Multi-source data-artificial intelligence fused building deviation correction settlement prediction method and system

Abstract

The invention relates to the field of building settlement monitoring, and discloses a multi-source data-artificial intelligence fused building deviation rectification settlement prediction method and system, and the method comprises the following steps: multi-source data collection: obtaining multi-class sensor data related to building settlement, the multiple types of sensors comprise a level gauge, a total station, an underground water level sensor and a soil body pressure sensor, and multi-modal data input is formed; the invention also discloses a multi-source data-artificial intelligence fused building deviation rectification settlement prediction system. The system comprises a data acquisition module, a data processing module, a deep learning prediction module, a model optimization module, an intelligent early warning module and a user interaction module. According to the invention, by dynamically adjusting the settlement risk threshold and introducing a multi-stage early warning signal and an intelligent decision support system, accurate early warning and real-time decision support of the settlement risk are realized, and the adaptability, accuracy and response efficiency of the system are significantly improved.

Description

Technical Field

[0001] The present invention relates to the field of building settlement monitoring, and specifically to a building correction settlement prediction method and system integrating multi-source data and artificial intelligence. Background Art

[0002] With the acceleration of urbanization, building settlement is a growing concern. Monitoring and early warning of settlement are particularly important for large buildings or high-risk areas. If settlement is not detected and addressed promptly, it can lead to serious structural safety issues and even catastrophic accidents.

[0003] Currently, the main technical solutions for subsidence monitoring rely on fixed thresholds and standardized warning signals. Some systems use static safety standards as settlement thresholds, triggering warning signals when monitored values ​​exceed these standards. Furthermore, some systems incorporate historical data for risk assessment to aid decision-making. These technical solutions can, to a certain extent, provide real-time monitoring and warning information, helping to promptly identify subsidence issues and enable appropriate measures to prevent escalation.

[0004] However, existing technologies have obvious shortcomings in practical applications. First, the use of fixed thresholds cannot effectively adapt to the dynamic changes in settlement conditions, resulting in frequent false alarms and missed reports. Settlement risks are often highly uncertain, and fixed standards are difficult to cope with settlement changes in complex environments or special climatic conditions. Secondly, most existing early warning systems use a single alarm mechanism, which cannot flexibly adjust the intensity of the warning signal according to different settlement levels. They lack multi-level warning functions and find it difficult to provide sufficient decision support. In addition, most existing technologies lack intelligent decision-making analysis functions. Managers can only respond based on a single alarm message and lack scientific risk assessment and disposal plans. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the present invention provides a building correction settlement prediction method and system that integrates multi-source data and artificial intelligence, which solves the shortcomings of the fixed threshold and single early warning mechanism in the existing technology; by dynamically calculating the settlement risk threshold, introducing multi-level early warning signals and providing intelligent decision-making support, it solves the problems of false alarms and missed alarms.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: a building correction settlement prediction method integrating multi-source data and artificial intelligence, comprising the following steps: Multi-source data acquisition: Acquire multiple types of sensor data related to building settlement, including levels, total stations, groundwater level sensors, and soil pressure sensors, to form multimodal data input; Data preprocessing: performing data cleaning, missing value filling, denoising, time alignment, and normalization on the multi-category sensor data to obtain high-quality standardized data; Deep learning prediction model construction: A deep learning model combining long short-term memory neural networks and convolutional neural networks is used to train preprocessed multi-class sensor data and build a building settlement prediction model; Model optimization and dynamic updating: Using incremental learning to continuously optimize the prediction model based on newly collected multimodal data; Intelligent early warning: Dynamically calculate the settlement risk threshold based on the prediction results. When the predicted value reaches the set safety warning range, different levels of early warning signals are triggered to provide settlement risk prompts and decision support.

[0007] Preferably, the data preprocessing includes: Missing value filling: fill in missing data through time series interpolation method; Denoising: Use signal filtering methods to filter out random noise in the data; Time alignment: Align data using interpolation methods based on the sampling time intervals of different sensors; Normalization processing: Normalize the data of each sensor to keep its value range consistent.

[0008] Preferably, the deep learning prediction model construction includes: Long short-term memory neural network is used for time series modeling to extract the long-term dependencies of building settlement data; Use convolutional neural networks to extract spatial features and analyze the correlation between different sensor data; The outputs of the long short-term memory neural network and the convolutional neural network are integrated to form a comprehensive settlement trend prediction model.

