Foundation pit safety monitoring method and device based on cooperation of unmanned aerial vehicle oblique photography and Internet of Things

Abstract

The invention discloses a foundation pit safety monitoring method and device based on cooperation of unmanned aerial vehicle oblique photography and Internet of Things, and relates to the technical field of data processing, and the method comprises the following steps: S1, collecting a three-dimensional image and spectral data of a foundation pit; s2, collecting data of the foundation pit through a monitoring sensor deployed in the foundation pit; s3, multi-source heterogeneous data weighted fusion is carried out; s4, calculating a three-dimensional deformation quantity; s5, performing crack detection on the spectral image acquired by the unmanned aerial vehicle; and S6, fusing the three-dimensional deformation quantity and an abnormal identification result to generate a risk signal, generating a digital twinborn body of the health state of the foundation pit, and realizing minute-level risk early warning. According to the invention, through cooperation of unmanned aerial vehicle oblique photography and a multispectral sensor, in combination with a foundation pit sensor network, the three-dimensional geometrical morphology, multi-physical field parameters and spectral feature data of the foundation pit are synchronously obtained. Compared with traditional total station single-point measurement, the monitoring precision is improved, and the coverage range is expanded to the whole foundation pit area from the local point location of the fixed sensor.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a foundation pit safety monitoring method and device using unmanned aerial vehicle (UAV) oblique photography in collaboration with the Internet of Things. Background Art

[0002] Foundation pit engineering is a high-risk link in civil engineering construction. The accuracy and real-time performance of its safety monitoring technology are directly related to engineering safety and construction efficiency. Traditional foundation pit monitoring technology mainly relies on the following methods:

[0003] Manual inspection and fixed-point measurement: Engineers conduct single-point measurements through visual inspection or using equipment such as total stations and levels. This method is limited by manual operation efficiency and can only cover a limited number of monitoring points. It also carries high operational risks in dangerous scenarios such as deep foundation pits and cannot achieve dynamic monitoring of the entire area.

[0004] Fixed sensor embedded monitoring: Data is collected by embedding stress gauges, displacement gauges and other sensors in the foundation pit support structure. However, this solution has the problems of high deployment cost and significant blind spots in the sensor network coverage. It is particularly difficult to obtain the overall deformation trend of the foundation pit, and data transmission relies on wired networks, which seriously lacks real-time performance.

[0005] Single-use drone 2D image monitoring: Utilizing only two-dimensional drone images, it lacks 3D modeling capabilities and multi-sensor data fusion mechanisms. Monitoring data requires manual interpretation and analysis, and the cycle from data collection to risk identification can take several hours, making it unable to meet the emergency response needs of deep foundation pit construction. Furthermore, existing drone solutions are unstable in environments with strong electromagnetic interference (such as those involving cranes on construction sites) or inclement weather (strong winds and heavy rain), often resulting in data loss or monitoring interruptions.

[0006] The core flaws of existing technologies are: a single data dimension, unable to simultaneously capture information on the three-dimensional deformation of foundation pits coupled with multiple physical fields (temperature, humidity, and stress); insufficient real-time early warning capabilities, high data transmission latency from fixed sensors, and time-consuming post-processing of drone imagery; low intelligence levels, relying on manual experience to identify anomalies such as cracks and leaks, which can easily miss potential risks; and poor adaptability to complex scenarios, making it difficult to operate stably in dynamic construction environments. Therefore, a monitoring technology that integrates multi-source data acquisition, real-time intelligent analysis, and an air-ground collaborative network is urgently needed to achieve comprehensive, automated early warning of foundation pit safety. Summary of the Invention

[0007] The purpose of the present invention is to provide a foundation pit safety monitoring method and device that cooperates with drone oblique photography and the Internet of Things to solve the problems raised in the above background technology.

