Landslide deformation monitoring data anomaly detection method and system based on real-time data

Abstract

The application belongs to the technical field of computer data processing and analysis, and specifically discloses a landslide deformation monitoring data anomaly detection method and system based on real-time data, which comprises the following steps: when key inducing factors such as heavy rainfall occur, automatic acquisition, self-calibration and quality evaluation are carried out to ensure that complete and accurate original information can be obtained at the moment when accurate data is most needed; deep learning and time-frequency analysis are used to deeply purify the data, and a dynamic calculation of a warning threshold is combined with a geomechanics model and historical data, so that the abnormality judgment standard can adapt to complex and changeable geological environment conditions in real time, and false positives and false negatives caused by dependence on static or subjective threshold values are avoided; through an integrated landslide type identification and trend prediction module, a leap from data anomaly detection to disaster evolution trend perception is realized; through seamless linkage of a warning mechanism and intelligent decision support, the efficiency and pertinence of emergency response are greatly improved.

Description

Technical Field

[0001] This invention belongs to the field of computer data processing and analysis technology, and relates to a method and system for detecting anomalies in landslide deformation monitoring data based on real-time data. Background Technology

[0002] Landslides are a common geological hazard. Their deformation monitoring involves deploying various sensors on the landslide body to collect real-time data on surface displacement, internal strain, and groundwater levels. This data is then analyzed using computer data processing technology to determine the stability of the landslide. This data-driven monitoring method is one of the core technologies for landslide prevention and mitigation. Its goal is to promptly detect abnormal deformations through data processing and analysis, providing a scientific basis for early warning and emergency response.

[0003] Currently, in landslide monitoring data processing practices, data is typically collected from each monitoring point at a fixed frequency, and then compared with pre-set fixed thresholds. An alarm is triggered when the monitored data exceeds the threshold. For data processing, some methods employ simple filtering algorithms to denoise the data, or manually remove obvious errors from the data. The entire process mainly relies on computer systems for data storage, basic processing, and simple threshold comparisons.

[0004] However, existing technologies have some shortcomings in data processing and analysis. First, the fixed data acquisition frequency is difficult to adapt to sudden disaster-causing factors such as rainfall, which may lead to the omission of key data details during the accelerated deformation stage of landslides. Second, the data processing methods are relatively simple and are not effective in handling complex environmental noise, equipment interference, and long-term baseline drift in multi-source monitoring data. In addition, the use of fixed, experience-based alarm thresholds cannot adapt to changes in different geological conditions and dynamic environments, which can easily lead to false alarms or missed alarms, reducing the reliability of early warnings. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art and to achieve the above objectives, the present invention proposes the following technical solution: a landslide deformation monitoring data anomaly detection method based on real-time data, comprising: S1, acquiring multi-dimensional monitoring data in real time through a multi-type sensor fusion array.

[0006] S2. Extract real-time rainfall and historical meteorological data, generate a strong rainfall response signal through a dynamic judgment function of instantaneous rainfall intensity and cumulative rainfall, increase the sensor acquisition frequency after activation, and start health diagnosis and data quality assessment, outputting multi-dimensional data with quality scores.

[0007] S3. Perform intelligent analysis on multi-dimensional data with quality scores to eliminate environmental noise and baseline drift, and generate purification time-series data.

[0008] S4. Combining purification time series data, real-time geological data, and historical data patterns, a dynamic threshold generation algorithm is used to calculate the dynamic threshold of landslide deformation with confidence interval.

[0009] S5. When the multi-dimensional monitoring data exceeds the dynamic threshold of landslide deformation of the corresponding data, the landslide type and deformation stage are identified by combining the historical case library, and deformation trend prediction results are generated, including the development trend of deformation in the future time period and the comprehensive probability of landslide occurrence.

[0010] S6. If the predicted probability of landslide occurrence based on deformation trends exceeds the weighted probability threshold set by experience, a graded early warning and intelligent decision-making process will be triggered, and early warning information and emergency response suggestions will be output.

[0011] The second aspect of the present invention provides a landslide deformation monitoring data anomaly detection system based on real-time data, comprising: a data acquisition module, which acquires multi-dimensional monitoring data in real time through a multi-type sensor fusion array.

[0012] The quality assessment module extracts real-time rainfall and historical meteorological data, generates a heavy rainfall response signal through a dynamic judgment function of instantaneous rainfall intensity and cumulative rainfall, increases the sensor acquisition frequency after activation, and initiates health diagnosis and data quality assessment, outputting multi-dimensional data with quality scores.

[0013] The data cleansing module intelligently analyzes multi-dimensional data with quality scores to eliminate environmental noise and baseline drift, generating cleansing time-series data.

[0014] The dynamic threshold generation module combines purification time-series data, real-time geological data, and historical data patterns, and uses a dynamic threshold generation algorithm to calculate the dynamic threshold of landslide deformation with confidence intervals.

[0015] The landslide deformation prediction module, when multi-dimensional monitoring data exceeds the corresponding dynamic threshold for landslide deformation, combines historical case libraries to identify landslide types and deformation stages, and generates deformation trend prediction results, including the development trend of deformation in the future time period and the overall probability of landslide occurrence.

[0016] The early warning and decision-making module triggers tiered early warning and intelligent decision-making if the predicted comprehensive landslide probability exceeds the weighted probability threshold set by experience, and outputs early warning information and emergency response suggestions.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) The present invention effectively improves the quality and timeliness of monitoring data by constructing an adaptive data acquisition and quality assessment mechanism. When key inducing factors such as heavy rainfall occur, the system can automatically encrypt the acquisition and perform self-calibration and quality assessment, ensuring that complete and accurate original information can be obtained at the moment when accurate data is most needed, thus guaranteeing the reliability of subsequent analysis from the source and overcoming the problem of information loss or distortion in the traditional fixed-frequency monitoring method when dealing with emergencies.

