity changes, main propagation segment extension changes, unconnected segment contraction changes, and critical bridging unit activity changes between the current and previous moments.

[0048] Based on changes in skeleton connectivity, extension of the main propagation segment, contraction of the segment to be connected, active changes of critical bridging units, and fission channel triggering state, a rock migration facies change state variable is constructed, which includes instantaneous evolution values ​​and historical memory values. The historical memory values ​​are updated cumulatively according to continuous time intervals, and the already formed connectivity extension state is maintained.

[0049] When the skeleton expands and strengthens, the bridging unit remains active, or the fission channel changes from non-triggered to triggered at the current moment, increase the historical memory value; when the current moment shows an abnormal decline but the section to be connected has not recovered or the fission channel is still in the triggered state, maintain the historical memory value; when the skeleton shrinks, the bridging unit becomes inactive, and the fission channel is in the non-triggered state for multiple consecutive moments, decrease the historical memory value.

[0050] The current rock movement evolution stage is determined based on the real-time evolution value and the historical memory value. The rock movement evolution stage includes the stable stage, the gestation stage, the penetration stage, and the instability stage. When the historical memory value reaches the corresponding stage condition, the rock movement evolution stage is increased. When the historical memory value does not drop to the corresponding release condition, the current rock movement evolution stage remains unchanged.

[0051] Based on the current stage of rock movement evolution, the corresponding surface rock movement disaster early warning results are output. The stable stage outputs normal monitoring results, the incubation stage outputs low-level early warning results, the penetration stage outputs medium-level early warning results, and the unstable stage outputs high-level early warning results.

[0052] The beneficial effects of this invention are:

[0053] This invention addresses the problem in existing technologies where multimodal monitoring data is only subjected to independent threshold judgment or simple fusion analysis, making it difficult to identify the evolution of anomalies from discrete distributions to spatial connections. By introducing an image modal anomaly evolution field and a structural projection mechanism for multimodal data, this invention effectively realizes the unified expression and correlation modeling of different modal anomaly information in spatial units. This allows anomalies to no longer be limited to single-point or single-source data, but to reflect their spatial propagation trends and connectivity, thereby improving the ability to identify precursors of structural instability.

[0054] This invention elevates anomaly analysis from traditional numerical judgment to structural evolutionary analysis by constructing a topological framework of rock movement precursors, identifying critical bridging units, and determining the triggering state of fission channels. It can accurately depict the key transition process from dispersion to connection between anomalous areas, making it more sensitive and discriminative to the key triggering conditions before the occurrence of surface rock movement disasters, and effectively reducing early warning lag and missed reporting.

[0055] This invention further constructs a rock movement phase change state variable with hysteresis memory to continuously model the abnormal evolution process over time. Even under abnormal short-term fluctuations or local declines, it can still maintain a stable judgment on the overall risk status, avoid frequent fluctuations in the warning level, and achieve smooth output and phased control of the warning results. Attached Figure Description

[0056] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0057] Figure 1 This is a flowchart of a surface rock movement disaster early warning method based on multimodal monitoring data proposed in this invention;

[0058] Figure 2 This is a schematic diagram of the dual-temporal crack image abnormality region extraction structure based on the improved PSPNet network, which is a surface rock movement disaster early warning method based on multimodal monitoring data proposed in this invention. Detailed Implementation

[0059] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0060] refer to Figure 1 and Figure 2 A method for early warning of surface rock movement disasters based on multimodal monitoring data, comprising:

[0061] Acquire multimodal monitoring data within the monitoring area, preprocess the multimodal monitoring data, and map various types of monitoring data to pre-divided rock movement influence units;

[0062] After registering the crack image data of the current time with the previous time, the data is input into the improved PSPNet network to extract the anomalous regions of each time phase, generate the temporal evolution results of the anomalous regions, and map them to the corresponding rock movement influence units to form an image modal anomalous evolution field.

[0063] Extract the unit anomaly intensity, anomaly direction, and anomaly duration corresponding to each modal data. Combine the image modal anomaly evolution field with the spatial adjacency relationship, geological continuity relationship, and consistency of anomaly propagation direction between rock movement-affected units to perform structural projection, obtain the anomaly transmission relationship, and generate the structural anomaly field.

[0064] Based on the structural anomaly field, the abnormal connectivity, abnormal propagation path, abnormal bifurcation structure and non-connected sections between each rock movement influence unit are extracted to construct the topological skeleton of rock movement precursors.

[0065] In the topological framework of rock movement precursors, rock movement influence units located between different anomalous branches and satisfying the conditions of consistent propagation direction, continuous geological structure, and enhanced local anomalousness are identified as critical bridging units. Based on the critical bridging units, it is determined whether a continuous through path is formed, thereby determining the triggering state of the fission channel.

[0066] Based on the structural changes of the topological framework of rock movement precursors, the activity level of critical bridging units, and the triggering state of fission channels, a rock movement phase transition state variable with hysteresis memory is constructed, and the current rock movement evolution stage is determined. According to the rock movement evolution stage, the corresponding surface rock movement disaster early warning results are output.

[0067] In this embodiment, the multimodal monitoring data includes surface displacement data, deep displacement data, remote sensing deformation data, remote sensing image data, and environmental triggering data.

