A sensor data-based landslide prediction method and system

Abstract

The application discloses a landslip prediction method and system based on sensor data, and relates to the technical field of landslip prediction of building engineering. The original data collected by a sensing device on a to-be-tested soil body is received, the original data is preprocessed and extracted, and a pressure change rate, a spatial pressure gradient, a strain accumulation amount and an inclination angle are obtained; a target weight coefficient is determined according to a current construction stage and real-time data characteristics; the pressure change rate, the inclination angle and the strain accumulation amount are calculated based on the target weight coefficient to obtain a current risk index; a time sequence in a preset historical time window is input into a preset long short-term memory network prediction model, and a risk probability is output; a dynamic early warning threshold is determined according to a soil body type and the current construction stage; a corresponding early warning level is determined by comprehensively considering a comparison result of the current risk index and the dynamic early warning threshold, the risk probability and the spatial pressure gradient, and an early warning signal is triggered based on the early warning level. By implementing the method, false positives in a disturbance period and false negatives in a high-risk period are reduced.

Description

Technical Field

[0001] This application relates to the field of building collapse prediction technology, specifically to a collapse prediction method and system based on sensor data. Background Technology

[0002] In the field of construction engineering, the excavation of deep foundation pits and narrow ditches is an extremely critical step. During the excavation process, the redistribution of internal soil stress and the continuous disturbance of external construction machinery can easily cause the soil piled on the slope to loosen, slide, or even collapse, posing a serious threat to the lives of on-site construction personnel and machinery.

[0003] Currently, intelligent sensing technology and the Internet of Things (IoT) have been widely applied in safety monitoring of civil engineering projects. Existing foundation monitoring systems mainly utilize wireless sensor networks deployed at construction sites to collect real-time basic physical data such as soil stress changes, displacement, and deformation using various embedded sensors. The system transmits the collected monitoring data to a back-end processing center via a wireless network and directly compares it with pre-set fixed safety thresholds. Once a monitoring value reaches or exceeds the set threshold, the system automatically triggers an alarm command to provide a safety warning.

[0004] However, existing technologies heavily rely on static warning thresholds and lack dynamic evolution prediction of multi-dimensional features. In actual engineering, the stress state, deformation tolerance, and potential risk characteristics of the soil undergo drastic dynamic changes depending on the construction stage of the foundation pit. Existing systems use unchanging static thresholds, which cannot adapt to dynamic construction conditions, leading to false alarms during normal disturbance periods and missed alarms during truly high-risk periods due to a single parameter not meeting the standard. Summary of the Invention

[0005] This application provides a landslide prediction method and system based on sensor data. The method adaptively adjusts the weights of each parameter according to the characteristics of real-time data and makes a comprehensive judgment on multiple factors, which effectively reduces the false alarm rate during the normal disturbance period and the false alarm rate during the real high-risk period.

[0006] Firstly, this application provides a landslide prediction method based on sensor data, applied in a foundation monitoring platform. The method includes: receiving raw data collected by a sensing device from the soil to be measured, the raw data including earth pressure data, strain data, and tilt angle data, the sensing device being deployed in the soil to be measured and penetrating the potential sliding surface; preprocessing the raw data to obtain preprocessed earth pressure data, preprocessed strain data, and preprocessed tilt angle data; calculating the pressure change rate and the spatial pressure gradient between adjacent sensing devices based on the preprocessed earth pressure data, calculating the cumulative strain based on the preprocessed strain data, and calculating the tilt angle based on the preprocessed tilt angle data; and according to the current... The target weighting coefficient is determined based on real-time data characteristics of the construction phase, pressure change rate, tilt angle, and strain accumulation. The current risk index is obtained by weighting the pressure change rate, tilt angle, and strain accumulation based on the target weighting coefficient. The time series of pressure change rate, tilt angle, and strain accumulation within a preset historical time window is input into a preset long short-term memory network prediction model for processing, outputting the probability of soil collapse in the future period. A dynamic early warning threshold is determined based on the soil type and current construction phase. The corresponding early warning level is determined by combining the comparison results of the current risk index and the dynamic early warning threshold, the risk probability, and the spatial pressure gradient, and the corresponding early warning signal is triggered based on the early warning level.

[0007] By employing the aforementioned technical solution, sensors penetrating the potential sliding surface are deployed in the soil to be tested. This allows for the simultaneous acquisition of multi-dimensional raw data, including soil pressure, strain, and tilt angle data. Preprocessing of the raw data, including outlier removal, temperature compensation, and sliding window denoising, effectively eliminates data distortion caused by environmental interference and sensor temperature drift. The system not only calculates single-point time-domain characteristics such as pressure change rate, strain accumulation, and tilt angle, but also introduces the spatial dimension of pressure gradient, enabling timely detection of pressure distribution anomalies in adjacent soil areas. Furthermore, by comprehensively considering the current construction... By adaptively determining the target weight coefficients of each monitoring parameter based on real-time data characteristics such as stage and pressure level, tilt angular velocity, and strain level, the risk assessment can dynamically respond to changes in construction progress and soil condition. Compared with traditional fixed-weight methods, this significantly improves the targeting and accuracy of early warnings. A dual assessment mechanism combining the current risk index and a long short-term memory network prediction model reflects both the immediate risk status at the current moment and predicts future risk evolution trends based on historical time-series data. Dynamic early warning thresholds are determined according to soil type and construction stage, avoiding false alarms or missed alarms caused by uniform thresholds. By comprehensively considering multi-dimensional information such as the current risk index, risk probability, and spatial pressure gradient to determine the early warning level, corresponding early warning signals and emergency measures can be triggered for different risk levels, providing a scientific and reliable landslide risk prevention and control plan for construction sites and effectively ensuring construction safety.

[0008] Optionally, the pressure change rate and spatial pressure gradient between adjacent sensing devices are calculated based on the preprocessed earth pressure data; the cumulative strain is calculated based on the preprocessed strain data; and the tilt angle is calculated based on the preprocessed tilt angle data. Specifically, this includes: acquiring preprocessed first earth pressure data and preprocessed second earth pressure data at the current sampling time and the previous adjacent sampling time; calculating the pressure difference between the preprocessed first earth pressure data and the preprocessed second earth pressure data; dividing the pressure difference by the time interval between the current sampling time and the previous adjacent sampling time to obtain the pressure change rate; and acquiring the preprocessed earth pressure at the same time for two sensing devices deployed at adjacent locations in the soil to be measured. The data is used to calculate the absolute value of the pressure difference between two sensing devices; the absolute value of the pressure difference is divided by the spatial arrangement distance between the two sensing devices to obtain the spatial pressure gradient; the preprocessed strain data at the initial moment is obtained as the reference strain value, and the preprocessed strain data at the current moment is subtracted from the reference strain value to obtain the cumulative strain; the X-axis tilt component and Y-axis tilt component are extracted from the preprocessed tilt angle data, the X-axis tilt component and Y-axis tilt component are converted to radians, the cosine value of the X-axis tilt component and the cosine value of the Y-axis tilt component are calculated, and the product is inverse cosineed to obtain the current tilt angle of the soil to be measured, where the tilt angle represents the axis of the sensing device relative to the direction of gravity.

[0009] By adopting the above technical solutions, in the calculation of the pressure change rate, the preprocessed first earth pressure data and preprocessed second earth pressure data of the current sampling time and the previous adjacent sampling time are obtained. The pressure difference between the two is calculated and divided by the time interval, which can accurately quantify the rate of change of the soil stress state. By obtaining the preprocessed earth pressure data of two sensors deployed at adjacent positions in the soil under test at the same time, the absolute value of the pressure difference is calculated and divided by the spatial deployment distance, which can effectively identify the non-uniformity of stress distribution and stress concentration areas inside the soil. In terms of strain accumulation calculation, by obtaining the preprocessed strain data at the initial time as the reference strain value, and subtracting the preprocessed strain data at the current time from the reference strain value, the overall deformation accumulation degree of the soil from the initial state of construction to the current time can be clearly reflected. For the calculation of the tilt angle, the mutually orthogonal X-axis tilt angle component and Y-axis tilt angle component are extracted from the preprocessed tilt angle data, and the product of their cosine values ​​is calculated to obtain the tilt angle index that can comprehensively reflect the three-dimensional spatial attitude change of the soil under test, avoiding the problem of missing key displacement information if only considering the tilt angle in a single direction.

[0010] Optionally, the target weight coefficient is determined based on the current construction stage and the real-time data characteristics of pressure change rate, tilt angle, and strain accumulation. Specifically, this includes: retrieving the foundation pressure rate weight corresponding to the pressure change rate, the foundation tilt angle weight corresponding to the tilt angle, and the foundation strain weight corresponding to the strain accumulation from a preset weight matching library, based on the current construction stage; using the pre-processed earth pressure data at the current moment as the current absolute earth pressure level, calculating the pressure ratio between the current absolute earth pressure level and the design earth pressure benchmark value, squaring the pressure ratio, multiplying it by the first sensitivity coefficient, and adding it to 1 to obtain the pressure level adjustment factor; and multiplying the foundation pressure rate weight by the pressure level adjustment factor to obtain the target weight for the pressure change rate. The tilt angle change rate is calculated based on the tilt angle at continuous sampling times. The absolute value of the tilt angle change rate is calculated as the tilt angle ratio to the angular velocity warning threshold. The tilt angle ratio is multiplied by the second sensitivity coefficient and added to 1 to obtain the tilt angular velocity adjustment factor. The basic tilt angle weight is multiplied by the tilt angular velocity adjustment factor to obtain the target weight of the tilt angle. The strain ratio of the accumulated strain to the strain limit threshold is calculated. The strain ratio is squared, multiplied by the third sensitivity coefficient, and added to 1 to obtain the strain level adjustment factor. The basic strain weight is multiplied by the strain level adjustment factor to obtain the target weight of the accumulated strain. The target weights of the pressure change rate, tilt angle, and accumulated strain are output as target weight coefficients.

[0011] By adopting the above technical solution, the weights of foundation pressure rate, foundation inclination angle, and foundation strain are retrieved from the preset weight matching library according to the current construction stage, ensuring that the weight system can reflect the differences in the importance of each monitoring parameter at different construction nodes. The pressure level adjustment factor calculates the pressure ratio between the current absolute earth pressure level and the design earth pressure benchmark value, squares this ratio, multiplies it by the first sensitivity coefficient, and adds it to 1, so that when the pressure on the soil approaches the design limit, the weight of the pressure change rate can be automatically amplified. The tilt angular velocity adjustment factor calculates the tilt angle change rate based on the tilt angle at continuous sampling time, and compares the absolute value of the tilt angle change rate with the tilt angular velocity warning threshold. The tilt angle ratio, multiplied by the second sensitivity coefficient and added to 1, can dynamically increase the weight of the tilt angle index when the soil undergoes rapid tilting movement. The strain level adjustment factor calculates the strain ratio of the cumulative strain to the strain limit threshold, squares the ratio, multiplies it by the third sensitivity coefficient, and adds it to 1. This ensures that the weight of the strain index is significantly enhanced when the cumulative soil deformation approaches the failure limit. By multiplying each basic weight with the corresponding adjustment factor, the target weights of the pressure change rate, tilt angle, and cumulative strain are obtained, forming target weight coefficients that can simultaneously reflect the characteristics of the construction stage and the real-time state of the soil, significantly improving the accuracy and adaptability of risk assessment.

[0012] Optionally, a dynamic early warning threshold is determined based on the soil type of the soil to be tested and the current construction stage. Specifically, this includes: obtaining the soil type of the soil to be tested; matching the corresponding foundation early warning threshold from a preset foundation threshold library based on the soil type, wherein the foundation early warning threshold is determined by back-calculation based on the building construction safety factor; obtaining the current construction stage; determining the corresponding construction risk coefficient from a preset risk coefficient library based on the current construction stage, wherein different construction nodes correspond to different construction risk coefficients; multiplying the construction risk coefficient by the foundation proportion coefficient and adding it to 1 to obtain a dynamic adjustment factor; and multiplying the dynamic adjustment factor by the foundation early warning threshold to obtain the dynamic early warning threshold.

[0013] By adopting the above technical solution, the soil type of the soil to be tested is obtained, and the corresponding foundation warning threshold is matched from the preset foundation threshold library according to the type. This fully considers the essential differences in mechanical properties, shear strength, and deformation characteristics of different soil materials. Since the foundation warning threshold is determined by back-calculation based on the building construction safety factor, the current construction stage is obtained, and the corresponding construction risk coefficient is determined from the preset risk coefficient library according to the stage. This can accurately reflect the significant differences in the stress state and stability of the soil at different construction nodes such as the initial stage of foundation pit excavation, mid-term support, deep excavation, bottom slab construction, and backfilling. Since different construction nodes correspond to different construction risk coefficients, the construction risk coefficient is multiplied by the foundation ratio coefficient and then added to 1 to obtain a dynamic adjustment factor. The dynamic adjustment factor is then multiplied by the foundation warning threshold to obtain the dynamic warning threshold. This realizes the adaptive adjustment of the threshold as the construction stage progresses, effectively avoiding a large number of false alarms caused by overly strict thresholds in low-risk construction stages, and also preventing the underreporting of real dangers caused by overly lenient thresholds in high-risk construction stages.

[0014] Optionally, the warning levels include Level 1, Level 2, and Level 3 warnings. The corresponding warning level is determined by comprehensively comparing the current risk index with the dynamic warning threshold, the risk probability, and the spatial pressure gradient. Based on the warning level, a corresponding warning signal is triggered. Specifically, if the current risk index is greater than the product of the dynamic warning threshold and a first proportional coefficient, or the risk probability is greater than a first probability threshold, or the current risk index is greater than the product of the dynamic warning threshold and a second proportional coefficient and the accumulated strain is greater than or equal to a preset strain threshold, then a Level 1 warning is determined, where the first proportional coefficient is greater than the second proportional coefficient. If the warning level is determined to be Level 1, a first warning signal is triggered, which includes continuously activating the on-site audible and visual alarm device, pushing an evacuation route map, and a shutdown command. If the triggering conditions for Level 1 warning are not met, and the current risk index is greater than the product of the dynamic warning threshold and the second proportional coefficient, then a Level 3 warning is triggered. If the product of the two proportional coefficients, or the risk probability is greater than the second probability threshold, or the spatial pressure gradient is greater than or equal to the preset gradient threshold and the tilt angle is greater than or equal to the preset tilt angle threshold, then a Level II warning is determined. Here, the first probability threshold is greater than the second probability threshold. If the warning level is determined to be Level II, a second warning signal is triggered, which includes activating the on-site audible and visual alarm device for intermittent alarms and sending a suspension of construction command. If the triggering conditions for Level I and Level II warnings are not met, and the current risk index is greater than the product of the dynamic warning threshold and the third proportional coefficient, and the risk probability is less than or equal to the third probability threshold, then a Level III warning is determined. The second proportional coefficient is greater than the third proportional coefficient, and the second probability threshold is greater than the third probability threshold. If the warning level is determined to be Level III, a third warning signal is triggered, which includes sending the risk location and generating a risk trend report.

[0015] By adopting the above technical solution, if multiple triggering conditions are met, such as the current risk index being greater than the product of the dynamic early warning threshold and the first proportional coefficient, the risk probability being greater than the first probability threshold, or the current risk index being greater than the product of the dynamic early warning threshold and the second proportional coefficient and the accumulated strain being greater than or equal to the preset strain threshold, a first early warning signal can be quickly triggered in an extremely dangerous state where soil collapse is imminent. This signal ensures that on-site personnel can immediately evacuate the danger zone and stop all construction work by continuously activating on-site audible and visual alarm devices, pushing evacuation route maps, and issuing shutdown commands, thus maximizing the safety of personnel and equipment. For situations where the first-level early warning conditions are not met but a high risk still exists, by combining conditions such as the current risk index exceeding a lower multiple threshold, the risk probability exceeding the second probability threshold, or the spatial pressure gradient and tilt angle simultaneously exceeding the standard, a medium-risk state of soil in the danger accumulation stage can be identified. The triggered second early warning signal activates on-site audible and visual alarm devices for intermittent alarms and pushes out suspension of construction commands, both reminding on-site personnel to pay attention to the risks and avoiding overreaction that could affect the construction progress. For situations where the risk level has not yet reached the medium or high level but abnormal trends have already emerged, the Level 3 early warning system identifies early signals of potential risks by requiring the current risk index to exceed the product of the dynamic early warning threshold and the third proportional coefficient, and the risk probability to be low. The triggered third early warning signal provides decision-making reference for construction management personnel by pushing the risk location and generating a risk trend report without affecting normal construction.

[0016] Optionally, the raw data is preprocessed to obtain preprocessed earth pressure data, preprocessed strain data, and preprocessed tilt angle data. Specifically, this includes: performing quality checks and anomaly identification on the raw data to obtain preliminary data; acquiring the current ambient temperature and calculating the temperature difference between the current ambient temperature and the initial calibration temperature of the sensing device; calculating the temperature compensation amount based on the preset temperature compensation coefficient of the sensing device and the temperature difference; algebraically superimposing the preliminary data and the temperature compensation amount to obtain the temperature-compensated data; dividing the temperature-compensated data into multiple sliding data windows according to the time series; calculating the arithmetic mean of the data in each sliding data window as the effective value at the corresponding time; using the effective value as the data after noise reduction to obtain the preprocessed earth pressure data, preprocessed strain data, and preprocessed tilt angle data.

[0017] By employing the above technical solutions, the raw data undergoes quality checks and anomaly identification processing. This distinguishes between false anomalies caused by instrument malfunctions and abrupt changes caused by genuine soil deformation. False anomalies are identified and filled in, while genuine abrupt changes are preserved and labeled with event identifiers, yielding preliminary data. This approach ensures the continuity of the data sequence while avoiding the impact of missing data on time-series analysis. In the temperature compensation stage, the current ambient temperature is acquired and its difference from the initial calibration temperature of the sensing equipment is calculated. The temperature compensation amount is then calculated based on a preset temperature compensation coefficient. Finally, the preliminary data and the temperature compensation amount are algebraically superimposed to obtain the temperature-compensated data, effectively eliminating the effects of diurnal temperature variations and seasonal changes. The systematic influence of temperature factors on the measurement accuracy of earth pressure sensors, strain gauges, and tilt sensors ensures that the monitoring data can truly reflect the mechanical state of the soil rather than the temperature effect. In the noise reduction process, by dividing the temperature-compensated data into multiple sliding data windows according to the time series, and calculating the arithmetic mean of the data in each sliding data window as the effective value at the corresponding time, random noise and high-frequency fluctuations can be smoothed, and effective trend signals in the data can be extracted. This results in preprocessed earth pressure data, preprocessed strain data, and preprocessed tilt angle data having higher signal-to-noise ratios and stability, significantly improving the accuracy of subsequent calculations of characteristic parameters such as pressure change rate and strain accumulation.

[0018] Optionally, the sensing device includes an earth pressure sensor, an upper strain gauge, a lower strain gauge, a biaxial tilt sensor, a data processing and communication module, an audible and visual alarm, and a rechargeable battery pack. The upper and lower strain gauges are respectively positioned in the upper and lower monitoring sections of the sensing device to measure the tensile and compressive deformation of the outer shell. The earth pressure sensor is embedded in a pre-drilled opening in the side wall of the outer shell to measure lateral earth pressure. The biaxial tilt sensor measures the tilt angle of the entire sensing device in two vertical directions. The rechargeable battery pack powers the sensing device. The data processing and communication module connects to and controls all sensors to perform data acquisition and transmission. The audible and visual alarm receives a warning signal and then responds with sound and flashing alarm light.

