A method for monitoring deformation of a mountain collapse disaster body

Abstract

The present application relates to the technical field of geological disaster monitoring, in particular to a mountain collapse disaster body deformation monitoring method, comprising: collecting multi-dimensional monitoring data of a target disaster body and performing synchronous processing and correlation analysis to obtain a multi-dimensional correlation data set; extracting feature parameters in the multi-dimensional correlation data set and performing dimension reduction processing to obtain a compressed feature set; performing feature fusion processing on the compressed feature set to obtain a real-time fusion feature set; if a deformation parameter in the real-time fusion feature set exceeds a preset deformation threshold, inputting the real-time fusion feature set into a pre-trained convolutional neural network to obtain a deformation trend sequence; based on the deformation trend sequence, combining historical deformation data and geological environment parameters of the target disaster body, using a pre-trained long short-term memory network to analyze trend change rules, obtaining a health state prediction value and performing early warning to complete deformation monitoring of the target disaster body. The present application realizes real-time monitoring, fusion analysis and early warning of multi-dimensional data of a disaster body.

Description

Technical Field

[0001] This invention relates to the field of geological disaster monitoring technology, and in particular to a method for monitoring the deformation of landslide disaster bodies. Background Technology

[0002] Landslides are sudden and highly destructive geological hazards, often triggered by geological tectonic activity, rainfall, weathering, or human engineering activities. Their deformation process is characterized by its hidden nature, nonlinearity, and multi-stage evolution; once unstable, it can cause significant casualties and property damage. Therefore, high-precision, real-time monitoring of the deformation of landslide hazard bodies and the realization of disaster early warning are key technological requirements in the field of disaster prevention and mitigation.

[0003] Currently, deformation monitoring for landslides mainly relies on: (1) Traditional surface displacement monitoring technologies, including total stations, levels, and GNSS (Global Navigation Satellite System), which rely on manual periodic measurement of surface displacement data. Although these methods are low-cost, they rely on manual operation, have poor real-time performance, and are difficult to capture deep deformation information, limiting their applicability in adverse weather or complex terrain conditions. (2) Remote sensing monitoring technologies, such as synthetic aperture radar interferometry (InSAR) and UAV aerial photography, which can acquire surface deformation data over a wide area. However, InSAR is easily affected by cloud cover and vegetation cover, resulting in low data update frequency; UAV monitoring is limited by endurance, making it difficult to achieve continuous observation around the clock. (3) Sensor network technology, including the deployment of sensors such as inclinometers, crack gauges, and microseismometers, which collect parameters such as local strain and vibration for judgment. However, sensors are easily affected by environmental interference, have insufficient long-term stability, and are expensive to deploy and maintain on large-scale mountains. (4) Exploration of new technologies: In recent years, technologies such as fiber optic sensing and three-dimensional laser scanning (LiDAR) have been gradually applied, but they are easily blocked in complex terrains, and the data processing algorithms are highly complex, making it difficult to meet the needs of real-time early warning.

[0004] Therefore, in response to the aforementioned technical shortcomings such as insufficient multi-source data fusion, inability to simultaneously achieve real-time performance and accuracy, poor environmental adaptability, and low level of intelligence, a deformation monitoring method for landslide disaster bodies is proposed. Summary of the Invention

[0005] The purpose of this invention is to provide a deformation monitoring method for landslide disaster bodies, which integrates multi-dimensional monitoring data, can adapt to complex environments, and improve the accuracy and timeliness of disaster early warning.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] A method for monitoring the deformation of a landslide hazard includes:

[0008] Collect multidimensional monitoring data of the target disaster body and perform synchronous processing and correlation analysis to obtain a multidimensional correlation dataset;

[0009] Extract the feature parameters from the multidimensional associated dataset and perform dimensionality reduction processing to obtain a compressed feature set;

[0010] The compressed feature set is subjected to feature fusion processing to obtain a real-time fused feature set;

[0011] If the deformation parameters in the real-time fused feature set exceed the preset deformation threshold, the real-time fused feature set is input into a pre-trained convolutional neural network to obtain a deformation trend sequence.

[0012] Based on the deformation trend sequence, combined with the historical deformation data and geological environmental parameters of the target disaster body, a pre-trained long short-term memory network is used to analyze the trend change pattern, obtain the health status prediction value and issue an early warning, thereby completing the deformation monitoring of the target disaster body.

[0013] Preferably, the collection of multidimensional monitoring data of the target disaster body includes:

[0014] GNSS displacement monitoring points, fiber optic strain sensors, and meteorological sensors are deployed on the surface of the target disaster body to collect real-time data on surface displacement, crack propagation, and temperature and humidity.