[0068] In this embodiment, the preprocessing of multimodal monitoring data, mapping various types of monitoring data to pre-divided rock movement influence units, includes:

[0069] The multimodal monitoring data is processed for time alignment and spatial coordinate unification. Abnormal data is denoised and smoothed. Rock movement influence units are divided according to the spatial range of the monitoring area. Multimodal monitoring data are mapped to the corresponding rock movement influence units according to the spatial location of each monitoring data.

[0070] In this embodiment, the step of generating the temporal evolution results of the anomalous region and mapping them to the corresponding rock movement influence unit to form an image modal anomaly evolution field includes:

[0071] Acquire the crack image data at the current moment and the crack image data at the previous moment. Perform denoising, grayscale normalization and spatial registration on the crack image data at the two moments in sequence to obtain the registered image at the current moment and the registered image at the previous moment. Establish the correspondence between the image pixel region and the rock movement influence unit according to the spatial range of the rock movement influence unit.

[0072] An improved PSPNet network is constructed, comprising a dual-temporal input layer, a high-resolution coding backbone layer, a differential guided fusion layer, a boundary refinement layer, a pyramid pooling layer, and a segmentation output layer, wherein:

[0073] The high-resolution coding backbone layer adopts the HRNet-W18 architecture. The first input branch receives the registered image at the current time step, and the second input branch receives the registered image at the previous time step. The two input branches are respectively processed by the first stage convolutional layer and the second stage high-resolution parallel convolutional layer to extract shallow texture features.

[0074] After the high-resolution parallel convolutional layer in the third stage, a difference-guided fusion layer is set up to concatenate the current time-time features, the previous time-time features, and the element-wise difference features through channel splicing and then compressing them through a one-to-one convolutional layer to output the fused features.

[0075] A boundary refinement layer is set after the high-resolution parallel convolutional layer in the fourth stage. The boundary refinement layer is composed of a 3x3 convolutional layer, a dilated convolutional layer, and a 1x1 convolutional layer connected in sequence. The boundary refinement layer extracts and smooths local neighborhood features through the 3x3 convolutional layer to enhance the continuous representation of the crack region. Then, the dilated convolutional layer expands the receptive range without reducing the resolution, enhancing the response to the extension direction of the crack edge and the narrow anomaly band. Finally, the 1x1 convolutional layer linearly combines and compresses the multi-channel features to highlight the boundary region features and suppress background noise, resulting in a clear boundary and directional continuity representation of the anomaly region features.

[0076] The pyramid pooling layer is connected after the boundary refinement layer. It uses four different pooling scales to perform context aggregation on the fused features. After receiving the feature map output by the boundary refinement layer, the pyramid pooling layer downsamples the features through four pooling branches of different scales to obtain context information in different spatial ranges. The pooling results of each scale are then mapped by the corresponding convolutional layer, and then upsampled to restore the original feature size. The results are then concatenated and fused with the original features. This process introduces multi-scale global context information while preserving detailed information, thereby enhancing the ability to express large-scale abnormal regions and their spatial distribution relationships.

[0077] The segmentation output layer is connected after the pyramid pooling layer, and outputs the current time-of-flight abnormal region segmentation map and the previous time-of-flight abnormal region segmentation map respectively. After receiving the multi-scale fusion features output by the pyramid pooling layer, the segmentation output layer compresses the channel dimension through a 1-to-1 convolutional layer and generates a classification feature map. The features corresponding to the current time-of-flight and the previous time-of-flight are processed through independent branches. Each branch goes through convolution mapping and upsampling operations in sequence to restore the feature map to the original image size. The category of each pixel is determined by a normalization function, and the current time-of-flight abnormal region segmentation map and the previous time-of-flight abnormal region segmentation map are obtained respectively.

[0078] A training sample set is constructed, including crack image samples from the current time step, crack image samples from the previous time step, and corresponding anomaly region annotation maps. This training sample set is then input into the improved PSPNet network. The improved PSPNet network is trained using a joint loss function that includes pixel-level segmentation constraints, boundary constraints, and bi-temporal consistency constraints, resulting in the trained improved PSPNet network.

[0079] Pixel-level segmentation constraints are calculated by comparing the current and previous abnormal region segmentation results obtained from the segmentation output layer with the corresponding abnormal region annotation maps pixel by pixel. The difference between the predicted category and the labeled category at each pixel position is statistically analyzed to form the overall segmentation error, which is used to constrain the overall recognition accuracy of the network for abnormal regions.

[0080] Boundary constraints are obtained by extracting boundary information from the anomaly region annotation map to generate a boundary annotation map, and comparing the corresponding boundary region in the segmentation result with the boundary annotation map. By strengthening the response consistency of the segmentation result at the boundary position, the ability of the PSPNet network to identify crack edges, narrow structures and irregular boundaries is improved.

[0081] The dual-temporal consistency constraint is obtained by performing a difference analysis on the segmentation results of the current time and the previous time. Combined with the registration relationship between the two temporal images, the segmentation result of the previous time is mapped to the spatial position of the current time and compared with the segmentation result of the current time. Consistency constraints are applied to the stable region outside the change region between the two temporal phases to ensure the continuity and stability of the segmentation results in the temporal dimension.

