Geological disaster multi-source sensing data fusion early warning method

Abstract

The present application belongs to the technical field of geological disaster monitoring data processing, and particularly relates to a geological disaster multi-source sensing data fusion early warning method, which comprises the following steps: obtaining pore water pressure data, surface displacement monitoring data, shallow soil volume water content data and rainfall data of a landslide body; calibrating the physical dissipation coefficient of the landslide soil body to determine the dissipation factor of the pore water pressure at each historical moment; calibrating the dynamic replenishment gain constant to determine the effective residual coefficient of the pore water pressure at each historical moment; determining the optimized input gain of each historical water pressure delay input node; obtaining the surface displacement prediction rate of the landslide body to generate a pre-early warning signal of instability and disaster. The present application introduces the pore pressure dissipation law and rainfall replenishment into the network input layer, reduces the contribution of long-term water pressure data, retains the historical water pressure information which still has a destructive effect under strong rainfall conditions, and improves the accuracy and timeliness of landslide early warning.

Description

Technical Field

[0001] This invention relates to the field of geological disaster monitoring data processing technology. More specifically, this invention relates to a method for early warning of geological disasters through multi-source sensor data fusion. Background Technology

[0002] With the increasing frequency of extreme heavy rainfall events, geological disasters such as landslides are becoming more frequent in some areas. Landslides are characterized by their suddenness, large destructive range, and hidden evolution process, which can easily pose a serious threat to surrounding infrastructure and the safety of people. Therefore, deploying surface displacement monitoring equipment on the surface of the landslide body and drilling piezometers at potential sliding shear surfaces to simultaneously monitor the surface deformation and deep pore water pressure of the landslide body has become an important technical means for landslide risk identification.

[0003] Dynamic time-delay neural networks have the ability to extract time-delay features and can establish a nonlinear mapping relationship between deep seepage state and surface displacement using pore water pressure data from multiple consecutive historical moments. Therefore, they can be used to predict the future surface deformation evolution trend of landslide bodies.

[0004] However, existing dynamic time-delay neural networks typically use the same input processing method for pore water pressure data from multiple consecutive sampling times when constructing historical water pressure delay input nodes. This does not consider the physical dissipation law of pore water pressure in soil pores diffusing and dissipating to low potential energy regions over time. Historical water pressure that is further away from the current time has usually completed a longer period of natural dissipation, and its weakening effect on the current landslide shear strength should be relatively smaller. If the same input contribution is still given to each historical node, the network is easily interfered with by long-term residual water pressure signals, which may lead to false alarms, missed alarms, or delayed alarms in landslide instability warnings.

[0005] In related technologies, for example, Chinese patent document CN110986747B, entitled "A Method and System for Predicting Combined Landslide Displacement," discloses a method for predicting landslide displacement by integrating satellite positioning monitoring data, rainfall data, and soil moisture content data. In the data processing stage, it first uses satellite positioning monitoring stations to acquire displacement data from various monitoring points on the landslide body, and then uses a Kalman filter to smooth the displacement data. Subsequently, it employs an ensemble empirical mode decomposition method to decompose the step landslide displacement data into a trend term displacement sequence and an induced term displacement sequence. A grey prediction model is used to obtain the predicted value of the trend term displacement. Simultaneously, rainfall data and soil moisture content data are input into an improved extreme learning machine to obtain the predicted value of the induced term displacement. Finally, the predicted values ​​of the trend term displacement and the induced term displacement are combined to obtain the predicted landslide displacement. The patent document, when predicting landslide displacement, mainly relies on the numerical decomposition results of surface displacement data and the data mapping relationship between rainfall, soil moisture content, and induced displacement. Its shortcomings are: it does not obtain pore water pressure data at the potential sliding shear surface, nor does it consider that pore water pressure at different historical moments will naturally diffuse and dissipate over time, and that under continuous rainfall conditions, it will have different degrees of retention due to external water supply. Therefore, it cannot dynamically adjust the input contribution of data at each historical moment according to the actual residual degree of pore water pressure. When facing continuous heavy rainfall or delayed dissipation of pore water pressure after rainfall stops, it is easy for long-term invalid monitoring data to continuously interfere with the prediction results, or for historical pore water pressure information that still has the driving force of instability to be not fully preserved. Its adaptability to complex rainfall conditions is insufficient. Summary of the Invention