[0019] By employing the above technical solution, upper and lower strain gauges are respectively arranged in the upper and lower monitoring sections of the sensing device to measure the tensile and compressive deformation of the shell. When the potential sliding surface undergoes shear displacement, the strain gauges located on the upper and lower sides of the sliding surface will produce significant differential strain responses, thereby achieving direct monitoring of the sliding surface's activity state. An earth pressure sensor is embedded in a pre-reserved opening in the side wall of the shell to measure lateral earth pressure, enabling real-time capture of changes and redistribution of stress within the soil. A dual-axis tilt sensor measures the overall tilt angle of the sensing device in two vertical directions, comprehensively reflecting the displacement and attitude changes of the soil in three-dimensional space, and promptly detecting slope tilting and slippage trends. The data processing and communication module, as the control core of the device, connects and controls all sensors to perform data acquisition and transmission. A rechargeable battery pack powers the sensing device, ensuring continuous and stable operation in the harsh environment of long-term burial within the soil, avoiding construction interference caused by frequent battery replacements. The audible and visual alarm is used to provide real-time on-site warnings through sound and flashing alarms after receiving early warning signals. When the monitoring platform determines a high-risk state and issues an early warning command, the sensing devices buried near the danger zone can immediately activate the audible and visual alarms to provide the most direct evacuation signal for on-site construction personnel.

[0020] The second aspect of this application provides a landslide prediction system based on sensor data. The system is located within a foundation monitoring platform and includes a data acquisition unit, a processing unit, an analysis unit, an evaluation unit, and an early warning unit. The data acquisition unit receives raw data collected by sensing devices from the soil under test. The raw data includes earth pressure data, strain data, and tilt angle data. The sensing devices are deployed in the soil under test and penetrate the potential sliding surface. The processing unit preprocesses the raw data to obtain preprocessed earth pressure data, preprocessed strain data, and preprocessed tilt angle data. The analysis unit calculates the rate of pressure change and the spatial pressure gradient between adjacent sensing devices based on the preprocessed earth pressure data, calculates the cumulative strain based on the preprocessed strain data, and calculates the preprocessed... The tilt angle is calculated from the tilt angle data; the evaluation unit determines the target weight coefficient based on the current construction stage and real-time data characteristics of pressure change rate, tilt angle, and strain accumulation; the pressure change rate, tilt angle, and strain accumulation are weighted and calculated based on the target weight coefficient to obtain the current risk index; the time series of pressure change rate, tilt angle, and strain accumulation within a preset historical time window is input into a preset long short-term memory network prediction model for processing, and the risk probability of soil collapse in the future period is output; the early warning unit determines the dynamic early warning threshold based on the soil type of the soil to be tested and the current construction stage; the corresponding early warning level is determined by comprehensively comparing the current risk index with the dynamic early warning threshold, the risk probability, and the spatial pressure gradient, and the corresponding early warning signal is triggered based on the early warning level.

[0021] In a third aspect, this application provides an electronic device including a processor, a memory, a user interface, and a network interface. The memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory, causing the electronic device to perform any of the methods described above in this application.

[0022] In a fourth aspect, this application provides a computer-readable storage medium storing instructions that, when executed, perform any of the methods described above in this application.

[0023] In summary, one or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: 1. Deploying sensors that penetrate potential sliding surfaces in the soil to be tested enables the simultaneous acquisition of multi-dimensional raw data, including soil pressure, strain, and tilt angle data. Preprocessing of the raw data, including outlier removal, temperature compensation, and sliding window denoising, effectively eliminates data distortion caused by environmental interference and sensor temperature drift. It not only calculates single-point time-domain features such as pressure change rate, strain accumulation, and tilt angle, but also introduces the spatial dimension of pressure gradient, enabling timely detection of pressure distribution anomalies in adjacent soil areas. By comprehensively considering the current construction stage and real-time data features such as pressure level, tilt angular velocity, and strain level, the target weight coefficients of each monitoring parameter are adaptively determined, allowing risk assessment to dynamically respond to changes in construction progress and soil condition. Compared to traditional fixed-weight methods, this significantly improves the targeting and accuracy of early warnings. A dual assessment mechanism combining the current risk index and a long short-term memory network prediction model reflects both the immediate risk status at the current moment and predicts future risk evolution trends based on historical time-series data. Dynamic early warning thresholds are determined according to soil type and construction stage, avoiding false alarms or missed alarms caused by uniform thresholds. By comprehensively considering multiple dimensions of information, such as the current risk index, risk probability, and spatial pressure gradient, the system can determine the early warning level and trigger corresponding early warning signals and emergency measures for different risk levels. This provides a scientific and reliable solution for preventing and controlling landslide risks at construction sites, effectively ensuring construction safety. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the first process of a landslide prediction method based on sensor data provided in an embodiment of this application; Figure 2 This is a cross-sectional schematic diagram of a landslide prediction method based on sensor data provided in an embodiment of this application; Figure 3 This is a schematic diagram of the overall structure of a landslide prediction method based on sensor data provided in an embodiment of this application; Figure 4 This is a schematic diagram of a landslide prediction method based on sensor data provided in an embodiment of this application; Figure 5 This is a schematic diagram of the second process of a landslide prediction method based on sensor data provided in an embodiment of this application; Figure 6 This is a schematic diagram of the structure of a landslide prediction system based on sensor data provided in an embodiment of this application; Figure 7 This is a schematic diagram of the structure of an electronic device disclosed in an embodiment of this application.

[0025] Explanation of reference numerals in the attached figures: 601, acquisition unit; 602, processing unit; 603, analysis unit; 604, evaluation unit; 605, early warning unit; 700, electronic equipment; 701, processor; 702, memory; 703, user interface; 704, network interface; 705, communication bus. Detailed Implementation

[0026] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0027] In the description of the embodiments of this application, the words "for example" or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design that is described as "for example" or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design options. Rather, the use of the words "for example" or "for instance" is intended to present the relevant concepts in a specific manner.

[0028] In the description of the embodiments of this application, the term "multiple" means two or more. For example, multiple systems means two or more systems, and multiple screen terminals means two or more screen terminals. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. The terms "comprising," "including," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0029] Therefore, effectively reducing the false alarm rate during normal disturbance periods and the false negative rate during periods of true high risk is a pressing issue. This application provides a landslide prediction method based on sensor data, applied to a foundation monitoring platform. The foundation monitoring platform of this application can provide landslide prediction services for excavation construction. Figure 1 This is a schematic diagram of the first process of a landslide prediction method based on sensor data provided in an embodiment of this application. (Refer to...) Figure 1 The method includes the following steps S101-S108.

[0030] S101: Receives raw data collected by the sensing device on the soil under test. The raw data includes soil pressure data, strain data, and tilt angle data.

[0031] In S101 above, at deep foundation construction sites, to achieve comprehensive perception of the stability of the soil to be tested, it is necessary to deploy sensing devices at the edge of the foundation pit to conduct multi-dimensional monitoring of the soil. The deployment location and insertion angle of the sensing devices are crucial to ensuring the effectiveness of monitoring. According to Bishop's circular arc sliding theory for slope stability analysis and the results of engineering geological surveys, the potential sliding surface of a deep foundation pit is usually circular, and its geometric characteristics are closely related to the depth of the foundation pit, the soil type, and the slope angle. For a foundation pit with a depth of H, the center of the potential sliding surface is usually located at a horizontal distance of 0.3H to 0.5H behind the top of the slope, the radius of the arc R is approximately 1.2H to 1.8H, and the tangent angle between the sliding surface and the horizontal plane gradually decreases along the depth direction from the β angle (slope angle) at the top of the slope to 10° to 20° near the toe of the slope. When deploying sensing devices, construction personnel should adopt the following standardized procedures to ensure that the devices can effectively penetrate the potential sliding surface. First, at the top of the slope at the edge of the excavation pit, monitoring sections are set up in a direction perpendicular to the edge of the pit. The spacing between the monitoring sections is determined according to the perimeter of the excavation pit, and is usually set to 1 to 1.5 times the depth of the excavation pit. For an excavation pit with a depth of 3 meters, the spacing between the monitoring sections is 3 to 4.5 meters. On each monitoring section, the insertion starting point is set 0.5 to 1.0 meters horizontally back from the edge of the slope towards the inside of the excavation pit. The purpose of setting this back distance is to avoid disturbing the surface soil at the top of the slope and to ensure that the sensing equipment has sufficient burial depth.

[0032] The insertion angle of the sensing device is determined based on the depth of the foundation pit and the soil type. For shallow foundation pits with a depth H of less than 5 meters, the insertion angle is set to be 45° to 55° downwards from the horizontal plane; for medium-deep foundation pits with a depth H between 5 and 10 meters, the insertion angle is set to 50° to 60°; and for deep foundation pits with a depth H greater than 10 meters, the insertion angle is set to 55° to 65°. The selection of this angle range is based on the following engineering considerations: First, to ensure that the sensing device can penetrate the surface disturbance layer and enter the undisturbed soil area. The thickness of the surface disturbance layer is usually 0.5 to 1.0 meters. When inserted at a 45° angle to a depth of 1.5 meters, the horizontal projection is about 1 meter, which can effectively penetrate the disturbance layer. Second, to ensure that the insertion trajectory of the sensing device forms an effective angle with the potential arc sliding surface in the key monitoring depth range. According to geometric analysis, when the sensing device is inserted at an angle of 50° ± 10°, in the key monitoring interval of depth 0.5H to 1.5H, the angle between the device trajectory and the tangent of the sliding surface can be maintained in the range of 25° to 65°. This angle range can ensure that when the sliding surface undergoes shear displacement, the sensing device can produce a significant strain response and tilt angle change.

[0033] The insertion depth is controlled using a relative depth method based on the pit depth, rather than an absolute depth method. For a pit with a depth of H, the insertion depth L of the sensing device is set to 1.5H to 2.0H. Taking a pit with a depth of 3 meters as an example, the sensing device is inserted at a 50° angle from 1 meter behind the top of the slope, and the insertion depth should reach 4.5 to 6 meters. At this point, the vertical burial depth of the bottom of the device is approximately 3.5 to 4.5 meters, and the horizontal extension is approximately 3 to 4 meters. This insertion depth design ensures that the sensing device can traverse the most likely sliding surface area. According to Bishop's circular arc sliding theory and statistics from numerous engineering cases, the potential sliding surface of deep pits usually develops in the range of depth 0.8H to 1.8H, and the sensing device with an insertion depth of 1.5H to 2.0H can cover this critical area.

[0034] In actual construction deployment, a portable drilling device is used to pre-drill a guide hole. The diameter of the guide hole is 5 to 10 mm larger than the outer diameter of the sensor, and the depth is consistent with the designed insertion depth. The angle is controlled within ±3° of the design angle using an angle measuring instrument. After the guide hole is drilled, the sensor is slowly pushed in along the guide hole, pausing for 10 seconds every 20 cm to allow the surrounding soil to fully contact the device. After the sensor is fully inserted, the guide hole opening is backfilled and compacted with fine sand or undisturbed soil similar in properties to the surrounding soil to ensure good mechanical coupling between the sensor and the surrounding soil.

[0035] To improve monitoring reliability and coverage, two to three sensors are typically deployed in a monitoring array at each monitoring section. The first sensor is inserted at a standard angle as the main monitoring unit; the second sensor is inserted at a horizontal distance of 1 to 2 meters from the first sensor at an angle 10° larger than the standard angle, serving as a shallow monitoring unit, focusing on monitoring shallow soil deformation; the third sensor is inserted at a horizontal distance of 2 to 3 meters from the first sensor at an angle 10° smaller than the standard angle, serving as a deep monitoring unit, focusing on monitoring deep slip surface activity. This layered monitoring array configuration can cover potential slip surfaces at different depths. When the actual slip surface location deviates from the theoretical prediction, at least one sensor can effectively penetrate the slip surface and capture shear displacement signals.

[0036] After the sensor equipment is deployed, an installation quality inspection is required. The initial tilt angle data of the sensor equipment is read to verify whether the actual insertion angle meets the design requirements, with an allowable deviation range of ±5°. The initial readings of the earth pressure sensor and strain gauge are read to determine the contact state between the sensor equipment and the surrounding soil. If the initial reading deviates from the theoretical earth pressure value at that depth (calculated based on soil unit weight and burial depth) by more than 20%, it indicates the presence of gaps or poor contact around the sensor equipment, requiring backfilling or redeployment. The first 24 hours after installation are the stabilization period, during which the sensor equipment reaches mechanical equilibrium with the surrounding soil. During the stabilization period, the readings of each sensor are recorded hourly. When the fluctuation range of pressure and strain readings is less than 5% for six consecutive hours, the installation is considered stable, and the average data from the last hour of the stabilization period is used as the monitoring baseline value.

[0037] For special working conditions, the deployment plan needs to be adjusted accordingly. When the foundation pit excavation encounters a weak interlayer or a high groundwater level, the potential sliding surface is more likely to develop along the weak interlayer or groundwater level. In this case, the insertion angle and depth of the sensing equipment should be adjusted according to the burial depth of the weak interlayer marked in the geological survey report, so that the equipment trajectory passes through the weak interlayer area. When the foundation pit is excavated in sections, independent monitoring sections should be deployed on the slope of each excavation section to avoid monitoring blind spots between different excavation sections. When there are building foundations or underground pipelines around the foundation pit, the insertion angle of the sensing equipment should avoid obstacles. If necessary, the insertion starting point position should be adjusted or a zigzag drilling technique should be used to bypass them.

[0038] In addition, the aforementioned sensing equipment includes an earth pressure sensor, an upper strain gauge, a lower strain gauge, a biaxial tilt sensor, a data processing and communication module, an audible and visual alarm, and a rechargeable battery pack. The upper and lower strain gauges are respectively located in the upper and lower monitoring sections of the sensing equipment to measure the tensile and compressive deformation of the outer shell. The earth pressure sensor is embedded in a pre-drilled opening in the side wall of the outer shell to measure lateral earth pressure. The biaxial tilt sensor measures the tilt angle of the entire sensing equipment in two vertical directions. The rechargeable battery pack powers the sensing equipment. The data processing and communication module connects to and controls all sensors to perform data acquisition and transmission. The audible and visual alarm receives warning commands from the ground monitoring platform and responds with sound and flashing alarm signals. The specific structure of the sensing equipment will be explained in detail below.

[0039] like Figure 2 As shown, Figure 2 A cross-sectional view of the sensing device is shown. The device uses a white plastic square tube as its outer casing, with an outer diameter of 50 mm × 50 mm and a wall thickness of 4 mm. Compared to traditional steel pipe solutions, plastic square tubes offer better processing convenience, facilitating the placement of various sensor components by slotting the tube wall. In the internal structural design of the sensing device, upper and lower strain gauges are symmetrically attached to opposite side walls of the internal cavity of the casing. The upper strain gauge is located in the upper half of the device, and the lower strain gauge in the lower half, with a spacing of 20 cm between them. This symmetrical arrangement is based on the fact that when the measured soil deforms, the soil above and below the sliding surface will experience relative displacement. The upper and lower strain gauges will respectively sense tensile or compressive strain, and the difference in strain between them can directly reflect the degree of movement of the sliding surface. Both the upper and lower strain gauges are RP-C7.6-ST-LF2 type strain gauges with an outer diameter of 7.6 mm. After being attached to the side walls of the casing, they are sealed with rubber gaskets to prevent moisture and impurities in the soil from affecting the resistance characteristics of the strain gauges. The working principle of the upper and lower strain gauges is to convert the tensile and compressive deformation of the shell into resistance changes, and then convert the resistance changes into strain data output through the data processing and communication module.

[0040] The earth pressure sensor is embedded in a pre-drilled opening in the side wall of the housing. The sensor is a 50 kPa range, 16 mm in diameter, and 6 mm thick earth pressure sensor. During installation, ensure the sensor's sensing surface is flush with the housing side wall and perpendicular to the direction of soil pressure transmission to guarantee accurate measurement of lateral earth pressure. To prevent soil particles from entering through gaps between the sensor and the housing wall and affecting measurement accuracy, after installation, adhesive is used to seal the gaps between the earth pressure sensor and the housing wall, while ensuring unobstructed direct contact between the sensor's sensing surface and the soil. The earth pressure sensor works by deforming the piezoresistive element inside the sensor when pressure is applied by the soil, causing a change in resistance. The data processing and communication module converts this resistance change into a voltage signal proportional to the soil pressure, ultimately outputting the earth pressure data. This direct contact measurement method, compared to indirect calculation methods, can accurately reflect the dynamic changes in the internal stress state of the soil in real time.

[0041] The dual-axis tilt sensor, model MCA426M, is installed centrally inside the sensor housing and measures the overall tilt angle of the sensor in two mutually perpendicular directions. The dual-axis tilt sensor outputs X-axis and Y-axis tilt components. The overall tilt angle of the sensor in three-dimensional space can be obtained by calculating the square root of the sum of the squares of these two components. When the measured soil slips or the slope becomes unstable, the sensor embedded in the soil will tilt along with the soil. The dual-axis tilt sensor can promptly capture this change in attitude, providing crucial information for assessing slope stability.

[0042] The data processing and communication module is the control core of the sensing device, integrating a microprocessor, signal conditioning circuitry, and wireless communication chip. This module establishes electrical connections with the upper strain gauge, lower strain gauge, earth pressure sensor, and dual-axis tilt sensor via power cables. The power cables are 30 cm long, 0.75 mm² DC5521 male-female connectors, used to connect the components in series to form a complete circuit system. The data processing and communication module controls each sensor to perform data acquisition at a sampling frequency of 1 Hz. After amplification, filtering, and analog-to-digital conversion of the acquired analog signals, it obtains digitized earth pressure, strain, and tilt angle data. To reduce the power consumption of wireless communication and improve data transmission efficiency, the data processing and communication module employs a data packetization mechanism, packaging the accumulated data into a data frame every 30 seconds and sending it to the edge gateway of the foundation monitoring platform via the wireless network. In terms of data transmission protocols, the system supports two wireless communication methods: ZigBee and LoRa. ZigBee networks are suitable for close-range scenarios with dense sensor deployments, and use a polling mechanism to avoid conflicts caused by multiple sensors sending data at the same time. LoRa gateways, on the other hand, are suitable for long-distance transmission, and synchronize the clock with the server every 5 minutes to ensure the accuracy of data timestamps.

[0043] The audible and visual alarm is integrated into the top of the sensor's housing. It uses a 12-volt, flashing buzzer with a 70mm outer diameter. When the data processing and communication module receives an early warning command from the ground monitoring platform, the alarm immediately activates, sending a warning signal to on-site personnel through sound and flashing lights. This localized early warning response mechanism offers a faster response time compared to relying solely on remote monitoring platform push notifications, providing valuable evacuation time for personnel in extremely dangerous situations. The rechargeable battery pack uses six 12950mWh pointed batteries to power all electronic components of the sensor. A single full charge allows for up to 10 days of continuous operation, meeting the needs of long-term construction monitoring. An opening is provided in the upper middle part of the sensor's housing for easy periodic removal of the rechargeable battery pack for charging or replacement, eliminating the need to remove the entire sensor from the soil and preventing damage to the soil structure and monitoring continuity from repeated insertion and removal.