[0082] During training, the initial learning rate was set to 0.0001, the batch size was set to eight, and the number of training rounds was set to one hundred and twenty. When the validation set segmentation accuracy reached the preset threshold, the improved PSPNet network was obtained after training.

[0083] The current and previous registered images are input into the trained improved PSPNet network, which outputs the current and previous abnormal region segmentation maps. The network extracts the boundary expansion, area increase / decrease, connectivity changes, and adjacent region expansion of the abnormal regions to form the temporal evolution results of the abnormal regions.

[0084] The temporal evolution results of the anomalous regions are mapped to the corresponding rock movement influence units according to the spatial location of the anomalous regions. The coverage area, expansion direction, connectivity changes, and continuous activity of the anomalous regions within each rock movement influence unit are statistically analyzed to generate image modal anomaly evolution values ​​corresponding to each rock movement influence unit, thereby forming an image modal anomaly evolution field. Specifically, the generation of image modal anomaly evolution values ​​corresponding to each rock movement influence unit is as follows:

[0085] Based on the temporal evolution results of the anomalous area, the spatial coverage ratio and the newly added expansion ratio of the anomalous area within each rock movement influence unit are calculated to characterize the degree of occupation and expansion intensity of the anomalous area in the rock movement influence unit.

[0086] Based on the expansion direction of the anomalous region within the rock movement influence unit and the spatial relationship between adjacent rock movement influence units, the dominant direction of anomalous expansion is determined. In conjunction with the connectivity changes of the anomalous region, directions with cross-unit connection trends are given higher weights to characterize the anomalous propagation trend.

[0087] By comprehensively considering the duration of activity, continuity of expansion, and degree of connectivity enhancement of the anomaly region within the rock movement influence unit, and by weighting and fusing the aforementioned spatial coverage and directional propagation characteristics, the image modal anomaly evolution value corresponding to the rock movement influence unit is obtained.

[0088] In this embodiment, obtaining the anomaly propagation relationship and generating the structural anomaly field includes:

[0089] Unit-level anomaly extraction was performed on surface displacement data, deep displacement data, remote sensing deformation data, and environmental triggering data to obtain the anomaly intensity, anomaly direction, and anomaly duration of each rock movement influence unit, and the image modal anomaly evolution value of each rock movement influence unit was obtained.

[0090] The abnormal intensity, abnormal direction, abnormal duration, and image modal abnormal evolution value of each rock movement influence unit are uniformly characterized to form the current abnormal state information of each rock movement influence unit. The current abnormal state information includes the abnormal activity level, abnormal expansion direction, and abnormal continuous change state of the rock movement influence unit.

[0091] Centered on each rock movement-affected unit, a candidate transmission unit set is established according to the spatially adjacent units. Based on the anomaly expansion direction in the current anomaly state information, the candidate transmission unit set is filtered by direction. Adjacent units located in the anomaly expansion forward region are retained as effective transmission units. The target units located in the anomaly edge expansion zone and the anomaly non-penetrating gap zone are identified by combining the image modal anomaly evolution value.

[0092] For each rock movement influence unit, the abnormal transmission value from the rock movement influence unit to each effective transmission unit is calculated. The abnormal transmission value is determined based on the abnormal intensity, abnormal duration, consistency between the abnormal expansion direction and the relative position of the target unit, the edge expansion activity corresponding to the image modality abnormal evolution value, and whether the target unit is located within the abnormal non-penetrating gap zone.

[0093] Anomaly transmission relationships are established based on the anomaly transmission values ​​between each rock movement influencing unit and its effective transmission units. The current anomaly state information of each rock movement influencing unit is fused with the corresponding anomaly transmission relationship to generate a structural anomaly field. The structural anomaly field refers to the spatial distribution state field constructed with rock movement influencing units as basic units. The state field comprehensively reflects the current anomaly state information of each rock movement influencing unit and its anomaly transmission relationship with adjacent units. It is used to characterize the spatial distribution degree, propagation direction and cross-unit connectivity trend of anomalies, forming a structured anomaly expression that can describe the evolution of anomalies from local occurrence to overall expansion.

[0094] In this embodiment, constructing the topological framework of rock migration precursors includes:

[0095] Rock movement influence units that reach a predefined level of anomalous activity in the structural anomaly field are identified as explicit anomaly units. Rock movement influence units located in the anomaly edge extension zone or the anomaly non-connected gap zone and exhibiting anomalous continuous change are identified as potential extension units. The predefined conditions are that the structural anomaly field value of the rock movement influence unit reaches a predefined anomaly intensity threshold and maintains an anomalous continuous state for multiple consecutive time periods. At the same time, there is a tendency for the rock movement influence unit to extend to adjacent rock movement influence units in the anomaly transmission relationship. When the above anomaly intensity condition, continuity condition, and propagation condition are met simultaneously, the rock movement influence unit is identified as an explicit anomaly unit.

[0096] Based on the anomalous transmission relationships between obvious anomalous units, obvious anomalous connected regions are constructed. Within each obvious anomalous connected region, the main propagation segment is extracted according to the anomalous transmission direction and intensity. The main propagation segment consists of consecutive adjacent rock movement influence units with consistent anomalous transmission directions. Specifically, the construction of obvious anomalous connected regions is as follows:

[0097] Starting with the visible anomaly unit, based on the established anomaly transmission relationship between each rock movement influence unit, the visible anomaly units that have anomaly transmission relationship and are adjacent to each other are extended and connected layer by layer. All visible anomaly units that can reach each other through anomaly transmission relationship are divided into the same connected set. Each connected set is regionally identified to form multiple visible anomaly connected regions.