[0006] To address the technical problem of insufficient accuracy in landslide instability early warning caused by the inability to combine the natural dissipation effect of pore water pressure with the same processing method for each historical water pressure delay input node in existing dynamic time-delay neural networks, this invention provides a multi-source sensor data fusion early warning method for geological disasters, including: S001, acquiring pore water pressure data, surface displacement monitoring data, shallow soil volumetric water content data, and rainfall data of the landslide body, and synchronizing the pore water pressure data, surface displacement monitoring data, shallow soil volumetric water content data, and rainfall data in time; S002, calibrating the physical dissipation coefficient of the landslide soil, and determining the pore water pressure at each historical moment based on the time step between the historical moment and the current moment corresponding to each historical water pressure delay input node, combined with the physical dissipation coefficient of the landslide soil. S003. The dynamic replenishment gain constant is calibrated. Based on the current shallow soil volumetric water content and the cumulative rainfall from each historical moment to the current moment, the effective residual coefficient of pore water pressure at each historical moment is determined, combining the dissipation factor and the dynamic replenishment gain constant. S004. Based on the proportion of the effective residual coefficient of pore water pressure at each historical moment to the sum of the effective residual coefficients of pore water pressure at all historical moments, the optimized input gain of each historical water pressure delay input node is determined. S005. The optimized input gain is used to weight the corresponding historical pore water pressure data. Based on the weighted historical pore water pressure data, the predicted surface displacement rate of the landslide body is obtained. Based on the predicted surface displacement rate, an instability and impending disaster warning signal is generated.

[0007] This invention reflects the degree of natural decline of historical pore water pressure over time through a dissipation factor. It compensates for the water-locking effect caused by heavy rainfall by using the change in the volumetric water content of shallow soil and the cumulative rainfall through an effective residual coefficient. Furthermore, it configures an optimized input gain based on the relative proportion of the effective residual coefficient in all historical nodes. This enables the dynamic time-delay neural network to reduce the contribution of long-term ineffective water pressure data and retain historical water pressure information that still has a real destructive effect under heavy rainfall conditions, thereby improving the accuracy of surface displacement prediction and instability warning.

[0008] Preferably, the method for obtaining the dissipation factor is as follows: calculate the time step difference between the historical time and the current time corresponding to each historical water pressure delay input node in the dynamic time delay neural network, perform maximum and minimum value normalization on the negative of the product of the time step difference and the physical dissipation coefficient of the landslide soil, and use the result as the dissipation factor.

[0009] This invention establishes an exponential decay relationship by utilizing the time step difference and the physical dissipation coefficient of the landslide soil, so that historical water pressure nodes that are far from the current time can obtain a smaller dissipation factor, thus avoiding continuous interference from long-term residual water pressure signals on the current stability assessment.

[0010] Preferably, the calibration of the physical dissipation coefficient of the landslide soil includes: collecting undisturbed soil samples from the potential sliding shear surface of the landslide body, performing a spontaneous dissipation test of pore water pressure under simulated principal stress of the overburden and without external water supply, recording the natural dissipation curve of pore water pressure decreasing over time, performing exponential decay fitting on the natural dissipation curve, and using the fitted decay index as the physical dissipation coefficient of the landslide soil.

[0011] Preferably, the method for obtaining the effective residual coefficient is as follows: calculate the rate of change of the volumetric water content of the shallow soil layer at the current time after maximum and minimum value normalization and the sum of the cumulative rainfall from a historical time to the current time after maximum and minimum value normalization; calculate the product of the sum and the dynamic replenishment gain constant; perform maximum and minimum value normalization on the product to obtain a first value; calculate the product of the first value and the dissipation factor; and use the obtained result as the effective residual coefficient.

[0012] This invention introduces the change rate of shallow soil volumetric water content and cumulative rainfall as dynamic replenishment correction on the basis of natural dissipation, so as to preserve the water pressure retention effect caused by continuous heavy rainfall and prevent the historical water pressure contribution from being excessively weakened according to the conventional dissipation law under extreme rainfall replenishment conditions.