[0015] A multi-parameter sensor chain, including a microseismometer, an inclinometer, and a pore water pressure gauge, is deployed in boreholes inside the target disaster body to acquire data on deep rock mass fracturing, dip angle changes, and groundwater activity.

[0016] Preferably, obtaining the multidimensional association dataset includes:

[0017] The multidimensional monitoring data is timestamped and spatially calibrated to obtain a synchronized dataset;

[0018] Multi-dimensional parameters are extracted based on the synchronized dataset, and correlation analysis is performed on the multi-dimensional parameters through algorithm kernel to obtain the multi-dimensional correlation dataset.

[0019] Preferably, extracting feature parameters from the multidimensional association dataset and performing dimensionality reduction processing to obtain a compressed feature set includes:

[0020] Feature parameters are extracted from the multidimensional associated dataset through edge computing nodes, and the feature parameters are cleaned to construct a feature set;

[0021] After performing feature decomposition and dimensionality reduction on the feature set using principal component analysis, the t-SNE algorithm is then used for compression to obtain the compressed feature set.

[0022] Preferably, performing feature fusion processing on the compressed feature set to obtain a real-time fused feature set includes:

[0023] A dynamic task scheduling strategy is used to allocate multi-node tasks to the compressed feature set to obtain the allocated task set.

[0024] The allocated task set is adjusted using load balancing technology to obtain a balanced task set;

[0025] In multi-node task processing, the balanced task set is executed to obtain the node processing feature set;

[0026] The node processing feature set is integrated and feature fusion is performed by data stream scheduling to obtain the real-time fused feature set.

[0027] Preferably, inputting the real-time fused feature set into a pre-trained convolutional neural network to obtain the deformation trend sequence includes:

[0028] The real-time fused feature set is input into the convolutional neural network (CNN) and the long short-term memory network (LSTM) respectively to obtain the first deformation trend classification result and the second deformation trend classification result.

[0029] Calculate the difference between the first deformation trend classification result and the second deformation trend classification result. If the difference exceeds a preset difference value, perform a weighted fusion process on the first deformation trend classification result and the second deformation trend classification result to obtain the final deformation trend sequence. If the difference does not exceed the preset difference value, use the first deformation trend classification result as the final deformation trend sequence.

[0030] Preferably, based on the deformation trend sequence, combined with the historical deformation data and geological environmental parameters of the target disaster body, a pre-trained long short-term memory network is used to analyze the trend change pattern and obtain the health status prediction value, including:

[0031] The deformation trend sequence and historical data are input into a pre-trained bidirectional long short-term memory network (BiLSTM) for variation pattern analysis to obtain preliminary health status values.

[0032] Based on the preliminary health status values ​​and geological environment parameters, environmental impact characteristics are obtained and the variation patterns of these environmental impact characteristics are analyzed to obtain a pattern sequence.

[0033] By combining the aforementioned pattern sequence with historical data, the stability of the health status is determined and the fluctuations in the health status are analyzed to obtain the final health status value.

[0034] Preferably, the method further includes:

[0035] Based on the predicted health status values, spatial distribution mapping and time series projection are performed to obtain the health status distribution map of the target disaster body;

[0036] Based on the health status distribution map, key area parameters are extracted. If the key area parameters are abnormal, the monitoring frequency is adjusted through a dynamic update algorithm to obtain an optimized monitoring scheme. The distributed sensor network settings are then updated according to the optimized monitoring scheme, and the sensor data acquisition and processing flow is adjusted.

[0037] The beneficial effects of this invention are as follows:

[0038] This invention collects multi-dimensional parameters such as displacement, tilt, strain, and cracks of landslide disaster bodies through a distributed sensor network and performs spatiotemporal synchronous processing to obtain a multi-dimensional correlated dataset. Subsequently, feature extraction and dimensionality reduction are performed on this dataset to obtain a compressed feature set, and real-time feature fusion is achieved through a parallel computing framework. When the deformation parameters in the fusion result exceed a preset threshold, a convolutional neural network is used to determine the deformation trend of the disaster body, and a long short-term memory network is combined to analyze historical data and geological environmental parameters to predict the health status of the disaster body and issue an early warning. This method realizes real-time monitoring, fusion analysis, and early warning of multi-dimensional data of disaster bodies, providing effective technical support for geological disaster prevention and control. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a flowchart of a deformation monitoring method for a landslide disaster subject according to an embodiment of the present invention. Detailed Implementation

[0041] 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 embodiments of the present invention, and not all embodiments. 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.