[0044] The sensor housing is also equipped with a reset switch and a push-button switch. The reset switch is a 16mm blue metal button used for hardware reset and restart when the sensor malfunctions. The push-button switch is a 16mm silver-green illuminated four-pin 2NO model, operating at 5 to 24 volts, used for manual control of the sensor's power on / off status. The entire assembly process of the sensor follows strict specifications, as assembly quality directly affects the contact state between the sensor and the soil, as well as signal transmission efficiency. During assembly, the upper and lower strain gauges are first precisely adhered to designated positions on the inner wall of the housing, using a special adhesive to ensure no air bubbles or delamination between the strain gauges and the housing. Then, the soil pressure sensor is embedded in the side wall opening and sealed with sol-gel. Next, the dual-axis tilt sensor, data processing and communication module, audible and visual alarm, and rechargeable battery pack are installed. Finally, all components are connected in series using power cables according to the circuit diagram. After assembly, functional testing is required. The sensing device is activated via a button switch to check if the data processing and communication modules can correctly collect data from each sensor and transmit it to the test terminal via the wireless network. Simultaneously, the response function of the audible and visual alarm is tested. Only after ensuring all components are functioning correctly can the device be deployed to the construction site. Figure 3 As shown, Figure 3 A physical image of the sensing device is displayed.

[0045] Furthermore, by deploying the assembled sensing equipment in the soil to be tested and penetrating the potential sliding surface, it is possible to achieve comprehensive perception of the multi-dimensional mechanical parameters of the soil, such as... Figure 4As shown, upper and lower strain gauges are respectively arranged in the upper and lower monitoring sections of the sensing device to capture the bending deformation of the sensing device under shear action of the sliding surface. The sensing device is a cylindrical rod structure. When buried inside the soil, the upper and lower ends are constrained by soil at different depths. When the potential sliding surface undergoes shear displacement, the soil on the upper and lower sides of the sliding surface moves horizontally relative to each other, and the rod of the sensing device is subjected to lateral shear force at the sliding surface position. Because the upper and lower ends of the sensing device are fixed by the surrounding soil, the rod cannot move as a whole with the soil, resulting in significant bending deformation in the area near the sliding surface.

[0046] According to the beam bending theory in mechanics of materials, when a rod-shaped member is subjected to a lateral force, a bending moment is generated in the cross-section of the rod. This bending moment causes the cross-section to rotate about the neutral axis. One side of the neutral axis experiences tensile strain, while the other side experiences compressive strain, with the strain increasing the further away from the neutral axis. For a circular cross-section rod, the neutral axis passes through the center of the circle, and the tension and compression sides are symmetrically distributed along the circumference. When the soil above the sliding surface moves horizontally towards the inside of the pit relative to the soil below, the sensing device experiences a lateral thrust pointing towards the pit in the area above the sliding surface. This causes the rod in that area to bend towards the pit. At this time, the outer shell surface facing the pit experiences tensile strain, while the outer shell surface facing away from the pit experiences compressive strain. In the area below the sliding surface, the bending direction of the rod is opposite, and the strain distribution is also reversed accordingly.

[0047] The upper strain gauge is positioned on the upper monitoring section of the sensing device, which extends downwards from the top of the device to approximately one-quarter of its total length. For example, for a 4-meter-long sensing device, the upper monitoring section is located within 0 to 1 meter from the top. The lower strain gauge is positioned on the lower monitoring section of the sensing device, which extends upwards from the bottom of the device to approximately one-quarter of its total length, corresponding to the 0 to 1-meter range from the bottom in the above example. The selection of the upper and lower monitoring sections is based on the fact that when the sliding surface is located near the middle of the device, the upper and lower sections are respectively located in the maximum bending moment regions above and below the sliding surface, enabling the capture of the most significant bending strain signals.

[0048] Strain gauges within each monitoring section are arranged in a circumferentially differential manner. Specifically, two strain gauges are positioned circumferentially at the same axial position in both the upper and lower monitoring sections, with a circumferential angle of 180 degrees between them, i.e., located at opposite ends of the diameter of the inner wall of the casing. Using the insertion direction of the sensing device as a reference, the strain gauge facing inwards from the pit is defined as strain gauge A, and the strain gauge facing away from the pit is defined as strain gauge B. When shear displacement occurs on the sliding surface, strain gauges A and B are located on the tension and compression sides of the bending section, respectively, generating strain signals of similar magnitude but opposite signs. The two strain gauges in the upper monitoring section are designated as upper strain gauge A and upper strain gauge B, and the two strain gauges in the lower monitoring section are designated as lower strain gauge A and lower strain gauge B. To simplify circuit design and data processing, in practical engineering applications, the upper strain gauge A and upper strain gauge B are usually connected in series as the upper strain gauge unit to output a comprehensive signal, and the lower strain gauge A and lower strain gauge B are connected in series as the lower strain gauge unit to output a comprehensive signal. This series differential configuration can automatically enhance bending strain and suppress common-mode interference (such as overall thermal strain caused by temperature changes).

[0049] The strain gauge is a metal foil strain gauge with a sensing grid length of 5 mm, a resistance of 120 ohms, and a sensitivity coefficient of 2.1. The strain gauge is bonded to the polished and cleaned surface of the inner wall of the housing using epoxy resin adhesive. After bonding, it is cured at room temperature for 24 hours to ensure reliable strain transfer between the strain gauge and the housing. The sensing grid direction of the strain gauge is parallel to the axis of the rod, enabling the measurement of axial strain, which is generated by bending deformation, specifically axial tensile and compressive strain. The housing material is stainless steel with an elastic modulus of 210 GPa and a yield strength of 500 MPa. Within the normal operating strain range (less than 1000 microstrains), it maintains linear elastic characteristics, and the strain value measured by the strain gauge is proportional to the actual stress.

[0050] When the sliding surface undergoes shear displacement, the bending deformation of the sensing device causes differential responses in the strain gauges. Assuming the sliding surface is located in the middle of the device, and the upper soil mass moves inwards towards the pit by a displacement δ. In the upper monitoring section, strain gauge A (facing the pit) experiences tensile stress, producing a positive strain εuA, while strain gauge B (facing away from the pit) experiences compressive stress, producing a negative strain εuB. The combined signal output by the upper strain gauge unit is εu, which equals εuA minus εuB, approximately twice the unilateral strain, reflecting the degree of bending. In the lower monitoring section, due to the opposite bending direction, strain gauge A experiences compressive stress, producing a negative strain εdA, while strain gauge B experiences tensile stress, producing a positive strain εdB. The combined signal output by the lower strain gauge unit is εd, which equals εdA minus εdB, a negative value with an absolute value approximately twice the unilateral strain. The feature extraction module calculates the strain difference Δε by comparing the signal differences between the upper and lower strain gauges. The absolute value of this difference reflects the degree of shear displacement of the sliding surface, while the sign of the difference indicates the shear direction (a positive value indicates that the upper soil moves towards the inside of the pit, and a negative value indicates that it moves towards the outside).

[0051] S102: Preprocess the raw data to obtain preprocessed earth pressure data, preprocessed strain data, and preprocessed tilt angle data.

[0052] In S102 above, at the deep foundation pit construction site, the sensing equipment continuously collects raw data of the soil to be measured at a sampling frequency of 1 Hz. This raw data includes soil pressure data, strain data, and tilt angle data. Due to the complex environment at the construction site, the raw data inevitably contains various types of interference signals. For example, the instantaneous impact caused by the sudden start or stop of construction machinery can cause spikes in the strain data; random noise generated by the sensor circuit in a strong electromagnetic environment can cause irregular fluctuations in the soil pressure data; and the diurnal variation of ambient temperature can cause systematic shifts in the zero point drift and sensitivity of the sensor. If these interference factors are not processed and are directly used for subsequent feature extraction and risk assessment, they will seriously affect the accuracy of the monitoring system's judgment on soil stability, and may even lead to false alarms or missed warning signals.

[0053] Therefore, the raw data is preprocessed to obtain preprocessed earth pressure data, preprocessed strain data, and preprocessed tilt angle data. Specifically, this includes: performing quality checks and anomaly identification on the raw data to obtain preliminary data; acquiring the current ambient temperature and calculating the temperature difference between the current ambient temperature and the initial calibration temperature of the sensing device; calculating the temperature compensation amount based on the preset temperature compensation coefficient of the sensing device and the temperature difference; algebraically superimposing the preliminary data and the temperature compensation amount to obtain the temperature-compensated data; dividing the temperature-compensated data into multiple sliding data windows according to the time series, calculating the arithmetic mean of the data in each sliding data window as the effective value at the corresponding time, and using the effective value as the data after noise reduction to obtain the preprocessed earth pressure data, preprocessed strain data, and preprocessed tilt angle data.

[0054] Specifically, the raw data undergoes quality checks and anomaly identification processing to obtain preliminary data. The processing distinguishes between two types of anomalies: one is spurious jumps caused by instrument malfunctions or communication errors, which are unrelated to the true state of the soil and need to be identified and addressed; the other is abrupt changes caused by actual soil deformation or construction activities, which reflect the true behavior of the monitored object and must be retained. Because soil monitoring data exhibits significant non-stationarity and non-normality, earth pressure data shows phased step changes during construction loading or unloading, strain data shows nonlinear cumulative growth during the development of the sliding surface, and dip angle data shows minimal changes in the early stages of slope deformation but may increase sharply before sliding. These genuine abrupt changes should not be simply judged as anomalies and removed, otherwise, crucial early warning information will be lost. Based on the above, the data quality check adopts a multi-level discrimination strategy: first, obvious instrument malfunction characteristics are identified; then, physical constraints and temporal logic are combined to determine whether suspected anomalies are actual deformations; finally, only confirmed spurious data is processed.

[0055] First, identify the instrument malfunction characteristics. Instrument malfunctions typically manifest in the following patterns: data suddenly becomes zero or exceeds the fixed extreme value of the sensor's range, for example, an earth pressure sensor with a range of 0 to 500 kPa suddenly displays negative values ​​or continuously shows 500 kPa; data shows a single-point spike followed by an immediate return to normal trends, for example, a strain sequence of 100, 102, 105 microstrains suddenly jumps to 800 microstrains, then returns to 108 microstrains; data exhibits high-frequency oscillations within a short period, with amplitudes far exceeding the normal fluctuation range, for example, the tilt angle repeatedly jumps between 2 degrees and 20 degrees more than 10 times within 1 minute; data remains unchanged for a long time, for example, earth pressure data remains exactly 85.3 kPa for 24 consecutive hours without any decimal changes. For these obvious instrument malfunction characteristics, they can be directly identified as false data and marked as anomalies to be processed without further judgment.

[0056] For data points not identified by obvious fault characteristics, a further rationality check based on physical constraints is conducted. This check is based on the physical laws of soil deformation and stress change, setting an acceptable upper limit for the rate of change. For earth pressure data, considering the construction loading rate and soil stress transmission characteristics, a pressure change rate threshold of 50 kPa per hour is set between adjacent sampling points (sampling interval of 1 hour). If the pressure change between two adjacent points exceeds 50 kPa, the data point is placed in a suspected anomaly queue for further evaluation. For strain data, considering the gradual nature of soil creep and slip surface development, a strain change rate threshold of 500 microstrains per hour is set. Data points exceeding this threshold are placed in a suspected anomaly queue. For dip angle data, considering the continuity of slope deformation, a dip angle change rate threshold of 5 degrees per hour is set. Data points exceeding this threshold are placed in a suspected anomaly queue. It should be noted that these thresholds are conservative upper limits determined based on engineering experience and statistical analysis of existing monitoring data, covering the fastest change rates under most real-world conditions, while also being able to identify spurious jumps that significantly exceed physical possibilities.

[0057] For data points entering the suspected anomaly queue, they are not immediately classified as anomalies, but rather a comprehensive judgment is made based on temporal logic. The core idea of ​​temporal logic verification is that genuine soil mutations are usually not isolated single-point phenomena, but rather continue or develop in subsequent moments, while false jumps caused by instrument malfunctions are usually single-point spikes that immediately return to the original trend after the jump. The judgment method is as follows: for a suspected anomaly point, check the data trend of its subsequent 3 sampling points (i.e., the next 3 hours). If the subsequent data continues the direction of change of the suspected anomaly point, for example, after the soil pressure suddenly increases by 30 kPa at the suspected anomaly point, the subsequent 3 points continue to maintain a high value or continue to increase slowly, then the suspected anomaly point is judged as a genuine mutation and is retained; if the subsequent data immediately returns to the trend before the suspected anomaly point, for example, the strain suddenly increases to 600 microstrain at the suspected anomaly point, but the subsequent 3 points immediately fall back to around 200 microstrain before the sudden increase, then the suspected anomaly point is judged as a false jump and marked as an anomaly point to be processed. In addition to temporal logic verification, a multi-sensor cross-validation mechanism is also introduced. Because the sensing equipment simultaneously monitors three types of data—earth pressure, strain, and inclination angle—these three types of data are physically correlated. Real soil deformation is usually reflected synchronously across multiple monitoring parameters, while a malfunction in a single sensor only affects the data from that sensor. The specific method for cross-validation is as follows: when a suspected anomaly appears in a monitoring parameter, check whether other monitoring parameters also show significant changes at the same time. If there is a sudden increase in earth pressure, and simultaneously a significant increase in strain or a significant change in inclination angle, it indicates a real change in soil condition, and the sudden increase in earth pressure is a genuine abrupt change. If there is a sudden increase in earth pressure, but strain and inclination angle do not change significantly at that time or in the preceding and following periods, it indicates that the sudden increase in earth pressure may be due to sensor malfunction, increasing the credibility of classifying it as a false jump. For cases where multi-sensor validation results are contradictory, a conservative strategy is adopted: suspected anomalies are retained without being removed to avoid mistakenly deleting real information.

[0058] After the above multi-level discrimination, the identified false anomalies need to be filled in with data. The data filling method is selected based on the duration of the anomaly and the characteristics of the surrounding data. For single points or short-term anomalies of fewer than three consecutive points, a time-weighted interpolation method is used. This method considers the trend information of the data before and after the anomaly. That is, the value of the missing point is equal to the value of the previous normal point plus the difference between the previous two normal points multiplied by the time weight coefficient plus the value of the next normal point plus the difference between the next two normal points multiplied by the time weight coefficient, and then divided by 2. The time weight coefficient is determined according to the time interval between the missing point and the previous and next normal points; the closer the distance, the greater the weight. For example, if the strain sequence is 100, 105, and 110 microstrains, with the fourth point being a spurious jump, the fifth point at 120, and the sixth point at 125 microstrains, then the interpolation for the fourth point is calculated as 110 plus the difference of 110 minus 105 multiplied by 5 (weight 0.6), plus 120 plus the difference of 125 minus 120 multiplied by 5 (weight 0.4), divided by 2, yielding approximately 115 microstrain. This interpolation considers both the upward trend in the earlier segment and the continuation of the trend in the later segment. For long-term anomalies of three or more consecutive points, it indicates a possible continuous sensor malfunction. In this case, interpolation is not performed; instead, this period is marked as missing data, and its weight is reduced or excluded in subsequent analyses to avoid spurious interpolation affecting the overall assessment.

[0059] For genuine abrupt changes, no numerical modifications are made, and the original data remains unchanged. However, a "mutation event" label is added to the data tag. This label includes information such as the time, magnitude, and type of the mutation, which is used by the subsequent risk assessment module. For example, if the dip angle suddenly changes from 0.8 degrees to 4.2 degrees at a certain moment, and this is confirmed as a genuine mutation through time-series logic verification and multi-sensor cross-validation, then the data point remains at 4.2 degrees. However, the data record is marked with "Dip angle mutation event, occurring on a certain day and time of a certain month of a certain year, with a change magnitude of 3.4 degrees, possibly related to the excavation of the foundation pit to the third layer." This labeling information is transmitted to the risk assessment module as an important basis for determining the collapse risk level, because a sharp increase in dip angle often indicates that the slope is entering a state of imminent slippage, which is the highest level of warning signal. After the above quality checks, anomaly identification, false data processing, and retention of genuine mutations, preliminary data is obtained. The difference between preliminary and raw data lies in the following: explicit instrument malfunction data are marked and filled in; short-term spurious jumps are smoothed; and genuine soil deformation mutations are fully preserved and marked with event identifiers. The overall reliability of the data is improved without weakening key early warning information. For example, a monitoring point collected 24 data points within 24 hours. One point, due to a communication error, was identified as a spurious anomaly and filled in (zero value); two points, showing a step increase in earth pressure, were identified as genuine mutations and preserved; one point, showing a single-point dip angle spike, was identified as a spurious anomaly and interpolated; the remaining 20 points were normal. The preliminary data then contains 24 complete data points, with two points marked as "mutation events." These markers will trigger more stringent risk assessment logic in subsequent analyses. After filling in spurious anomalies and preserving genuine mutation points, a continuous and complete preliminary data sequence is obtained.

[0060] In the temperature compensation stage, it is necessary to eliminate the systematic offset caused by changes in ambient temperature to the sensor measurement results. The measurement principle of the sensor determines that its output characteristics are affected by the operating temperature. For strain gauges, temperature changes cause changes in the resistivity and geometry of the strain gauge, resulting in false strain outputs even when no external force is applied; this phenomenon is called temperature drift. For earth pressure sensors, temperature changes affect the characteristic curve of the internal piezoresistive element, causing the output voltage value to change with temperature under the same pressure. If temperature effects are not compensated for, in construction environments with large diurnal temperature differences or significant seasonal variations, the sensor output data will contain a large number of false signals introduced by temperature changes, severely interfering with the judgment of the true stress and strain state of the soil.

[0061] To achieve accurate temperature compensation, a preset temperature compensation coefficient for the sensing device needs to be determined. This coefficient requires a specialized temperature calibration experiment before the sensing device leaves the factory or is deployed. In practice, the sensing device is placed in a precisely temperature-controlled high and low temperature test chamber, with the chamber's temperature range covering extreme temperature conditions that may occur at the construction site, from -10°C to +60°C. A temperature test point is set every 5°C within this temperature range, for a total of 15 test points. At each test point, after the temperature inside the test chamber stabilizes, the sensing device is controlled to collect a set of temperature-strain data pairs under static conditions without external force loading, recording the current ambient temperature and the corresponding strain measurement value. For strain gauges, the strain measurement value should be zero at different temperatures; however, due to temperature drift effects, the actual measured strain value will deviate from zero as the temperature changes. After collecting the complete temperature-strain data pair sequence, a linear regression method is used to fit the relationship curve between temperature and strain measurement values. The goal of linear regression is to find the best-fitting straight line that minimizes the sum of the squared distances of all data points to that line. The slope of the fitted straight line is the preset temperature compensation coefficient kT. Its calculation formula is: kT equals the difference in strain measurements at different temperatures divided by the corresponding temperature difference; that is, kT equals εT2 minus εT1 divided by T2 minus T1, where T1 and T2 are any two different test temperatures, and εT1 and εT2 are the strain values ​​measured at the corresponding temperatures. By performing linear regression fitting on the data from all temperature test points, the obtained temperature compensation coefficient kT reflects the sensitivity of the sensor's strain output to temperature changes. After calculation, this coefficient needs to be entered into the system parameter library of the ground monitoring platform and stored long-term as an inherent characteristic parameter of the sensing device.