[0013] Preferably, the calibration of the dynamic replenishment gain constant includes: setting up an artificial rainfall test area at the landslide site, conducting artificial rainfall tests according to multiple preset rainfall intensities, simultaneously measuring the increase in volumetric water content of the shallow soil layer, the cumulative rainfall, and the pore water pressure retention, performing regression fitting on the response relationship between the increase in volumetric water content of the shallow soil layer and the cumulative rainfall and the pore water pressure retention, and using the fitted gain parameter as the dynamic replenishment gain constant.

[0014] Preferably, the method for obtaining the optimized input gain of each historical water pressure delay input node is as follows: calculate the ratio of the effective residual coefficient of pore water pressure at a historical moment to the sum of the effective residual coefficients of pore water pressure at all historical moments, and use the result as the optimized input gain of each historical water pressure delay input node.

[0015] Preferably, the step of weighting the corresponding historical pore water pressure data using the optimized input gain includes: multiplying the pore water pressure value at each historical moment with the corresponding optimized input gain according to the one-to-one correspondence between historical time and historical water pressure delay input nodes, and combining the products in chronological order to obtain a time-series input vector.

[0016] This invention configures the input gain of each historical water pressure delay input node by the relative proportion of the effective residual coefficient, so that the input layer weight corresponds to the actual seepage state, which both compresses redundant historical information and highlights the effective historical information that still has the driving force of instability.

[0017] Preferably, the step of generating an instability and impending disaster warning signal based on the predicted surface displacement rate includes: performing a limit equilibrium analysis based on the cohesion and internal friction angle obtained from the direct shear test of the undisturbed soil sample of the landslide body, combined with the geometry of the landslide body and the groundwater level conditions, determining the surface displacement rate when the landslide body reaches a critical state, obtaining a surface displacement rate warning threshold, generating an instability and impending disaster warning signal in response to the predicted surface displacement rate being greater than the surface displacement rate warning threshold, and controlling the disconnection of the on-site construction power supply and sending emergency evacuation information to a preset receiving terminal.

[0018] Preferably, the acquisition of pore water pressure data, surface displacement monitoring data, shallow soil volumetric water content data, and rainfall data of the landslide body includes: collecting pore water pressure data using borehole piezometers deployed at the potential sliding shear surface of the landslide body; collecting surface displacement monitoring data using GPS monitoring stations and fixed surface crack gauges deployed in the surface deformation monitoring area of ​​the landslide body; collecting shallow soil volumetric water content data using volumetric water content sensors deployed in the shallow surface layer of the landslide; and collecting rainfall data using rain gauges.

[0019] This invention uses multiple types of sensors to collaboratively collect pore water pressure, surface displacement, shallow soil volumetric water content, and rainfall, enabling the network input data to simultaneously cover deep seepage response, surface deformation results, and external recharge conditions, providing a complete data foundation for seepage stress coupling analysis.

[0020] Preferably, the time synchronization of pore water pressure data, surface displacement monitoring data, shallow soil volumetric water content data, and rainfall data includes: controlling borehole piezometers, GPS monitoring stations, fixed surface crack gauges, volumetric water content sensors, and rain gauges to perform data acquisition under the same hardware clock synchronization signal, and aligning the four types of data according to the sampling time to form a multi-source synchronous time-series data stream.

[0021] The beneficial effects of this invention are as follows:

[0022] This invention introduces the natural dissipation law of pore water pressure in landslide soil into the input layer of a dynamic time-delay neural network, so that each historical water pressure delay input node no longer uses the same input contribution without difference, which can reduce the interference of historical water pressure data that is far away from the current time and has completed sufficient natural dissipation on the prediction results.

[0023] This invention further utilizes the change rate of shallow soil volumetric water content and cumulative rainfall to dynamically replenish and correct the dissipation factor, thereby quantifying the external replenishment and water pressure retention effect generated by continuous heavy rainfall, and avoiding the erroneous weakening of historical water pressure data according to the dissipation law without replenishment under extreme rainfall conditions.