[0042] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0043] This embodiment provides a method for monitoring the deformation of a landslide disaster body, such as... Figure 1As shown, it includes:

[0044] Collect multidimensional monitoring data of the target disaster body and perform synchronous processing and correlation analysis to obtain a multidimensional correlation dataset;

[0045] Extract the feature parameters from the multidimensional associated dataset and perform dimensionality reduction processing to obtain a compressed feature set;

[0046] The compressed feature set is subjected to feature fusion processing to obtain a real-time fused feature set;

[0047] If the deformation parameters in the real-time fused feature set exceed the preset deformation threshold, the real-time fused feature set is input into a pre-trained convolutional neural network to obtain a deformation trend sequence.

[0048] Based on the deformation trend sequence, combined with the historical deformation data and geological environmental parameters of the target disaster body, a pre-trained long short-term memory network is used to analyze the trend change pattern, obtain the health status prediction value and issue an early warning, thereby completing the deformation monitoring of the target disaster body.

[0049] Specifically, this embodiment collects multi-dimensional parameters such as displacement, tilt, strain, and cracks of landslide disaster bodies through a distributed sensor network and performs spatiotemporal synchronous processing to obtain a multi-dimensional correlated dataset. Subsequently, feature extraction and dimensionality reduction are performed on this dataset to obtain a compressed feature set, and real-time feature fusion is achieved through a parallel computing framework. When the deformation parameters in the fusion result exceed a preset threshold, a convolutional neural network is used to determine the deformation trend of the disaster body, and a long short-term memory network is combined to analyze historical data and geological environmental parameters to predict the health status of the disaster body and issue an early warning. This method realizes real-time monitoring, fusion analysis, and early warning of multi-dimensional data of disaster bodies, providing effective technical support for geological disaster prevention and control.

[0050] Furthermore, obtaining the multidimensional association dataset includes:

[0051] The multidimensional monitoring data is timestamped and spatially calibrated to obtain a synchronized dataset;

[0052] Multi-dimensional parameters are extracted based on the synchronized dataset, and correlation analysis is performed on the multi-dimensional parameters through algorithm kernel to obtain the multi-dimensional correlation dataset.

[0053] The multidimensional monitoring data collected from the target disaster body includes:

[0054] GNSS displacement monitoring points, fiber optic strain sensors, and meteorological sensors are deployed on the surface of the target disaster body to collect real-time data on surface displacement, crack propagation, and temperature and humidity.

[0055] A multi-parameter sensor chain, including a microseismometer, an inclinometer, and a pore water pressure gauge, is deployed in boreholes inside the target disaster body to acquire data on deep rock mass fracturing, dip angle changes, and groundwater activity.

[0056] Specifically, in this embodiment, sensors are deployed in a distributed network to collect multidimensional monitoring data of the target disaster body to construct an original dataset; the synchronization set is labeled with timestamps to obtain time-consistent data, and the time-consistent data is processed by a coordinate system mapping method to obtain spatial calibration data, thus constructing a synchronization dataset; multidimensional parameters are extracted from the synchronization dataset, and the correlation analysis of the multidimensional parameters is performed by an algorithm to obtain a multidimensional correlation dataset.

[0057] In synchronous data processing, a time reference unification algorithm is first applied to acquire time signals accurate to the nanosecond level through the Global Positioning System (GPS), ensuring that the timestamps of all sensors are aligned under a unified time reference. For example, the Network Time Protocol (NTP) is used to calibrate sensor time, with the error controlled within 1 millisecond. Next, a spatial reference consistency algorithm is used to match the location information of each sensor with a high-precision geographic coordinate system (such as WGS84). A Kalman filter algorithm is used to perform spatial coordinate mapping on the sensor data, eliminating deviations caused by installation errors and environmental interference, with spatial errors controlled within 5 millimeters. Subsequently, a multiple linear regression analysis algorithm is used to perform correlation analysis on the time-aligned and spatially mapped data to extract multidimensional features. For example, by analyzing the correlation between displacement, tilt, strain, and crack data, a dynamic response model is established, and the least squares method is used to optimize the model parameters, ensuring a goodness of fit of over 95%. Finally, the processed multidimensional correlation dataset is imported into a distributed database, and spatiotemporal indexing technology is used to achieve efficient querying and analysis, supporting subsequent real-time monitoring and prediction tasks.