[0009] Preferably, the model optimization and dynamic update include: Dynamic optimization based on incremental learning automatically adjusts model parameters when new data is input, so that the prediction model can adapt to the new settlement environment; Through the adaptive adjustment mechanism, the learning weights of the model are automatically corrected according to the changes in the building settlement trend; A feedback learning mechanism is used to continuously optimize the model structure and parameters based on historical prediction errors.

[0010] Preferably, the intelligent early warning includes: Calculate dynamic subsidence risk thresholds based on historical data and adjust warning trigger conditions in real time; Early warning signal classification includes three levels: low risk, medium risk, and high risk, each corresponding to different safety alert measures; Early warning information is visualized and based on digital twin technology, the building settlement status is displayed to assist managers in making decisions.

[0011] Preferably, the dynamic update includes: Adaptive learning rate adjustment: Dynamically adjust the learning rate based on the prediction error generated by the model in real-time applications; Incremental learning technology: When new data arrives, the existing model is optimized using incremental learning methods, allowing the model to continuously adapt to new data changes without complete retraining; Performance feedback mechanism: Each feedback of new data will monitor the performance of the model, and the reduction of prediction error will guide the direction of model parameter adjustment.

[0012] Preferably, the long short-term memory neural network comprises: Multi-layer LSTM network structure: The long short-term memory neural network is composed of multiple stacked LSTM layers; Bidirectional LSTM design: uses a bidirectional long short-term memory network to consider both future and past information; Forget gate optimization: Based on the traditional LSTM network, the forget gate is optimized so that the model can flexibly adjust the memory length according to changes in input data and optimize the information retention process; Time-weighted mechanism: A time-weighted mechanism is introduced to dynamically adjust the model's predictive ability according to the weights of different time steps during model training.

[0013] Preferably, the convolutional neural network includes: Convolutional layer design: A multi-layer convolutional neural network structure is used to gradually extract local and global features from settlement data, helping the model identify the relationship between building settlement and changes in the external environment; Adaptive convolution kernel adjustment: By adaptively adjusting the size and step size of the convolution kernel, the network can automatically select the most appropriate convolution configuration based on the spatial distribution characteristics of different data; Combining pooling layer and convolution layer: Set the pooling layer after the convolution layer, using the maximum pooling or average pooling method to reduce the feature dimension and avoid overfitting; Combining convolutional networks with LSTM networks: Convolutional neural networks are combined with long short-term memory neural networks to simultaneously extract temporal and spatial features from data.

[0014] Preferably, the data cleaning includes: Missing data detection and filling: Statistical methods are used to detect and fill missing values ​​in sensor data; Outlier detection: Use Z-Score and IQR methods to detect outliers in the data and correct or delete them; Data denoising: Apply signal processing methods such as wavelet transform or Kalman filtering to filter out noise in the data; Deduplication of data: Deduplication of duplicated data that is likely to exist in sensor data is performed to avoid repeated use of data at the same time; Data time synchronization: For data from different sensors, time alignment technology is used to ensure that all data is analyzed within a unified time dimension to avoid errors caused by time differences.

[0015] The present invention also provides a building correction settlement prediction system integrating multi-source data and artificial intelligence, comprising: A data acquisition module is used to collect multi-source data related to building settlement, including data from a level, a total station, a groundwater level sensor, and a soil pressure sensor, and transmit the collected data to a data processing module; A data processing module, connected to the data acquisition module, is used to receive the multi-source data, clean, denoise, fill missing values, normalize and time-align the data to form standardized data input, and transmit the processed data to the deep learning prediction module; The deep learning prediction module is connected to the data processing module and is used to receive standardized data input and build a settlement prediction model based on long short-term memory neural network and convolutional neural network to predict the future settlement trend of buildings. The prediction results are then transmitted to the model optimization module and the intelligent early warning module; A model optimization module, connected to the deep learning prediction module, is used to receive prediction results and continuously optimize the deep learning prediction model using incremental learning, adaptive adjustment and reinforcement learning mechanisms to adapt it to long-term sedimentation changes, and feed back the optimized model parameters to the deep learning prediction module to update the prediction model; The intelligent early warning module is connected to the deep learning prediction module and the user interaction module respectively. It is used to receive the settlement prediction results and calculate the dynamic early warning threshold based on historical data. When the settlement prediction result exceeds the warning threshold, it triggers an early warning signal according to different risk levels and sends an early warning information to the user interaction module. The user interaction module is connected to the intelligent early warning module, and is used to receive early warning information and provide users with real-time monitoring of settlement status, historical data query and early warning feedback functions. Users can adjust the early warning threshold or input additional information through the user interaction module. The input information can be transmitted to the intelligent early warning module to optimize the early warning mechanism.