[0018] (2) This invention significantly improves the accuracy of anomaly detection by introducing a multi-technology integrated intelligent data analysis and dynamic threshold generation method. The system uses deep learning and time-frequency analysis to deeply purify the data, and combines geomechanical models and historical data to dynamically calculate the early warning threshold, so that the anomaly judgment criteria can adapt to complex and ever-changing geological environmental conditions in real time, avoiding false alarms and missed alarms caused by relying on static or subjective thresholds, and enhancing the robustness and intelligence level of the monitoring system.

[0019] (3) This invention achieves a leap from data anomaly detection to disaster evolution awareness through an integrated landslide type identification and trend prediction module. When data anomalies occur, the system can not only issue an alarm, but also automatically diagnose the inherent type and deformation stage of the landslide, and predict its future development trend, providing managers with in-depth and forward-looking decision-making information, realizing a precise profile of the risk, and changing the previous extensive early warning mode that simply relied on deformation exceeding the limit.

[0020] (4) This invention significantly improves the efficiency and targeting of emergency response through the seamless linkage of early warning mechanism and intelligent decision support. The system combines the early warning information obtained from in-depth analysis with geographic information system and emergency resource information to automatically generate and issue emergency response suggestions containing specific evacuation routes and resource allocation plans. This realizes an intelligent closed loop from monitoring data to specific action instructions, automates and optimizes the decision-making process, and ensures the speed, accuracy and coordination of emergency actions. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of 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.

[0022] Figure 1 This is a schematic diagram of the implementation steps of the method of the present invention.

[0023] Figure 2 This is a schematic diagram of the system module connections of the present invention. Detailed Implementation

[0024] 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.

[0025] Example 1

[0026] Please see Figure 1 As shown, the landslide deformation monitoring data anomaly detection method based on real-time data proposed in this invention includes: S1, collecting multi-dimensional monitoring data in real time through a multi-type sensor fusion array, including rainfall, geological structure, groundwater level, etc.

[0027] The deployment of this multi-type sensor fusion array aims to achieve real-time and accurate monitoring of landslide deformation by comprehensively monitoring multi-dimensional information of the landslide body (such as displacement, groundwater level, soil moisture, and rainfall), providing a scientific basis for disaster early warning and emergency decision-making. Its core is to capture early signs of landslide deformation through the fusion and analysis of sensor data, and dynamically adjust monitoring strategies to adapt to environmental changes. Specifically, it includes rain gauges, displacement sensors, soil moisture meters, groundwater level monitors, ground temperature sensors, inclinometers, and accelerometers.

[0028] S2. Extract real-time rainfall and historical meteorological data, generate a strong rainfall response signal through a dynamic judgment function of instantaneous rainfall intensity and cumulative rainfall, increase the sensor acquisition frequency after activation, and start health diagnosis and data quality assessment, outputting multi-dimensional data with quality scores.

[0029] In a preferred embodiment, the step of extracting real-time rainfall and historical meteorological data, generating a strong rainfall response signal through a dynamic judgment function of instantaneous rainfall intensity and cumulative rainfall, and increasing the sensor acquisition frequency after activation includes: acquiring real-time rainfall data, including instantaneous rainfall intensity and cumulative rainfall, through a rain gauge.

[0030] Real-time rainfall data is input into a dynamically set conditional judgment function, which simultaneously retrieves historical meteorological data for the region from a local database as a reference benchmark.

[0031] The dynamically set condition judgment function is implemented through a comprehensive rainfall event index, which combines the two dimensions of rainfall intensity and duration and compares them with historical statistical patterns. This rainfall event index can be expressed as: ,in This represents the calculated real-time rainfall event index; This indicates the instantaneous rainfall intensity extracted from real-time rainfall data; This indicates the cumulative rainfall within a specific time window in the past; and It represents the instantaneous rainfall intensity and the corresponding cumulative rainfall threshold for typical meteorological conditions (such as high temperature, high humidity, or hail) statistically represented from historical meteorological data. and These are preset weighting coefficients, which sum to one, used to adjust the relative importance of instantaneous rainfall intensity and cumulative rainfall in the judgment.

[0032] When the output value of the function exceeds the preset trigger threshold, a heavy rainfall response trigger signal is generated.

[0033] The heavy rainfall response trigger signal includes two parallel commands: the first is a sensor array acquisition frequency adjustment command, which represents a digital command containing new, higher acquisition frequency parameters, sent to the control modules of all sensors within the monitoring area. This switches them from the conventional low-frequency monitoring mode to a high-frequency emergency monitoring mode. The conventional low-frequency monitoring mode acquires data, for example, every 30 minutes or 1 hour, while the high-frequency emergency monitoring mode acquires data, for example, every 1 minute or even shorter intervals. The second is a self-calibration command, which triggers the sensor's internal self-test and calibration procedures. For example, for a displacement sensor, it might recalibrate the zero point to eliminate drift caused by sudden temperature changes; for a soil moisture meter, it might perform a remeasurement of the reference dielectric constant. These two commands are triggered synchronously, ensuring that the entire sensor network is in optimal working condition before high-frequency data acquisition begins.

[0034] This invention's pre-judgment and triggering mechanism provides an intelligent "switch" for the adaptiveness of the entire monitoring system. By introducing historical meteorological data as a dynamic benchmark, it enables response conditions to accurately match local climate characteristics and disaster patterns, avoiding misjudgments or omissions caused by using universal static thresholds. More importantly, it cleverly binds and synchronizes the two actions of increasing the acquisition frequency and initiating sensor self-calibration. This synergy ensures that during the most critical periods of heavy rainfall, the system not only captures more data details (high frequency) but also guarantees the accuracy and reliability of this key data (self-calibration). It fundamentally solves the problem of traditional monitoring methods being "visible but inaccurate" or "accurate but incomplete" when dealing with sudden environmental changes, providing the highest quality source data input for all subsequent data analysis and early warning decisions.

[0035] In a further preferred embodiment, the step of initiating health diagnosis and data quality assessment and outputting multi-dimensional data with quality scores includes: performing integrity verification and outlier detection on the multi-dimensional monitoring data to generate preliminary verification data.