[0008] To achieve the above-mentioned purpose, the present invention provides the following technical solution: a foundation pit safety monitoring method using drone oblique photography and the Internet of Things, comprising the following steps:

[0009] S1, collects three-dimensional images and spectral data of foundation pits through the oblique photography camera and multispectral sensor carried by the UAV;

[0010] S2, collects stress, axial force, water level, displacement, inclinometer, static leveling, inclination and acceleration data of the foundation pit through monitoring sensors deployed in the foundation pit;

[0011] S3, based on the edge computing gateway, performs multi-source heterogeneous data weighted fusion on the 3D image data collected by the drone in steps S1 and S2 and the multi-physical field data collected by the foundation pit monitoring sensors, assigns weights based on sensor accuracy, and adjusts the weight coefficients in real time through a dynamic evaluation system;

[0012] S4, based on the fused data, a multivariate regression equation of foundation pit deformation and multiple environmental parameters is established to perform deformation analysis and calculate the three-dimensional deformation variable;

[0013] S5, performs crack detection on the spectral images collected by the drone, automatically identifying cracks and leakage anomalies, including inputting spectral images, enhancing crack feature extraction, synchronously detecting crack areas, and outputting detection results;

[0014] S6, integrates three-dimensional deformation variables and anomaly recognition results to generate risk signals, and transmits the analysis results to the cloud-based early warning platform in real time through the 5G network, generating a digital twin of the foundation pit health status and achieving minute-level risk warnings.

[0015] Preferably, in the drone data collection in step S1, a "spiral + grid" composite route planning is adopted, and a low-altitude circumferential scan with a height of less than 20m is performed near the foundation pit excavation surface. The route planning covers the foundation pit slope, support structure and surrounding surface area. The oblique photography camera collects five-view images at an inclination angle of 45°-60°, and the multispectral sensor simultaneously obtains visible light band and near-infrared band data to obtain spectral images containing mineral composition information for feature identification of leakage areas.

[0016] Preferably, the method for laying out the sensor network in step S2 is: implanting a fiber optic Bragg grating sensor array in the support pile body, the fiber optic Bragg grating sensor including but not limited to a displacement meter, an inclinometer, and an inclination sensor, wherein the displacement meter is arranged in groups every 18-22 m along the foundation pit, the spacing within the group is one tenth of the foundation pit depth, and the depth is less than 2 m, the inclinometer is installed to a depth of at least 5 m below the potential slip surface, and the inclination sensor is laid out on the top crown beam and the middle waist beam of the support pile.

[0017] Preferably, the dynamic evaluation system in step S3 includes:

[0018] Quantitative model of environmental interference factors: electromagnetic interference intensity, temperature and humidity fluctuations, and vibration amplitude are used as evaluation parameters to establish a three-dimensional weight correction matrix. When a certain type of interference parameter exceeds the threshold, the formula Dynamically adjust the corresponding sensor weight, where α is the interference sensitivity coefficient, D is the real-time interference value, and D th is the preset threshold;

[0019] Sensor health scoring mechanism: Generates a health index H=0.5H by monitoring sensor data jump frequency, data loss rate, and calibration deviation value f +0.3H m +0.2H c , where H f Frequency stability score, H m Score data integrity, H c For calibration consistency scoring, when H<60 points, a protection mechanism of 30% weight reduction will be automatically triggered;

[0020] Strong electromagnetic interference emergency mechanism: When the electromagnetic interference is greater than 100dB, the lidar auxiliary scanning is activated and the data is fused through Kalman filtering;

[0021] Multi-scale fusion strategy: Within the minute-level data fusion cycle, high-frequency vibration data is denoised using wavelet transform and then individually weighted, while low-frequency environmental parameters are processed using sliding window mean filtering to form a weight distribution matrix with differentiated time scales.

[0022] Collaborative module with LSTM model: The real-time corrected weight coefficient is used as the input feature vector of the LSTM model. Through the attention mechanism, the model pays more attention to sensor data with large weight fluctuations, forming a closed-loop optimization system of "evaluation-fusion-feedback", improving the deformation prediction error to ≤5%.

[0023] Preferably, in step S4, the three-dimensional shape variable equation is:

[0024] ΔD=k1σ+k2β+k3γ+k4∈

[0025] Among them, ΔD is the three-dimensional deformation variable, σ, β, and γ are sensor data, ∈ is the UAV monitoring data, and k1, k2, k3, and k4 are the dynamic correlation coefficients updated through LSTM model learning.

[0026] Preferably, the process of learning and updating the dynamic correlation coefficient of the LSTM model in step S4 includes the following specific steps:

[0027] S41 constructs a multidimensional input vector containing the historical sequence of three-dimensional deformation variables, fluctuation characteristics of multi-sensor data, environmental interference factors, and real-time sensor weight coefficients. It captures the temporal dependencies of data through a bidirectional gated recurrent unit combined with an attention mechanism, and assigns a weight factor of 1.5-2 to the deformation data of key locations such as pit corners and support nodes.