[0016] The present invention provides a building correction settlement prediction method and system that integrates multi-source data and artificial intelligence. It has the following beneficial effects: 1. This invention utilizes a method for dynamically calculating settlement risk thresholds, enabling real-time adjustment of warning criteria based on actual prediction results, thereby improving the system's flexibility and adaptability. Compared to existing technologies, this solution can more accurately identify building settlement risks and respond promptly to changing environmental conditions, avoiding the false alarms and missed alarms associated with fixed thresholds in traditional early warning systems.

[0017] 2. This invention introduces a multi-level early warning signal triggering mechanism, enabling the settlement monitoring system to issue corresponding early warning prompts at different risk levels. Compared with the single alarm method used in traditional technologies, the multi-level early warning of this invention can better reflect the severity of settlement changes, avoid information overload or oversimplified risk warnings, and thus improve the system's early warning accuracy and reliability.

[0018] 3. The decision support system of this invention provides intelligent intervention recommendations based on real-time data and historical information, helping managers make informed decisions. Compared to existing solutions that lack effective decision support, this invention offers more targeted risk assessments and solutions, avoiding engineering safety hazards caused by human misjudgment.

[0019] 4. This invention combines incremental learning with intelligent early warning, enabling the system to self-optimize and continuously improve prediction accuracy. Unlike traditional methods that rely on manual intervention, this invention implements intelligent adaptive adjustment, effectively avoiding the lag and deviation problems that occur during manual model updates, thereby making settlement risk prediction more accurate and real-time. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 Schematic diagram of the method flow of the present invention; Figure 2 The following is a schematic diagram of the system construction of the present invention. DETAILED DESCRIPTION

[0021] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the present specification. 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.

[0022] Please see the attached Figure 1 , an embodiment of the present invention provides a building correction settlement prediction method integrating multi-source data and artificial intelligence, comprising the following steps: S1. Multi-source data acquisition: Acquire multiple types of sensor data related to building settlement, including levels, total stations, groundwater level sensors, and soil pressure sensors, to form multimodal data input; To accurately predict building settlement, a multi-source data acquisition approach was adopted, acquiring data from various sensor types. These sensor data sources accurately captured the entire building settlement process and the factors influencing it, providing reliable input for subsequent data processing, deep learning prediction, and other steps.

[0023] In this embodiment, the main function of the data acquisition module is to obtain multiple types of sensor data related to building settlement. Specifically, it includes the following types of sensors: Levels: Used to detect changes in building height, they can accurately measure settlement or displacement. Levels provide precise settlement values, making them particularly useful for detecting subtle changes in a building. Level data can help models capture overall building settlement trends.

[0024] Total Station: A total station is a high-precision ground-based measurement instrument that provides spatial location information on building settlement. Using three-dimensional coordinate measurement, it efficiently captures changes in building settlement points. Total station data not only reflects building settlement but also provides information on other geometric changes, providing spatial characteristics for subsequent settlement prediction.

[0025] Groundwater level sensors: Groundwater level fluctuations are correlated with building settlement, especially for underground structures or structures significantly affected by hydrological conditions. Groundwater level sensors can monitor groundwater level fluctuations in real time and predict the sensitivity of settlement to groundwater fluctuations. The data they provide reflects the dynamics of groundwater levels, providing time series features for settlement prediction models.

[0026] Soil pressure sensors: Soil pressure sensors measure pressure changes in the soil, which are closely related to building settlement. Changes in soil pressure often accompany building settlement, so incorporating this sensor data can provide key insights into soil deformation and stress distribution for settlement models.