[0036] Integrity verification identifies data gaps in the time series by comparing the timestamps of data frames with a preset data collection frequency. Outlier detection employs statistical methods, such as calculating the mean and standard deviation of data within a set time window, and marking data points exceeding a predetermined multiple of the standard deviation (e.g., three times the standard deviation) as outliers. This step aims to identify missing points and statistically significant outliers in the data stream, providing clear targets for subsequent data correction and filtering, and avoiding unnecessary processing of normal data.

[0037] Through environmental adaptive adjustment and self-calibration functions, noise filtering and signal enhancement are performed on the preliminary verification data to generate calibrated monitoring data.

[0038] This process employs an adaptive filtering algorithm that dynamically adjusts the filter coefficients based on real-time environmental parameters such as temperature and humidity changes, effectively suppressing noise introduced by environmental factors. Simultaneously, for marked outliers, the system performs smooth replacement or interpolation repair based on the trends of neighboring data points, using methods such as linear interpolation, multinomial interpolation, or spline interpolation. For identified missing data, methods such as spline interpolation are used to fill in the gaps. Spline interpolation constructs a smooth curve to connect known data points before and after the missing data point, thus obtaining a reasonable estimate, thereby enhancing the effective deformation signal and ensuring the continuity and smoothness of the data sequence.

[0039] By combining the calibrated monitoring data, real-time environmental parameters, and sensor status parameters, a quantitative data quality score is generated. This score is then combined with the calibrated monitoring data to generate multi-dimensional monitoring data with a quality score.

[0040] The self-state parameters are provided by the sensor's self-diagnostic program and cover key performance indicators such as sensor operating voltage and internal temperature.

[0041] The data quality scoring process is implemented using a weighted model, as shown in the following formula: ,in This represents the final data quality score, which is a standardized numerical value. The data integrity score is calculated from the data missing rate, which refers to the percentage of outlier data points detected within the expected time window out of the total expected data points. The signal-to-noise ratio (SNR) score is quantified by evaluating the change in data variance before and after filtering. , and This means calculating the variance of the data before and after filtering, and mapping the signal-to-noise ratio to a score range of 0 to 1; The health status score represents the equipment's health status, which is assessed based on the deviation between the sensor's own status parameters and standard operating parameters. The deviation can be quantified by calculating the absolute or relative value of the difference between the actual value and the standard value, and mapped to a score range of 0 to 1 according to the magnitude of the deviation. , and This represents the weighting coefficients for each score item, which sum to one. These coefficients are preset based on the importance of each factor in different monitoring scenarios to ensure the scoring accurately reflects the overall data quality. For example, in environments with significant environmental interference, the weight of the signal-to-noise ratio score may need to be increased; while in applications with extremely high data integrity requirements, the weight of the data integrity score may need to be increased. This quality score not only quantifies the reliability of the current data but also provides a dynamic weighting basis for subsequent multi-source data fusion analysis.

[0042] This invention, through the synergistic effect of the aforementioned steps, achieves a dynamic, closed-loop data quality assurance process. It not only passively cleans and repairs data but also proactively adapts processing strategies based on environmental and equipment conditions, quantifying the processing results into quality scores. This step-by-step refinement process, from initial data acquisition to quality assessment, ensures that the data input to subsequent analysis modules is high-quality and carries a credibility label. This significantly improves the robustness and accuracy of the entire landslide deformation monitoring data anomaly detection method, effectively avoiding the problem of minor issues in the source data having a significant impact on the results. Therefore, it lays a solid data foundation for achieving accurate early warning.

[0043] S3. Perform intelligent analysis on multi-dimensional data with quality scores to eliminate environmental noise and baseline drift, and generate purification time-series data.

[0044] In a preferred embodiment, the intelligent analysis of multi-dimensional data with quality scores to eliminate environmental noise and baseline drift and generate purification time-series data involves the following steps: multi-source heterogeneous data fusion of the multi-dimensional monitoring data after quality assessment to generate standardized fused data containing data quality score weights.

[0045] Specifically, since the multi-dimensional monitoring data comes from different types of sensors, their physical dimensions, numerical ranges, and statistical distributions vary, standardization is necessary. For each dimension of data, the Z-score standardization method is used to transform it to a uniform dimensionless scale, as calculated below. ,in This represents the standardized data points; Represents the original data points; This represents the average value of the data in this dimension over a sliding time window. The size of the sliding time window can be adjusted according to actual needs, such as setting it to data from the past 1 hour or 24 hours, in order to capture local features of the data. This represents the standard deviation within the same time window. During the standardization process, based on the previously generated data quality scores, corresponding weights are assigned to each dimension of the data. Dimensions with higher data quality scores have greater weights in the fused data, and vice versa. The standardized data for each dimension are multiplied by their corresponding data quality score weights, and then summed to generate standardized fused data that includes the data quality score weights. This step ensures that the fused data not only includes monitoring information from each sensor but also reflects the quality status of the data.

[0046] A trained deep learning denoising model is used to process the standardized fused data and generate a denoised data sequence.

[0047] The deep learning denoising model described is a denoising model based on an autoencoder neural network. This model consists of an encoder and a decoder. The encoder is responsible for compressing the input data into a low-dimensional representation, while the decoder is responsible for reconstructing the low-dimensional representation back to the original data dimensions. The model is trained on a large amount of historical data, covering monitoring data under different operating conditions and noise environments. The model parameters are adjusted using optimization algorithms (such as stochastic gradient descent) so that the model can learn to extract core signal features from noisy input data and reconstruct a noise-free, clean signal by the decoder. After processing by the deep learning denoising model, most common noise in the data is effectively suppressed. However, some specific types of noise may still exist, possessing unique frequency characteristics, requiring further targeted processing.

[0048] High-frequency components and potential abnormal patterns in the denoised data sequence are identified by time-frequency joint analysis, and purified time-series data are generated by combining adaptive filtering and dynamic baseline correction techniques.