[0058] This embodiment uses timestamps to label the synchronization set, ensuring that data collected by different sensors remains consistent in the time dimension. For example, in disaster monitoring, multiple sensors may have slight time deviations due to transmission delays or equipment differences. Timestamps record the acquisition time of each data set with millisecond precision, and through comparison and adjustment, time-consistent data is generated. This method avoids analysis errors caused by time misalignment. In this embodiment, assuming that the timestamp recorded by the displacement sensor is 10:00:00.125 in a certain monitoring, while that of the crack sensor is 10:00:00.130, the two sets of data are unified to the same time, such as 10:00:00.100, through a timestamp alignment algorithm, ensuring the reliability of subsequent processing. A coordinate system mapping method is used to process the time-consistent data, aiming to correlate the data with spatial locations and form spatial calibration data. Specifically, latitude and longitude information provided by the global navigation satellite system can be used to map the position of each sensor to a unified geographic coordinate system. For example, the original location data of a sensor is 120.5 degrees east longitude and 30.2 degrees north latitude. After being adjusted to x = 500 meters and y = 300 meters in a standard grid coordinate system using a coordinate transformation algorithm, the error is controlled within the centimeter level. This spatial calibration provides a precise spatial reference for subsequent multi-dimensional analysis. Multi-dimensional parameters are extracted from the spatial calibration data to obtain parameter set data, which can be understood as separating independent variables such as displacement, tilt, and strain from the original data. For example, in one monitoring session, the displacement parameter shows that a point moved by 3 millimeters, the tilt parameter records an angle change of 0.2 degrees, and the strain value increases by 10 microstrains. These parameters are integrated into a multidimensional dataset. This extraction process lays the foundation for correlation analysis. Correlation analysis is performed on the parameter set data using algorithms to generate characteristic correlation data, aiming to discover potential relationships between different parameters. In this embodiment, the analysis found that as displacement increases, crack width also expands, showing a positive correlation trend with a correlation coefficient of 0.85. This correlation reveals the stress change pattern within the disaster body, providing a basis for risk assessment.

[0059] Further, extracting feature parameters from the multidimensional association dataset and performing dimensionality reduction processing to obtain a compressed feature set includes:

[0060] Feature parameters are extracted from the multidimensional associated dataset through edge computing nodes, and the feature parameters are cleaned to construct a feature set;

[0061] After performing feature decomposition and dimensionality reduction on the feature set using principal component analysis, the t-SNE algorithm is then used for compression to obtain the compressed feature set.

[0062] Specifically, in this embodiment, feature parameters are extracted from multidimensional data through edge computing nodes to obtain an initial feature set; the initial feature set is cleaned using a preprocessing method to obtain a standardized feature set; a feature decomposition method is applied to the standardized feature set to obtain a decomposed feature set; and a dimensionality reduction algorithm is used to process the decomposed feature set to obtain a compressed feature set.

[0063] In disaster monitoring scenarios, multidimensional correlated datasets contain various parameters such as displacement, tilt, strain, and cracks. These data are preprocessed using edge computing nodes. First, Principal Component Analysis (PCA) is used to decompose the high-dimensional data, reducing 10-dimensional data to 3-dimensionality, retaining 95% of the original information and reducing data redundancy. By calculating the covariance matrix and eigenvalue decomposition, key eigenvectors are extracted; for example, the first principal component explains 60% of the variance, the second principal component explains 30%, and the third principal component explains 5%. Next, the t-SNE algorithm is used to further compress the dimensionality-reduced data, mapping the 3-dimensional data to a 2-dimensional space while preserving the local structural relationships between data points. For example, in a 2D plane, the distribution of displacement and strain exhibits a clear clustering trend.

[0064] An initial feature set is obtained by extracting feature parameters from multidimensional data using edge computing nodes. For example, in disaster monitoring, edge computing nodes can extract parameters such as displacement, strain, and tilt angle from raw data collected by sensors. For instance, data obtained by a node from a sensor on a mountain surface includes a displacement of 2 mm, a strain of 8 microstrains, and a tilt angle of 1.5 degrees; these parameters constitute the initial feature set. This approach moves the computational task to the edge, reducing the data transmission burden. The initial feature set is then cleaned using preprocessing methods to obtain a standardized feature set. Specifically, data can be cleaned by removing noise and outliers. For example, if a sensor's displacement value suddenly jumps to 50 mm, significantly deviating from the normal range, the preprocessing layer will discard it. In one possible implementation, data normalization maps parameters such as displacement and strain to a range of 0 to 1, facilitating subsequent analysis. This cleaning ensures data quality and provides a reliable foundation for feature decomposition. For the standardized feature set, feature decomposition methods are applied to obtain the decomposed feature set. It should be noted that feature decomposition can break down complex data into independent components. For example, displacement and strain may be affected by multiple factors, and decomposition can separate the main trend and secondary disturbance components. In this embodiment, the displacement of a certain mountain area is decomposed into a 70% trend component and a 30% noise component. This decomposition helps to focus on key information. The decomposed feature set is processed by a dimensionality reduction algorithm to obtain a compressed feature set. Specifically, principal component analysis can be used to reduce high-dimensional data to a few main dimensions. For example, data that originally contained multiple dimensions such as displacement, strain, and tilt angle can still represent 90% of the information after dimensionality reduction by retaining only two principal components. This compression reduces redundancy and improves processing efficiency.