[0024] This invention generates optimized input gains for each historical water pressure delay input node based on the relative proportion of the effective residual coefficient. It combines physical dissipation laws, external supply status, and network input feature configuration, enabling the dynamic time delay neural network to focus on historical pore water pressure information that still has real instability driving force. This improves the accuracy of surface displacement prediction rate, the timeliness of early warning, and the robustness under complex meteorological conditions. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating a method for early warning of geological disasters using multi-source sensor data fusion according to the present invention; Figure 2 This is a comparison chart of existing technologies and the present invention regarding the predicted rate of surface displacement. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0028] This invention discloses a method for early warning of geological disasters by fusing multi-source sensor data, referring to... Figure 1 This includes steps S001 to S005: S001. Obtain pore water pressure data, surface displacement monitoring data, shallow soil volumetric water content data, and rainfall data of the landslide body.

[0029] Specifically, pore water pressure data is collected using borehole piezometers deployed at the potential sliding shear surface of the landslide; surface displacement monitoring data is collected using GPS monitoring stations and fixed surface crack gauges deployed in the landslide surface deformation monitoring area; volumetric water content data of the shallow soil layer is collected using volumetric water content sensors deployed in the shallow surface layer of the landslide; and rainfall data is collected using rain gauges. The borehole piezometers, GPS monitoring stations, fixed surface crack gauges, volumetric water content sensors, and rain gauges are all controlled to perform data acquisition under the same hardware clock synchronization signal. The four types of data collected are timestamped according to the sampling time to form a multi-source synchronous time-series data stream. In this embodiment, the collection frequency of pore water pressure data, surface displacement monitoring data, shallow soil volume water content data, and rainfall data is set to once per minute. In other embodiments, the data collection frequency can be adjusted according to the landslide deformation rate, the duration of the rainfall process, and the communication bandwidth. When the monitoring area enters the heavy rainfall warning period, the data collection frequency can be appropriately increased to improve the ability to capture the rapid rise of pore water pressure.

[0030] S002. The physical dissipation coefficient of the landslide soil is calibrated to determine the dissipation factor of pore water pressure at each historical moment.

[0031] It should be noted that, since existing dynamic time-delay neural networks use the same input processing method for each historical water pressure delay input node, they do not consider the natural diffusion and discharge of historical pore water pressure to low potential energy regions over time. As a result, historical water pressure data that is far from the current time and has been sufficiently dissipated continues to interfere with the current deformation prediction. Therefore, this invention combines the time step difference between the historical time and the current time with the physical dissipation coefficient of the landslide soil to determine the dissipation factor, which is used to characterize the residual weakening effect of historical pore water pressure on the current shear strength of the landslide body under the condition of no external continuous replenishment.

[0032] Specifically, the physical dissipation coefficient of the landslide soil is calibrated, including: Uncirculated soil samples were collected from the potential sliding shear surface of the landslide body. The uncirculated soil samples were placed in a triaxial high-pressure permeameter for soil and rock. Under the condition of simulating the principal stress of the overburden and without external water supply, the pore water pressure was increased to a preset peak value. Then, water supply was stopped and the natural dissipation curve of pore water pressure decreasing over time was recorded. The natural dissipation curve was fitted with an exponential decay, and the decay index obtained by fitting was used as the physical dissipation coefficient of the landslide soil. In this embodiment, the physical dissipation coefficient of the landslide soil was set to 0.025.

[0033] Specifically, the method for obtaining the dissipation factor is as follows: calculate the time step difference between the historical moment and the current moment corresponding to each historical water pressure delay input node in the dynamic time delay neural network, perform maximum and minimum value normalization on the negative of the product of the time step difference and the physical dissipation coefficient of the landslide soil, and use the result as the dissipation factor.

[0034] Specifically, the product of the time step difference and the physical dissipation coefficient of the landslide soil is used to jointly characterize the degree of natural dissipation of historical pore water pressure before the current moment. The larger the time step difference, the longer the corresponding historical pore water pressure has undergone diffusion and discharge. The larger the physical dissipation coefficient of the landslide soil, the faster the natural dissipation rate of pore water pressure in the landslide soil. An increase in either of these factors will increase the absolute value of the negative exponential term, thereby reducing the dissipation factor.

[0035] It should be noted that the number of historical moments selected for the input layer of the dynamic time delay neural network is determined based on the average complete period required for the pore water pressure to naturally dissipate from the peak value to the background value in the geological exploration data of the project site. In this embodiment, the average complete period is 72 hours and the single time step is 1 minute. Therefore, the number of historical moments selected for the input layer of the dynamic time delay neural network is set to 720.