[0049] Specifically, the process first employs continuous wavelet transform for joint time-frequency analysis. Wavelet transform decomposes a one-dimensional time-series signal into two dimensions: time and frequency. This clearly reveals the evolution characteristics of high-frequency components and anomalous patterns over time, which is crucial for identifying weak high-frequency vibration signals in the early stages of accelerated landslide deformation. In actual landslide monitoring environments, different types of noise often have specific frequency ranges. For example, mechanical vibration noise may be concentrated in a relatively high frequency band, while environmental interference noise (such as wind noise, electromagnetic interference, etc.) may be distributed in other frequency bands. By carefully analyzing the time-frequency diagrams and combining known noise source characteristics with practical monitoring experience, these specific noise frequency bands can be accurately identified.

[0050] Building upon this foundation, adaptive filtering technology is employed to design and apply targeted digital filters based on specific noise frequency bands identified through time-frequency analysis. These filters effectively eliminate noise interference. Simultaneously, dynamic baseline correction technology is used to eliminate baseline drift caused by seasonal variations or instrument aging during long-term monitoring. This technology extracts the long-term trend term of the signal—the dynamic baseline—by performing low-pass filtering or long-period moving average. This dynamic baseline is then subtracted from the denoised data sequence to obtain accurate purification time-series data reflecting the true short-term deformation of the landslide.

[0051] In this invention, the intelligent real-time data analysis engine organically combines data fusion, deep learning noise reduction, and time-frequency analysis with dynamic baseline correction to achieve a progressively layered processing flow from multi-source to single-source, from coarse to fine, and from signals containing noise and trends to clean, deformed signals. This multi-technology collaborative processing approach not only overcomes the bottleneck of limited processing capabilities of single technologies but also significantly enhances the ability to capture weak anomaly signals in complex backgrounds through the introduction of deep learning and time-frequency analysis. It transforms raw, mixed monitoring data into a highly purified, information-concentrated time-series data stream, thereby greatly improving the accuracy and sensitivity of subsequent dynamic threshold determination and landslide trend identification, providing high-quality data support for reliable early warning.

[0052] S4. Combining purification time series data, real-time geological data, and historical data patterns, a dynamic threshold generation algorithm is used to calculate the dynamic threshold of landslide deformation with confidence interval.

[0053] In a preferred embodiment, the step of combining purification time-series data, real-time geological data, and historical data patterns, and using a dynamic threshold generation algorithm to calculate the dynamic threshold of landslide deformation with confidence intervals includes: calculating and generating an initial threshold range based on geologically relevant multi-dimensional data and a geomechanical model.

[0054] The geologically relevant multi-dimensional data includes geological structure data, slope data, aspect data, groundwater level change data, soil moisture data, ground temperature data, and rainfall intensity and pattern data for short-term heavy rainfall. Specifically, geological structure data is acquired using specialized geological exploration equipment (such as ground-penetrating radar and borehole samplers) to clarify key information such as soil and rock types, bedding structures, and fault distribution; slope and aspect data are measured using high-precision slope meters and aspect meters; groundwater level change data, soil moisture data, and ground temperature data are collected in real time by deploying groundwater level monitoring instruments, soil moisture sensors, and ground temperature sensors; and rainfall intensity and pattern data for short-term heavy rainfall are recorded by setting up rainfall monitoring stations.

[0055] The geomechanical model is essentially a landslide stability assessment function based on physical principles. This function comprehensively considers multiple factors such as geological structure, geometric parameters, and hydrogeological conditions. Based on physical principles, it describes the stability state of the landslide body under various forces. It uses multi-dimensional geological data as input parameters to calculate the theoretical value of the deformation or deformation rate corresponding to the landslide body reaching the critical instability state under current conditions, using this as an initial threshold. Furthermore, based on the uncertainties of the model calculations and the complexity of actual geological conditions, a reasonable initial threshold range is set. For example, the calculated initial threshold... Using the center value as the reference, the minimum value of the range boundary is determined by fluctuating it up or down by a certain percentage (e.g., 10%-20%). and maximum value .

[0056] The landslide stability assessment function can be expressed as follows: In the formula Indicates the initial threshold; Represents a geomechanical model function; It represents a set of geological parameters, including shear strength indices of soil and rock mass extracted from geological structure data, as well as geometric parameters such as slope and aspect; This represents a set of hydrogeological parameters, including pore water pressure calculated from groundwater level and soil moisture, and surface infiltration determined by rainfall intensity and pattern data. Specifically, the shear strength of the soil and rock mass is determined by a combination of indoor geotechnical tests (such as direct shear tests and triaxial compression tests) and in-situ tests (such as vane shear tests); pore water pressure is calculated based on groundwater level and soil saturation using Terzaghi's effective stress principle; and surface infiltration is estimated based on rainfall intensity and pattern, combined with soil infiltration capacity curves (such as the Horton infiltration curve).

[0057] Calculate the first and second time derivatives of the purification time series data to obtain deformation rate and deformation acceleration characteristics.

[0058] Specifically, numerical differentiation methods (such as the central difference method) are used to calculate: first-order time derivative (deformation rate). In the formula Indicates time Deformation rate at time, and Indicates time Time and Deformation amount at time For time step; second time derivative (deformation acceleration). In the formula Indicates time The deformation acceleration at different times. By analyzing the changes in deformation rate and deformation acceleration, the deformation trend of the landslide body can be grasped in real time.

[0059] By using the calculated first and second time derivatives as key input parameters and combining them with historical data patterns, the initial threshold range is dynamically adjusted to generate a dynamic threshold boundary.

[0060] Preferably, data mining algorithms (such as similarity-based retrieval algorithms) are used to retrieve working conditions in historical databases that are similar to the current geological and hydrological conditions. Similarity judgment criteria include similarity in geological structure (such as rock and soil type, fault distribution, etc.), similarity in geometric parameters (slope, aspect, etc.), and similarity in hydrogeological parameters (groundwater level, soil moisture, rainfall intensity, and rain pattern, etc.). Then, the actual deformation data under the selected similar working conditions are analyzed in depth to statistically analyze the range and trend of deformation amount or deformation rate when abnormal deformation occurs under these working conditions. For example, by plotting the deformation amount over time, the temporal and deformation characteristics before the occurrence of abnormal deformation can be analyzed.