[0028] S42 uses a sliding time window to continuously import real-time data streams, and uses the adaptive moment estimation optimizer every 10 minutes with a mean square error loss function:

[0029] Iterative optimization parameters, where n is the number of samples and i is the sample index;

[0030] S43 uses wavelet decomposition to preprocess high-frequency vibration data and sliding average filtering to reduce the dimension of low-frequency environmental parameters, and realizes multi-scale feature fusion through a hierarchical LSTM network;

[0031] S44, when the sensor data missing rate is greater than 20%, activating the interpolation prediction module based on Bayesian optimization to estimate missing period parameters.

[0032] Preferably, the crack detection method in step S5 includes: using a pixel-level fusion algorithm of visible light images and thermal infrared images to identify leakage hazards by calculating the dual criteria of temperature anomaly gradient and multispectral reflectance anomaly; introducing an attention mechanism to enhance crack edge feature extraction.

[0033] Preferably, the method for risk signal generation and early warning in step S6 includes: constructing a four-dimensional monitoring database based on a space-time cube model, performing time-series differential comparison of the three-dimensional deformation data and crack detection results every 10 minutes with the BIM model, and dynamically optimizing the early warning threshold through a reinforcement learning dynamic strategy algorithm; utilizing the network slicing technology of the 5G network to realize the priority transmission of monitoring data, controlling the end-to-end delay of the risk warning information within 35 seconds, and synchronously generating a three-dimensional visualization report containing the risk evolution trend.

[0034] The foundation pit safety monitoring device that combines drone oblique photography with the Internet of Things includes:

[0035] A drone module equipped with an oblique camera and multispectral sensor for aerial data collection;

[0036] Foundation pit sensor network: including but not limited to multiple types of sensors such as strain gauges and axial force gauges, to synchronously collect multi-physics field data;

[0037] The data processing unit is used to perform weighted fusion of multi-source heterogeneous data on the three-dimensional image data collected by the drone and the data collected by the sensor group, establish a multivariate regression equation for deformation analysis, and perform crack detection on the spectral image;

[0038] Edge computing unit, integrated with LSTM model, is used to dynamically update the correlation coefficient in the multivariate regression equation;

[0039] 5G transmission module, used to transmit processed monitoring data to the cloud early warning platform in real time;

[0040] The cloud-based early warning platform generates a digital twin of the foundation pit health status based on the space-time cube model, achieving minute-level early warning.

[0041] Preferably, the data processing unit includes a multi-source data fusion module, a deformation analysis module and a crack detection module. The multi-source data fusion module dynamically allocates weights according to the sensor accuracy. The deformation analysis module quantifies the contribution of each factor to the deformation based on the multivariate regression equation. The crack detection module enhances the crack characteristics and identifies the crack area through deep learning.

[0042] The technical effects and advantages of the present invention are as follows:

[0043] (1) This invention addresses the problem of "single data dimension" in traditional technologies. By using drone oblique photography (45°-60° five-view imaging) in conjunction with multispectral sensors and combining with foundation pit sensor networks (stress gauges, inclinometers, etc.), the three-dimensional geometric shape, multi-physics field parameters, and spectral characteristic data of the foundation pit are simultaneously acquired. Compared with the single-point measurement of traditional total stations, in order to improve monitoring accuracy, the coverage range is expanded from the local point of fixed sensors to the entire foundation pit area, and dynamic cloud maps such as slope displacement and support pile inclination are generated through three-dimensional model differential analysis, realizing the transition from "single-point discrete monitoring" to "global three-dimensional modeling";

[0044] (2) This invention addresses the shortcoming of "lack of real-time performance" by processing drone images and sensor data in real time through an edge computing gateway, combined with 5G network slicing technology (end-to-end latency ≤ 35 seconds), shortening the traditional "data collection-manual processing-early warning" cycle of more than 2 hours to less than 1 minute. For example, when the three-dimensional deformation variable ΔD ≥ 2.8mm or the stress mutation exceeds the design value by 12%, the system triggers a red alert in real time through a threshold mechanism dynamically optimized by reinforcement learning, and simultaneously generates a three-dimensional visualization report containing the risk evolution trend, meeting the minute-level emergency response requirements of deep foundation pit construction;