[0098] Based on the abnormal transmission relationship between potential extension units and adjacent explicit abnormal connected regions, potential extension paths located at the edge of explicit abnormal connected regions and distributed along the extension direction of the main propagation segment are extracted. The potential extension paths are determined as segments to be connected, and the connection positions between each segment to be connected and the corresponding explicit abnormal connected region are recorded.

[0099] Branch propagation paths in each visible anomaly connected region are identified. When the same rock movement influence unit is connected to two or more anomaly transmission paths at the same time, the rock movement influence unit is identified as a bifurcation unit. A hierarchical connection relationship is established by combining the main propagation segment, the segment to be connected and the bifurcation unit. The visible anomaly connected region and the main propagation segment constitute the visible skeleton, and the segment to be connected constitutes the hidden extension skeleton.

[0100] The explicit framework, implicit extension framework, bifurcation units, and unconnected sections are integrated to construct the topological framework of rock movement precursors.

[0101] In this embodiment, determining whether a continuous through-path is formed based on critical bridging units, thereby determining the fission channel triggering state, includes:

[0102] Obtain the topological skeleton of rock movement precursors, extract the visible anomaly connected regions and their corresponding main propagation segments, and determine the unconnected segments and potential extension paths between adjacent visible anomaly connected regions.

[0103] In the unconnected sections and potential extension paths, rock movement influence units that simultaneously meet the following requirements—being consistent with the main propagation section of the adjacent visible anomaly connected area, having an anomaly transmission relationship with the adjacent anomaly unit, and having a continuous anomaly change state—are selected as candidate bridging units.

[0104] For each candidate bridging unit, the corresponding anomaly intensity value, anomaly duration value, and image modal anomaly evolution value are obtained. Combined with the spatial positional relationship and anomaly propagation relationship with the adjacent explicit anomaly connected regions, the bridging potential value of the candidate bridging unit is determined. The bridging potential value is jointly determined by the anomaly intensity, anomaly duration, the consistency of anomaly propagation direction, and the degree of edge expansion activity of the anomaly region.

[0105] Based on the bridging potential value and spatial distribution relationship of each candidate bridging unit, rock movement influence units that meet the preset conditions and form a continuous connection relationship in space are selected as critical bridging units. The critical bridging units are associated with the corresponding obvious anomaly connected regions. The preset conditions are that the bridging potential value of the candidate bridging unit reaches the preset bridging strength threshold, and the candidate bridging unit maintains a stable abnormal active state for multiple consecutive time periods. At the same time, there is a continuous anomaly transmission relationship between the candidate bridging unit and the adjacent obvious anomaly connected regions, and it is located at the connection position between two or more obvious anomaly connected regions in space. When the bridging strength condition, the continuity condition, and the spatial connection condition are all met, the candidate bridging unit is determined as a critical bridging unit.

[0106] Based on the connection relationship between the critical bridging unit and the explicit anomalous connected region, it is determined whether a continuous through-path exists, consisting of the explicit anomalous connected region, the main propagation segment, and the critical bridging unit. If a continuous through-path exists, the fission channel is determined to be in a triggered state; otherwise, it is determined to be in a non-triggered state. Specifically, determining whether a continuous through-path exists involves the explicit anomalous connected region, the main propagation segment, and the critical bridging unit:

[0107] Starting from the main propagation segment in the obvious abnormal connected region, the critical bridging units that are adjacent to each other and have an abnormal transmission relationship are sequentially searched along the main propagation direction to determine whether the main propagation segment can extend to another obvious abnormal connected region through the critical bridging unit.

[0108] Detect whether a continuous spatial adjacency relationship is formed between the main propagation segment and the critical bridging unit, as well as between the critical bridging unit and the adjacent explicit abnormal connected region, to confirm that there is an abnormal transmission relationship between each connecting unit and that the propagation direction is consistent.

[0109] The overall connectivity of the path is verified. If a continuous path exists when starting from an obvious abnormal connected region, it can be continuously reached from the main propagation segment and the critical bridging unit to another obvious abnormal connected region, and there are no interruption units in the path; otherwise, it is determined that there is no continuous path.

[0110] In this embodiment, the construction of rock movement phase transition state variables with hysteresis memory, the determination of the current rock movement evolution stage, and the output of corresponding surface rock movement disaster early warning results based on the rock movement evolution stage include:

[0111] Obtain the topological skeleton of rock movement precursors, as well as the triggering states of critical bridging units and fission channels. Extract the skeleton connectivity changes, main propagation segment extension changes, unconnected segment contraction changes, and critical bridging unit activity changes between the current and previous moments.