[0027] By combining these sensors, the system can capture multimodal data inputs encompassing diverse physical quantities and environmental conditions. This multimodal data input incorporates multidimensional information across time, space, and environment, providing rich features for subsequent deep learning prediction modules. The data from these sensors not only exists independently but also exhibits interdependencies across time and space, providing a more comprehensive perspective on changes in building settlement.

[0028] During the implementation of this invention, the multi-source data acquisition module first acquires real-time data from various sensors. This data includes key indicators such as building settlement and displacement, groundwater level changes, and soil pressure. Key points in the data acquisition process are ensuring that each sensor has a consistent data acquisition frequency and that the collected data is aligned with data from other sensors.

[0029] Specifically, the collected data is transmitted to the data processing module. During transmission, the data needs to be properly labeled and encoded so that the data processing module can accurately identify the source of each type of data. For example, level data might be labeled "level_001," and total station data might be labeled "totalstation_001," etc., to ensure data uniqueness during transmission and processing.

[0030] In some embodiments, the system can configure different communication protocols to ensure that data from different types of sensors can be effectively connected to the same data network. This allows for real-time transmission of sensor data, not only ensuring data integrity but also improving the real-time nature of data collection and reducing latency during the data collection process.

[0031] As an option, collected data can be timestamped, typically with millisecond accuracy, to ensure data timeliness. During preprocessing in the data processing module, time alignment techniques can be used to consolidate data from different sensors onto the same time scale, avoiding time deviations caused by different sensor sampling frequencies.

[0032] The input data generated by the multi-source data acquisition module forms a multidimensional matrix. Each row represents data from different sensors at a specific point in time, while each column represents a specific type of data. For example, the first column might represent settlement values ​​from a level, the second column might represent 3D coordinates from a total station, the third column might represent groundwater levels, and the fourth column might represent soil pressure. This multidimensional data input structure provides an excellent foundation for subsequent data processing and model training.

[0033] In some embodiments, to ensure the quality of data input, data verification techniques may be employed during the data collection process to further enhance data reliability. For example, by comparing measurement results from different sensors at the same location, consistency of measurement results can be ensured and outliers can be corrected.

[0034] After the multi-source data acquisition module completes its work, all collected data is transmitted via the network to the data processing module. This module performs data cleaning, denoising, and missing value filling to ensure that the data meets the input requirements of the deep learning prediction model. To avoid model training deviations caused by data quality issues, the connection between the data acquisition and data processing modules must ensure key factors such as data format, type, and time alignment.

[0035] In some possible implementations, the data acquisition module can also be linked with other external systems (such as weather monitoring systems, architectural design systems, etc.) to obtain data support for more environmental factors. This supplementary external data can further enhance the model's predictive capabilities.

[0036] In this embodiment, the type and number of sensors are not limited to levels, total stations, groundwater level sensors, and soil pressure sensors. Other types of sensors, such as temperature sensors and displacement sensors, can be selected based on actual needs. Different types of sensors can provide data inputs of varying dimensions, further enhancing the accuracy of settlement prediction.

[0037] Furthermore, as technology evolves, data collection methods and equipment can be continuously updated. For example, with the development of the Internet of Things (IoT), more new sensors will be able to transmit data in real time, and the application of 5G communication technology will further enhance the speed and reliability of data transmission. System design should consider compatibility with new technologies to ensure smooth operation.

[0038] S2. Data preprocessing: Data cleaning, missing value filling, denoising, time alignment and normalization are performed on multi-class sensor data to obtain high-quality standardized data; In the multi-source data-AI-integrated building correction settlement prediction method, data preprocessing is a key step in ensuring that the subsequent deep learning model can obtain high-quality, standardized data. The data preprocessing module cleans, denoises, fills missing values, time-aligns, and normalizes the data collected from multiple sensor types to ensure that the data meets the input requirements of the deep learning model, thereby improving the accuracy and robustness of settlement prediction.

[0039] In this embodiment, the data preprocessing module receives multiple types of sensor data from the multi-source data acquisition module. This data often contains issues such as incompleteness, high noise, or inconsistent sampling times. To effectively utilize this data in subsequent deep learning models, it must first undergo a series of preprocessing operations. These operations include data cleaning, missing value filling, denoising, time alignment, and normalization.