[0045] (3) This invention addresses the problem of "strong reliance on manual interpretation and high missed detection rate" by dynamically updating the deformation correlation coefficient through the LSTM model and combining it with the crack detection algorithm of the U-Net attention mechanism to achieve automatic identification of abnormalities such as cracks and leaks. At the same time, the drone is linked with the autonomous charging station and the sensor network to form a 24-hour unmanned monitoring system, eliminating the blind spots of manual inspections and transforming the traditional "passive identification" that relies on manual experience into "active intelligent early warning", thereby reducing the missed detection rate of potential risks.

[0046] (4) The present invention addresses the problem of “poor stability under strong electromagnetic interference or inclement weather” by using an electromagnetic interference emergency mechanism in the dynamic evaluation system (starting lidar auxiliary scanning when >100dB) and a temperature and humidity fluctuation weight correction matrix to ensure the continuity of data collection in complex environments such as construction site tower crane operation areas;

[0047] (5) This invention quantifies the contribution of different factors to foundation pit deformation (such as stress, displacement, and the dynamic correlation coefficient of drone monitoring data) through a multi-source heterogeneous data weighted fusion mechanism (dynamically assigning weights based on sensor accuracy) and time series feature learning of the LSTM model. Compared with traditional single data source analysis, this solution achieves collaborative modeling of multi-physics field data, reduces deformation prediction errors, provides a quantitative basis for emergency decision-making, and avoids the risk of misjudgment due to one-sided data. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is a schematic block diagram of the monitoring method of the present invention.

[0049] Figure 2 This is a schematic block diagram of the monitoring method structure of the present invention. DETAILED DESCRIPTION

[0050] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0051] The present invention provides Figure 1-2The foundation pit safety monitoring method shown in the figure, which is coordinated by the oblique photography of the drone and the Internet of Things, includes the following steps: S1, collecting three-dimensional images and spectral data of the foundation pit through the oblique photography camera and multispectral sensor carried by the drone; S2, collecting the stress, axial force, water level, displacement, inclinometer, static leveling, inclination and acceleration data of the foundation pit through the monitoring sensors deployed in the foundation pit; S3, based on the edge computing gateway, performing multi-source heterogeneous data weighted fusion on the three-dimensional image data collected by the drone in steps S1 and S2 and the multi-physical field data collected by the foundation pit monitoring sensors, assigning weights according to the sensor accuracy, and adjusting the weight coefficients in real time through the dynamic evaluation system; S4, establishing a multivariate regression equation between the foundation pit deformation and multiple environmental parameters based on the fused data Perform deformation analysis and calculate three-dimensional deformation variables; S5, perform crack detection on the spectral images collected by the drone, and automatically identify cracks and leakage anomalies, including inputting spectral images, strengthening crack feature extraction, synchronously detecting crack areas and outputting detection results; S6, fuses three-dimensional deformation variables and anomaly identification results to generate risk signals, and transmits the analysis results to the cloud-based early warning platform in real time through the 5G network to generate a digital twin of the health status of the foundation pit and realize minute-level risk warning. The drone's tilted photography camera (five viewing angles of 45°-60°) and multi-spectral sensors collaborate to collect three-dimensional images and spectral data. At the same time, the foundation pit sensor network (stress gauges, inclinometers, etc.) synchronously obtains stress, displacement and other multi-physical field parameters to realize "air-ground" multi-source data fusion.

[0052] Among them, in the drone data collection in step S1, the "spiral + grid" composite route planning is adopted, and a low-altitude circumferential scan with a height of less than 20m is performed near the foundation pit excavation surface. The route planning covers the foundation pit slope, support structure and surrounding surface area. The oblique photography camera collects five-view images at an inclination angle of 45°-60°. The multispectral sensor simultaneously obtains visible light band and near-infrared band data, and obtains spectral images containing mineral composition information for feature identification of leakage areas. The "spiral + grid" composite route planning performs a circumferential scan at a low altitude of 20m above the foundation pit excavation surface. Combined with the mineral composition identification capability of the multispectral sensor, it reduces the impact of bad weather on image quality, and the data loss rate is reduced compared with traditional drone solutions.