[0036] If the dissipation factor is uncertain, the network will continue to assign the same input contribution to the long-term historical water pressure peak as to the recent water pressure. Under the condition of no rain for a long time and the pore pressure has naturally fallen, it is easy to misjudge the old peak as the current high-risk seepage state, which will lead to false alarms of instability risk. Therefore, determining the dissipation factor can suppress invalid historical information in the long term.

[0037] S003. The dynamic replenishment gain constant is calibrated to determine the effective residual coefficient of pore water pressure at each historical moment.

[0038] It should be noted that continuous heavy rainfall will lead to a sustained external infiltration recharge, which will cause the volumetric water content of the shallow soil layer to increase rapidly and restrict the normal discharge of pore water pressure. As a result, the historical pore water pressure still retains a high instability driving force at the current moment. Therefore, this invention combines the change rate of the volumetric water content of the shallow soil layer with the cumulative rainfall to dynamically recharge the dissipation factor and determine the effective residual coefficient, which is used to characterize the actual residual degree of historical pore water pressure at the current moment after the combined effect of natural dissipation and external recharge.

[0039] Specifically, the dynamic replenishment gain constant is calibrated, including: A 1-square-meter artificial rainfall test area was set up at the landslide site. Artificial rainfall tests were conducted at rainfall intensities of 10 mm / h, 30 mm / h, 50 mm / h, and 80 mm / h, respectively. The increase in volumetric water content of the shallow soil layer, the cumulative rainfall, and the pore water pressure retention were measured and recorded simultaneously. The response relationship between the increase in volumetric water content of the shallow soil layer and the cumulative rainfall after normalization to the maximum and minimum values ​​and the pore water pressure retention was regressed and fitted. The fitted gain parameter was used as the dynamic replenishment gain constant. In this embodiment, the dynamic replenishment gain constant was set to 0.65. The value of the dynamic replenishment gain constant should make the correction amplitude of historical pore water pressure under no-rain conditions close to zero, and the effective residual coefficient under continuous heavy rainfall conditions should be able to reflect the measured pore water pressure retention ratio. Under the test conditions of this embodiment, its value range was set to 0.4 to 0.8. A larger value was taken when the landslide soil permeability was weak and the water-locking effect was obvious, and a smaller value was taken when the landslide soil permeability was strong and the drainage channel was unobstructed.

[0040] Specifically, the effective residual coefficient is obtained as follows: calculate the rate of change of the volumetric water content of the shallow soil layer at the current time after maximum and minimum value normalization and the sum of the cumulative rainfall from a historical time to the current time after maximum and minimum value normalization; calculate the product of the sum and the dynamic replenishment gain constant; perform maximum and minimum value normalization on the product to obtain a first value; calculate the product of the first value and the dissipation factor; and use the result as the effective residual coefficient.

[0041] Specifically, the rate of change of shallow soil volumetric water content is obtained by dividing the difference between the current shallow soil volumetric water content and the previous monitoring time by the sampling time interval. Cumulative rainfall is obtained by summing the rainfall amounts for each sampling period from a historical time to the current time. The dissipation factor characterizes the natural residual degree of historical pore water pressure under conditions of no external continuous recharge. The rate of change of shallow soil volumetric water content characterizes the rate of change in the water content state of the shallow soil caused by rainfall infiltration at the current time. Cumulative rainfall characterizes the intensity of continuous external water recharge from the corresponding historical time to the current time. A larger rate of change of shallow soil volumetric water content indicates a faster rate of rainfall infiltration into the landslide soil mass. A larger cumulative rainfall indicates a greater total amount of water recharge received by the landslide soil mass, thus resulting in a higher effective residual coefficient. The dynamic recharge gain constant is used to characterize the enhanced effect of continuous rainfall recharge on the residual historical pore water pressure. It is used to adjust the influence of the rate of change of shallow soil volumetric water content and cumulative rainfall on the effective residual coefficient. The larger the value, the more sensitive the landslide soil is to rainfall infiltration recharge, and the larger the dynamic recharge correction term is for the same rate of change of shallow soil volumetric water content and cumulative rainfall. When the rate of change of shallow soil volumetric water content or cumulative rainfall is close to zero, the dynamic recharge correction effect weakens, and the effective residual coefficient is mainly determined by the dissipation factor. When continuous heavy rainfall causes both the rate of change of shallow soil volumetric water content and cumulative rainfall to increase, the dynamic recharge correction term applies a positive gain correction to the dissipation factor, increasing the effective residual coefficient. This avoids underestimating the actual residual effect of historical pore water pressure under continuous rainfall conditions at the current monitoring time based solely on natural dissipation laws.