[0040] During data cleaning, the first step is to detect and clean any abnormal data that may have occurred during the acquisition process. Sensors often experience external interference when collecting data, resulting in some data not conforming to the normal range or data formatting issues. For example, when collecting data using a total station, equipment failure or environmental changes may lead to errors in some measurement results. To prevent these abnormal data from impacting subsequent processing and predictive models, outliers must be detected and cleaned at this stage.

[0041] Specifically, outlier detection methods can be performed using methods based on Z-score or IQR (interquartile range). For example, for data x from a certain sensor, we can use the Z-score formula to detect whether it is an outlier: Where μ is the mean of the data, σ is the standard deviation of the data, and x is the sensor data. Generally, when the Z value is greater than 3 or less than -3, the data point is considered an outlier. Outliers are then replaced with appropriate estimates or deleted.

[0042] Filling missing data is another important step in data preprocessing. In building settlement monitoring, sensor data at certain points in time may be lost due to sensor failure, data transmission problems, or external factors. To ensure data integrity and continuity, we need to fill missing values.

[0043] In some embodiments, missing values ​​can be filled using time series interpolation methods. For example, missing data for a sensor at time t can be filled using the mean of the preceding and following data points, linear interpolation, or more complex interpolation algorithms. In particular, in multimodal data scenarios, joint interpolation techniques can be used to fill missing data based on the correlation between sensors.

[0044] For example, if sensor A loses data at time t1, we can use the interpolation algorithm to predict the value of A(t1) based on the values ​​of sensors B and C at the same time. Specifically, linear interpolation can be performed according to the following formula: Among them, A(t0) and A(t2) are the data at time t0 and t2 respectively, t1 is the time to be filled, and A is the sensor.

[0045] Data denoising aims to improve data quality and reduce the impact of noise on model training. In sensor data, noise often manifests as irregular fluctuations or sudden anomalies. Common denoising methods include signal filtering, Kalman filtering, and wavelet transforms.

[0046] Specifically, high-frequency noise can be removed through signal filtering methods, such as low-pass filtering. When processing building settlement data, low-frequency components often contain valuable settlement trends, while high-frequency components are often noise. Therefore, an appropriate filter can be designed to remove these higher-frequency signals, thereby extracting the building settlement trends.

[0047] For example, assuming that the data x(t) contains noise n(t), the output after passing through the low-pass filter is: Where h(τ) is the impulse response function of the filter, x(t) is the original signal at time t, y(t) is the denoised signal, τ is the integral variable representing the time delay, and dτ is the differential element in the integral symbol, representing the integral of the variable τ.

[0048] Time alignment in multi-source data addresses the potential inconsistencies in sampling frequencies and timestamps between different sensors. In multimodal data processing, time alignment is a necessary step to ensure that data from different sensors can be analyzed in the same time dimension.

[0049] Typically, different sensors collect data at different time intervals. To eliminate these time deviations, interpolation is used to align the data to the same time interval. This can be achieved using linear interpolation, spline interpolation, or other suitable interpolation algorithms.

[0050] For example, suppose the data of sensor A is collected at time t1, t2, t3, ..., and the data of sensor B is collected at t ′ 1,t ′ 2,t ′ To align the data, the data of B can be interpolated at times t1, t2, t3, ... to make them consistent with the timestamp of the data of A.

[0051] To ensure consistent numerical ranges for different sensor data and prevent certain sensor data from significantly impacting model training, all sensor data is normalized. The primary purpose of normalization is to map data to a uniform numerical range. Common normalization methods include min-max normalization and Z-score normalization.

[0052] In some embodiments, the normalization process can be implemented by the following formula: Among them, x is the data of the sensor, x ′The data is normalized, min(x) and max(x) are the minimum and maximum values ​​in the data x, respectively. After normalization, the data will be mapped to the range of [0, 1], which helps improve the training efficiency and stability of the model.

[0053] The data preprocessing module is closely integrated with the aforementioned data acquisition module. While the data acquisition module provides data from various sensor types, the data preprocessing module is responsible for standardizing and cleaning this raw data to ensure that it meets the input requirements of the deep learning model. This processed data is then transmitted to the deep learning prediction module for further predictive modeling.