[0062] During actual construction monitoring, the sensing equipment collects raw data such as soil pressure, strain, and tilt angle data, while simultaneously acquiring the current ambient temperature data. The current ambient temperature is measured by a temperature sensor integrated inside or near the sensing equipment. This temperature data shares the same timestamp as the soil pressure, strain, and tilt angle data, ensuring accurate time correspondence in temperature compensation calculations. The initial calibration temperature of the sensing equipment is read from the parameter library; this temperature is typically set to 25 degrees Celsius, which is the reference temperature during sensor factory calibration. The preprocessing module calculates the temperature difference between the current ambient temperature and the initial calibration temperature using the formula: the temperature difference equals the current ambient temperature Tmeasured minus 25 degrees Celsius. Based on the preset temperature compensation coefficient kT and the temperature difference, the temperature compensation amount is calculated using the formula: the temperature compensation amount equals kT multiplied by the temperature difference, i.e., kT multiplied by Tmeasured minus 25 degrees Celsius. The physical meaning of the temperature compensation amount is the spurious strain or pressure value introduced due to temperature drift effect at the current ambient temperature.

[0063] To eliminate the influence of temperature drift, the preliminary data is algebraically superimposed with the temperature compensation amount to obtain the temperature-compensated data. For strain data, the temperature-compensated strain value εcompensated is equal to the original measured strain value εmeasured in the preliminary data minus the temperature compensation amount. Subtraction is used here because the temperature compensation amount represents the spurious component introduced by temperature drift, which needs to be subtracted from the measured value to obtain the true strain. The original measured strain value εmeasured is the uncompensated data directly collected from the sensor, containing both the strain component generated by the actual deformation of the measured soil and the spurious strain component caused by temperature changes. The temperature-compensated strain value εcompensated is the true strain value after eliminating temperature interference, reflecting only the deformation state of the measured soil itself. For earth pressure data, the principle and calculation method of temperature compensation are similar to those for strain data; the true earth pressure value is obtained by subtracting the pressure offset caused by the temperature effect.

[0064] After completing the temperature compensation calculation, the compensation effect needs to be verified to ensure the accuracy of the temperature compensation coefficient and the effectiveness of the compensation algorithm. The verification standard is to continuously collect 10 strain data points under the same temperature conditions and calculate the fluctuation range of the 10 temperature-compensated strain values ​​εcompensated. If the fluctuation range is less than or equal to 5 microstrains, it indicates that the temperature interference has been effectively eliminated and the temperature compensation has achieved the expected effect. 5 microstrains is a threshold determined according to the accuracy requirements of strain monitoring. This threshold reflects that after temperature drift is fully suppressed, the random noise of the strain data should be controlled within this range to meet the data quality requirements of subsequent risk assessment. If the verification finds that the fluctuation range exceeds 5 microstrains, it indicates that the currently used temperature compensation coefficient kT may not be accurate enough or the temperature characteristics of the sensor have changed. In this case, the system needs to trigger a recalibration process, notifying maintenance personnel to perform offline temperature calibration of the sensing equipment and update the temperature compensation coefficient in the system parameter library.

[0065] In the sliding window denoising stage, further denoising processing is applied to the temperature-compensated data to suppress residual high-frequency random noise. Although the data quality has been significantly improved after outlier removal and temperature compensation, the temperature-compensated data still exhibits a certain degree of high-frequency fluctuations due to unavoidable circuit thermal noise, quantization errors, and random interference within the sensor during the sampling process. While the amplitude of these high-frequency fluctuations is relatively small, if directly used to calculate derivative-type characteristic parameters such as pressure change rate, they will be amplified by differential operations, causing severe jitter in the characteristic parameters and seriously affecting the stability of risk assessment. Moving average filtering is a classic time-domain denoising method. Its basic principle is to replace the original value of a data point with the average value of the data point and its neighboring points, smoothing out high-frequency random noise through averaging while retaining the low-frequency effective signal in the data.

[0066] The temperature-compensated data are arranged in time series order, and the data series is scanned point by point using a fixed-size sliding data window. The initial size of the sliding data window is set to 5 sampling points, meaning each window contains 5 consecutive data points. For each position in the data series, 5 data points are extracted, centered at that position and preceding and following it, for a total of 5 data points, as the current sliding data window. The arithmetic mean of these 5 data points is calculated within the window using the formula εfiltered = the sum of ε1 + ε2 + ε3 + ε4 + ε5 divided by 5, where ε1, ε2, ε3, ε4, and ε5 represent the values ​​of the first to fifth strain sampling points within the window, and εfiltered is the filtered strain value. The calculated arithmetic mean is taken as the effective value at the center position of the window, and this effective value is the denoised data. As the sliding window moves forward one sampling point along the time series, the arithmetic mean calculation of the data within the window is repeated, generating the denoised data series point by point.

[0067] The sliding data window size of 5 sampling points is determined based on industry practice in strain monitoring and the actual needs of this application. Strain monitoring falls under the category of low-frequency, slow-changing signal monitoring. The change process of soil stress and strain is relatively slow, with the main frequency components concentrated in the low-frequency band, while random noise introduced during sampling is mainly concentrated in the high-frequency band. The 5-point moving average filter can effectively filter out high-frequency random noise such as electromagnetic interference, circuit jitter, and instantaneous sampling errors during the sampling process, while not producing significant hysteresis or ambiguity effects on the real strain signal, such as small structural deformations and slow stress changes. It can retain the strain dynamic characteristics at the second level to meet the requirements of real-time monitoring. From the perspective of computational efficiency, the calculation of the average of 5 data points is minimal and requires no complex calculations. It can be adapted to the limited computing power of edge monitoring devices such as microcontrollers and IoT modules, meeting the response speed requirements of real-time monitoring. Each filtering calculation can be completed within milliseconds. Based on the analysis of error statistics, according to the principles of mathematical statistics, the mean of 5 consecutive sampling points can reduce the standard deviation of random noise by about √5, or 2.2 times. This can suppress the random noise of strain sampling from 10 to 15 microstrains to within 5 microstrains. This noise reduction effect just matches the accuracy threshold set by the temperature compensation circuit and meets the minimum requirements for monitoring accuracy.

[0068] After completing the moving average filtering calculation, the filtering effect is quantitatively evaluated to confirm whether high-frequency noise has been effectively suppressed. The evaluation method is to calculate the variance of the data sequences before and after filtering, and judge the noise reduction effect by comparing the percentage reduction in variance. Variance is a core statistical indicator for measuring the degree of data dispersion. The variance before filtering represents the total dispersion of the original data plus noise, while the variance after filtering represents the dispersion of the effective signal plus residual noise. The percentage reduction in variance directly reflects the degree to which high-frequency noise is suppressed. The filtering effect judgment standard is set as follows: if the variance of the filtered data is reduced by more than or equal to 30% compared to before filtering, the noise reduction effect is judged to be satisfactory, and the current filtering result is retained for subsequent processing. The 30% threshold is determined based on a large amount of engineering measured data verification and industry standards. When the variance reduction reaches 30%, the strain fluctuation caused by high-frequency noise can be stably controlled within 5 micro-strains, which fully meets the data quality requirements for subsequent feature extraction and risk assessment. If the variance reduction after filtering is less than 30%, it means that the noise reduction capability of the current 5-point window is insufficient to effectively suppress the noise components in the data. At this time, the sliding data window size is temporarily adjusted to 7 sampling points, and the moving average filtering calculation is re-executed.

[0069] Adjusting the window size from 5 to 7 uses the principle of minimum increment. The noise reduction effect of the 7-point window is approximately 1.18 times better than that of the 5-point window (√7 / √5). This effectively addresses the issue of noise exceeding the standard even after 5-point filtering, without causing excessive signal smoothing or time lag due to an excessively large window. At a sampling frequency of 1 Hz, the 7-point window lags by only 1 second more than the 5-point window, still meeting the timeliness requirements of real-time monitoring, while preventing the over-filtering of effective signals such as sudden, minute structural deformations. This window adjustment mechanism only makes a single adjustment to the current data segment, rather than permanently modifying the window size parameter. The core of the single-adjustment constraint strategy is to avoid long-term over-filtering caused by local noise fluctuations. If the noise is only temporarily enhanced at a certain moment, after processing that segment using 7-point filtering, the 5-point window is used again to process subsequent data, preventing the loss of detailed features of the strain signal, such as slow stress accumulation and small deformation fluctuations, due to the continuous increase in window size.

[0070] After the sliding window denoising process is completed and validated, the denoised data is marked as preprocessed and stored according to sensor type. Data from earth pressure sensors is marked as preprocessed earth pressure data; data from upper and lower strain gauges is marked as preprocessed strain data; and data from dual-axis tilt sensors is marked as preprocessed tilt angle data. This preprocessed data is then pushed to the cloud server in the data processing layer for analysis by subsequent feature extraction and risk assessment modules.

[0071] S103: Calculate the rate of change of pressure and the spatial pressure gradient between adjacent sensing devices based on the preprocessed earth pressure data, calculate the cumulative strain based on the preprocessed strain data, and calculate the tilt angle based on the preprocessed tilt angle data.

[0072] In S103 above, the preprocessed earth pressure data, preprocessed strain data, and preprocessed tilt angle data obtained after the preprocessing stage have eliminated interference factors such as outliers, temperature drift, and high-frequency noise. However, these data are still raw physical quantities directly measured by sensors and cannot be directly used to assess the stability risk of the soil. Soil collapse is a complex mechanical process, and its occurrence is often not determined by the instantaneous value of a single physical quantity, but by the combined characteristics of the changes of multiple physical quantities over time and space.

[0073] Therefore, the pressure change rate and spatial pressure gradient between adjacent sensing devices are calculated based on the preprocessed earth pressure data; the cumulative strain is calculated based on the preprocessed strain data; and the tilt angle is calculated based on the preprocessed tilt angle data. Specifically, this includes: acquiring the preprocessed first earth pressure data and preprocessed second earth pressure data at the current sampling time and the previous adjacent sampling time; calculating the pressure difference between the preprocessed first earth pressure data and the preprocessed second earth pressure data; dividing the pressure difference by the time interval between the current sampling time and the previous adjacent sampling time to obtain the pressure change rate; and acquiring the preprocessed earth pressure at the same time for two sensing devices deployed at adjacent locations in the soil to be measured. The data is used to calculate the absolute value of the pressure difference between two sensing devices; the absolute value of the pressure difference is divided by the spatial arrangement distance between the two sensing devices to obtain the spatial pressure gradient; the preprocessed strain data at the initial moment is obtained as the reference strain value, and the preprocessed strain data at the current moment is subtracted from the reference strain value to obtain the cumulative strain; the X-axis tilt component and Y-axis tilt component are extracted from the preprocessed tilt angle data, the X-axis tilt component and Y-axis tilt component are converted to radians, the cosine value of the X-axis tilt component and the cosine value of the Y-axis tilt component are calculated, and the product is inverse cosineed to obtain the current tilt angle of the soil to be measured, where the tilt angle represents the axis of the sensing device relative to the direction of gravity.

[0074] Specifically, preprocessed earth pressure data for the current sampling time and the previous adjacent sampling time are read from the time-series database. Since sampling is performed at a frequency of 1 Hz, the time interval between the current sampling time and the previous adjacent sampling time is 1 second. Assuming the current sampling time is t, the previous adjacent sampling time is t-1 seconds. The feature extraction module extracts the preprocessed earth pressure data corresponding to time t from the database, labeling it as the first preprocessed earth pressure data (Pcurrent), and simultaneously extracts the preprocessed earth pressure data corresponding to time t-1 seconds, labeling it as the second preprocessed earth pressure data (Pprevious). Both of these earth pressure data are high-quality data processed by outlier removal, temperature compensation, and sliding window denoising, accurately reflecting the earth pressure state at the corresponding time.

[0075] After acquiring the earth pressure data at these two moments, the feature extraction module calculates the pressure difference between the preprocessed first earth pressure data Pcurrent and the preprocessed second earth pressure data Pprevious. The formula for calculating the pressure difference is that the pressure difference ΔP equals Pcurrent minus Pprevious. The physical meaning of this pressure difference is the absolute change in earth pressure within one second, expressed in kilopascals (kPa). A positive pressure difference indicates an increase in earth pressure, which may be due to increased superstructure load or stress redistribution within the soil, causing the area to bear greater pressure. A negative pressure difference indicates a decrease in earth pressure, which may be due to localized soil loosening or stress transfer to other areas. Regardless of the sign of the pressure difference, its absolute value reflects the magnitude of the pressure change.

[0076] After calculating the pressure difference, to obtain a standardized rate of change index for easier comparison across different time periods and sensors, the feature extraction module divides the pressure difference by the time interval between the current sampling time and the previous adjacent sampling time. Since the sampling frequency is fixed at 1 Hz and the time interval is constant at 1 second, the formula for calculating the pressure change rate v is v = ΔP divided by 1 second, i.e., v = Pcurrent minus Pprevious divided by 1 second, in kilopascals per second. To maintain consistency with the threshold units used in the subsequent risk assessment stage, the unit of the pressure change rate is converted to kilopascals per minute. The conversion formula is v = ΔP divided by 1 second multiplied by 60 seconds per minute, i.e., v = 60 multiplied by ΔP, in kilopascals per minute. The calculated pressure change rate v is a scalar value, and its magnitude directly reflects the rate of change of soil pressure. Based on extensive engineering measurement data and statistical analysis of collapse cases, when the pressure change rate v exceeds 0.5 kilopascals per minute, it indicates that the soil pressure is changing rapidly, which is usually a precursor signal of soil internal structural instability and requires high vigilance.

[0077] The pressure change rate is continuously calculated at a frequency of 1 Hz for each sampling moment, forming a continuous time series of pressure change rates. This time series is written to a time-series database in real time for subsequent risk assessment modules. To improve the accuracy of risk identification, the average pressure change rate over a longer time window is calculated, for example, the average pressure change rate of the most recent 30 minutes (1800 sampling points) is used as an assessment indicator of the medium-term pressure change trend. The formula for calculating the average pressure change rate vavg over the 30-minute time window is: vavg = the absolute value of the pressure Pstart at the beginning of the 30-minute time window minus the absolute value of the pressure Pend at the end of the window, divided by 30 minutes, i.e., vavg = |Pend - Pstart| divided by 30, in kilopascals per minute. This average pressure change rate reflects the overall trend of soil pressure change, filtering out short-term random fluctuations and providing more stable and reliable input parameters for risk assessment.

[0078] For example, a foundation pit is 3 meters deep, and five sensors, labeled Sensor 1 to Sensor 5, are installed along its edge at 3-meter intervals. Sensor 3 is located in the middle of the pit. At 10:00:00 AM on a certain day, the calculation of various characteristic parameters of Sensor 3 at that moment begins. The preprocessed earth pressure data Pcurrent at the current time 10:00:00 is read from the time-series database, which is 25.8 kPa, and the preprocessed earth pressure data Pprevious at the previous sampling time 09:59:59 is 25.2 kPa. The pressure difference ΔP between the two times is calculated to be 0.6 kPa. Since the sampling time interval is 1 second, the instantaneous pressure change rate is 0.6 kPa per second. Converting this rate to kPa per minute and multiplying by 60, we get v = 0.6 multiplied by 60 = 36 kPa per minute. This instantaneous pressure change rate is very high, far exceeding the warning threshold of 0.5 kPa per minute, indicating that the soil pressure is rising rapidly at that moment. To determine whether this is a momentary fluctuation or a continuous trend, the average pressure change rate over the past 30 minutes was calculated. Earth pressure data from sensor 3, retrieved from the database, was 18.5 kPa at 09:30:00 and 25.8 kPa at 10:00:00. The 30-minute average pressure change rate, vavg, was 0.24 kPa per minute. Although the 30-minute average rate of 0.24 kPa per minute is below the threshold of 0.5 kPa per minute, it is still at a relatively high level, indicating that the soil pressure has been slowly increasing over the past half hour. Combined with the abnormally high instantaneous rate of 36 kPa per minute, it can be determined that the soil pressure change has accelerated significantly recently, requiring close monitoring.

[0079] Next, it is necessary to determine which sensors belong to adjacent sensor pairs. In the sensor array deployment scheme, sensors are evenly distributed along the edge of the foundation pit at a certain interval. The sensor spacing is usually set to 1 to 1.5 times the depth of the foundation pit. For example, for a foundation pit with a depth of 2 meters, the sensor spacing is set to 2 to 3 meters. During the configuration phase, the spatial coordinate position information of each sensor is entered into the database. By calculating the Euclidean distance between the sensors, it can be determined which sensor pairs belong to the adjacent relationship. Usually, the two closest sensors are defined as a pair of adjacent sensors. For a linearly arranged sensor array, the adjacent sensor pairs are determined according to the deployment order, with the i-th sensor and the (i+1)-th sensor forming a pair of adjacent sensors. After determining the adjacent sensor pairs, the feature extraction module extracts the preprocessed earth pressure data of the two sensors deployed at adjacent positions in the soil to be measured at the same time from the time series database. Assuming that the two adjacent sensors are labeled as sensor A and sensor B, at the current time t, the preprocessed earth pressure data measured by sensor A is denoted as PA, and the preprocessed earth pressure data measured by sensor B is denoted as PB. These two pressure data points must be collected at the same time to ensure that the calculated pressure gradient reflects the true spatial distribution characteristics at that moment, rather than a spurious gradient caused by time differences.

[0080] After acquiring soil pressure data from two sensors, the feature extraction module calculates the absolute value of the pressure difference between the two sensors. The formula for calculating the absolute value of the pressure difference is ΔPadj, which equals the absolute value of PA minus PB, i.e., ΔPadj equals |PA minus PB|. Absolute value calculation is used because the spatial pressure gradient is concerned with the magnitude of the pressure difference rather than its direction. Regardless of whether the pressure of sensor A is higher than that of sensor B or vice versa, a significant difference indicates stress inhomogeneity in the region. The unit of the absolute value of the pressure difference is kilopascals (kPa), and a larger value indicates a more significant pressure difference between adjacent locations. After calculating the absolute value of the pressure difference, to obtain a standardized gradient index and eliminate the influence of different sensor spacings on the results, the feature extraction module divides the absolute value of the pressure difference by the spatial deployment distance between the two sensors. The spatial deployment distance LAB is actually measured during the sensor array deployment phase, and this distance information has been entered into the system database. The formula for calculating the spatial pressure gradient G is G equal to ΔPadj divided by LAB, i.e., G equals |PA minus PB| divided by LAB, with the unit being kilopascals per meter. The physical meaning of spatial pressure gradient is the average change in soil pressure per meter along the spatial direction. This parameter reflects the non-uniformity of the stress field within the soil. According to soil mechanics principles and engineering experience, when the spatial pressure gradient exceeds a certain threshold, such as 10 kPa per meter, it indicates a significant difference in stress state between adjacent areas. Stress concentration or localized loosening may exist within the soil. This stress non-uniformity can cause high-pressure areas to reach their strength limit first, leading to local instability and subsequently triggering large-scale collapse. The spatial pressure gradient is calculated for each adjacent sensor pair in the sensor array, forming a spatial pressure gradient distribution map covering the entire monitoring area. This distribution map visually displays which areas have abnormally high stress gradients, helping construction managers quickly locate potentially high-risk areas.