[0053] The sensor network deployment method in step S2 is as follows: an array of fiber grating sensors is implanted in the support pile body. The fiber grating sensors include but are not limited to displacement meters, inclinometers, and tilt sensors. The displacement meters are arranged in groups every 18-22 m along the foundation pit. The spacing within the group is one tenth of the foundation pit depth, and the depth is less than 2 m. The inclinometer is installed at a depth of at least 5 m below the potential slip surface. The tilt sensors are deployed on the top crown beam and the middle waist beam of the support pile.

[0054] The dynamic evaluation system in step S3 includes: environmental interference factor quantification model: taking electromagnetic interference intensity, temperature and humidity fluctuation, and vibration amplitude as evaluation parameters, a three-dimensional weight correction matrix is ​​established. When a certain type of interference parameter exceeds the threshold, the following formula is used: Dynamically adjust the corresponding sensor weight, where α is the interference sensitivity coefficient (electromagnetic interference is 0.3, temperature and humidity is 0.15), D is the real-time interference value, and D th is the preset threshold; sensor health scoring mechanism: by monitoring the sensor data jump frequency, data loss rate, and calibration deviation value to generate a health index H = 0.5H f +0.3H m +0.2H c , where H f Frequency stability score, H m Score data integrity, H c To score the calibration consistency, a protection mechanism that automatically reduces the weight by 30% is triggered when H is less than 60 points; a strong electromagnetic interference emergency mechanism: when the electromagnetic interference is greater than 100dB, a lidar-assisted scan is initiated, and data is fused through a Kalman filter; a multi-scale fusion strategy: within the minute-level data fusion cycle, high-frequency vibration data (>10Hz) is denoised using a wavelet transform and then individually weighted, and low-frequency environmental parameters (temperature and humidity) are processed using a sliding window mean filter to form a time-scale differentiated weight distribution matrix to improve dynamic response accuracy; a collaborative module with the LSTM model: the real-time corrected weight coefficient is used as the input feature vector of the LSTM model, and the attention mechanism is used to enable the model to pay more attention to sensor data with large weight fluctuations, forming an "evaluation-fusion-feedback" closed-loop optimization system, improving the deformation prediction error to ≤5%. The dynamic evaluation system includes a strong electromagnetic interference emergency mechanism (lidar-assisted scanning is initiated when the electromagnetic interference is greater than 100dB) and a temperature and humidity fluctuation weight correction matrix. The sensor credibility is dynamically adjusted through the three-dimensional weight correction matrix to ensure data continuity.

[0055] Among them, in step S4, the three-dimensional deformation variable equation is: ΔD = k1σ + k2β + k3γ + k4∈, where ΔD is the three-dimensional deformation variable, σ, β, and γ are sensor data, ∈ is the UAV monitoring data, and k1, k2, k3, and k4 are dynamic correlation coefficients updated through learning of the LSTM model. Through the weighted fusion mechanism of multi-source heterogeneous data, multidimensional data such as three-dimensional deformation variables, stress, and spectral characteristics are incorporated into a unified analysis framework, and a multivariate regression equation is constructed to quantify the contribution of multiple factors, completely breaking through the dimensional limitation of traditional single-point measurement.