[0065] Further, the compressed feature set is subjected to feature fusion processing to obtain a real-time fused feature set, including:

[0066] A dynamic task scheduling strategy is used to allocate multi-node tasks to the compressed feature set to obtain the allocated task set.

[0067] The allocated task set is adjusted using load balancing technology to obtain a balanced task set;

[0068] In multi-node task processing, the balanced task set is executed to obtain the node processing feature set;

[0069] The node processing feature set is integrated and feature fusion is performed by data stream scheduling to obtain the real-time fused feature set.

[0070] Specifically, this implementation uses a parallel computing framework to allocate tasks to the compressed feature set to obtain a preliminary task set; it then uses load balancing technology to adjust the preliminary task set to obtain a balanced task set; the balanced task set is executed in multi-node processing to obtain a node processing feature set; the node processing feature set is integrated through data flow scheduling to obtain a scheduling feature set; and a fusion algorithm is applied to the scheduling feature set to obtain a real-time fused feature set.

[0071] Within the distributed computing framework, a dynamic task scheduling strategy is employed to distribute the compressed feature set across multiple computing nodes. For example, Spark's RDD partitioning mechanism divides the data into eight parallel blocks, each processing approximately 15% of the data volume. A consistent hashing algorithm ensures uniform data distribution, keeping node load differences within 5%. In the feature fusion stage, a MapReduce model is used to perform weighted calculations on parameters such as displacement and tilt angle. The weighting coefficients are set based on sensor accuracy; for example, displacement weight is 4, tilt angle weight is 3, strain weight is 2, and crack weight is 1. A sliding window mechanism is used to dynamically aggregate time-series data, with a window length of 10 seconds and a step size of 5 seconds to ensure data continuity. During the fusion process, a Kalman filter algorithm is used to eliminate noise, reducing displacement measurement errors from ±2 mm to ±3 mm. For real-time performance assurance, a delayed scheduling strategy optimizes the task queue, automatically triggering load migration when a node's processing time exceeds 50 milliseconds. The final fusion result is output in JSON format, including statistics such as mean and variance. For example, the displacement fusion value is 8 mm (variance 9) and the tilt angle fusion value is 2 degrees (variance 4). The result is pushed to the downstream via a message middleware (such as Kafka).

[0072] When allocating tasks to compressed feature sets using a parallel computing framework, the core principle is to rationally split data processing tasks to improve efficiency. For example, in a disaster monitoring scenario, the compressed feature set may contain key parameters such as displacement and strain, assuming a data volume of 100,000 records. The parallel computing framework will divide the data into five equal parts, each containing 20,000 records, based on the number of nodes, such as five edge computing nodes, thus initially forming an allocated task set. This approach fully utilizes the computing power of multiple nodes to ensure parallel task execution. In this embodiment, when adjusting the initially allocated task set using load balancing technology, the performance differences between nodes must be considered. If a node has a slower processing speed, its task volume can be dynamically reduced to 15,000 records, while high-performance nodes can have their tasks increased to 25,000 records, forming a balanced task set. Specifically, task allocation can be adjusted by monitoring the real-time load of nodes, such as CPU utilization or memory usage. This method avoids overloading any node and ensures balanced processing efficiency. When multiple nodes process a balanced task set, the process of obtaining the node's processed feature set emphasizes independence. For example, each node extracts the mean displacement and peak strain from its assigned 20,000 data points. Node 1 obtains a mean displacement of 3 mm and a peak strain of 10 microstrain, while node 2 obtains 4 mm and 12 microstrain. This distributed processing can quickly generate local feature sets, providing a foundation for subsequent integration. It's important to note that when integrating feature sets from nodes in the data stream scheduling, the focus is on temporal consistency. In this embodiment, the feature sets of the five nodes can be sorted by acquisition time using timestamp alignment to generate a scheduled feature set. For example, displacement data streams are integrated at a sampling frequency of one second to ensure that the temporal relationship between data is not lost. This scheduling method facilitates maintaining data integrity during subsequent analysis. When applying a fusion algorithm to the scheduled feature set, a weighted average method can be used to generate a real-time fused feature set. For example, the mean displacements of the five nodes are weighted and fused, with weights allocated according to the amount of data at each node, ultimately yielding a global mean displacement of 3.8 mm. This fusion can integrate multi-source data and improve the representativeness of the features.

[0073] Furthermore, inputting the real-time fused feature set into a pre-trained convolutional neural network to obtain the deformation trend sequence includes:

[0074] The real-time fused feature set is input into the convolutional neural network (CNN) and the long short-term memory network (LSTM) respectively to obtain the first deformation trend classification result and the second deformation trend classification result.