[0061] Furthermore, the initial threshold is corrected by setting a dynamic adjustment factor, which is a function of historical data patterns and current trends, so that the threshold boundary can reflect the dynamic response characteristics of the landslide in real time. For example, the dynamic adjustment factor... ,in and These represent the deformation rate and deformation acceleration at the current moment, respectively. and These represent the average deformation rate and average deformation acceleration under similar historical working conditions, respectively. and This is a weighting coefficient, which can be adjusted according to the actual situation. Then, the dynamic threshold boundary is obtained by multiplying the dynamic adjustment factor by the initial threshold.

[0062] An uncertainty quantification method is used to evaluate the confidence level of the dynamic threshold boundary and generate a dynamic threshold with confidence intervals.

[0063] Specifically, sensor measurements, model parameters, and the physical model itself all contain uncertainties. For example, during the measurement process, sensors are affected by environmental factors (such as temperature, humidity, electromagnetic interference, etc.), instrument accuracy, and stability, leading to certain errors in the measurement data. By using methods such as Bayesian inference, all input parameters and model parameters are treated as probability distributions rather than deterministic values, thereby calculating the posterior probability distribution of the dynamic threshold.

[0064] Simultaneously, based on historical data and expert experience, the prior probability distribution of each parameter is determined. For example, the internal friction angle in the shear strength index of soil and rock mass... Assuming it follows a normal distribution ,in and These are the mean and variance, respectively. The mean and variance can be determined through statistical analysis of historical test data.

[0065] Furthermore, from the generated dynamic threshold samples, confidence intervals are calculated based on the required confidence level (e.g., 95%). For example, for a 95% confidence interval, the sample data are arranged in ascending order, and the 2.5th and 97.5th quantiles are used as the lower and upper limits of the confidence interval. When monitoring data enters this confidence interval, the system issues a warning signal, prompting staff to closely monitor the deformation of the landslide; when the monitoring data exceeds the upper boundary of the interval, a higher-level alarm is triggered, and corresponding preventative measures are taken promptly.

[0066] This invention fundamentally changes the traditional approach of setting thresholds based on static experience by combining physical models with data-driven methods. First, a scientifically based initial threshold is established. Then, historical data and real-time dynamics are used to intelligently fine-tune it. Finally, uncertainty quantification is introduced to give the threshold a probabilistic meaning. This progressive generation mechanism ensures that the final threshold is not only dynamically changing but also robust and reliable. It can accurately adapt to the unique characteristics of each monitoring point under different times and environmental conditions, thereby greatly improving the accuracy of anomaly detection and effectively reducing false alarms and missed alarms caused by inappropriate thresholds. This provides core technical support for accurate early warning of landslide disasters.

[0067] S5. When the multi-dimensional monitoring data exceeds the dynamic threshold of landslide deformation of the corresponding data, the landslide type and deformation stage are identified by combining the historical case library, and deformation trend prediction results are generated, including the development trend of deformation in the future time period and the comprehensive probability of landslide occurrence.

[0068] In a preferred embodiment, the step of combining a historical case database to identify landslide types and deformation stages, and generating deformation trend prediction results includes: obtaining deformation rate characteristics reflecting deformation speed and deformation mode characteristics indicating whether the deformation is uniform, accelerating, or decelerating based on the first and second time derivatives of the purified time series data; forming soil and rock parameter characteristics by combining the physical and mechanical parameters of the soil and rock mass obtained from the original multi-dimensional monitoring data, such as cohesion and internal friction angle; and integrating these three types of features into a multi-feature vector.

[0069] Multiple feature vectors are input into a trained classification model, which compares them with feature patterns in a historical landslide case library to output the landslide type and deformation stage.

[0070] Specifically, the historical landslide case database stores a large amount of complete monitoring data on past landslide events, corresponding multi-features, and landslide type and deformation stage labels determined by experts. The matching process is implemented through a pre-trained deep learning classification model, such as a convolutional neural network or support vector machine, which learns the feature vector patterns corresponding to different landslide types and deformation stages. Inputting the current multi-feature vectors into the model, the model outputs the most probable landslide type, such as shove-type, traction-type, or a combination thereof, and its current deformation stage, such as the initial creep stage, the accelerated deformation stage, or the imminent landslide stage.

[0071] Based on the identified deformation stages, appropriate dynamic trend prediction techniques are selected to analyze the purification time series data and generate deformation trend prediction results.

[0072] The selection of appropriate dynamic trend prediction techniques includes, for example, employing a time series prediction model based on Long Short-Term Memory (LSTM) networks if the deformation is in an accelerated phase. LSTM models can capture long-term dependencies in data and are suitable for predicting nonlinear time series. This model takes recent cleanup time series data as input and predicts the deformation trend over a future period, such as predicting the peak deformation that may be reached in the next few hours or days and its spatial expansion range, and outputs the probability of landslide occurrence based on the dynamic trend prediction model.

[0073] Meanwhile, by analyzing the proportion or frequency of landslides that actually occurred under similar historical conditions, a landslide occurrence probability based on historical data patterns is obtained; a landslide stability assessment function based on current geological structure, slope, groundwater level, and other conditions is evaluated by a geomechanical model, which serves as the landslide occurrence probability based on current geological conditions; and then, these three landslide occurrence probabilities are weighted and summed to obtain the comprehensive landslide occurrence probability.

[0074] This invention tightly couples feature extraction, pattern recognition, and dynamic prediction, enabling not only simple anomaly detection but also a deep dive into the underlying physical meaning and evolutionary patterns of the data to provide a clear picture of the intrinsic mechanisms and future behavior of landslides. This combination of intelligent identification and dynamic prediction based on multi-feature fusion allows the early warning system to provide decision support information far exceeding traditional methods. It enables decision-makers to clearly understand the type of landslide, its current dangerous stage, and its most likely development trend, thereby allowing them to formulate more targeted, timely, and effective emergency plans. This represents a shift from passive response to proactive prediction, significantly improving the scientific rigor and foresight of landslide disaster early warning systems.