[0112] Based on changes in skeleton connectivity, extension of the main propagation segment, contraction of the segment to be penetrated, active changes in critical bridging units, and fission channel triggering states, a rock-movement facies transition state variable is constructed, including instantaneous evolution values ​​and historical memory values. Historical memory values ​​are updated cumulatively over consecutive time intervals, and the established connectivity extension states are preserved. Specifically:

[0113] The instant evolution value is obtained by comprehensively extracting the structural changes of the topological skeleton of the rock movement precursor at the current moment. Specifically, based on the changes in the number of connected regions of the skeleton at the current moment, the extension of the main propagation segment, the contraction of the segment to be connected, and the activity of the critical bridging unit, the changes of each are uniformly characterized to form an instant evolution value that reflects the intensity of the abnormal expansion and the propagation trend at the current moment.

[0114] The historical memory value is obtained by accumulating and updating the instantaneous evolution value at multiple consecutive moments. Specifically, the instantaneous evolution value at the current moment is merged with the historical memory value at the previous moment. The accumulated result is strengthened or suppressed according to the triggering state of the fission channel. When the fission channel is in the triggered state or the skeleton connectivity is continuously enhanced, the historical memory value is increased. When an anomaly occurs and the connectivity structure is maintained, the historical memory value is maintained. When the skeleton shrinks and the bridging unit is deactivated at multiple consecutive moments, the historical memory value is reduced, thus forming a historical memory value with hysteresis characteristics.

[0115] When the skeleton expands and strengthens, the bridging unit remains active, or the fission channel changes from non-triggered to triggered at the current moment, increase the historical memory value; when the current moment shows an abnormal decline but the section to be connected has not recovered or the fission channel is still in the triggered state, maintain the historical memory value; when the skeleton shrinks, the bridging unit becomes inactive, and the fission channel is in the non-triggered state for multiple consecutive moments, decrease the historical memory value.

[0116] The current rock movement evolution stage is determined based on the instantaneous evolution value and the historical memory value. The rock movement evolution stage includes a stable stage, a gestation stage, a penetration stage, and an instability stage. When the historical memory value reaches the corresponding stage condition, the rock movement evolution stage is increased; when the historical memory value does not decrease to the corresponding release condition, the current rock movement evolution stage remains unchanged. Specifically, determining the current rock movement evolution stage based on the instantaneous evolution value and the historical memory value involves:

[0117] The instantaneous evolution value is compared with the preset stable stage determination threshold. When the instantaneous evolution value is less than the stable stage determination threshold and the historical memory value is less than the corresponding initial memory threshold, it is determined to be a stable stage. When the instantaneous evolution value is greater than the stable stage determination threshold and continues to grow for multiple consecutive time periods, and the historical memory value is greater than the initial memory threshold, it is determined to be a gestation stage.

[0118] When the instantaneous evolution value reaches the preset threshold for determining the breakthrough stage and the historical memory value is greater than the corresponding threshold for the breakthrough stage memory, and there is a cross-unit connectivity trend in the structural anomaly field, the current state is upgraded from the incubation stage to the breakthrough stage; when the instantaneous evolution value reaches the preset threshold for determining the instability stage and the historical memory value is greater than the corresponding threshold for the instability stage memory, and the fission channel is in the triggered state, the current state is determined to be the instability stage.

[0119] When the instantaneous evolution value decreases, the current historical memory value is compared with the release threshold of the corresponding stage. When the historical memory value is greater than the release threshold of the stage, the current rock movement evolution stage remains unchanged. When the historical memory value is less than the release threshold of the stage, the rock movement evolution stage is gradually reduced in a preset order.

[0120] The corresponding stage conditions refer to the pre-set entry conditions for different rock movement evolution stages. The entry conditions are determined by the instant evolution value threshold and the historical memory value threshold, and are used to determine whether the current state meets the judgment requirements for entering the stable stage, the gestation stage, the penetration stage or the unstable stage.

[0121] The corresponding release conditions refer to the exit conditions pre-set for each rock movement evolution stage. The exit conditions are determined by the release threshold of the historical memory value and are used to determine whether the current state allows the transfer from the current rock movement evolution stage to a lower-level rock movement evolution stage.

[0122] Based on the current stage of rock movement evolution, the corresponding surface rock movement disaster early warning results are output. The stable stage outputs normal monitoring results, the incubation stage outputs low-level early warning results, the penetration stage outputs medium-level early warning results, and the unstable stage outputs high-level early warning results.

[0123] Example 1: To verify the feasibility of this invention in practice, it was applied to an open-pit mining area in a mountainous region. This area has significant topographic relief, fractured rock mass, numerous joints and fissures, and is subject to long-term rainfall and mining disturbances, resulting in a high risk of surface rock movement. A multi-source monitoring system has been deployed in this area, including surface GNSS displacement monitoring points, deep displacement monitoring boreholes, an InSAR remote sensing deformation monitoring system, and fixed crack image acquisition equipment. Rain gauges and groundwater level monitoring equipment have also been deployed. Existing technologies primarily rely on threshold judgments of displacement monitoring data or simple overlay analysis of multi-source data. These methods fail to identify the spatial connectivity between crack propagation, deformation enhancement, and displacement changes, making timely early warning difficult before anomalies lead to overall instability.