[0042] Without dynamic replenishment correction, under conditions of continuous rainfall or extreme rainstorms, the algorithm will significantly reduce the input contribution of pore water pressure from several days ago according to the conventional dissipation law without external replenishment. This will cause the network to underestimate the actual residual water pressure on the current sliding surface, resulting in an underestimation of the surface displacement prediction rate and underreporting of impending disasters. Therefore, determining the effective residual coefficient can prevent misjudgment of the mechanical state under extreme rainfall conditions.

[0043] S004. Determine the optimized input gain for each historical water pressure delay input node.

[0044] It should be noted that the effective residual coefficient reflects both the natural dissipation of historical pore water pressure and the retention caused by external rainfall. The larger the effective residual coefficient, the greater the instability driving force of the pore water pressure at the corresponding historical moment at the current moment. Therefore, this invention determines the optimized input gain of each historical water pressure delay input node based on the proportion of each effective residual coefficient in the total effective residual coefficient, which is used to realize the relative allocation of effective historical information in the network input layer.

[0045] Specifically, the optimized input gain of each historical water pressure delay input node is obtained by calculating the ratio of the effective residual coefficient of pore water pressure at a historical moment to the sum of the effective residual coefficients of pore water pressure at all historical moments, thus obtaining the optimized input gain of each historical water pressure delay input node.

[0046] Specifically, the effective residual coefficient of pore water pressure at a historical moment is positively correlated with the optimized input gain of the corresponding historical water pressure delay input node. By calculating the proportion of each effective residual coefficient in the total effective residual coefficient, the input contribution ratio of each historical water pressure delay input node is determined. The larger the effective residual coefficient, the larger the corresponding optimized input gain, which makes the dynamic time delay neural network pay more attention to the historical pore water pressure data that still has a large residual effect. At the same time, the sum of each optimized input gain is always 1, which can avoid the total input gain from increasing with the number of historical water pressure delay input nodes.

[0047] If the optimized input gain is not configured according to the effective residual coefficient, the dynamic time delay neural network will still be unable to distinguish the importance of information among multiple historical nodes. The long-term redundant water pressure signal will enter the hidden layer together with the recent effective water pressure signal, causing the network feature extraction direction to shift. Therefore, determining the optimized input gain of each historical water pressure delay input node can substantially apply the aforementioned physical analysis results to the network input structure, rather than forming isolated indicators that are not used subsequently.

[0048] S005. Obtain the predicted surface displacement rate of the landslide body and generate an early warning signal for instability and impending disaster.

[0049] Specifically, generating early warning signals for instability and impending disaster includes: A dynamic time-delay neural network model is constructed in a cloud server. The input layer includes 720 historical water pressure delay input nodes. The number of neurons in the hidden layer is determined by solving the principal component eigenvalues ​​of the covariance matrix of the historical monitoring data on site and selecting the number of eigenvalues ​​with a cumulative contribution rate of 95%. In this embodiment, the hidden layer is set to 32 neurons and the output layer is set to 1 node, which is used to output the predicted surface displacement rate of the landslide body. Based on the one-to-one correspondence between historical time and historical water pressure delay input node, the pore water pressure value at each historical time is multiplied by the corresponding optimized input gain. The products are combined in chronological order to obtain the time-series input vector, and each element in the time-series input vector is input to the corresponding historical water pressure delay input node. During the network training phase, the time-series input vector obtained in the historical period is used as the training sample input, and the actual surface displacement rate corresponding to the next sampling time is used as the training label. The mean square error is used as the loss function for backpropagation training. After the dynamic time delay neural network is trained, the time-series input vector corresponding to the current time is input into the trained dynamic time delay neural network to obtain the predicted surface displacement rate of the landslide. Based on the landslide body survey results, a slope profile model is established. The cohesion, internal friction angle, and soil weight obtained from the direct shear test are assigned to the corresponding soil layers. The pore water pressure on the sliding surface is calculated based on the groundwater level. Then, the Bishop method is used to search for potential sliding surfaces and calculate the safety factor. The working parameters such as the groundwater level are gradually adjusted. When the safety factor of the most dangerous sliding surface decreases to 1, this state is determined as the critical state of the landslide body. Then, the displacement change per unit time is calculated from the surface displacement monitoring sequence of the corresponding time period. The obtained displacement rate is used as the surface displacement rate when the landslide body reaches the critical state, that is, the surface displacement rate warning threshold. In this embodiment, the surface displacement rate warning threshold is set to 5 mm / h. In response to the predicted surface displacement rate exceeding the surface displacement rate warning threshold, an instability and impending disaster warning signal is generated, and the on-site construction power supply is disconnected and emergency evacuation information is sent to a preset receiving terminal.