[0075] Calculate the difference between the first deformation trend classification result and the second deformation trend classification result. If the difference exceeds a preset difference value, perform a weighted fusion process on the first deformation trend classification result and the second deformation trend classification result to obtain the final deformation trend sequence. If the difference does not exceed the preset difference value, use the first deformation trend classification result as the final deformation trend sequence.

[0076] Specifically, in this embodiment, when the deformation parameters in the real-time fused feature set exceed a preset threshold (e.g., displacement exceeds 15 mm or tilt angle exceeds 5 degrees), the deformation trend analysis process is automatically triggered. First, the fused deformation time-series data is input into a pre-trained convolutional neural network (CNN) model. This network structure contains three convolutional layers (kernel size 3×3, stride 1, activation function ReLU) and two fully connected layers (128 and 64 nodes respectively). The input layer receives time-series data segments with a sliding window length of 20 seconds and a stride of 10 seconds. The CNN output layer uses the Softmax function for classification, dividing the deformation trend into three states: slow growth (probability < 3), accelerated growth (3 ≤ probability < 7), and abrupt change (probability ≥ 7). For example, when the input displacement sequence is [12mm, 15mm, 18mm, 22mm], the model outputs an accelerated growth probability of 68, triggering a warning signal.

[0077] To enhance the reliability of trend judgment, a Long Short-Term Memory (LSTM) network is simultaneously used to perform time-dimensional analysis on the same data. The number of hidden layer units is set to 32, and the time step is 8. When the difference between the CNN and LSTM results exceeds 15% (e.g., CNN outputs 68 while LSTM outputs 52), adaptive weighted fusion is initiated, with weight coefficients of CNN6 and LSTM4. The final generated deformation trend sequence is updated every 10 seconds. A GRU neural network predicts the deformation change curve within the next 30 seconds, for example, outputting a predicted value sequence [25mm, 28mm, 32mm], with a mean square error controlled within 3mm.

[0078] Furthermore, based on the deformation trend sequence, combined with the historical deformation data and geological environmental parameters of the target disaster body, a pre-trained long short-term memory network is used to analyze the trend change patterns and obtain the health status prediction value, including:

[0079] The deformation trend sequence and historical data are input into a pre-trained bidirectional long short-term memory network (BiLSTM) for variation pattern analysis to obtain preliminary health status values.

[0080] Based on the preliminary health status values ​​and geological environment parameters, environmental impact characteristics are obtained and the variation patterns of these environmental impact characteristics are analyzed to obtain a pattern sequence.

[0081] By combining the aforementioned pattern sequence with historical data, the stability of the health status is determined and the fluctuations in the health status are analyzed to obtain the final health status value.

[0082] Specifically, this embodiment uses a long short-term memory network to analyze the change patterns through deformation trend sequences and historical data to obtain a preliminary health status value. For the preliminary health status value, geological environmental parameters are integrated to obtain environmental impact characteristics, resulting in adjusted status characteristics. If the adjusted status characteristics exceed a preset threshold, trend analysis is used to determine the change patterns, resulting in a pattern sequence. Based on the pattern sequence, historical data is processed using a data integration method to obtain an integrated feature set. The stability of the health status is determined by integrating the feature set and environmental impact characteristics, resulting in a stability sequence. For the stability sequence, a state judgment method is used to analyze sequence fluctuations to obtain a final predicted value. Based on the final predicted value and the change patterns, trend evolution characteristics are obtained to determine the evolution direction.

[0083] Based on the deformation trend sequence, the system extracts historical displacement data (e.g., [8mm, 10mm, 12mm, 15mm, 18mm]) and geological environmental parameters (including rock mass strength 35MPa, joint density 5 joints / meter, and groundwater permeability coefficient 2×10⁻⁶) from the past 30 days. -5 A multi-source feature matrix was constructed using a bidirectional long short-term memory (BiLSTM) network. The network structure consists of two layers of LSTM units (64 neurons per layer), with a time step of 5, an input window length of 15 minutes, and a sliding step of 5 minutes. During model training, the mean squared error loss function was used, with a learning rate of 0.01 and a batch size of 32. After 100 iterations, the validation set error stabilized within 1 mm. For a real-time input deformed sequence [16 mm, 18 mm, 20 mm], the model outputs a predicted health status value of 82 for the next hour (threshold range: 0-1, above 8 indicates high risk), and simultaneously generates the contribution of each influencing factor (rock mass strength 28%, joint density 35%, groundwater 22%). An attention mechanism is introduced during the prediction process, automatically weighting key time nodes (e.g., weights of 18 and 21 at the 3rd and 8th minutes, respectively). To eliminate data noise, wavelet transform (db4 wavelet basis, decomposition level 3) was used to filter the original signal, and the root mean square error of the reconstructed displacement sequence was reduced to 8 mm.