[0124] In this scenario, the method described in this invention is used to process multimodal monitoring data. First, surface displacement data, deep displacement data, remote sensing deformation data, and crack image data undergo unified preprocessing to achieve time synchronization, spatial coordinate unification, and noise suppression. Based on the slope structure, rock movement influence units are divided, and various monitoring data are mapped to corresponding units. Subsequently, crack image data acquired at two consecutive time points are spatially registered and input into an improved PSPNet network. A dual-temporal input structure and difference-guided fusion module are used to extract crack regions and their expansion, obtaining the distribution and variation characteristics of anomalous areas, and forming an image modal anomaly evolution field. Unit-level anomaly intensity, direction of change, and duration are extracted from displacement and deformation data, and structural projection is performed with the image modal anomaly evolution field to generate a structural anomaly field that reflects the anomaly propagation trend.

[0125] Based on the structural anomaly field, the anomalous connectivity regions and main propagation paths are further extracted, and potential extension units located at the anomaly edges and incomplete segments are identified to construct a topological framework for rock movement precursors. By analyzing the connectivity between different anomaly branches, critical bridging units are identified, and it is determined whether a continuous through-path is formed, thereby confirming whether fission channels have been triggered. Based on the changes in the topological framework, the activity level of bridging units, and the state of fission channels, rock movement phase transition state variables with hysteresis memory are constructed to determine the rock movement evolution stage and output corresponding early warning results.

[0126] In actual operation, continuous observation of typical monitoring units revealed that before crack propagation caused significant displacement abrupt changes, an enhanced trend of anomalous connectivity emerged in the structural anomaly field. Potential through-paths gradually formed within the topological framework, and multiple critical bridging units appeared. The method of this invention can identify the trend of anomalies transforming from discrete distribution to a connected structure at this stage and output early warning signals in advance. In contrast, traditional methods only trigger warnings after a significant increase in displacement, exhibiting a significant lag. Even in the later stages of anomaly development, although local displacements may temporarily decrease, the method of this invention maintains a high level of warning due to the introduction of a hysteresis memory mechanism, avoiding the problem of frequent fluctuations in warnings.

[0127] Table 1 Comparison of Multimodal Anomaly Evolution and Early Warning Results of Rock Movement Monitoring Unit

[0128]

[0129] As shown in Table 1, with the gradual increase of displacement rate and crack propagation, the structural connectivity and bridging unit activity exhibit a synchronous upward trend, indicating that the anomaly is gradually evolving from a locally discrete state to a spatially connected state. In the U1 and U2 stages, although the displacement rate and crack propagation have increased to some extent, the structural connectivity remains at a low level, and the anomaly has not yet formed a clear propagation path. Traditional methods do not provide early warnings, while this invention outputs a low-level early warning based on crack propagation and structural connectivity trends in the U2 stage, demonstrating sensitivity to early anomaly evolution.

[0130] In stages U3 to U5, the displacement rate and crack propagation significantly increase, while the structural connectivity and bridging unit activity rapidly increase, indicating that relatively stable propagation paths have formed between anomalies and are gradually developing towards a connected state. In this invention, the warning level is raised from medium to high at this stage, consistent with the increase in structural connectivity. Traditional methods, however, primarily rely on changes in displacement amplitude, resulting in a relatively delayed increase in the warning level and failing to fully reflect the evolution of the anomalous structure.

[0131] In stages U6 and U7, although the displacement rate decreased, the crack propagation remained at a high level, and the structural connectivity and bridging unit activity remained strong, indicating that the abnormal structure had not been resolved and the risk of further penetration remained. Traditional methods often resulted in a decrease in the warning level or even a low-level warning at this stage. However, this invention, by incorporating the activity of bridging units and the structural connectivity for comprehensive judgment, maintained a high-level warning, avoiding misjudgments caused by local numerical declines, thus demonstrating stronger stability and reliability.

[0132] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for early warning of surface rock movement disasters based on multimodal monitoring data, characterized in that, include: Acquire multimodal monitoring data within the monitoring area, preprocess the multimodal monitoring data, and map various types of monitoring data to pre-divided rock movement influence units; After registering the crack image data of the current time with the previous time, the data is input into the improved PSPNet network to extract the anomalous regions of each time phase, generate the temporal evolution results of the anomalous regions, and map them to the corresponding rock movement influence units to form an image modal anomalous evolution field. Extract the unit anomaly intensity, anomaly direction, and anomaly duration corresponding to each modal data. Combine the image modal anomaly evolution field with the spatial adjacency relationship, geological continuity relationship, and consistency of anomaly propagation direction between rock movement-affected units to perform structural projection, obtain the anomaly transmission relationship, and generate the structural anomaly field. Based on the structural anomaly field, the abnormal connectivity, abnormal propagation path, abnormal bifurcation structure and non-connected sections between each rock movement influence unit are extracted to construct the topological skeleton of rock movement precursors. In the topological framework of rock movement precursors, rock movement influence units located between different anomalous branches and satisfying the conditions of consistent propagation direction, continuous geological structure, and enhanced local anomalousness are identified as critical bridging units. Based on the critical bridging units, it is determined whether a continuous through path is formed, thereby determining the triggering state of the fission channel. Based on the structural changes of the topological framework of rock movement precursors, the activity level of critical bridging units, and the triggering state of fission channels, a rock movement phase transition state variable with hysteresis memory is constructed, and the current rock movement evolution stage is determined. According to the rock movement evolution stage, the corresponding surface rock movement disaster early warning results are output. The construction of the rock movement precursor topological framework includes: Rock movement influence units that reach the preset conditions in the structural anomaly field are extracted as explicit anomaly units, and rock movement influence units located in the anomaly edge extension zone or the anomaly non-connected gap zone and have an abnormal continuous change state are extracted as potential extension units. Based on the anomalous transmission relationship between obvious anomalous units, obvious anomalous connected regions are constructed. In each obvious anomalous connected region, the main propagation segment is extracted according to the anomalous transmission direction and the order of anomalous transmission strength. The main propagation segment is composed of continuous adjacent rock movement influence units with consistent anomalous transmission direction. Based on the abnormal transmission relationship between potential extension units and adjacent explicit abnormal connected regions, potential extension paths located at the edge of explicit abnormal connected regions and distributed along the extension direction of the main propagation segment are extracted. The potential extension paths are determined as segments to be connected, and the connection positions between each segment to be connected and the corresponding explicit abnormal connected region are recorded. Branch propagation paths in each visible anomaly connected region are identified. When the same rock movement influence unit is connected to two or more anomaly transmission paths at the same time, the rock movement influence unit is identified as a bifurcation unit. A hierarchical connection relationship is established by combining the main propagation segment, the segment to be connected and the bifurcation unit. The visible anomaly connected region and the main propagation segment constitute the visible skeleton, and the segment to be connected constitutes the hidden extension skeleton. The explicit framework, implicit extension framework, bifurcation units, and unconnected sections are integrated to construct the topological framework of rock movement precursors.