[0075] In a further preferred embodiment, the further identification steps of the landslide type and deformation stage are as follows: based on different feature combinations in multiple feature vectors, a pre-trained classifier is used to output a landslide type identifier indicating the landslide type in parallel, including lateral, traction, and composite types.

[0076] Specifically, the classifier learns from a large number of multi-feature vectors in a historical landslide case database and their corresponding expert-annotated landslide types. The feature vectors of scour landslides typically show significant deep shear surface activity, while traction landslides show obvious tension characteristics in the upper part of the slope. Composite landslides, on the other hand, usually exhibit a mixture of scour and traction features in their feature vectors. This classifier analyzes the combination and weights of each feature component in the current multi-feature vector, mapping it to the most probable landslide type category to generate a landslide type identifier.

[0077] Based on the temporal evolution of deformation rate features and deformation mode features in multiple feature vectors, the classifier outputs a deformation stage identifier in parallel, which indicates the deformation stage, including the initial creep stage, the accelerated deformation stage, and the slip stage.

[0078] This task relies not only on the instantaneous values ​​of multiple feature vectors, but also on their temporal evolution. The system continuously tracks the deformation rate and its acceleration: the initial creep stage is characterized by a low and relatively stable deformation rate, with the acceleration value fluctuating slightly around zero; when the deformation rate begins to increase continuously and non-linearly, and the acceleration is a significant positive value, the landslide is determined to have entered the accelerated deformation stage; furthermore, the system calculates a key stage characteristic parameter, namely the reciprocal of the deformation rate. In the pre-slide stage of a landslide, the reciprocal of the deformation rate decreases linearly with time and tends to zero. When the system detects this clear linear trend, it determines that the landslide has entered the pre-slide stage.

[0079] The landslide type identifier and deformation stage identifier are integrated to form the final identification result of landslide type and deformation stage.

[0080] This invention provides unprecedented diagnostic depth to the entire early warning system through refined and parallel identification of landslide types and deformation stages. Clearly defining whether a landslide is lateral or traction-driven directly relates to its instability mechanism and potential impact range, providing crucial information for emergency response force deployment and evacuation area delineation. Simultaneously, accurately classifying deformation stages, particularly identifying the transition from accelerated deformation to the imminent landslide stage, provides decisive scientific support for ensuring the timeliness of early warnings. Combining type identification with stage classification creates a synergistic effect, enabling the system to more accurately assess landslide hazards. This represents a qualitative leap from "detecting anomalies" to "understanding the nature and urgency of the anomalies," significantly enhancing the scientific rigor and relevance of early warning decisions.

[0081] S6. If the predicted probability of landslide occurrence based on deformation trends exceeds the weighted probability threshold set by experience, a graded early warning and intelligent decision-making process will be triggered, and early warning information and emergency response suggestions will be output.

[0082] In a preferred embodiment, the triggering of graded early warning and intelligent decision-making includes: determining the early warning level in a risk matrix based on the comprehensive landslide occurrence probability predicted by the deformation trend and the identified deformation stage, and generating a graded early warning signal.

[0083] Specifically, this process is achieved through a pre-defined two-dimensional risk matrix. The two dimensions of the matrix are the deformation stage (e.g., initial creep, accelerated deformation, imminent landslide) and the probability of landslide occurrence (e.g., low, medium, high). Each cell in the matrix is ​​predefined with a specific warning level, such as "blue, yellow, orange, red". For example, when the system identifies that a landslide is in the "accelerated deformation" stage and predicts its probability of occurrence as "high", it will trigger an "orange" warning level and generate a digital graded warning signal containing this level information.

[0084] Based on the graded early warning signals, geographic information system technology and emergency resource distribution information are invoked to automatically generate evacuation route plans and rescue force deployment schemes.

[0085] In this process, the system first marks the potential landslide impact area on the GIS map based on the deformation trend prediction results, that is, the spatial location and expansion range where the deformation may reach its peak in the next few hours or days. Then, using a dynamic path planning algorithm, after excluding roads within the impact area, the system calculates the optimal evacuation route to the preset safe zone for the residential areas within the impact area. At the same time, the system queries the emergency resource database, which contains the location, status, and equipment information of surrounding fire, medical, and engineering rescue teams. Based on the principle of "proximity and professional matching," the system automatically generates a rescue force deployment plan, suggesting which teams to send to which assembly point.

[0086] Through multiple communication channels, early warning information, including the warning level, as well as emergency response suggestions, including evacuation route planning and rescue force deployment plans, are simultaneously sent to relevant departments and personnel.

[0087] Specifically, the system integrates warning information, including alert levels, with emergency response recommendations containing specific plans. This information is then disseminated through various communication channels, such as SMS messages, dedicated app push notifications, and emergency broadcast systems, to pre-defined decision-makers in relevant departments, emergency responders, and the public in affected areas. The information sent to decision-makers and professionals includes complete evacuation route plans and detailed rescue force deployment schemes, while the information sent to the public consists of concise warning levels and specific evacuation instructions.

[0088] This invention achieves closed-loop management from monitoring data to emergency action by seamlessly integrating three stages: early warning level determination, intelligent solution generation, and multi-channel information dissemination. It goes beyond simply issuing an alarm; through intelligent connectivity, it tightly binds abstract hazard levels with specific, actionable geospatial solutions. This deep integration of early warning and decision support eliminates reliance on manual on-site judgment and information aggregation for emergency response. Instead, it automatically generates a scientific and customized action plan the moment a hazard is identified. This significantly shortens the delay from early warning to response, improves the efficiency and targeting of emergency response, and directly transforms the value of data analysis into practical action to protect life and property.

[0089] In a further preferred embodiment, the emergency response suggestion is further generated in the following steps: based on the warning level and the acquired real-time environmental and terrain data, emergency evacuation area identifiers are dynamically delineated and generated in the geographic information system.