[0084] When analyzing deformation trends and historical data to identify patterns of change, Long Short-Term Memory (LSTM) networks excel at capturing long-term dependencies in time series. In geological deformation monitoring, assuming 30 consecutive days of displacement data for a certain area as input, the network analyzes daily displacement changes through 32 hidden units to initially derive a health status value. For example, if displacement slowly increases from 5 mm to 10 mm, the network outputs a health status value of 75, indicating that the structure is still stable. When integrating geological environmental parameters with the initial health status value, environmental impact characteristics such as rainfall and temperature can be introduced. Assuming a day with 50 mm of rainfall and a sudden 10-degree drop in temperature, combining these parameters with the health status value and adjusting the status characteristic might reduce it from 75 to 60, reflecting the environmental impact on the structure. It should be noted that this adjustment more accurately reflects the actual health status. If the adjusted status characteristic exceeds a preset threshold, such as a health value below 65, trend analysis is used to determine the pattern of change. Specifically, analyzing displacement data from the past 15 days, such as an increase from 10 mm to 18 mm, shows a pattern where the daily increase rises from 0.5 mm to 1 mm, indicating accelerated change. When integrating historical data based on regular sequences, such as combining displacement records from the past year, the data integration method can use weighted averaging to obtain an integrated feature set. For example, with a weight of 0.7 for recent data and 0.3 for longer-term data, the integrated feature set might show an average displacement rate of 0.8 mm / day. When judging stability by integrating the feature set and environmental impact features, in one embodiment, if the integrated feature set shows a rate of 0.8 mm / day, and the rate suddenly increases to 1.5 mm / day after rainfall, the stability sequence might be [stable, slightly fluctuating, unstable], reflecting dynamic changes in health status. Specifically, this judgment can promptly identify potential risks. When analyzing fluctuations in the stability sequence, it is understood that the state judgment method can use a sliding window to statistically analyze variance. For example, with a window length of 5 days, if the sequence variance increases from 0.2 to 0.8, the final predicted value might be 55, indicating a deterioration in health status. This method can effectively quantify the degree of fluctuation. When obtaining trend evolution characteristics based on the final predicted value and the pattern of change, for example, a predicted value of 55 combined with the pattern sequence [0.5, 0.7, 1.0] can infer that the evolution direction is accelerated deterioration. This analysis can provide a basis for early warning and avoid sudden collapse.

[0085] Furthermore, the method also includes:

[0086] Based on the predicted health status values, spatial distribution mapping and time series projection are performed to obtain the health status distribution map of the target disaster body;

[0087] Based on the health status distribution map, key area parameters are extracted. If the key area parameters are abnormal, the monitoring frequency is adjusted through a dynamic update algorithm to obtain an optimized monitoring scheme. The distributed sensor network settings are then updated according to the optimized monitoring scheme, and the sensor data acquisition and processing flow is adjusted.

[0088] Specifically, in this embodiment, the initial distribution characteristics are obtained by using spatial distribution analysis methods based on health status and predicted values. For these initial distribution characteristics, mapping techniques are applied to obtain spatial distribution mapping results. Based on the spatial distribution mapping results, time series characteristics are obtained through time series analysis. For these time series characteristics, projection techniques are used to obtain health map projection data. If the health map projection data exceeds a preset threshold, the distribution characteristics are determined through disaster body state distribution analysis. Based on the changes in distribution characteristics, a data integration method is used to obtain the adjusted state distribution. Through the adjusted state distribution and the acquisition method, the disaster body health status distribution map is determined.

[0089] Key area parameters are obtained from the health status distribution map. Preset thresholds are used to determine whether the parameters are abnormal, resulting in an anomaly assessment. If an anomaly is found, a frequency adjustment value is calculated using a dynamic update algorithm to determine a new monitoring frequency. Based on the new monitoring frequency, the adjusted monitoring time interval is obtained, resulting in optimized time series data. Data analysis methods are used to extract trends in health status changes and identify change characteristics. If the change characteristics deviate from the normal range, an anomaly detection algorithm is used to identify anomaly patterns and determine the anomaly type. Based on the anomaly type, a frequency adjustment strategy is used to update the monitoring plan, resulting in an optimized monitoring plan.

[0090] An initial data stream is acquired through a sensor network, and a distributed configuration is used to divide the collection area into zones, resulting in zoned monitoring data. Based on the zoned monitoring data and preset thresholds, data acquisition anomalies are identified. If anomalies are detected, the sensor network is adjusted through a feedback mechanism to obtain an updated data stream. For the updated data stream, a processing flow is used to separate noise signals, resulting in filtered monitoring data. The filtered monitoring data is then used in conjunction with an optimization scheme to calculate the trend, resulting in trend analysis results. Based on the trend analysis results, if the trend exceeds a preset range, the data acquisition frequency is adjusted through a control flow to obtain frequency-optimized data. For the frequency-optimized data, a machine learning algorithm is used to detect potential abnormal patterns, resulting in anomaly detection results. Based on the anomaly detection results and configuration management, the distributed configuration is updated to obtain continuously optimized monitoring data.