2. The method for early warning of surface rock movement disasters based on multimodal monitoring data according to claim 1, characterized in that, The multimodal monitoring data includes surface displacement data, deep displacement data, remote sensing deformation data, remote sensing image data, and environmental triggering data.

3. The method for early warning of surface rock movement disasters based on multimodal monitoring data according to claim 1, characterized in that, The preprocessing of multimodal monitoring data, mapping various types of monitoring data to pre-divided rock movement influence units, includes: The multimodal monitoring data is processed for time alignment and spatial coordinate unification. Abnormal data is denoised and smoothed. Rock movement influence units are divided according to the spatial range of the monitoring area. Multimodal monitoring data are mapped to the corresponding rock movement influence units according to the spatial location of each monitoring data.

4. The method for early warning of surface rock movement disasters based on multimodal monitoring data according to claim 1, characterized in that, The generation of the temporal evolution results of the anomalous region and mapping them to the corresponding rock movement influence units to form an image modal anomaly evolution field includes: Acquire the crack image data at the current moment and the crack image data at the previous moment. Perform denoising, grayscale normalization and spatial registration on the crack image data at the two moments in sequence to obtain the registered image at the current moment and the registered image at the previous moment. Establish the correspondence between the image pixel region and the rock movement influence unit according to the spatial range of the rock movement influence unit. An improved PSPNet network is constructed, comprising a dual-temporal input layer, a high-resolution coding backbone layer, a differential guided fusion layer, a boundary refinement layer, a pyramid pooling layer, and a segmentation output layer, wherein: The high-resolution coding backbone layer adopts the HRNet-W18 architecture. The first input branch receives the registered image at the current time step, and the second input branch receives the registered image at the previous time step. The two input branches are respectively processed by the first stage convolutional layer and the second stage high-resolution parallel convolutional layer to extract shallow texture features. After the high-resolution parallel convolutional layer in the third stage, a difference-guided fusion layer is set up to concatenate the current time-time features, the previous time-time features and their element-wise difference features through channel splicing and then compressing them through a one-to-one convolutional layer to output the fused features. A boundary refinement layer is set after the high-resolution parallel convolutional layer in the fourth stage. The boundary refinement layer is composed of a 3x3 convolutional layer, a dilated convolutional layer, and a 1x1 convolutional layer connected in sequence. The pyramid pooling layer is connected after the boundary refinement layer, and uses four different pooling scales to perform contextual aggregation on the fused features; The segmentation output layer is connected to the pyramid pooling layer and outputs the anomaly region segmentation map at the current time and the anomaly region segmentation map at the previous time, respectively. A training sample set is constructed, which includes crack image samples at the current time, crack image samples at the previous time, and corresponding abnormal region annotation maps. The training sample set is input into the improved PSPNet network, and the improved PSPNet network is trained using a joint loss function that includes pixel-level segmentation constraints, boundary constraints, and dual temporal consistency constraints, to obtain the trained improved PSPNet network. The current and previous registered images are input into the trained improved PSPNet network, which outputs the current and previous abnormal region segmentation maps. The network extracts the boundary expansion, area increase / decrease, connectivity changes, and adjacent region expansion of the abnormal regions to form the temporal evolution results of the abnormal regions. The temporal evolution results of the anomalous region are mapped to the corresponding rock movement influence unit according to the spatial location of the anomalous region. The coverage range, expansion direction, connectivity changes and continuous activity of the anomalous region in each rock movement influence unit are statistically analyzed to generate the image modal anomalous evolution value corresponding to each rock movement influence unit, thereby forming the image modal anomalous evolution field.