[0090] The environmental data includes wind direction and wind speed; the terrain data includes digital elevation models (DEMs), which accurately represent the undulations of the Earth's surface and its associated features (such as bare land after buildings and vegetation have been removed) in a digital form.

[0091] This process is completed on a Geographic Information System (GIS) platform. The system first marks the predicted landslide impact area on a map, and then applies a dynamically wide buffer zone to the impact area based on the warning level. More importantly, the system uses topographic data to perform flow path analysis, simulating the potential spread paths of subsequent debris flows triggered by the landslide, and including these path areas in the risk zone. Simultaneously, combined with real-time wind direction data, it delineates the dust diffusion impact area that may be generated by the landslide. The union of these areas constitutes a clearly defined, geographically defined emergency evacuation zone marker visualized on the map.

[0092] Based on evacuation route planning and real-time updated rescue force deployment information, a resource allocation plan is generated.

[0093] The information on the deployment of rescue forces includes the real-time location, number of personnel, type of equipment, and availability of each emergency team.

[0094] Generating a resource scheduling plan is a multi-objective optimization process. For each designated refuge location and key traffic nodes along evacuation routes, the system assigns the optimal team from available rescue resources in the surrounding area. The optimization objective function comprehensively considers factors such as shortest arrival time, best matching of professional capabilities, and most balanced task load. For example, the system prioritizes assigning the nearest traffic police team to congestion points along evacuation routes to manage traffic, and assigns search and rescue teams equipped with life detectors to stand by in densely populated areas closest to the landslide. The final output resource scheduling plan is a detailed set of instructions, specifying which team, when, via which route, to where, and what task to perform.

[0095] The system integrates emergency evacuation zone markings, evacuation route planning, and resource allocation schemes to generate emergency response recommendations.

[0096] Specifically, the system integrates the emergency evacuation zone markers generated in the first step with the evacuation routes generated in the preceding steps and the resource allocation plan generated in the second step at the data level, packaging them into a structured, machine-readable emergency response suggestion data. This data not only includes text and map information for human reading, but more importantly, it is a standardized data package that can be directly parsed and executed by other subsystems of the emergency command platform or mobile terminals, achieving seamless information flow and automated processing.

[0097] This invention, through the aforementioned steps, decomposes a macro-level emergency response recommendation into three specific, executable, quantifiable, and dynamically generated components. By introducing real-time environmental and terrain data, it transforms the delineation of evacuation zones from static and empirical to dynamic, scientific, and precise. By combining route planning with optimized resource scheduling, it ensures that the two core issues of "where to evacuate" and "who will guarantee the evacuation" are addressed simultaneously and collaboratively. The resulting emergency response recommendation data package achieves seamless integration from decision-making to execution. This mechanism transforms emergency response from a vague instruction into a digital operational plan precise down to spatial locations, timeframes, and responsible units, significantly improving the accuracy, timeliness, and coordination of emergency response.

[0098] Example 2

[0099] Please see Figure 2As shown, based on Embodiment 1, the second aspect of the present invention provides a landslide deformation monitoring data anomaly detection system based on real-time data, comprising: a data acquisition module, a quality assessment module, a data purification module, a dynamic threshold generation module, a landslide deformation prediction module, and an early warning decision module.

[0100] The data acquisition module, quality assessment module, data purification module, dynamic threshold generation module, landslide deformation prediction module, and early warning decision module are connected in sequence.

[0101] The data acquisition module collects multi-dimensional monitoring data in real time through a multi-type sensor fusion array.

[0102] The quality assessment module extracts real-time rainfall and historical meteorological data, generates a strong rainfall response signal through a dynamic judgment function of instantaneous rainfall intensity and cumulative rainfall, increases the sensor acquisition frequency after activation, and initiates health diagnosis and data quality assessment, outputting multi-dimensional data with quality scores.

[0103] The data purification module intelligently analyzes multi-dimensional data with quality scores to eliminate environmental noise and baseline drift, generating purification time-series data.

[0104] The dynamic threshold generation module combines purification time-series data, real-time geological data, and historical data patterns, and uses a dynamic threshold generation algorithm to calculate the dynamic threshold of landslide deformation with confidence intervals.

[0105] The landslide deformation prediction module, when the multi-dimensional monitoring data exceeds the corresponding dynamic threshold for landslide deformation, combines the historical case library to identify the landslide type and deformation stage, and generates a deformation trend prediction result, including the deformation development trend and the overall probability of landslide occurrence in the future time period.

[0106] If the predicted probability of a landslide due to deformation trends exceeds the weighted probability threshold set by experience, the early warning and decision-making module will trigger tiered early warning and intelligent decision-making, and output early warning information and emergency response suggestions.

[0107] The above content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined by the present invention, and all such modifications and additions should fall within the protection scope of the present invention.

Claims

1. A landslide deformation monitoring data anomaly detection method based on real-time data, characterized in that, The method comprises the following steps: S1, real-time collection of multi-dimensional monitoring data through a multi-type sensor fusion array; S2, extraction of real-time rainfall and historical meteorological data, generation of a strong rainfall response signal through a dynamic judgment function of instantaneous rainfall intensity and cumulative rainfall, activation of the sensor to increase the collection frequency, and initiation of health diagnosis and data quality assessment, output of multi-dimensional data with quality scores; S3, intelligent analysis of the multi-dimensional data with quality scores to eliminate environmental noise and baseline drift, and generation of purified time series data; S4, combination of the purified time series data, real-time geological data, and historical data rules, and use of a dynamic threshold generation algorithm to calculate a landslide deformation dynamic threshold with a confidence interval; S5, when the multi-dimensional monitoring data exceeds the corresponding data of the landslide deformation dynamic threshold, identification of the landslide type and deformation stage in combination with a historical case library, generation of a deformation trend prediction result, and inclusion of the development trend of the deformation amount in the future time period and the comprehensive landslide occurrence probability; S6, if the comprehensive landslide occurrence probability of the deformation trend prediction exceeds an empirically set weighted probability threshold, triggering of a hierarchical early warning and intelligent decision, and output of early warning information and emergency disposal suggestions. The extraction of real-time rainfall and historical meteorological data, and the generation of a strong rainfall response signal through a dynamic judgment function of instantaneous rainfall intensity and cumulative rainfall, and the activation to increase the sensor collection frequency comprise the following steps: acquisition of real-time rainfall data, including instantaneous rainfall intensity and cumulative rainfall, through a rain gauge; ​ ​ ​ ​ ​ ​ ​ ​ 2. The real-time data based landslide deformation monitoring data anomaly detection method according to claim 1, characterized in that, ​ ​ The real-time rainfall data is input into a dynamic setting condition judgment function, which simultaneously calls historical meteorological data of the region from a local database as a reference benchmark; When the output value of the function exceeds a preset trigger threshold, a heavy rainfall response trigger signal is generated; The heavy rainfall response trigger signal contains two parallel instructions: the first is a sensor array acquisition frequency adjustment instruction, which is sent to the control modules of all sensors in the monitoring area, causing them to switch from a regular low-frequency monitoring mode to a high-frequency emergency monitoring mode; the second is a self-calibration instruction, which is used to trigger the self-checking and calibration procedures inside the sensors.