[0081] For example, sensor 3's adjacent sensors are identified as sensor 2 and sensor 4. At time 10:00:00, the earth pressure data for sensor 2 (P2) is 24.2 kPa, for sensor 3 (P3) is 25.8 kPa, and for sensor 4 (P4) is 27.5 kPa. The absolute value of the pressure difference ΔP23 between sensor 2 and sensor 3 is calculated to be 1.6 kPa. The installation distance between sensor 2 and sensor 3 (L23) is 3 meters, and the spatial pressure gradient G23 is 0.53 kPa / m. The absolute value of the pressure difference ΔP34 between sensor 3 and sensor 4 is calculated to be 1.7 kPa, and the spatial pressure gradient G34 is 0.57 kPa / m. Both of these spatial pressure gradient values ​​are relatively small, far below the warning threshold of 10 kPa / m, indicating that the pressure distribution around sensor 3 is relatively uniform, and there is no significant stress concentration. To more comprehensively assess the stress distribution of the entire monitoring area, the spatial pressure gradients between sensor 1 and sensor 2, and between sensor 4 and sensor 5, were also calculated. Assume the earth pressure at sensor 1 is 23.1 kPa and at sensor 5 is 35.2 kPa. The absolute value of the pressure difference ΔP45 between sensor 4 and sensor 5 is 7.7 kPa, and the spatial pressure gradient G45 is 2.57 kPa / m. Although this gradient value does not reach the threshold of 10 kPa / m, it is significantly higher than that of other sensor pairs, indicating that the pressure in the area near sensor 5 is significantly higher than that in adjacent areas, and there is a certain degree of stress non-uniformity. G45 is marked as a feature point of interest, and the risk weight of this area will be appropriately increased during risk assessment.

[0082] A new initial moment is determined as the reference moment for calculating the cumulative strain, and the strain value corresponding to this initial moment is used as the reference strain value. The selection of the reference moment is crucial; it should be the moment when the sensor is installed and the soil is in a relatively stable state. The initial stabilization period can be defined as 24 hours after sensor installation. During these 24 hours, the sensor is in full contact with the surrounding soil and reaches mechanical equilibrium, allowing the measured strain value to accurately reflect the initial stress state of the soil. The feature extraction module calculates the mean of the preprocessed strain data from all sampling points within this 24-hour period after sensor installation, and uses this mean as the reference strain value εstatic. The formula for calculating the reference strain value is: εstatic equals the sum of all strain sampling points within 24 hours divided by the total number of sampling points. The total number of sampling points equals 24 hours multiplied by 3700 seconds per hour multiplied by a sampling frequency of 1 Hz, i.e., 86400 sampling points. The calculated reference strain value εstatic represents the initial strain state of the soil at the sensor location, and this value is permanently stored in the database as a reference for all subsequent calculations of the cumulative strain of this sensor.

[0083] After determining the baseline strain value, the feature extraction module calculates the cumulative strain at each subsequent sampling time. Specifically, it retrieves the preprocessed strain data εcurrent from the time-series database and subtracts it from the baseline strain value εstatic to obtain the cumulative strain εacc. The formula for calculating the cumulative strain is εacc = εcurrent minus εstatic. The physical meaning of this formula is how much the strain value at the current time has changed relative to the strain value in the initial steady state; this change represents the total deformation of the soil from the initial time to the current time. The unit of the cumulative strain is microstrain, and the value can be positive or negative. A positive value indicates that the soil has undergone tensile deformation relative to its initial state, i.e., an increase in strain, which usually corresponds to an increase in compressive stress or outward expansion of the soil. A negative value indicates that the soil has undergone compressive deformation, i.e., a decrease in strain, which usually corresponds to a decrease in compressive stress or inward contraction of the soil. Regardless of the sign of the cumulative strain, the absolute value reflects the severity of the cumulative soil deformation.

[0084] For example, the baseline strain value εstatic measured by sensor 3 during the 24-hour stabilization period after installation is equal to 120 microstrain. At the current time 10:00:00, the preprocessed strain data εcurrent of sensor 3, read from the database, is equal to 685 microstrain. The cumulative strain εacc is equal to 565 microstrain. This cumulative strain of 565 microstrain exceeds the precursor threshold of 500 microstrain, indicating that the soil at the location of sensor 3 has accumulated significant plastic deformation from its initial stable state to the current time, and the internal structure of the soil has undergone irreversible deterioration, approaching the critical state of instability. The system marks this moment as a characteristic point where the cumulative strain exceeds the standard. Further calculation of the 30-minute cumulative strain is performed. The strain data εstart of sensor 3 at 09:30:00, 30 minutes ago, is equal to 610 microstrain, and the strain data εend at the current time is equal to 685 microstrain. The 30-minute cumulative strain εacc30 is equal to 75 microstrain. Although 75 microstrain is below the rapid accumulation threshold of 100 microstrain, considering that the total accumulated strain has reached 565 microstrain, which is at a high level, the fact that 75 microstrain is added within 30 minutes indicates that the soil deformation has not stopped and the risk continues to accumulate.

[0085] A dual-axis tilt sensor is used to monitor the tilt state of the sensor tube in real time. A dual-axis tilt sensor, such as the MCA426M model, can simultaneously measure the tilt angle components of the sensor relative to the horizontal plane in two mutually orthogonal directions. According to the definition of the sensor coordinate system, the axial direction of the sensor tube is defined as the Z-axis, and the two orthogonal directions perpendicular to the Z-axis are defined as the X-axis and Y-axis, respectively. The dual-axis tilt sensor outputs two measured values, the X-axis tilt component θx and the Y-axis tilt component θy, in degrees. The X-axis tilt component θx represents the tilt angle of the sensor tube relative to the vertical direction in the XZ plane, and the Y-axis tilt component θy represents the tilt angle of the sensor tube relative to the vertical direction in the YZ plane. These two tilt components are the measurement results directly output after internal signal processing by the sensor and are already included in the pre-processed tilt angle data.

[0086] The feature extraction module extracts the X-axis tilt component θx and Y-axis tilt component θy from the preprocessed tilt angle data. The X-axis tilt component represents the pitch angle of the sensing device's rotation around the Y-axis, and the Y-axis tilt component represents the roll angle of the sensing device's rotation around the X-axis. The dual-axis tilt sensor is based on the principle of a MEMS accelerometer, calculating the tilt angle by measuring the components of gravitational acceleration in two orthogonal directions. When the sensing device tilts relative to the horizontal plane, the gravitational acceleration component sensed by the internal accelerometer changes, and the sensor outputs the corresponding tilt angle value through inverse trigonometric function calculation.

[0087] To obtain the total tilt angle of the sensing device relative to the vertical direction, the tilt components of the X and Y axes need to be spatially synthesized. This synthesis is not a simple square root operation, but a rigorous calculation based on the rotational geometry of three-dimensional space. The feature extraction module first converts the tilt components from angles to radians, with the formulas θxrad equal to θx multiplied by π divided by 180°, and θyrad equal to θy multiplied by π divided by 180°. Then, the cosine values ​​of the X-axis tilt component (cos(θxrad)) and Y-axis tilt component (cos(θyrad)) are calculated, and multiplying the two cosine values ​​yields the projection component of the sensing device's axis in the direction of gravity. The physical meaning of this projection component is the cosine of the angle between the device's axis and the vertical direction when the sensing device tilts simultaneously in two orthogonal directions. Finally, the arccosine operation is performed on the projection component to obtain the angle θ between the axis of the sensing device and the direction of gravity, in radians. Then, it is converted back to degrees to obtain the final tilt angle. The calculation formula is θ = arccos(cos(θxrad) * cos(θyrad)) * 180 / π.

[0088] The mathematical principle behind this calculation method is based on the direction cosine theorem in three-dimensional space. In a three-dimensional Cartesian coordinate system, the cosine of the angle between a vector and a coordinate axis is called the direction cosine, and the sum of the squares of the three direction cosines equals 1. For a sensing device, let its axis be the positive Z-axis. When the device tilts around the Y-axis by an angle θx, the projected length of the Z-axis in the vertical direction becomes cos(θx). When it tilts around the X-axis by an angle θy, this projected length is shortened again to cos(θx) multiplied by cos(θy). Therefore, the cosine of the angle between the sensing device's axis and the direction of gravity is cos(θx) * cos(θy), and the angle can be obtained through inverse cosine calculation. For example, at a certain moment, the dual-axis tilt sensor of sensor 3 measures the X-axis tilt component θx as 0.6 degrees and the Y-axis tilt component θy as 0.4 degrees. First, converting to radians, θx (rad) equals 0.6 multiplied by π divided by 180, approximately 0.01047 radians, and θy (rad) equals 0.4 multiplied by π divided by 180, approximately 0.00698 radians. Calculating the cosine values, cos(0.01047) is approximately 0.99995, and cos(0.00698) is approximately 0.99998. Their product is 0.99993. Performing an inverse cosine operation, arccos(0.99993) is approximately 0.01257 radians. Converting to degrees, 0.01257 multiplied by 180 divided by π is approximately 0.72 degrees. Therefore, the total tilt angle of the soil at the location of sensor 3 is 0.72 degrees.

[0089] The calculated tilt angle θ is a scalar value representing the degree to which the sensing device deviates from the vertical direction; a larger value indicates a more severe soil tilt. This tilt angle parameter is stored in a time-series database for subsequent use by the risk assessment module. Based on extensive engineering measurement data, when the tilt angle exceeds 0.5 degrees, it indicates that the soil has undergone noticeable geometric deformation, requiring attention; when the tilt angle exceeds 1.0 degrees, it indicates that the soil is experiencing significant lateral displacement, with a high risk of collapse; when the tilt angle exceeds 2.0 degrees, the soil is in a critical instability state, and immediate reinforcement measures must be taken.

[0090] S104: Determine the target weighting coefficient based on the current construction stage and real-time data characteristics of pressure change rate, tilt angle, and strain accumulation.

[0091] In S104 above, the traditional static weighting method typically uses pre-set fixed weight values ​​based on the construction stage. For example, during the excavation stage, the weight for the rate of pressure change is set to 0.4, the weight for the inclination angle is set to 0.3, and the weight for the cumulative strain is set to 0.3. While this fixed weighting method is simple and easy to use, it has significant drawbacks. Soil collapse is a dynamic evolution process, and the degree of danger and early warning value of various physical parameters changes constantly at different times. For example, when the internal pressure of the soil is close to its limit state, the early warning value of pressure-related parameters should be significantly increased, while when the soil deformation rate increases sharply, strain-related parameters should receive higher weights. If fixed weights are always used, the risk assessment will be slow to respond to the actual risk state, unable to capture changes in key risk signals in a timely manner, and reducing the accuracy and timeliness of early warnings.

[0092] Therefore, this application retrieves the foundation weights from the preset weight library as initial values ​​based on the current construction stage, and then calculates the corresponding adjustment factors based on the real-time data characteristics of each characteristic parameter. The foundation weights are then dynamically corrected using these adjustment factors to obtain the final target weight coefficient. The target weight coefficient is determined based on the current construction stage and the real-time data characteristics of the pressure change rate, tilt angle, and strain accumulation. Specifically, this includes: retrieving the foundation pressure rate weight corresponding to the pressure change rate, the foundation tilt angle weight corresponding to the tilt angle, and the foundation strain weight corresponding to the strain accumulation from the preset weight matching library based on the current construction stage; using the pre-processed earth pressure data at the current moment as the current absolute earth pressure level, calculating the pressure ratio between the current absolute earth pressure level and the design earth pressure benchmark value, squaring the pressure ratio, multiplying it by the first sensitivity coefficient, and adding it to 1 to obtain the pressure level adjustment factor; multiplying the foundation pressure rate weight by the pressure level adjustment factor to obtain the target weight for the pressure change rate; and so on. The tilt angle change rate is calculated at continuous sampling times. The absolute value of the tilt angle change rate is calculated as the tilt angle ratio to the angular velocity warning threshold. The tilt angle ratio is multiplied by the second sensitivity coefficient and added to 1 to obtain the tilt angular velocity adjustment factor. The basic tilt angle weight is multiplied by the tilt angular velocity adjustment factor to obtain the target weight of the tilt angle. The strain ratio of the accumulated strain to the strain limit threshold is calculated. The strain ratio is squared, multiplied by the third sensitivity coefficient, and added to 1 to obtain the strain level adjustment factor. The basic strain weight is multiplied by the strain level adjustment factor to obtain the target weight of the accumulated strain. The target weights of the pressure change rate, tilt angle, and accumulated strain are output as target weight coefficients.

[0093] Specifically, after acquiring the pressure change rate v, tilt angle θ, and strain accumulation εacc data, the basic weights are first determined as the starting point for dynamic adjustment. A pre-set weight matching library is built-in, storing the basic weight parameters corresponding to different construction stages. The current construction scenario is divided into three construction stages, with one weight combination corresponding to each stage. The pre-set weight matching library contains weight combinations for the three construction stages: for the excavation stage, the basic pressure rate weight αbase equals 0.4, the basic tilt angle weight βbase equals 0.3, and the basic strain weight γbase equals 0.3; for the support stage, the basic pressure rate weight αbase equals 0.3, the basic tilt angle weight βbase equals 0.4, and the basic strain weight γbase equals 0.3; and for the backfilling stage, the basic pressure rate weight αbase equals 0.2, the basic tilt angle weight βbase equals 0.3, and the basic strain weight γbase equals 0.5. These basic weight combinations are the optimal initial weight configurations obtained through gradient descent training optimization based on data from similar deep foundation pit collapse cases over the past 5 years, reflecting the average importance of each parameter in different construction stages. The risk assessment module reads the current construction stage identifier from the system configuration parameters and calls the corresponding basic weight combination from the preset weight matching library based on the identifier.

[0094] After obtaining the basic weights, the pressure level adjustment factor used to correct the pressure change rate is calculated. When the absolute pressure level inside the soil is close to or reaches the soil's ultimate strength, even if the pressure change rate is the same, the degree of danger is much higher than under low pressure levels because the soil's safety margin is very small under high pressure; any small pressure increment can lead to instability. Therefore, the weight of the pressure change rate is amplified according to the current absolute pressure level; the higher the pressure level, the larger the amplification factor. The preprocessed earth pressure data Pcurrent at the current moment is used as the current absolute earth pressure level, and the design earth pressure benchmark value Pdesign for the construction area is read from the parameter library. The design earth pressure benchmark value is the normal working pressure value calculated according to soil type and pit depth based on soil mechanics principles, representing the pressure level that the soil should withstand in a stable state. The pressure ratio RP between the current absolute earth pressure level and the design earth pressure benchmark value is calculated as Pcurrent divided by Pdesign. When the pressure ratio RP equals 1, it indicates that the current pressure is within the normal design range; when RP is greater than 1, it indicates that the current pressure has exceeded the design benchmark and entered a high-risk area; the larger the RP, the more severe the pressure exceedance. To quantify the impact of this pressure level as a weighting adjustment factor, the pressure ratio RP is squared to obtain the square of RP. Squaring amplifies the impact of high pressure ratios, making the adjustment factor more sensitive to pressure exceeding limits. Then, the square of RP is multiplied by a first sensitivity coefficient k1, a dynamic parameter reflecting the degree of unevenness in spatial stress distribution. The calculation method for the first sensitivity coefficient will be detailed later. The result of multiplying the square of RP by k1 is added to 1, yielding the pressure level adjustment factor FP, which equals 1 plus the square of RP multiplied by k1. This formula is designed so that when the pressure ratio RP equals 1 (i.e., normal pressure), the adjustment factor FP is approximately equal to 1 plus k1, providing a moderate adjustment to the base weight; when RP is much greater than 1, the adjustment factor FP increases significantly, greatly increasing the weight of the pressure change rate. Multiplying the base pressure rate weight αbase by the pressure level adjustment factor FP yields the preliminary target weight αtarget for the pressure change rate.

[0095] In addition, the first sensitivity coefficient k1 reflects the amplifying effect of the non-uniformity of spatial stress distribution on pressure risk. When the pressure difference between adjacent areas is large, it indicates that there is stress concentration, and the pressure parameter is more dangerous.

[0096] The spatial ratio of the spatial pressure gradient to the safe spatial gradient benchmark value is calculated, and the spatial ratio is used as the first sensitivity coefficient. The full spatial gradient benchmark value Gsafe is determined based on the homogeneity of the soil, and the typical value is 5 kPa per meter.

[0097] Next, the tilt angular velocity adjustment factor is calculated to correct the weight of the tilt angle. The rate of soil tilt, i.e., angular velocity, is a key indicator for judging the urgency of soil slippage. If the tilt angle is large but changes slowly, it indicates that the soil is in a slow creep stage and there is still time to deal with it. However, if the tilt angle increases rapidly, it indicates that the soil is accelerating its slippage and collapse is imminent, requiring immediate warning. Therefore, the tilt angle weight needs to be dynamically adjusted according to the tilt angle change rate; the faster the angular velocity, the greater the weight. The tilt angle change rate is calculated based on the tilt angle data from continuous sampling times. The tilt angle θcurrent at the current time t and the tilt angle θprevious at the previous sampling time t-1 seconds are read from the time series database. The tilt angle change rate dθdt is calculated as θcurrent minus θprevious divided by 1 second, in degrees per second. The tilt angle change rate may be positive, indicating that the tilt is intensifying, or negative, indicating that the tilt is slowing down. The absolute value of the tilt angle change rate |dθdt| is calculated to eliminate the influence of the sign. The angular velocity warning threshold dθlimit is read from the system parameter library. This threshold is determined based on the soil type and the geometry of the foundation pit, representing the critical angular velocity value at which the soil enters a rapid sliding state; a typical value is 0.01 degrees per second. The ratio of the absolute value of the rate of change of the inclination angle to the inclination angle of the angular velocity warning threshold, Rθ, is calculated as |dθdt| divided by dθlimit. When the inclination angle ratio Rθ is less than 1, it indicates that the angular velocity is still within a safe range; when Rθ is greater than or equal to 1, it indicates that the angular velocity has reached or exceeded the warning value, and the soil sliding is accelerating. The inclination angle ratio Rθ is multiplied by the second sensitivity coefficient k2 and then added to 1 to obtain the inclination angular velocity adjustment factor Fθ, which is equal to 1 plus Rθ multiplied by k2. The second sensitivity coefficient k2 reflects the amplifying effect of the fluctuation of the inclination angle itself on the risk. The basic inclination angle weight βbase is multiplied by the inclination angular velocity adjustment factor Fθ to obtain the preliminary target weight βtarget of the inclination angle.

[0098] Furthermore, the second sensitivity coefficient k2 reflects the amplifying effect of tilt angle fluctuation on angular velocity risk; severe tilt angle fluctuation indicates that the soil is in an unstable state. The system acquires a sequence of historical tilt angle data within a preset time period (which can be set to 10 minutes) prior to the current moment. The standard deviation of the historical tilt angle data sequence is calculated as the current tilt angle fluctuation rate. This current tilt angle fluctuation rate is divided by the baseline tilt angle fluctuation rate to obtain the fluctuation ratio. This fluctuation ratio is used as the second sensitivity coefficient k2. The baseline tilt angle fluctuation rate is the fluctuation level measured during the 24-hour stabilization period after sensor installation.

[0099] Finally, a strain level adjustment factor is calculated to correct the weight of the accumulated strain. The transition of soil from the elastic stage to the plastic stage until failure is a process of continuous strain accumulation. When the accumulated strain approaches the soil's ultimate strain, the internal structure of the soil has severely deteriorated and is likely to become unstable at any time. At this point, the warning value of the accumulated strain parameter should be significantly increased. The current accumulated strain εacc is compared with the strain limit threshold εlimit. The strain limit threshold is obtained from a geotechnical test database based on the soil type and represents the critical strain value at which this type of soil reaches the failure state; a typical value is 800 microstrains. The strain ratio Rε, calculated as εacc divided by εlimit, is used to calculate the ratio. When the strain ratio Rε is less than 1, it indicates that the strain has not yet reached its limit; when Rε is close to or exceeds 1, it indicates that the soil is about to enter or has already entered the failure stage. The square of the strain ratio Rε is obtained by squaring Rε. This squaring operation produces a strong weight amplification effect when the strain approaches its limit. Multiplying the square of Rε by the third sensitivity coefficient k3 and adding it to 1 yields the strain level adjustment factor Fε, which is equal to 1 plus the square of Rε multiplied by k3. The third sensitivity coefficient k3 reflects the dynamic amplification effect of soil deformation rate on risk. Multiplying the base strain weight γbase by the strain level adjustment factor Fε yields the preliminary target weight γtarget for the cumulative strain.