[0056] Furthermore, the process of learning and updating the dynamic correlation coefficient of the LSTM model in step S4 includes the following specific steps: S41, constructing a multidimensional input vector including a historical sequence of three-dimensional deformation variables, fluctuation characteristics of multi-sensor data, environmental interference factors and real-time weight coefficients of sensors, capturing the temporal dependency of data through a bidirectional gated recurrent unit combined with an attention mechanism, and assigning a 1.5-2 times weight factor to the deformation data of key parts such as the positive corners of the foundation pit and support nodes (the weight factor is determined by generating a part importance matrix through supervised training of a historical accident case library); S42, continuously importing real-time data streams using a sliding time window (window length 72 hours), and using an adaptive moment estimation optimizer every 10 minutes with a mean square error loss function: Iterative optimization parameters, where n is the number of samples, representing the total number of data samples used for model training or validation, that is, the number of deformation data points involved in error calculation. For example, if monitoring data within 72 hours is selected, n can represent the number of samples in that time period (such as once every 10 minutes, n = 432), and i is the sample index, representing the sequence number of the i-th sample (i = 1, 2, ..., n), which is used to traverse each set of differences between the predicted value and the true value. For example, ΔD pred,i is the predicted value of the three-dimensional shape of the i-th sample, ΔD true,i The update step size is ≤0.01, and gradient clipping is used to prevent gradient explosion. The threshold is set to 5.0. S43: high-frequency vibration data is preprocessed by wavelet decomposition, and low-frequency environmental parameters are reduced by sliding average filtering. Multi-scale feature fusion is achieved through a hierarchical LSTM network. S44: when the sensor data missing rate is greater than 20%, the interpolation prediction module based on Bayesian optimization is activated to estimate the parameters of the missing period to ensure the continuity of deformation analysis. The interpolation error is ≤3%. Through closed-loop feedback with the dynamic evaluation system, the deformation prediction error is ≤4.5%. The LSTM model uses a 72-hour sliding window to continuously import real-time data streams, and iteratively optimizes the deformation prediction parameters every 10 minutes to ensure real-time tracking of dynamic deformation trends. The LSTM model combines the attention mechanism to give 1.5-2 times the weight to the deformation data of key parts such as the positive corners of the foundation pit. The dynamic warning threshold is trained through the historical accident case library (such as ΔD ≥ 2.8mm triggers a red warning) to replace traditional manual experience judgment.

[0057] Among them, the crack detection method in step S5 includes: using a pixel-level fusion algorithm of visible light images and thermal infrared images to identify leakage risks by calculating the dual criteria of temperature anomaly gradient (ΔT≥2.5℃) and multispectral reflectivity anomaly (near-infrared band reflectivity>0.35); introducing an attention mechanism to enhance crack edge feature extraction, embedding a residual attention module in the encoder-decoder structure of the U-Net network, assigning a 1.8-fold weight factor to the grayscale mutation area in the spectral image, combining morphological operations to eliminate artifacts, so that the crack width detection accuracy reaches 0.1mm, and the leakage identification accuracy is increased to 98.5%. By embedding the residual attention module in the U-Net network and combining the visible light-thermal infrared dual criteria (ΔT≥2.5℃+near-infrared reflectivity>0.35), automatic analysis of crack width detection at the 0.1mm level and leakage identification rate of 98.5% is achieved.

[0058] Among them, the method for risk signal generation and warning in step S6 includes: building a four-dimensional monitoring database based on the Time-SpaceCube model, performing time-series difference comparison on the three-dimensional deformation data and crack detection results every 10 minutes with the BIM model, and dynamically optimizing the warning threshold through the reinforcement learning dynamic strategy algorithm (a red warning is triggered when the three-dimensional deformation variable ΔD ≥ 2.8 mm and the stress mutation exceeds the design value by 12%); using the network slicing technology of the 5G network to realize the priority transmission of monitoring data, control the end-to-end delay of risk warning information within 35 seconds, and simultaneously generate a three-dimensional visualization report containing risk evolution trends to support real-time dynamic adjustment of the construction plan, and process drone images and sensor data in real time through the edge computing gateway to avoid the transmission delay of traditional cloud-based centralized processing; using 5G network slicing technology to control the end-to-end delay of risk warning information within 35 seconds, and the full process response time from data collection to warning is ≤ 1 minute.

[0059] A foundation pit safety monitoring device that collaborates with drone oblique photography and the Internet of Things includes: a drone module equipped with an oblique photography camera and a multispectral sensor for aerial data collection; a foundation pit sensor network: multiple types of sensors including but not limited to strain gauges and axial force gauges for synchronously collecting multi-physical field data; a data processing unit for weighted fusion of multi-source heterogeneous data on three-dimensional image data collected by the drone and data collected by the sensor group, establishing a multivariate regression equation for deformation analysis, and detecting cracks in spectral images; an edge computing unit integrating an LSTM model for dynamically updating the correlation coefficient in the multivariate regression equation; a 5G transmission module for transmitting the processed monitoring data to a cloud-based early warning platform in real time; and a cloud-based early warning platform for generating a digital twin of the foundation pit health status based on a space-time cube model (corresponding to the four-dimensional database of claim 8) to achieve minute-level early warning.