3. The real-time data based landslide deformation monitoring data anomaly detection method of claim 1, wherein, The health diagnosis and data quality assessment are started, and multi-dimensional data with quality scores are output, including: Integrity checking and outlier detection are performed on the multi-dimensional monitoring data to generate preliminary checked data; Through environmental self-adaptive adjustment and self-calibration functions, noise filtering and signal enhancement are performed on the preliminary checked data to generate calibrated monitoring data; The calibrated monitoring data, real-time environmental parameters, and sensor state parameters are integrated to generate a quantitative data quality score, and the score is combined with the calibrated monitoring data to generate multi-dimensional monitoring data with quality scores.

4. The real-time data based landslide deformation monitoring data anomaly detection method according to claim 3, characterized in that, The historical case library is combined to identify the landslide type and deformation stage, and generate a deformation trend prediction result, including: Based on the first-order time derivative and the second-order time derivative of the purified time series data, the deformation rate characteristics reflecting the deformation speed and the deformation mode characteristics representing whether the deformation is uniform speed, acceleration, or deceleration are obtained; combined with the rock-soil physical and mechanical parameters obtained from the original multi-dimensional monitoring data, the rock-soil parameter characteristics are formed; and the three types of characteristics are integrated into a multi-feature vector; The multi-feature vector is input into a trained classification model, which compares the feature patterns in the historical landslide case library and outputs the landslide type and deformation stage; According to the identified deformation stage, the corresponding dynamic trend prediction technology is selected to analyze the purified time series data and generate a deformation trend prediction result.

5. The real-time data based landslide deformation monitoring data anomaly detection method according to claim 4, characterized in that, The further identification steps of the landslide type and deformation stage are as follows: According to different feature combinations in the multi-feature vector, a pre-trained classifier is used to output a landslide type identifier indicating the landslide type, including push type, pull type, and composite type; According to the time evolution of the deformation rate characteristics and the deformation mode characteristics in the multi-feature vector, the classifier is used to output a deformation stage identifier indicating the deformation stage, including the initial creep stage, the accelerated deformation stage, and the impending slide stage; The landslide type identifier and the deformation stage identifier are integrated to serve as the final identification result of the landslide type and deformation stage.

6. The real-time data based landslide deformation monitoring data anomaly detection method of claim 1, wherein, The trigger of the hierarchical warning and intelligent decision-making includes: According to the comprehensive landslide occurrence probability of the deformation trend prediction and the identified deformation stage, the warning level is determined in a risk matrix, and a hierarchical warning signal is generated; According to the hierarchical warning signal, geographic information system technology and emergency resource distribution information are called to automatically generate evacuation route planning and rescue force deployment schemes; Through multiple communication channels, warning information containing the warning level and emergency disposal suggestions containing evacuation route planning and rescue force deployment schemes are sent to relevant departments and personnel simultaneously.

7. The real-time data based landslide deformation monitoring data anomaly detection method according to claim 6, characterized in that, The further generation steps of the emergency disposal suggestion are as follows: According to the early warning level and the obtained real-time environmental data and terrain data, an emergency refuge area identifier is dynamically delimited and generated in a geographic information system; According to the evacuation route planning and the real-time updated rescue force deployment information, a resource scheduling scheme is generated; The emergency refuge area identifier, the evacuation route planning and the resource scheduling scheme are integrated to generate an emergency disposal suggestion.

8. A landslide deformation monitoring data anomaly detection system based on real-time data, which is suitable for the landslide deformation monitoring data anomaly detection method based on real-time data according to any one of claims 1-7, characterized in that, It comprises: A data acquisition module that acquires multi-dimensional monitoring data in real time through a multi-type sensor fusion array; A quality assessment module that extracts real-time rainfall and historical meteorological data, generates a heavy rainfall response signal through a dynamic judgment function of instantaneous rainfall intensity and cumulative rainfall, activates the sensor acquisition frequency, and starts health diagnosis and data quality assessment, and outputs multi-dimensional data with quality scores; A data purification module that intelligently analyzes the multi-dimensional data with quality scores to eliminate environmental noise and baseline drift, and generates purified time series data; A dynamic threshold generation module that combines the purified time series data, real-time geological data and historical data rules, and uses a dynamic threshold generation algorithm to calculate the landslide deformation dynamic threshold with a confidence interval; A landslide deformation prediction module that, when the multi-dimensional monitoring data exceeds the corresponding landslide deformation dynamic threshold, identifies the landslide type and deformation stage in combination with a historical case library, and generates a deformation trend prediction result including the deformation amount development trend and the comprehensive landslide occurrence probability in the future time period; An early warning decision module that, if the comprehensive landslide occurrence probability of the deformation trend prediction exceeds the weighted probability threshold set by experience, triggers a hierarchical early warning and intelligent decision, and outputs early warning information and an emergency disposal suggestion.

Images