[0100] Furthermore, the third sensitivity coefficient k3 reflects the amplifying effect of the strain change rate on strain risk; rapid strain accumulation indicates that the soil is accelerating its instability. The strain accumulation at the current sampling time and the previous adjacent sampling time is obtained, and the difference between the strain accumulation at the current sampling time and the previous adjacent sampling time is calculated. This difference is then divided by the time interval between the current sampling time and the previous adjacent sampling time to obtain the current strain rate. The ratio of the current strain rate to the elastic deformation rate threshold is calculated, and this ratio is used as the third sensitivity coefficient. The elastic deformation rate threshold represents the normal deformation rate of the soil in the elastic stage, with a typical value of 10 microstrains per second.

[0101] After calculating the adjustment factors described above, three preliminary target weights, αtarget, βtarget, and γtarget, are obtained. However, the sum of these three weights is usually not equal to 1, and normalization is required to ensure that the total weight sum is 1. The sum of the three preliminary target weights, S = αtarget + βtarget + γtarget, is calculated. Then, each preliminary target weight is divided by the sum of the weights to obtain the final normalized target weights. The final target weight αfinal for the pressure change rate is equal to αtarget divided by S, the final target weight βfinal for the tilt angle is equal to βtarget divided by S, and the final target weight γfinal for the strain accumulation is equal to γtarget divided by S. αfinal, βfinal, and γfinal are then output as target weight coefficients to the risk index calculation module.

[0102] Taking a real-world scenario of excavation in a sandy soil foundation pit as an example, the basic weights for the excavation stage are retrieved from the preset weight library: αbase = 0.4, βbase = 0.3, and γbase = 0.3. The current absolute pressure Pcurrent = 28 kPa, the design baseline Pdesign = 20 kPa, the pressure ratio RP = 1.4, the spatial gradient G = 2.5 kPa / m, the safety baseline Gsafe = 5 kPa / m, the first sensitivity coefficient k1 = 0.5, the pressure level adjustment factor FP = 1 + 1.4 squared multiplied by 0.5 = 1.98, and the initial pressure rate weight αtarget = 0.4 multiplied by 1.98 = 0.792. The rate of change of tilt angle |dθdt| equals 0.008 degrees per second, the warning threshold dθlimit equals 0.01 degrees per second, the tilt angle ratio Rθ equals 0.8, the tilt angle fluctuation ratio k2 equals 1.2, the tilt angle velocity adjustment factor Fθ equals 1 plus 0.8 multiplied by 1.2 equals 1.96, and the initial tilt angle weight βtarget equals 0.3 multiplied by 1.96 equals 0.588. The cumulative strain εacc equals 620 microstrains, the limiting threshold εlimit equals 800 microstrains, the strain ratio Rε equals 0.775, the strain rate ratio k3 equals 1.5, the strain level adjustment factor Fε equals 1 plus 0.775 squared multiplied by 1.5 equals 1.9, and the initial strain weight γtarget equals 0.3 multiplied by 1.9 equals 0.57. The initial weight sum S equals 0.792 + 0.588 + 0.57, which equals 1.95. After normalization, the final target weights αfinal are approximately 0.406, βfinal approximately 0.302, and γfinal approximately 0.292, with their sum being exactly 1. This dynamic weighting increases the pressure parameter weight from the basic 0.4 to 0.406, reflecting the danger of the current high-pressure state and achieving adaptive adjustment of the weights to real-time risk characteristics.

[0103] S105: The current risk index is obtained by weighting the pressure change rate, tilt angle and strain accumulation based on the target weight coefficient.

[0104] In step S105 above, after obtaining the dynamically adjusted target weight coefficients, the target weight coefficients are weighted and calculated with the corresponding characteristic parameters to comprehensively assess the real-time collapse risk level of the soil and output a quantified risk index value. The risk index is the core output indicator of the entire early warning system. It integrates monitoring data from three different physical dimensions—pressure change rate, tilt angle, and strain accumulation—into a unified risk metric through a scientific weighting mechanism. This value can intuitively reflect the current stability state and collapse hazard level of the soil.

[0105] The target weight determination module receives the final target weights αfinal for the pressure change rate, βfinal for the tilt angle, and γfinal for the strain accumulation. Simultaneously, the feature extraction module receives the real-time values ​​of the pressure change rate v, tilt angle θ, and strain accumulation εacc. These three feature parameters have undergone a complete data preprocessing workflow, including outlier removal, temperature compensation, and moving average filtering, ensuring reliable and interference-free data quality. Numerical normalization is performed on these three feature parameters because their physical dimensions and numerical ranges are completely different: the pressure change rate v is in kilopascals per minute with a typical range of 0 to 2; the tilt angle θ is in degrees with a typical range of 0 to 5; and the strain accumulation εacc is in microstrain with a typical range of 0 to 1000. Directly weighting these raw values ​​of different magnitudes would lead to parameters with larger values ​​dominating the risk index while parameters with smaller values ​​have a weak impact, rendering the weighting coefficients meaningless. To eliminate the influence of dimensions and numerical ranges, a maximum value normalization method is used to map each feature parameter to a unified interval of 0 to 1. The maximum reference values ​​for each characteristic parameter are read from the parameter configuration library. The maximum reference value for the pressure change rate, vmax, is determined based on the soil type: 2.0 kPa / min for sandy soil and 1.5 kPa / min for cohesive soil. This maximum value represents the limit of the pressure change rate that this type of soil may experience before collapse. The maximum reference value for the tilt angle, θmax, is determined based on the geometric parameters of the foundation pit, with a typical value of 5 degrees, representing the limit angle of the pipe's tilt. The maximum reference value for the cumulative strain, εmax, is determined based on the soil's mechanical properties: 1000 microstrain for sandy soil and 1200 microstrain for cohesive soil, representing the limit strain at which the soil reaches the failure state. The normalized pressure change rate, vnorm, is calculated as v divided by vmax. When the actual pressure change rate v equals 0, vnorm equals 0, indicating no risk. When v equals vmax, vnorm equals 1, indicating extremely high risk. When v is between 0 and vmax, vnorm is proportionally mapped to the 0-1 interval. The normalized tilt angle θnorm is calculated by dividing θ by θmax, and the normalized strain accumulation εnorm is calculated by dividing εacc by εmax. All three normalized parameters are uniformly mapped to the dimensionless interval of 0 to 1, eliminating the influence of differences in the original numerical range, so that the weight coefficients in the subsequent weighted calculation can truly reflect the importance of each parameter.

[0106] After normalization, a weighted calculation is performed to construct the current risk index. The risk index is calculated using a linear weighted summation formula, which has the advantages of clear physical meaning, high computational efficiency, and strong interpretability. The current risk index is calculated as αfinal multiplied by vnorm plus βfinal multiplied by θnorm plus γfinal multiplied by εnorm. The physical meaning of this formula is that it linearly combines the three normalized characteristic parameters according to their actual contribution to the risk. The contribution of each parameter is equal to the normalized value multiplied by the corresponding target weight, and the sum is finally obtained to obtain the comprehensive risk index. Since the values ​​of the three normalized parameters are all between 0 and 1, and the sum of the three target weights is exactly equal to 1 after normalization, the theoretical range of the calculated current risk index, Risk Index, is also between 0 and 1. 0 represents a completely risk-free state, and 1 represents a state of extreme risk, i.e., collapse is imminent. Values ​​between 0 and 1 proportionally reflect the level of risk. This standardized risk index of 0 to 1 has strong engineering practicality, and construction and management personnel can intuitively understand the meaning of the risk index. For example, a risk index of 0.3 indicates a risk level of 30%, while a risk index of 0.8 indicates a risk level of 80% requiring high vigilance. The calculated current risk index value is stored in a time-series database, with timestamps accurate to the second, facilitating the subsequent generation of risk trend curves and historical data analysis.

[0107] Continuing with the previous case study of the sandy soil foundation pit excavation stage, the target weights αfinal have been calculated to be approximately 0.406, βfinal approximately 0.302, and γfinal approximately 0.292. Assume the actual monitoring data at the current moment are: pressure change rate v = 0.8 kPa / min, tilt angle θ = 0.8 degrees, and strain accumulation εacc = 620 microstrains. For sandy soil, the maximum reference value for pressure change rate vmax = 2.0 kPa / min, the maximum reference value for tilt angle θmax = 5 degrees, and the maximum reference value for strain accumulation εmax = 1000 microstrains. The normalized pressure change rate vnorm is calculated to be 0.4, the normalized tilt angle θnorm to be 0.16, and the normalized strain accumulation εnorm to be 0.62. After performing a weighted calculation, the current risk index is 0.406*0.4+0.302*0.16+0.292*0.62=0.39176, which indicates that the overall risk level of the soil is currently at 39.1%.

[0108] S106: Input the time series of pressure change rate, tilt angle and strain accumulation within the preset historical time window into the preset long short-term memory network prediction model for processing, and output the risk probability of soil collapse in the future period.

[0109] In S106 above, after calculating the current risk index, the trend of soil collapse risk in future periods is further predicted, achieving a leap from passive monitoring to proactive early warning. The current risk index only reflects the immediate state of the soil, but soil collapse is a gradual and cumulative dynamic evolution process, often requiring tens of minutes or even hours to develop from the initial risk to the final collapse. If early warning is based solely on the current risk index, the soil may already be in a critical instability state when the risk index exceeds the threshold, leaving extremely limited emergency response time for on-site personnel, making timely evacuation and effective handling difficult. Therefore, this application uses a pre-set long short-term memory network prediction model to predict risks in future periods, captures the development trend of risks by analyzing the temporal evolution patterns of historical data, and triggers early warnings when the risk has not yet reached the danger threshold but has already shown a deteriorating trend, thus gaining valuable emergency response time for on-site personnel.

[0110] Before inputting the time series data into the pre-defined Long Short-Term Memory (LSTM) network prediction model, the model must first be constructed. The data processing layer extracts historical monitoring data from the InfluxDB time series database as training samples for the model. The sample construction follows a sliding window approach, with each training sample containing the input sequence and corresponding label results. The input sequence consists of time series data of the pressure change rate v, tilt angle θ, and strain accumulation εacc continuously collected over the past hour. Since the sampling frequency is 1Hz, a total of 3700 sampling moments are generated within one hour. Each sampling moment contains the values ​​of the three feature parameters, thus a single input sequence constitutes a 3700 x 3 two-dimensional data matrix. The label results are binary labels indicating whether soil collapse will occur within 30 minutes after the corresponding moment of the input sequence; a collapse is marked as 1, and no collapse is marked as 0. Historical monitoring data from similar deep foundation pit construction projects over the past five years were extracted from the database. A dataset containing complete collapse event records was selected, and tens of thousands of training samples were generated using the sliding window method. Positive samples (collapse precursor samples) accounted for approximately 10% of the total samples, while negative samples (normal construction samples) accounted for approximately 90%. To avoid sample imbalance that could lead to a bias in predicting the negative class, the system performed oversampling on positive samples or undersampling on negative samples, adjusting the ratio of positive to negative samples to a reasonable range of 1:3.

[0111] The model's network structure employs a deep learning architecture with multiple stacked Long Short-Term Memory (LSTM) networks. The input layer receives a 3700 x 3 dimensional temporal data matrix, inputting the evolution sequence of the three feature parameters over time. The first LSM layer contains 64 neurons, each processing the input sequence through three gating mechanisms: a forget gate, an input gate, and an output gate. The forget gate determines which information to forget from the previous cell state, the input gate determines which information from the current input needs to be stored in the cell state, and the output gate determines which information to output from the cell state to the hidden state. This gating mechanism enables the LSM network to effectively capture long-term dependencies in temporal data, overcoming the gradient vanishing problem in traditional recurrent neural networks when processing long sequences. The output of the first LSM layer serves as the input to the second LSM layer, which also contains 64 neurons. This two-layer stacking enhances the network's ability to learn complex temporal patterns. The output of the second LSM layer is connected to a Dropout layer with a Dropout ratio of 0.3, randomly discarding 30% of neuron connections during training to prevent overfitting and improve generalization performance. The dropout layer output is connected to a fully connected output layer. The output layer contains one neuron and uses the sigmoid activation function to map the network's computation results to the interval between 0 and 1. The output value is the probability of soil collapse within the next 30 minutes.

[0112] The model training employed the Adam optimizer and the binary cross-entropy loss function. The constructed training sample set was divided into a training set and a validation set in an 8:2 ratio. The training set was used for model parameter learning, while the validation set was used to evaluate model performance and prevent overfitting. During training, the batch size was set to 32, the initial learning rate to 0.001, and the number of training epochs to 100. After each training epoch, the accuracy and loss value on the validation set were calculated. An early stopping mechanism was triggered to terminate training when the validation set accuracy failed to improve for five consecutive epochs, preventing overtraining and performance degradation. After training, the model achieved an accuracy of over 90% on the validation set and a recall rate (the recognition rate of real collapse events) of over 85%, meeting engineering application requirements. The trained model was then output as a pre-defined long short-term memory network prediction model.

[0113] The prediction module of the data processing layer reads real-time monitoring data on pressure change rate, tilt angle, and cumulative strain from the time-series database for the past hour prior to the current moment. This data has undergone a complete preprocessing process, including outlier removal, temperature compensation, and moving average filtering, ensuring the reliability of the input data quality. The 3D feature data from 3700 time points are organized into a 3700x3 input matrix and fed into a pre-trained Long Short-Term Memory (LSTM) network prediction model for forward propagation calculation. The input data first enters the first layer of the LSM network, which processes the input sequence step by step. In each time step, the gating mechanism updates the cell and hidden states based on the current input and the hidden state of the previous time step, gradually extracting dynamic features and evolutionary patterns from the time-series data. The output sequence of the first layer is passed to the second layer of the LSM network for deeper feature learning. The final hidden state of the second layer contains a comprehensive representation of the information from the entire one-hour time-series data. This hidden state is then fed into the output layer after passing through a Dropout layer. The sigmoid function of the output layer calculates a value between 0 and 1, which represents the probability of soil collapse within the next 30 minutes. The calculated risk probability values ​​are stored in a time-series database and pushed to the application-layer monitoring interface in real time, where the risk probability is displayed as a percentage.

[0114] S107: Determine the dynamic early warning threshold based on the soil type of the soil to be tested and the current construction stage.

[0115] In S107 above, after calculating the current risk index and predicting future risk probabilities, a reasonable dynamic early warning threshold needs to be determined as the criterion for triggering an early warning. The setting of the early warning threshold directly affects the effectiveness and reliability of the early warning system. Setting the threshold too high can lead to the failure to promptly warn of real risks, causing safety accidents; setting the threshold too low can frequently trigger false alarms, interfering with normal construction and reducing system credibility. Traditional monitoring systems generally use fixed static thresholds, such as uniformly stipulating that an early warning is triggered when the risk index exceeds 0.5. This simplistic approach ignores the essential impact of different soil types and construction stages on collapse risk. Sandy soil and cohesive soil have significantly different mechanical properties. Sandy soil has a large internal friction angle, low cohesion, and a fast instability rate, while cohesive soil has high cohesion but is sensitive to moisture and deforms slowly. The same risk index value represents completely different levels of actual danger in different soil types. Similarly, the risk of collapse is highest during the excavation stage due to severe stress redistribution; the risk is reduced during the support stage due to the additional constraints provided by the support structure; and the risk is lowest during the backfilling stage as the soil gradually stabilizes. Fixed thresholds cannot adapt to the dynamic changes in risk levels during construction.

[0116] Therefore, this application employs a mechanism for dynamically adjusting the early warning threshold based on soil type and construction stage. This mechanism allows the threshold to adapt to different working conditions, using a lower threshold in high-risk stages to improve early warning sensitivity and a higher threshold in low-risk stages to avoid over-warning, thus achieving an optimal balance between early warning accuracy and construction efficiency. The dynamic early warning threshold is determined based on the soil type and current construction stage, specifically including: obtaining the soil type of the soil to be tested; matching the corresponding foundation early warning threshold from a preset foundation threshold library based on the soil type, wherein the foundation early warning threshold is determined by back-calculation based on the building construction safety factor; obtaining the current construction stage; determining the corresponding construction risk coefficient from a preset risk coefficient library based on the current construction stage, wherein different construction nodes correspond to different construction risk coefficients; multiplying the construction risk coefficient by the foundation proportional coefficient and adding it to 1 to obtain the dynamic adjustment factor; and multiplying the dynamic adjustment factor by the foundation early warning threshold to obtain the dynamic early warning threshold.

[0117] Specifically, information on the soil type of the soil to be tested is obtained. During the project initialization phase, the construction unit determines the soil type within the foundation pit area through geological survey reports or on-site sampling tests. The soil is mainly divided into two categories: sandy soil and cohesive soil. Sandy soil includes granular, dispersed soils such as sand, silt, and gravelly sand. Its characteristics include a relatively large internal friction angle (typically 28 to 35 degrees), extremely low cohesion (close to zero), high permeability, rapid drainage consolidation, slow strength recovery after disturbance, and rapid, sudden collapse. Cohesive soil includes soils with high clay content such as silty clay, clay, and silty mud. Its characteristics include relatively high cohesion (typically 10 to 40 kPa), a relatively small internal friction angle (15 to 25 degrees), poor permeability, slow drainage consolidation, sensitivity to changes in water content (softening upon contact with water significantly reduces strength), and relatively slow, gradual collapse. Construction personnel select the soil type of the current foundation pit in the system configuration interface of the monitoring terminal, store the soil type identifier in the parameter configuration library, and associate it with the specific sensor unit to ensure that each monitoring point uses the soil type parameter corresponding to its soil layer. When performing dynamic threshold calculation, the threshold calculation module of the data processing layer reads the soil type identifier of the soil to be measured from the parameter configuration library, and matches the corresponding basic warning threshold T0 from the preset basic threshold library based on the identifier. The basic threshold library stores the scientifically calibrated basic warning threshold values ​​for different soil types. These values ​​are determined by back-calculation based on the building construction safety factor and statistical analysis of a large number of historical collapse cases. According to the requirements of building construction specifications, the soil stability safety factor should not be less than 1.2, that is, the anti-sliding moment of the soil should be greater than or equal to 1.2 times the sliding moment to meet the safety requirements. Through back-calculation, it is determined at what level the risk index reaches that the soil safety factor drops to the critical value of 1.2, and this risk index value is used as the basic warning threshold. For cohesive soils, numerous engineering cases have verified that when the risk index reaches 0.5, the soil safety factor drops to approximately 1.2, indicating a critical stable state. Therefore, the foundation warning threshold T0 for cohesive soils is set at 0.5. For sandy soils, since their cohesion is almost zero and they mainly rely on the internal friction angle to provide anti-sliding force, the instability speed is faster. When the risk index reaches 0.4, collapse may be triggered. Therefore, the foundation warning threshold T0 for sandy soils is set at 0.4.