[0060] Among them, the data processing unit includes a multi-source data fusion module, a deformation analysis module and a crack detection module. The multi-source data fusion module dynamically allocates weights according to the sensor accuracy. The deformation analysis module quantifies the contribution of each factor to the deformation based on the multivariate regression equation. The crack detection module enhances the crack characteristics and identifies the crack area through deep learning. The data processing unit also includes a residual attention module, which is used to eliminate artifacts.

[0061] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A foundation pit safety monitoring method based on the collaboration of UAV oblique photography and the Internet of Things, characterized by: The following steps are involved: S1, collects three-dimensional images and spectral data of foundation pits through the oblique photography camera and multispectral sensor carried by the UAV; S2, collects stress, axial force, water level, displacement, inclinometer, static leveling, inclination and acceleration data of the foundation pit through monitoring sensors deployed in the foundation pit; S3, based on the edge computing gateway, performs weighted fusion of multi-source heterogeneous data on the 3D image data collected by the drone in steps S1 and S2 and the multi-physical field data collected by the foundation pit monitoring sensors. Weights are assigned according to sensor accuracy, and the weight coefficients are adjusted in real time through a dynamic evaluation system. S4, based on the fused data, a multivariate regression equation of foundation pit deformation and multiple environmental parameters is established to perform deformation analysis and calculate the three-dimensional deformation variable; S5, performs crack detection on the spectral images collected by the drone, automatically identifying cracks and leakage anomalies, including inputting spectral images, enhancing crack feature extraction, synchronously detecting crack areas, and outputting detection results; S6, integrates three-dimensional deformation variables and anomaly recognition results to generate risk signals, and transmits the analysis results to the cloud-based early warning platform in real time through the 5G network, generating a digital twin of the foundation pit health status and achieving minute-level risk warnings.

2. The foundation pit safety monitoring method using drone oblique photography and the Internet of Things in collaboration according to claim 1 is characterized in that: During drone data collection in step S1, a "spiral + grid" composite route planning was used to perform a low-altitude circumferential scan at an altitude of less than 20 meters near the foundation pit excavation surface. The route planning covered the foundation pit slope, support structure, and surrounding surface area. The oblique camera collected five-view images at an inclination angle of 45°-60°. The multispectral sensor simultaneously acquired visible light and near-infrared band data, obtaining spectral images containing mineral composition information for feature identification of leakage areas.

3. The foundation pit safety monitoring method using drone oblique photography and the Internet of Things in collaboration according to claim 1 is characterized in that: The sensor network deployment method in step S2 is as follows: an array of fiber Bragg grating sensors is implanted in the support pile. The fiber Bragg grating sensors include but are not limited to displacement meters, inclinometers, and tilt sensors. The displacement meters are arranged in groups every 18-22 m along the foundation pit. The spacing within the groups is one-tenth of the foundation pit depth and the depth is less than 2 m. The inclinometer is installed at a depth of at least 5 m below the potential slip surface. The tilt sensors are deployed on the top crown beam and the middle waist beam of the support pile.

4. The foundation pit safety monitoring method using drone oblique photography and the Internet of Things in collaboration according to claim 1 is characterized in that: The dynamic evaluation system in step S3 includes: Quantitative model of environmental interference factors: electromagnetic interference intensity, temperature and humidity fluctuations, and vibration amplitude are used as evaluation parameters to establish a three-dimensional weight correction matrix. When a certain type of interference parameter exceeds the threshold, the formula Dynamically adjust the corresponding sensor weight, where α is the interference sensitivity coefficient, D is the real-time interference value, and D th is the preset threshold; Sensor health scoring mechanism: Generates a health index H=0.5H by monitoring sensor data jump frequency, data loss rate, and calibration deviation value f +0.3H m +0.2H c , where H f Frequency stability score, H m Score data integrity, H c For calibration consistency scoring, when H<60 points, a protection mechanism of 30% weight reduction will be automatically triggered; Strong electromagnetic interference emergency mechanism: When the electromagnetic interference is greater than 100dB, the lidar auxiliary scanning is activated and the data is fused through Kalman filtering; Multi-scale fusion strategy: Within the minute-level data fusion cycle, high-frequency vibration data is denoised using wavelet transform and then individually weighted, while low-frequency environmental parameters are processed using sliding window mean filtering to form a weight distribution matrix with differentiated time scales. Collaborative module with LSTM model: The real-time corrected weight coefficient is used as the input feature vector of the LSTM model, and the attention mechanism is used to make the model pay more attention to sensor data with large weight fluctuations.