[0050] like Figure 2 As shown in the figure, this diagram illustrates the correspondence between the actual surface displacement rate, the surface displacement prediction rate using existing technology, the surface displacement prediction rate of this invention, the trend of residual water pressure influence after rainfall, and the early warning threshold for surface displacement rate within the same monitoring period. In the figure, the horizontal axis represents monitoring time, and the vertical axis represents the predicted surface displacement rate. In the early monitoring phase, all curves show a slow upward trend. After approximately 48 hours, during the accelerated deformation phase, the actual surface displacement rate increases rapidly, reaching the early warning threshold at 58 hours. The curve corresponding to the predicted surface displacement rate of this invention closely follows the actual changes, reaching the threshold at 58.5 hours, while the curve corresponding to the predicted surface displacement rate using existing technology grows relatively slowly, only reaching the threshold at 63 hours. This figure visually verifies that this invention weakens long-term ineffective pore water pressure information through a dissipation factor and retains historical pore pressure data that still has an impact under continuous rainfall conditions using an effective residual coefficient, thereby adjusting the input gain of each historical water pressure delay input node accordingly. The results show that this invention can more accurately track the accelerated deformation process of landslides, significantly shorten the early warning lag time, and thus improve the accuracy and timeliness of instability and impending disaster early warning under complex rainfall conditions.

Claims

1. A method for early warning of geological disasters through multi-source sensor data fusion, characterized in that, include: S001. Acquire pore water pressure data, surface displacement monitoring data, shallow soil volumetric water content data, and rainfall data of the landslide body, and synchronize the pore water pressure data, surface displacement monitoring data, shallow soil volumetric water content data, and rainfall data in time. S002. The physical dissipation coefficient of the landslide soil is calibrated. Based on the time step between the historical moment and the current moment corresponding to each historical water pressure delay input node, and combined with the physical dissipation coefficient of the landslide soil, the dissipation factor of pore water pressure at each historical moment is determined. S003. The dynamic replenishment gain constant is calibrated. Based on the current shallow soil volumetric water content and the cumulative rainfall from each historical moment to the current moment, the effective residual coefficient of pore water pressure at each historical moment is determined by combining the dissipation factor and the dynamic replenishment gain constant. S004. Determine the optimized input gain of each historical water pressure delay input node based on the proportion of the effective residual coefficient of pore water pressure at each historical moment to the total effective residual coefficient of pore water pressure at all historical moments. S005. The corresponding historical pore water pressure data is weighted using the optimized input gain. The predicted surface displacement rate of the landslide body is obtained based on the weighted historical pore water pressure data. An instability and impending disaster warning signal is generated based on the predicted surface displacement rate.

2. The method for early warning of geological disasters by multi-source sensor data fusion according to claim 1, characterized in that, The method for obtaining the dissipation factor is as follows: calculate the time step difference between the historical time and the current time corresponding to each historical water pressure delay input node in the dynamic time delay neural network, perform maximum and minimum value normalization on the negative of the product of the time step difference and the physical dissipation coefficient of the landslide soil, and use the result as the dissipation factor.