[0118] After obtaining the basic early warning threshold, information on the current construction stage is acquired, and the corresponding construction risk coefficient is determined. The foundation pit construction process is typically divided into three main stages: excavation, support, and backfilling. The stress state and collapse risk level of the soil differ significantly at each stage. During the excavation stage, soil unloading leads to the most severe stress redistribution, disrupting the original equilibrium state and causing a sharp increase in lateral earth pressure. Furthermore, the long exposure time of the excavation face makes it highly susceptible to rainfall and groundwater influences, resulting in the highest collapse risk at this stage. In the support stage, after the support structure is completed, the support system provides additional lateral constraint on the soil, stabilizing the soil stress state. The collapse risk is significantly reduced compared to the excavation stage, but continuous monitoring is still necessary. During the backfilling stage, soil is backfilled layer by layer inside the foundation pit. The backfill soil exerts counter-pressure on the sidewalls, further enhancing soil stability. Additionally, the excavation face is covered, eliminating external environmental interference, minimizing the collapse risk at this stage. Construction management personnel update the current construction stage identifier in real time on the monitoring terminal and match the corresponding construction risk coefficient S from the preset risk coefficient database based on this stage identifier. The risk coefficient database stores standard values ​​for risk coefficients in three construction stages. A risk coefficient S of 0.5 for the excavation stage indicates a high risk level requiring increased early warning sensitivity. A risk coefficient S of 0.2 for the support stage indicates a reduced risk but still requires vigilance. A risk coefficient S of 0.1 for the backfilling stage indicates a lower risk level, allowing for a more relaxed threshold. After reading the construction risk coefficient S corresponding to the current construction stage, S is multiplied by the foundation proportionality coefficient 0.1 to obtain the risk adjustment amount. The foundation proportionality coefficient of 0.1 is set based on engineering experience to ensure a moderate threshold adjustment range that reflects the differences between construction stages without causing excessive threshold fluctuations. The dynamic adjustment factor is calculated as 1 plus the construction risk coefficient S multiplied by 0.1. The dynamic adjustment factor is a multiplier coefficient that adjusts the construction risk level proportionally based on the foundation threshold. Finally, the dynamic adjustment factor is multiplied by the foundation early warning threshold T0 to obtain the dynamic early warning threshold T adapted to the current soil type and construction stage.

[0119] Taking the excavation stage of a sandy soil foundation pit as an example, the soil type is identified as sandy soil, and the threshold value T0 is obtained from the foundation threshold library, which equals 0.4. The construction stage is identified as excavation stage, and the risk coefficient value S is obtained from the risk coefficient library, which equals 0.5. The dynamic adjustment factor is calculated as 1 + 0.1 multiplied by 0.5, which equals 1.05. The dynamic early warning threshold T is calculated as 0.4 multiplied by 1.05, which equals 0.42. When the sandy soil area enters the support stage, the construction risk coefficient is updated to S equals 0.2, the dynamic adjustment factor is updated to 1.02, and the dynamic early warning threshold is updated to T equals 0.408. After entering the backfilling stage, the construction risk coefficient is updated to S equals 0.1, the dynamic adjustment factor is updated to 1.01, and the dynamic early warning threshold is updated to T equals 0.404. Through this dynamic adjustment mechanism, the early warning threshold gradually increases as the construction risk decreases. A lower threshold of 0.42 is used in the high-risk stage of excavation to ensure timely early warning, while the threshold is increased to 0.404 in the low-risk stage of backfilling to avoid excessive early warning. This achieves a precise match between the early warning sensitivity and the actual construction needs, significantly improving the practicality and reliability of the early warning system.

[0120] S108: Determine the corresponding warning level by combining the comparison results of the current risk index and the dynamic warning threshold, the risk probability, and the spatial pressure gradient, and trigger the corresponding warning signal based on the warning level.

[0121] In S108 above, after calculating the current risk index, predicting future risk probabilities, and determining the dynamic early warning threshold, the early warning level is determined by comprehensively considering multi-dimensional monitoring parameters, and corresponding early warning response measures are triggered. The core value of the early warning system lies in implementing differentiated emergency response strategies based on the severity of the risk. This ensures that genuine high-risk situations trigger timely emergency evacuations to avoid casualties, while also preventing overreactions to low-level risks that could lead to losses in construction efficiency and a decline in system credibility. Traditional early warning systems generally use a single-parameter threshold determination method, such as deciding whether to issue an early warning based solely on whether the risk index exceeds a fixed threshold. This simplistic logic has two serious flaws. First, when a monitoring parameter experiences short-term abnormal fluctuations due to sensor drift or local construction disturbances, a single-parameter determination will immediately trigger an early warning, resulting in false alarms. Frequent false alarms can lead to early warning fatigue and complacency among construction personnel, causing them to neglect real risks and delay emergency response when they materialize. On the other hand, soil collapse is a complex process involving multiple coupled factors, including the accumulation of internal stress and changes in geometric shape. A single parameter is difficult to fully reflect the true state of risk. There may be situations where a certain parameter does not reach the warning threshold, but multiple parameters are close to the critical value at the same time. In this case, the soil is actually in a high-risk state, but the warning is not triggered, resulting in a missed report.

[0122] Therefore, this application provides an early warning mechanism based on multi-parameter fusion judgment and graded response. By comprehensively analyzing multi-dimensional information such as the current risk index, future risk probability, and spatial pressure gradient, three early warning levels are set, each corresponding to a different degree of risk severity, and differentiated emergency response measures are configured for each early warning level. The early warning levels include Level 1, Level 2, and Level 3. The corresponding early warning level is determined by comprehensively comparing the current risk index with the dynamic early warning threshold, the risk probability, and the spatial pressure gradient. Based on the early warning level, a corresponding early warning signal is triggered. Specifically, if the current risk index is greater than the product of the dynamic early warning threshold and a first proportional coefficient, or the risk probability is greater than a first probability threshold, or the current risk index is greater than the product of the dynamic early warning threshold and a second proportional coefficient and the accumulated strain is greater than or equal to a preset strain threshold, then it is determined to be a Level 1 early warning, where the first proportional coefficient is greater than the second proportional coefficient. If the early warning level is determined to be Level 1, then a first early warning signal is triggered, which includes continuously activating the on-site audible and visual alarm device, pushing an evacuation route map, and a shutdown command. If the triggering conditions for Level 1 are not met, and the current risk index is greater than the product of the dynamic early warning threshold and the second proportional coefficient, then a Level 1 early warning is triggered. If the product of the coefficients, or the risk probability is greater than the second probability threshold, or the spatial pressure gradient is greater than or equal to the preset gradient threshold and the tilt angle is greater than or equal to the preset tilt angle threshold, then a Level II warning is determined, where the first probability threshold is greater than the second probability threshold. If the warning level is determined to be Level II, then a second warning signal is triggered, which includes activating the on-site audible and visual alarm device for interval alarms and pushing a stop construction command. If the triggering conditions for Level I and Level II warnings are not met, and the current risk index is greater than the product of the dynamic warning threshold and the third proportional coefficient, and the risk probability is less than or equal to the third probability threshold, then a Level III warning is determined, where the second proportional coefficient is greater than the third proportional coefficient and the second probability threshold is greater than the third probability threshold. If the warning level is determined to be Level III, then a third warning signal is triggered, which includes pushing the risk location and generating a risk trend report.

[0123] Specifically, the system acquires the current Risk Index, dynamic warning threshold T, risk probability Prob for the next 30 minutes, and spatial dimension characteristic parameters, including the pressure difference ΔPadj between adjacent sensing units and the pipe tilt angle θ. Simultaneously, it reads the cumulative strain εacc as an auxiliary judgment parameter. The warning judgment module sequentially determines whether the triggering conditions for Level 1, Level 2, and Level 3 warnings are met in descending order of priority. Once the conditions for a certain level are met, the warning level is determined, and the subsequent judgment process terminates. This priority judgment mechanism ensures that the highest risk level is identified and responded to first.

[0124] First, determine if the triggering conditions for a Level 1 warning (red alert) are met. A Level 1 warning indicates that the soil has entered an extremely dangerous state and a large-scale collapse could occur at any time, requiring immediate evacuation of personnel and cessation of all construction work. The triggering conditions for a Level 1 warning are set with three parallel decision branches; meeting any one of these branches triggers a Level 1 warning. The first decision branch checks if the current risk index is greater than the product of the dynamic warning threshold and a first proportional coefficient. The first proportional coefficient is set to 1.5, meaning that a Level 1 warning is triggered when the current risk index exceeds the dynamic threshold by 50%. This condition reflects that the current stress state of the soil is severely excessive, indicating that the internal stress concentration of the soil far exceeds the safe tolerance range and is on the verge of instability. The second decision branch checks if the risk probability in the next 30 minutes is greater than a first probability threshold, set at 0.9 (90%). This condition reflects that based on historical data evolution patterns, a collapse is highly likely in the future, even if the current risk index has not yet reached its highest level, but the development trend is already extremely dangerous. The third judgment branch adopts a combination condition judgment, which requires that the current risk index is greater than the product of the dynamic warning threshold and the second proportional coefficient, and the cumulative strain is greater than or equal to the preset strain threshold. The second proportional coefficient is set to 1.3, that is, the current risk index exceeds the dynamic threshold by 30%, and the preset strain threshold is set to 800 microstrain. This condition reflects that the soil has both a high instantaneous stress level and significant cumulative deformation. The double exceedance indicates that the soil has been subjected to continuous stress for a long time, resulting in serious deterioration of the internal structure. Although a single parameter may not reach the extreme value, the combination of the two constitutes an extremely high risk.

[0125] Once the early warning judgment module confirms that the conditions for a Level 1 early warning are met, it immediately sends a Level 1 early warning command to the application layer and marks the warning level as red. Upon receiving the Level 1 early warning command, the application layer triggers the first early warning signal to execute the emergency response process. The first early warning signal includes three linked actions: First, the on-site audible and visual alarm device is continuously activated, entering a high-frequency continuous alarm mode without intervals. The piercing buzzer and flashing red warning light raise the alert level of all personnel on site, ensuring timely detection even in noisy construction environments. Second, an evacuation route map is immediately pushed to all relevant personnel via a mobile app. Based on the BIM 3D map of the construction area, the evacuation route map automatically generates the optimal evacuation path, marking the locations of risk sensors and the extent of danger zones, instructing on-site personnel to quickly evacuate to a safe area along the shortest path. Simultaneously, it automatically dials emergency contact numbers according to preset priorities: first the project manager, then the construction team leader, and finally the emergency management department, ensuring that management can grasp the situation and activate the emergency plan immediately. Finally, the edge gateway sends a shutdown command to the construction equipment control system, automatically cutting off the power supply to the risk area and its surrounding areas, forcibly stopping the operation of large construction equipment such as excavators and cranes, and preventing equipment vibration or increased load from further accelerating soil instability.

[0126] If none of the triggering conditions for a Level 1 warning are met, the system continues to determine whether the triggering conditions for a Level 2 warning (yellow warning) are met. A Level 2 warning indicates a significantly increased soil risk level, but not yet reaching an extremely dangerous state, requiring the suspension of construction work in the risk area and the implementation of reinforcement measures. The triggering conditions for a Level 2 warning also have three parallel decision branches. The first decision branch checks whether the current risk index is greater than the product of the dynamic warning threshold and the second proportional coefficient 1.3, indicating that the current stress level exceeds the safety threshold by 30% and has entered the warning zone. The second decision branch checks whether the risk probability in the next 30 minutes is greater than the second probability threshold, which is set at 0.7 (70%), indicating a high probability of future collapse requiring early intervention. The third decision branch uses a combination of conditions requiring the spatial pressure gradient ΔPadj to be greater than or equal to the preset gradient threshold of 15 kPa and the tilt angle θ to be greater than or equal to the preset tilt angle threshold of 0.8 degrees. This condition reflects a significant uneven stress distribution between adjacent monitoring points and a noticeable tilt of the sensor tube, indicating that the soil in a local area may be undergoing slippage deformation; these dual spatial anomalies predict a local collapse risk.

[0127] Once the conditions for a Level 2 warning are confirmed, the application layer triggers a second warning signal to execute a control response process. The second warning signal includes two main actions: First, the on-site audible and visual alarm device is activated for intermittent alarms, sounding every 30 seconds to alert on-site personnel to the risk without causing panic. Simultaneously, the edge gateway sends a command to all sensor units to increase the data acquisition frequency from 1 Hz to 2 Hz, increasing the density of monitoring data acquisition for more precise tracking of the risk evolution process. A stop-work order is pushed to the construction manager via a mobile app, along with slope reinforcement suggestions, such as sandbag reinforcement or shotcrete slope protection within a 5-meter radius of the risk sensor. The name of the person in charge and the start time are recorded to establish a responsibility traceability mechanism. After reinforcement measures are implemented, the risk index is recalculated every 5 minutes. If the risk index is less than or equal to the dynamic threshold for three consecutive times (15 minutes), the risk is considered effectively controlled, and the warning level is downgraded to Level 3. If the risk index remains more than 1.3 times the dynamic threshold, the reinforcement measures are deemed insufficient and the risk is still worsening, and the warning is immediately upgraded to Level 1, initiating an emergency evacuation process.

[0128] If neither the triggering conditions for a Level 1 nor Level 2 warning are met, the final determination is whether the triggering conditions for a Level 3 warning (blue warning) are met. A Level 3 warning indicates that the soil risk level has slightly increased but is still within a controllable range, requiring strengthened monitoring to observe the risk development trend. The triggering conditions for a Level 3 warning require that the current risk index is greater than the product of the dynamic warning threshold and the third proportional coefficient, and the risk probability in the next 30 minutes is less than or equal to the third probability threshold. The third proportional coefficient is set to 1.1, meaning the current risk index exceeds the dynamic threshold by 10%, and the third probability threshold is set to 0.5, or 50%. This condition reflects a slight exceedance of the current stress level but a low probability of future collapse, constituting an early risk warning signal. Once the Level 3 warning conditions are confirmed to be met, the application layer triggers the third warning signal and executes the prompting response procedure. The third early warning signal includes pushing the risk location and current risk index value to the construction manager, and marking abnormal characteristics such as the pressure in sandy soil areas exceeding the threshold of 12%. At the same time, the system maintains a continuous monitoring frequency of 1 Hz and automatically generates a risk trend report every 10 minutes, including the risk index change curve, and pushes it to the monitoring terminal. The construction manager can determine whether preventive reinforcement measures need to be taken based on the trend report. If there is no manual intervention, the system will continue to maintain the three-level early warning status to closely track the risk evolution.

[0129] Taking the excavation stage of a sandy soil foundation pit as an example, the dynamic early warning threshold T for this foundation pit has been calculated to be 0.42 based on the soil type and construction stage. At a certain moment, the system monitors the current risk index of the sensor unit as 0.58, the risk probability (Prob) for the next 30 minutes as 0.78, the cumulative strain (εacc) as 700 microstrains, the pressure difference (ΔPadj) between adjacent sensor units as 12 kPa, and the tilt angle (θ) as 0.6 degrees. First, the first-level early warning conditions are checked. The calculated dynamic threshold of 1.5 times is 0.63. The current risk index of 0.58 does not exceed 0.63, which does not meet the first branch. The risk probability of 0.78 does not exceed 0.9, which does not meet the second branch. The calculated dynamic threshold of 1.3 times is 0.546. The current risk index of 0.58 exceeds this value, but the cumulative strain of 700 microstrains does not reach 800 microstrains, which does not meet the third branch. Therefore, the first-level early warning is not triggered. The secondary warning conditions were further checked. The current risk index of 0.58 exceeded 1.3 times the dynamic threshold of 0.546, meeting the first branch. The warning level was determined to be secondary, and the second warning signal was triggered. The on-site audible and visual alarms started sounding every 30 seconds. The mobile app pushed a suggestion to the construction supervisor to suspend construction and reinforce the slope. The data collection frequency was increased to 2 Hz. The construction team then reinforced the area with sandbags. Twenty minutes after reinforcement, the risk index was monitored and found to have dropped to 0.40, which was below the dynamic threshold of 0.42. Three consecutive calculations confirmed that the risk index was below the threshold, so the warning level was downgraded to secondary. The audible and visual alarms stopped, but risk trend reports continued to be pushed for close monitoring. Normal construction could only resume after the risk was effectively controlled. The entire warning response process achieved accurate risk identification and efficient handling.

[0130] like Figure 5As shown, the process begins with raw data collected from sensors, acquiring multi-dimensional raw data on soil stress, deformation, and tilt angle from real-time monitoring by soil pressure sensors, strain gauges, and tilt sensors. In the data preprocessing stage, the collected raw data undergoes three processes: outlier removal, temperature compensation, and moving average filtering. Anomalies caused by sensor malfunctions or construction disturbances are identified and removed using the 3σ criterion combined with a sliding window verification. A linear compensation formula is used to eliminate the drift effect of temperature changes on strain measurements. Finally, a 5-point moving average window is used to suppress high-frequency random noise, ensuring the continuity and accuracy of the output data. After preprocessing, the process moves to feature extraction and risk assessment. The filtered strain values ​​are converted into soil pressure, and the pressure change rate and strain accumulation within a 30-minute time window are calculated as time-dimensional features. Simultaneously, the pressure gradient of adjacent sensor units and the pipe tilt angle are extracted as spatial-dimensional features. Based on the extracted multi-dimensional feature parameters, the current risk index is calculated using a weighted formula and input into an LSTM neural network model to predict the probability of collapse risk in the next 30 minutes. The warning threshold is dynamically adjusted according to the construction stage. The process then moves to the core early warning determination stage. This stage comprehensively compares multiple parameters, including the current risk index, future risk probability, and spatial pressure gradient, and determines whether the triggering conditions for a Level 1, Level 2, or Level 3 early warning are met, according to priority from highest to lowest. If the Level 1 warning conditions are met (the highest risk level), a red alert is immediately triggered, and an emergency evacuation response is implemented. If the Level 2 warning conditions are met, a yellow alert is triggered, and a control response involving suspension of construction and slope reinforcement is implemented. If only the Level 3 warning conditions are met, a blue alert is triggered, and an alert requiring enhanced monitoring is implemented. After each level of warning is triggered, continuous monitoring begins. Based on changes in the risk index from subsequent monitoring data, it is determined whether the risk is under control, thus deciding whether to downgrade the warning; or if the risk continues to worsen, the warning is upgraded, forming a dynamic closed-loop risk management process to ensure the safety of the construction site.

[0131] This application also provides a landslide prediction system based on sensor data. Figure 6 This is a schematic diagram of a landslide prediction system based on sensor data provided in an embodiment of this application. (Refer to...) Figure 6 The system is located in a ground-based monitoring platform and also includes a data acquisition unit 601, a processing unit 602, an analysis unit 603, an evaluation unit 604, and an early warning unit 605. The acquisition unit 601 receives raw data collected by the sensing device from the soil to be measured. The raw data includes earth pressure data, strain data and tilt angle data. The sensing device is deployed in the soil to be measured and penetrates the potential sliding surface. Processing unit 602 preprocesses the raw data to obtain preprocessed earth pressure data, preprocessed strain data, and preprocessed tilt angle data. Analysis unit 603 calculates the rate of change of pressure and the spatial pressure gradient between adjacent sensing devices based on the pre-processed earth pressure data, calculates the cumulative strain based on the pre-processed strain data, and calculates the tilt angle based on the pre-processed tilt angle data. Evaluation unit 604 determines the target weight coefficient based on the current construction stage and the real-time data characteristics of pressure change rate, tilt angle and strain accumulation; it calculates the current risk index by weighting the pressure change rate, tilt angle and strain accumulation based on the target weight coefficient; it inputs the time series of pressure change rate, tilt angle and strain accumulation within the preset historical time window into the preset long short-term memory network prediction model for processing, and outputs the risk probability of soil collapse in the future period. The early warning unit 605 determines the dynamic early warning threshold based on the soil type and the current construction stage of the soil to be tested; it determines the corresponding early warning level by comprehensively considering the comparison results between the current risk index and the dynamic early warning threshold, the risk probability, and the spatial pressure gradient, and triggers the corresponding early warning signal based on the early warning level.