5. The foundation pit safety monitoring method using drone oblique photography and the Internet of Things in collaboration according to claim 1 is characterized in that: In step S4, the three-dimensional shape variable equation is: ΔD=k1σ+k2β+k3γ+k4∈ Among them, ΔD is the three-dimensional deformation variable, σ, β, and γ are sensor data, ∈ is the UAV monitoring data, and k1, k2, k3, and k4 are the dynamic correlation coefficients updated through LSTM model learning.

6. The foundation pit safety monitoring method using drone oblique photography and the Internet of Things in collaboration according to claim 5 is characterized in that: The process of learning and updating the dynamic correlation coefficient of the LSTM model in step S4 includes the following specific steps: S41 constructs a multidimensional input vector containing the historical sequence of three-dimensional deformation variables, fluctuation characteristics of multi-sensor data, environmental interference factors, and real-time sensor weight coefficients. It captures the temporal dependencies of data through a bidirectional gated recurrent unit combined with an attention mechanism, and assigns a weight factor of 1.5-2 to the deformation data of key locations such as pit corners and support nodes. S42 uses a sliding time window to continuously import real-time data streams, and uses the adaptive moment estimation optimizer every 10 minutes with a mean square error loss function: Iterative optimization parameters, where n is the number of samples and i is the sample index; S43 uses wavelet decomposition to preprocess high-frequency vibration data and sliding average filtering to reduce the dimension of low-frequency environmental parameters, and realizes multi-scale feature fusion through a hierarchical LSTM network; S44, when the sensor data missing rate is greater than 20%, activating the interpolation prediction module based on Bayesian optimization to estimate the missing period parameters.

7. The foundation pit safety monitoring method using drone oblique photography and the Internet of Things in collaboration according to claim 1 is characterized in that: The crack detection method in step S5 includes: using a pixel-level fusion algorithm of visible light images and thermal infrared images to identify leakage hazards by calculating the dual criteria of temperature anomaly gradient and multispectral reflectivity anomaly; and introducing an attention mechanism to enhance crack edge feature extraction.

8. The foundation pit safety monitoring method using drone oblique photography and the Internet of Things in collaboration according to claim 1 is characterized in that: The method for risk signal generation and early warning in step S6 includes: constructing a four-dimensional monitoring database based on the space-time cube model, performing time-series differential comparison of the three-dimensional deformation data and crack detection results every 10 minutes with the BIM model, and dynamically optimizing the early warning threshold through a reinforcement learning dynamic strategy algorithm; utilizing the network slicing technology of the 5G network to realize the priority transmission of monitoring data, controlling the end-to-end delay of risk warning information within 35 seconds, and synchronously generating a three-dimensional visualization report containing the risk evolution trend.

9. The foundation pit safety monitoring device using drone oblique photography and the Internet of Things is characterized by: include: A drone module equipped with an oblique camera and multispectral sensor for aerial data collection; Foundation pit sensor network: including but not limited to multiple types of sensors such as strain gauges and axial force gauges, to synchronously collect multi-physics field data; The data processing unit is used to perform weighted fusion of multi-source heterogeneous data on the three-dimensional image data collected by the drone and the data collected by the sensor group, establish a multivariate regression equation for deformation analysis, and perform crack detection on the spectral image; Edge computing unit, integrated with LSTM model, is used to dynamically update the correlation coefficient in the multivariate regression equation; 5G transmission module, used to transmit processed monitoring data to the cloud early warning platform in real time; The cloud-based early warning platform generates a digital twin of the foundation pit health status based on the space-time cube model, achieving minute-level early warning.

10. The foundation pit safety monitoring device combining drone oblique photography with the Internet of Things according to claim 9, characterized in that: The data processing unit includes a multi-source data fusion module, a deformation analysis module and a crack detection module. The multi-source data fusion module dynamically assigns weights according to sensor accuracy. The deformation analysis module quantifies the contribution of each factor to deformation based on a multivariate regression equation. The crack detection module enhances crack characteristics and identifies crack areas through deep learning.

Images