5. The method for early warning of surface rock movement disasters based on multimodal monitoring data according to claim 1, characterized in that, The process of obtaining the anomaly propagation relationship and generating the structural anomaly field includes: Unit-level anomaly extraction was performed on surface displacement data, deep displacement data, remote sensing deformation data, and environmental triggering data to obtain the anomaly intensity, anomaly direction, and anomaly duration of each rock movement influence unit, and the image modal anomaly evolution value of each rock movement influence unit was obtained. The abnormal intensity, abnormal direction, abnormal duration, and image modal abnormal evolution value of each rock movement-affected unit are uniformly characterized to form the current abnormal state information of each rock movement-affected unit. Centered on each rock movement-affected unit, a candidate transmission unit set is established according to the spatially adjacent units. Based on the anomaly expansion direction in the current anomaly state information, the candidate transmission unit set is filtered by direction. Adjacent units located in the anomaly expansion forward region are retained as effective transmission units. The target units located in the anomaly edge expansion zone and the anomaly non-penetrating gap zone are identified by combining the image modal anomaly evolution value. For each rock movement influence unit, the abnormal transmission value from the rock movement influence unit to each effective transmission unit is calculated. The abnormal transmission value is determined based on the abnormal intensity, abnormal duration, consistency between the abnormal expansion direction and the relative position of the target unit, the edge expansion activity corresponding to the image modality abnormal evolution value, and whether the target unit is located within the abnormal non-penetrating gap zone. Anomaly transmission relationships are established based on the anomalous transmission values ​​between each rock movement influence unit and its effective transmission unit. The current anomalous state information of each rock movement influence unit is then fused with the corresponding anomalous transmission relationships to generate a structural anomalous field.

6. The method for early warning of surface rock movement disasters based on multimodal monitoring data according to claim 1, characterized in that, The determination of whether a continuous through-path is formed based on critical bridging units, thereby determining the fission channel triggering state, includes: Obtain the topological skeleton of rock movement precursors, extract the visible anomaly connected regions and their corresponding main propagation segments, and determine the unconnected segments and potential extension paths between adjacent visible anomaly connected regions. In the unconnected sections and potential extension paths, rock movement influence units that simultaneously meet the following requirements—being consistent with the main propagation section of the adjacent visible anomaly connected area, having an anomaly transmission relationship with the adjacent anomaly unit, and having a continuous anomaly change state—are selected as candidate bridging units. For each candidate bridging unit, the corresponding anomaly intensity value, anomaly duration value, and image modal anomaly evolution value are obtained. Combined with the spatial positional relationship and anomaly propagation relationship with the adjacent explicit anomaly connected regions, the bridging potential value of the candidate bridging unit is determined. The bridging potential value is jointly determined by the anomaly intensity, anomaly duration, the consistency of anomaly propagation direction, and the degree of edge expansion activity of the anomaly region. Based on the bridging potential value and spatial distribution relationship of each candidate bridging unit, rock movement influence units that meet the preset conditions and form a continuous connection relationship in space are selected as critical bridging units, and the critical bridging units are associated with the corresponding obvious anomaly connected areas. Based on the connection relationship between the critical bridging unit and the obvious abnormal connected region, it is determined whether there is a continuous through path consisting of the obvious abnormal connected region, the main propagation segment and the critical bridging unit. When a continuous through path exists, the fission channel is determined to be in the triggered state. When no continuous through path exists, the fission channel is determined to be in the untriggered state.

7. The method for early warning of surface rock movement disasters based on multimodal monitoring data according to claim 1, characterized in that, The process involves constructing hysteresis memory-based phase transition state variables for rock movement, determining the current rock movement evolution stage, and outputting corresponding surface rock movement hazard early warning results based on the rock movement evolution stage, including: Obtain the topological skeleton of rock movement precursors, as well as the triggering states of critical bridging units and fission channels. Extract the skeleton connectivity changes, main propagation segment extension changes, unconnected segment contraction changes, and critical bridging unit activity changes between the current and previous moments. Based on changes in skeleton connectivity, extension of the main propagation segment, contraction of the segment to be connected, active changes of critical bridging units, and fission channel triggering state, a rock migration facies change state variable is constructed, which includes instantaneous evolution values ​​and historical memory values. The historical memory values ​​are updated cumulatively according to continuous time intervals, and the already formed connectivity extension state is maintained. When the skeleton expands and strengthens, the bridging unit remains active, or the fission channel changes from non-triggered to triggered at the current moment, increase the historical memory value; when the current moment shows an abnormal decline but the section to be connected has not recovered or the fission channel is still in the triggered state, maintain the historical memory value; when the skeleton shrinks, the bridging unit becomes inactive, and the fission channel is in the non-triggered state for multiple consecutive moments, decrease the historical memory value. The current rock movement evolution stage is determined based on the real-time evolution value and the historical memory value. The rock movement evolution stage includes the stable stage, the gestation stage, the penetration stage, and the instability stage. When the historical memory value reaches the corresponding stage condition, the rock movement evolution stage is increased. When the historical memory value does not drop to the corresponding release condition, the current rock movement evolution stage remains unchanged. Based on the current stage of rock movement evolution, the corresponding surface rock movement disaster early warning results are output. The stable stage outputs normal monitoring results, the incubation stage outputs low-level early warning results, the penetration stage outputs medium-level early warning results, and the unstable stage outputs high-level early warning results.

Images