[0091] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method of monitoring deformation of a mountain collapse disaster body, characterized by, The method comprises the following steps: Collecting multi-dimensional monitoring data of a target disaster body and performing synchronous processing and correlation analysis to obtain a multi-dimensional correlation data set; Extracting feature parameters in the multi-dimensional correlation data set and performing dimension reduction processing to obtain a compressed feature set; Performing feature fusion processing on the compressed feature set to obtain a real-time fusion feature set; If the deformation parameter in the real-time fusion feature set exceeds a preset deformation threshold, inputting the real-time fusion feature set into a pre-trained convolutional neural network to obtain a deformation trend sequence; Inputting the real-time fusion feature set into a pre-trained convolutional neural network to obtain a deformation trend sequence comprises: Respectively inputting the real-time fusion feature set into a convolutional neural network (CNN) and a long short-term memory network (LSTM) to obtain a first deformation trend classification result and a second deformation trend classification result; Calculating a difference value of the first deformation trend classification result and the second deformation trend classification result, if the difference value exceeds a preset difference value, performing weighted fusion processing on the first deformation trend classification result and the second deformation trend classification result to obtain a final deformation trend sequence, if not, taking the first deformation trend classification result as the final deformation trend sequence; Based on the deformation trend sequence, combining historical deformation data and geological environment parameters of the target disaster body, using a pre-trained long short-term memory network to analyze the trend change rule, obtaining a health state prediction value and performing early warning to complete deformation monitoring of the target disaster body; Based on the deformation trend sequence, combining historical deformation data and geological environment parameters of the target disaster body, using a pre-trained long short-term memory network to analyze the trend change rule, obtaining a health state prediction value comprises: Inputting the deformation trend sequence and historical data into a pre-trained bidirectional long short-term memory network (BiLSTM) for change rule analysis to obtain a preliminary health state value; Based on the preliminary health state value and the geological environment parameters, obtaining environmental influence features and analyzing the change rule of the environmental influence features to obtain a rule sequence; Combining the rule sequence with the historical data, judging the stability of the health state and analyzing the fluctuation of the health state to obtain a final health state value.

2. The method of claim 1, wherein Collecting multi-dimensional monitoring data of the target disaster body comprises: Laying out GNSS displacement monitoring points, optical fiber strain sensors and weather sensors on the surface of the target disaster body to collect real-time surface displacement, crack expansion and temperature and humidity data; Laying out multi-parameter sensing chains in the internal borehole of the target disaster body, including microseismographs, inclinometers and pore water pressure gauges, to obtain deep rock mass rupture, inclination change and groundwater activity data.

3. The method of claim 1, wherein, Obtaining the multi-dimensional correlation data set comprises: Timestamp labeling and spatial coordinate calibration of the multi-dimensional monitoring data to obtain a synchronous data set; Based on the synchronous data set, extracting multi-dimensional parameters and checking the multi-dimensional parameters through an algorithm for correlation analysis to obtain the multi-dimensional correlation data set.

4. The method of claim 1, wherein, Extracting feature parameters in the multi-dimensional correlation data set and performing dimension reduction processing to obtain a compressed feature set comprises: Extracting feature parameters from the multi-dimensional correlation data set through an edge computing node and cleaning the feature parameters to construct a feature set; The feature set is subjected to feature decomposition and dimension reduction processing by principal component analysis, and then subjected to compression by t-SNE algorithm to obtain the compressed feature set.

5. The method of claim 1, wherein, The compressed feature set is subjected to feature fusion processing to obtain a real-time fusion feature set, including: The compressed feature set is subjected to multi-node task allocation by a dynamic task scheduling strategy to obtain an allocated task set; The allocated task set is adjusted by a load balancing technique to obtain a balanced task set; The balanced task set is executed in multi-node task processing to obtain a node processing feature set; The node processing feature set is integrated and subjected to feature fusion processing by data flow scheduling to obtain the real-time fusion feature set.

6. The method of monitoring the deformation of a mountain collapse disaster body according to any one of claims 1 to 5, characterized in that, The method further includes: Based on the health state prediction value, a spatial distribution map and a time series projection are performed to obtain a health state distribution map of the target disaster body; Based on the health state distribution map, key area parameters are extracted, and if the key area parameters are abnormal, a dynamic updating algorithm is used to adjust the monitoring frequency, an optimized monitoring scheme is obtained, and the distributed sensor network setting is updated according to the optimized monitoring scheme to adjust the sensor data acquisition and processing flow.

Images