[0132] In one possible implementation, the acquisition unit 601 is used to acquire preprocessed first earth pressure data and preprocessed second earth pressure data at the current sampling time and the previous adjacent sampling time, and calculate the pressure difference between the preprocessed first earth pressure data and the preprocessed second earth pressure data; the analysis unit 603 is used to divide the pressure difference by the time interval between the current sampling time and the previous adjacent sampling time to obtain the pressure change rate; acquire the preprocessed earth pressure data of two sensing devices deployed at adjacent locations in the soil to be measured at the same time, calculate the absolute value of the pressure difference between the two sensing devices; and convert the absolute value of the pressure difference into the absolute value of the pressure difference. The spatial pressure gradient is obtained by dividing the value by the spatial distance between the two sensing devices; the preprocessed strain data at the initial moment is obtained as the reference strain value, and the preprocessed strain data at the current moment is subtracted from the reference strain value to obtain the strain accumulation; the X-axis and Y-axis tilt components are extracted from the preprocessed tilt angle data, converted to radians, and the product of the cosine of the X-axis and Y-axis tilt components is calculated. The product is then subjected to inverse cosine operation to obtain the current tilt angle of the soil to be measured, where the tilt angle represents the axis of the sensing device relative to the direction of gravity.

[0133] In one possible implementation, the evaluation unit 604 is used to, based on the current construction stage, retrieve from a preset weight matching library the foundation pressure rate weight corresponding to the pressure change rate, the foundation inclination angle weight corresponding to the inclination angle, and the foundation strain weight corresponding to the strain accumulation; use the preprocessed earth pressure data at the current moment as the current absolute earth pressure level, calculate the pressure ratio between the current absolute earth pressure level and the design earth pressure benchmark value, square the pressure ratio, multiply it by a first sensitivity coefficient, and add it to 1 to obtain the pressure level adjustment factor; multiply the foundation pressure rate weight by the pressure level adjustment factor to obtain the target weight of the pressure change rate; and calculate based on the inclination angle at continuous sampling moments. The tilt angle change rate is calculated by dividing the absolute value of the tilt angle change rate by the tilt angle ratio of the angular velocity warning threshold. This tilt angle ratio is multiplied by a second sensitivity coefficient and then added to 1 to obtain the tilt angular velocity adjustment factor. The basic tilt angle weight is multiplied by the tilt angular velocity adjustment factor to obtain the target weight of the tilt angle. The strain accumulation is calculated by dividing the strain ratio by the strain limit threshold. This strain ratio is squared, multiplied by a third sensitivity coefficient, and then added to 1 to obtain the strain level adjustment factor. The basic strain weight is multiplied by the strain level adjustment factor to obtain the target weight of the strain accumulation. The target weights of the pressure change rate, tilt angle, and strain accumulation are output as target weight coefficients.

[0134] In one possible implementation, the acquisition unit 601 is used to acquire the soil type of the soil to be tested, and match the corresponding foundation early warning threshold from a preset foundation threshold library according to the soil type, wherein the foundation early warning threshold is determined by back-calculation based on the building construction safety factor; acquire the current construction stage, and determine the corresponding construction risk coefficient from a preset risk coefficient library according to the current construction stage, wherein different construction nodes correspond to different construction risk coefficients; the early warning unit 605 is used to multiply the construction risk coefficient by the foundation ratio coefficient and add it to 1 to obtain a dynamic adjustment factor; and multiply the dynamic adjustment factor by the foundation early warning threshold to obtain a dynamic early warning threshold.

[0135] In one possible implementation, the early warning unit 605 is configured to determine a Level 1 early warning if the current risk index is greater than the product of the dynamic early warning threshold and a first proportional coefficient, or the risk probability is greater than a first probability threshold, or the current risk index is greater than the product of the dynamic early warning threshold and a second proportional coefficient and the accumulated strain is greater than or equal to a preset strain threshold, wherein the first proportional coefficient is greater than the second proportional coefficient; if the early warning level is determined to be Level 1, a first early warning signal is triggered, the first early warning signal including continuously activating the on-site audible and visual alarm device, pushing an evacuation route map, and a shutdown command; if the triggering conditions for Level 1 early warning are not met, and the current risk index is greater than the product of the dynamic early warning threshold and the second proportional coefficient, or the risk probability is greater than the second probability threshold, or the spatial pressure gradient is greater than or equal to... If a level 2 warning is determined when the preset gradient threshold and the tilt angle is greater than or equal to the preset tilt angle threshold, the first probability threshold is greater than the second probability threshold. If the warning level is determined to be level 2, a second warning signal is triggered, which includes activating the on-site audible and visual alarm device for interval alarm and pushing a stop construction command. If the triggering conditions for level 1 and level 2 warnings are not met, and the current risk index is greater than the product of the dynamic warning threshold and the third proportional coefficient, and the risk probability is less than or equal to the third probability threshold, a level 3 warning is determined when the second proportional coefficient is greater than the third proportional coefficient and the second probability threshold is greater than the third probability threshold. If the warning level is determined to be level 3, a third warning signal is triggered, which includes pushing the risk location and generating a risk trend report.

[0136] In one possible implementation, the processing unit 602 is used to perform quality checks and anomaly identification on the raw data to obtain preliminary data; acquire the current ambient temperature and calculate the temperature difference between the current ambient temperature and the initial calibration temperature of the sensing device; calculate the temperature compensation amount according to the preset temperature compensation coefficient of the sensing device and the temperature difference; algebraically superimpose the preliminary data and the temperature compensation amount to obtain the temperature-compensated data; divide the temperature-compensated data into multiple sliding data windows according to the time series, calculate the arithmetic mean of the data in each sliding data window as the effective value at the corresponding time, and use the effective value as the data after noise reduction to obtain the pre-processed earth pressure data, pre-processed strain data, and pre-processed tilt angle data.

[0137] In one possible implementation, the sensing device includes an earth pressure sensor, an upper strain gauge, a lower strain gauge, a biaxial tilt sensor, a data processing and communication module, an audible and visual alarm, and a rechargeable battery pack. The upper and lower strain gauges are respectively arranged in the upper and lower monitoring sections of the sensing device to measure the tensile and compressive deformation of the outer shell. The earth pressure sensor is embedded in a pre-reserved opening in the side wall of the outer shell to measure lateral earth pressure. The biaxial tilt sensor measures the tilt angle of the entire sensing device in two vertical directions. The rechargeable battery pack powers the sensing device. The data processing and communication module connects to and controls all sensors to perform data acquisition and transmission. The audible and visual alarm receives a warning signal and then responds with sound and flashing alarm light.

[0138] It should be noted that the system provided in the above embodiments is only illustrated by the division of the above functional modules. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the system and method embodiments provided in the above embodiments belong to the same concept, and the specific implementation process can be found in the method embodiments, which will not be repeated here.

[0139] This application also discloses an electronic device. (See reference...) Figure 7 , Figure 7 This application provides a schematic diagram of the structure of an electronic device. The electronic device 700 may include: at least one processor 701, at least one network interface 704, a user interface 703, a memory 702, and at least one communication bus 705.

[0140] The communication bus 705 is used to enable communication between these components.

[0141] The user interface 703 may include a display screen and a camera. Optionally, the user interface 703 may also include a standard wired interface and a wireless interface.

[0142] The network interface 704 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0143] The processor 701 may include one or more processing cores. The processor 701 connects to various parts of the server using various interfaces and lines, and performs various server functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in memory 702, and by calling data stored in memory 702. Optionally, the processor 701 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). The processor 701 may integrate one or a combination of several of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and application requests; the GPU is responsible for rendering and drawing the content required for display; and the modem handles wireless communication. It is understood that the modem may also be implemented as a separate chip without being integrated into the processor 701.

[0144] The memory 702 may include random access memory (RAM) or read-only memory. Optionally, the memory 702 may include a non-transitory computer-readable storage medium. The memory 702 can be used to store instructions, programs, code, code sets, or instruction sets. The memory 702 may include a program storage area and a data storage area. The program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch functionality, sound playback functionality, image playback functionality, etc.), instructions for implementing the various method embodiments described above, etc.; the data storage area may store data involved in the various method embodiments described above, etc. Optionally, the memory 702 may also be at least one storage device located remotely from the aforementioned processor 701.

[0145] like Figure 7 As shown, the memory 702, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and an application for landslide prediction based on sensor data.

[0146] exist Figure 7In the electronic device 700 shown, the user interface 703 is mainly used to provide an interface for users to input data and obtain user input data; while the processor 701 can be used to call the application program stored in the memory 702 based on the collapse prediction of sensor data. When executed by one or more processors, the electronic device performs one or more of the methods described in the above embodiments.

[0147] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0148] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0149] In the several embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some service interfaces; indirect couplings or communication connections between devices or units may be electrical or other forms.

[0150] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0151] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0152] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as USB flash drives, portable hard drives, magnetic disks, or optical disks.

[0153] The foregoing description is merely an exemplary embodiment of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure. Those skilled in the art will readily conceive of other embodiments of this disclosure upon considering the specification and practical application disclosed herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described in this disclosure.

Claims

1. A landslide prediction method based on sensor data, characterized in that, When applied to a ground-based monitoring platform, the method includes: The system receives raw data collected by a sensing device on the soil under test. The raw data includes earth pressure data, strain data, and tilt angle data. The sensing device is deployed in the soil under test and penetrates the potential sliding surface. The original data is preprocessed to obtain preprocessed earth pressure data, preprocessed strain data, and preprocessed tilt angle data. The pressure change rate and the spatial pressure gradient between adjacent sensing devices are calculated based on the preprocessed earth pressure data. The cumulative strain is calculated based on the preprocessed strain data. The tilt angle is calculated based on the preprocessed tilt angle data. The target weighting coefficient is determined based on the current construction stage and the real-time data characteristics of the pressure change rate, the tilt angle, and the strain accumulation. The current risk index is obtained by weighting the pressure change rate, the tilt angle, and the cumulative strain based on the target weighting coefficient. The time series of the pressure change rate, the tilt angle and the cumulative strain within the preset historical time window are input into the preset long short-term memory network prediction model for processing, and the probability of soil collapse in the future period is output. A dynamic early warning threshold is determined based on the soil type of the soil to be tested and the current construction stage. The corresponding warning level is determined by combining the comparison results of the current risk index and the dynamic warning threshold, the risk probability, and the spatial pressure gradient, and the corresponding warning signal is triggered based on the warning level.

2. The method according to claim 1, characterized in that, The calculation of the pressure change rate and spatial pressure gradient between adjacent sensing devices based on the preprocessed earth pressure data, the calculation of the cumulative strain based on the preprocessed strain data, and the calculation of the tilt angle based on the preprocessed tilt angle data specifically includes: Obtain the preprocessed first earth pressure data and the preprocessed second earth pressure data at the current sampling time and the previous adjacent sampling time, and calculate the pressure difference between the preprocessed first earth pressure data and the preprocessed second earth pressure data. The pressure difference is divided by the time interval between the current sampling time and the previous adjacent sampling time to obtain the pressure change rate. The preprocessed earth pressure data of two sensing devices deployed at adjacent locations in the soil to be tested are obtained at the same time, and the absolute value of the pressure difference between the two sensing devices is calculated. The spatial pressure gradient is obtained by dividing the absolute value of the pressure difference by the spatial arrangement distance between the two sensing devices. The preprocessed strain data at the initial moment is obtained as the reference strain value. The preprocessed strain data at the current moment is subtracted from the reference strain value to obtain the cumulative strain. The X-axis and Y-axis tilt components are extracted from the preprocessed tilt angle data. The X-axis and Y-axis tilt components are converted to radians. The product of the cosine of the X-axis tilt component and the cosine of the Y-axis tilt component is calculated, and the product is subjected to inverse cosine operation to obtain the current tilt angle of the soil to be measured. The tilt angle represents the axis of the sensing device relative to the direction of gravity.

3. The method according to claim 2, characterized in that, The determination of the target weighting coefficient based on the current construction stage and real-time data characteristics of the pressure change rate, the tilt angle, and the accumulated strain specifically includes: Based on the current construction stage, the foundation pressure rate weight corresponding to the pressure change rate, the foundation inclination angle weight corresponding to the inclination angle, and the foundation strain weight corresponding to the strain accumulation are retrieved from the preset weight matching library. The preprocessed earth pressure data at the current moment is taken as the current absolute earth pressure level. The pressure ratio between the current absolute earth pressure level and the design earth pressure benchmark value is calculated. The pressure ratio is squared, multiplied by the first sensitivity coefficient, and added to 1 to obtain the pressure level adjustment factor. Multiply the base pressure rate weight by the pressure level adjustment factor to obtain the target weight of the pressure change rate; The tilt angle change rate is calculated based on the tilt angle at continuous sampling time. The absolute value of the tilt angle change rate is calculated as the tilt angle ratio to the angular velocity warning threshold. The tilt angle ratio is multiplied by the second sensitivity coefficient and then added to 1 to obtain the tilt angular velocity adjustment factor. Multiply the base tilt angle weight by the tilt angular velocity adjustment factor to obtain the target tilt angle weight; Calculate the strain ratio between the accumulated strain and the strain limit threshold, square the strain ratio, multiply it by the third sensitivity coefficient, and add it to 1 to obtain the strain level adjustment factor. The target weight of the cumulative strain is obtained by multiplying the basic strain weight by the strain level adjustment factor, and the target weight of the pressure change rate, the target weight of the tilt angle, and the target weight of the cumulative strain are output as the target weight coefficient.

4. The method according to claim 1, characterized in that, The determination of the dynamic early warning threshold based on the soil type of the soil to be tested and the current construction stage specifically includes: Obtain the soil type of the soil to be tested, and match the corresponding foundation warning threshold from the preset foundation threshold library according to the soil type. The foundation warning threshold is determined by back-calculation based on the building construction safety factor. The current construction stage is obtained, and the corresponding construction risk coefficient is determined from the preset risk coefficient library based on the current construction stage, wherein different construction nodes correspond to different construction risk coefficients; Multiply the construction risk coefficient by the foundation ratio coefficient and add it to 1 to obtain the dynamic adjustment factor; The dynamic adjustment factor is multiplied by the basic early warning threshold to obtain the dynamic early warning threshold.

5. The method according to claim 4, characterized in that, The warning levels include Level 1, Level 2, and Level 3. The corresponding warning level is determined by comprehensively considering the comparison between the current risk index and the dynamic warning threshold, the risk probability, and the spatial pressure gradient. Based on the warning level, a corresponding warning signal is triggered, specifically including: If the current risk index is greater than the product of the dynamic warning threshold and the first proportional coefficient, or the risk probability is greater than the first probability threshold, or the current risk index is greater than the product of the dynamic warning threshold and the second proportional coefficient and the accumulated strain is greater than or equal to the preset strain threshold, then it is determined to be the first-level warning, wherein the first proportional coefficient is greater than the second proportional coefficient. If the warning level is determined to be the Level 1 warning, then the first warning signal is triggered, which includes continuously activating the on-site audible and visual alarm device, pushing the evacuation route map, and the shutdown command. If the triggering conditions for the first-level warning are not met, and the current risk index is greater than the product of the dynamic warning threshold and the second proportional coefficient, or the risk probability is greater than the second probability threshold, or the spatial pressure gradient is greater than or equal to the preset gradient threshold and the tilt angle is greater than or equal to the preset tilt angle threshold, then it is determined to be the second-level warning, wherein the first probability threshold is greater than the second probability threshold. If the warning level is determined to be the Level II warning, a second warning signal is triggered. The second warning signal includes activating the on-site audible and visual alarm device for interval alarm and sending a suspension of construction command. If the triggering conditions for the first-level warning and the second-level warning are not met, and the current risk index is greater than the product of the dynamic warning threshold and the third proportional coefficient, and the risk probability is less than or equal to the third probability threshold, then it is determined to be the third-level warning, where the second proportional coefficient is greater than the third proportional coefficient, and the second probability threshold is greater than the third probability threshold. If the warning level is determined to be Level 3, a third warning signal is triggered, which includes pushing the risk location and generating a risk trend report.

6. The method according to claim 1, characterized in that, The preprocessing of the original data to obtain preprocessed earth pressure data, preprocessed strain data, and preprocessed tilt angle data specifically includes: The raw data is subjected to quality checks and anomaly identification processing to obtain preliminary data; Obtain the current ambient temperature and calculate the temperature difference between the current ambient temperature and the initial calibration temperature of the sensing device; The temperature compensation amount is calculated based on the preset temperature compensation coefficient of the sensing device and the temperature difference. The preliminary data and the temperature compensation amount are then algebraically superimposed to obtain the temperature-compensated data. The temperature-compensated data is divided into multiple sliding data windows according to the time series. The arithmetic mean of the data in each sliding data window is calculated as the effective value at the corresponding time. The effective value is used as the data after noise reduction to obtain the pre-processed earth pressure data, the pre-processed strain data, and the pre-processed tilt angle data.

7. The method according to claim 1, characterized in that, The sensing device includes an earth pressure sensor, an upper strain gauge, a lower strain gauge, a biaxial tilt sensor, a data processing and communication module, an audible and visual alarm, and a rechargeable battery pack. The upper and lower strain gauges are respectively arranged in the upper and lower monitoring sections of the sensing device to measure the tensile and compressive deformation of the outer shell. The earth pressure sensor is embedded in a pre-drilled opening in the side wall of the outer shell to measure lateral earth pressure. The biaxial tilt sensor measures the tilt angle of the entire sensing device in two vertical directions. The rechargeable battery pack powers the sensing device. The data processing and communication module connects to and controls all sensors to perform data acquisition and transmission. The audible and visual alarm receives a warning signal and then responds with sound and flashing alarm light.

8. A landslide prediction system based on sensor data, characterized in that, The system is located within a ground-based monitoring platform, and it further includes a data acquisition unit, a processing unit, an analysis unit, an evaluation unit, and an early warning unit. The acquisition unit receives raw data collected by the sensing device from the soil to be tested. The raw data includes soil pressure data, strain data, and tilt angle data. The sensing device is deployed in the soil to be tested and penetrates the potential sliding surface. The processing unit preprocesses the raw data to obtain preprocessed earth pressure data, preprocessed strain data, and preprocessed tilt angle data. The analysis unit calculates the rate of change of pressure and the spatial pressure gradient between adjacent sensing devices based on the preprocessed earth pressure data, calculates the cumulative strain based on the preprocessed strain data, and calculates the tilt angle based on the preprocessed tilt angle data. The evaluation unit determines the target weighting coefficient based on the current construction stage and the real-time data characteristics of the pressure change rate, the tilt angle, and the strain accumulation; and performs a weighted calculation on the pressure change rate, the tilt angle, and the strain accumulation based on the target weighting coefficient to obtain the current risk index. The time series of the pressure change rate, the tilt angle and the cumulative strain within the preset historical time window are input into the preset long short-term memory network prediction model for processing, and the probability of soil collapse in the future period is output. The early warning unit determines a dynamic early warning threshold based on the soil type of the soil to be tested and the current construction stage. The corresponding warning level is determined by combining the comparison results of the current risk index and the dynamic warning threshold, the risk probability, and the spatial pressure gradient, and the corresponding warning signal is triggered based on the warning level.

9. An electronic device, characterized in that, The device includes a processor, a memory, a user interface, and a network interface. The memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory to cause the electronic device to perform the method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed, perform the method as described in any one of claims 1-7.

Images