3. The method for early warning of geological disasters by multi-source sensor data fusion according to claim 1, characterized in that, The calibration of the physical dissipation coefficient of landslide soil includes: collecting undisturbed soil samples from the potential sliding shear surface of the landslide body; conducting a spontaneous dissipation test of pore water pressure under simulated principal stress of the overburden and without external water supply; recording the natural dissipation curve of pore water pressure decreasing over time; performing exponential decay fitting on the natural dissipation curve; and using the obtained decay index as the physical dissipation coefficient of the landslide soil.

4. The method for early warning of geological disasters by multi-source sensor data fusion according to claim 1, characterized in that, The method for obtaining the effective residual coefficient is as follows: calculate the rate of change of the volumetric water content of the shallow soil layer at the current time after maximum and minimum value normalization and the sum of the cumulative rainfall from a historical time to the current time after maximum and minimum value normalization; calculate the product of the sum and the dynamic replenishment gain constant; perform maximum and minimum value normalization on the product to obtain a first value; calculate the product of the first value and the dissipation factor; and use the result as the effective residual coefficient.

5. The method for early warning of geological disasters by multi-source sensor data fusion according to claim 1, characterized in that, The calibration of the dynamic replenishment gain constant includes: setting up an artificial rainfall test area at the landslide site, conducting artificial rainfall tests according to multiple preset rainfall intensities, simultaneously measuring the increase in volumetric water content of the shallow soil layer, the cumulative rainfall, and the pore water pressure retention, performing regression fitting on the response relationship between the increase in volumetric water content of the shallow soil layer and the cumulative rainfall and the pore water pressure retention, and using the fitted gain parameter as the dynamic replenishment gain constant.

6. The method for early warning of geological disasters by multi-source sensor data fusion according to claim 1, characterized in that, The method for obtaining the optimized input gain of each historical water pressure delay input node is as follows: calculate the ratio of the effective residual coefficient of pore water pressure at a historical moment to the sum of the effective residual coefficients of pore water pressure at all historical moments, and use the result as the optimized input gain of each historical water pressure delay input node.

7. The method for early warning of geological disasters by fusion of multi-source sensor data according to claim 1, characterized in that, The step of using the optimized input gain to weight the corresponding historical pore water pressure data includes: according to the one-to-one correspondence between historical time and historical water pressure delay input nodes, multiplying the pore water pressure value at each historical time with the corresponding optimized input gain, and combining the products in chronological order to obtain the time-series input vector.

8. The method for early warning of geological disasters by fusion of multi-source sensor data according to claim 1, characterized in that, The process of generating an instability and impending disaster warning signal based on the predicted surface displacement rate includes: performing a limit equilibrium analysis based on the cohesion and internal friction angle obtained from the direct shear test of the undisturbed soil sample of the landslide body, combined with the geometry of the landslide body and the groundwater level conditions, to determine the surface displacement rate when the landslide body reaches a critical state, obtaining a surface displacement rate warning threshold, and generating an instability and impending disaster warning signal in response to the predicted surface displacement rate being greater than the surface displacement rate warning threshold, and controlling the disconnection of the on-site construction power supply and sending emergency evacuation information to a preset receiving terminal.

9. The method for early warning of geological disasters by fusion of multi-source sensor data according to claim 1, characterized in that, The acquisition of pore water pressure data, surface displacement monitoring data, shallow soil volumetric water content data, and rainfall data of the landslide body includes: collecting pore water pressure data using borehole piezometers deployed at the potential sliding shear surface of the landslide body; collecting surface displacement monitoring data using GPS monitoring stations and fixed surface crack gauges deployed in the surface deformation monitoring area of ​​the landslide body; collecting shallow soil volumetric water content data using volumetric water content sensors deployed in the shallow surface layer of the landslide; and collecting rainfall data using rain gauges.

10. The method for early warning of geological disasters by fusion of multi-source sensor data according to claim 1, characterized in that, The time synchronization of pore water pressure data, surface displacement monitoring data, shallow soil volumetric water content data, and rainfall data includes: controlling borehole piezometers, GPS monitoring stations, fixed surface crack gauges, volumetric water content sensors, and rain gauges to perform data acquisition under the same hardware clock synchronization signal, and aligning the four types of data according to the sampling time to form a multi-source synchronous time-series data stream.

Images