[0054] In some embodiments, data preprocessing can be enhanced by combining other technical means. For example, statistical analysis techniques can be used to identify trends and periodicity in the data, further improving data usability. Furthermore, with the development of big data and artificial intelligence technologies, more intelligent algorithms can be integrated into the data preprocessing process to achieve automated data cleaning and processing, further improving the accuracy and efficiency of the system.

[0055] S3. Deep learning prediction model construction: A deep learning model combining long short-term memory neural network and convolutional neural network is used to train preprocessed multi-class sensor data and build a building settlement prediction model; In building settlement prediction methods that integrate multi-source data and artificial intelligence, the construction of a deep learning prediction model is a key step in achieving efficient settlement prediction. By combining long-short-term memory neural networks (LSTMs) and convolutional neural networks (CNNs), features can be effectively extracted in both spatial and temporal dimensions, thereby constructing an accurate building settlement prediction model. This model not only reflects building settlement trends but also provides effective predictions of future settlement behavior.

[0056] In this embodiment, the construction of the deep learning prediction model includes the following key steps: first, the time dependency of the data is modeled using the LSTM network to extract the long-term dependency features in the settlement data; second, the spatial features in the sensor data are extracted through CNN to analyze the relationship between different sensors; finally, the output results of LSTM and CNN are combined to construct a comprehensive settlement prediction model to predict the future trend of building settlement.

[0057] Long short-term memory (LSTM) neural networks are neural network models that can capture long-term dependencies in time series data. In building settlement prediction tasks, LSTM networks can effectively capture the temporal dynamics of building settlement data, especially when modeling periodic settlement changes.

[0058] Specifically, an LSTM network consists of multiple stacked LSTM layers, each of which controls the flow of information through a forget gate, an input gate, and an output gate. Through this mechanism, LSTM can avoid the vanishing gradient problem in traditional recurrent neural networks (RNNs), thereby better learning the dependencies of long time series.

[0059] In this embodiment, the LSTM network is used to model time series data of building settlement. For example, assuming that sensor A collects settlement data of a building at different time points, the basic operation of the LSTM network can be described by the following formula: f t =σ(W f ·[h t-1 ,x t ]+b f ); i t =σ(W i ·[h t-1 ,x t ]+b i ); o t =σ(W o ·[h t-1 ,x t ]+b o ); h t =o t *tanh(C t ); Among them, f t 、i t 、o t are the activation values ​​of the forget gate, input gate, and output gate respectively, C t is the cell state, h t is the hidden state, x t is the input data at the current moment, W f ,W i ,W C ,W o , is the weight matrix, b f ,b i ,b C ,b o , is the bias term, σ is the Sigmoid activation function, and tanh is the hyperbolic tangent activation function.

[0060] In this way, LSTM is able to learn the long-term dependencies in building settlement data and extract the trend characteristics of settlement in the model.

[0061] Convolutional neural networks (CNNs) are neural network models that excel at extracting spatial features. In building settlement prediction, CNNs are primarily used to analyze the spatial relationships between different sensor data. Through convolution operations, CNNs can automatically extract useful spatial features from multidimensional input data, which is crucial for analyzing differences in building settlement at different locations.

[0062] In this embodiment, CNN processes multi-source sensor data by stacking multiple convolutional layers and pooling layers. Specifically, each convolutional layer processes the data locally using a convolution kernel, gradually extracting features from local to global perspectives.

[0063] Assuming that the data matrix of a certain sensor is X, the convolution operation can be expressed as: Y=X*K+b; Among them, Y is the convolution result, K is the convolution kernel, b is the bias term, and * represents the convolution operation.

[0064] By stacking multiple convolutional layers, CNN can effectively extract spatial patterns in building settlement data and reduce data dimensions through pooling operations to avoid overfitting.

[0065] In this example, a combination of LSTM and CNN is used to build a comprehensive settlement prediction model. Specifically, the CNN first extracts spatial features from the input data, which are then passed as input to the LSTM network. The LSTM network further learns the temporal dependencies in the data and ultimately outputs a predicted value for building settlement.

[0066] This combination can fully utilize the advantages of LSTM in time series modeling, while also leveraging the capabilities of CNN in spatial feature extraction to fully capture the dynamic process of building settlement.

[0067] In some embodiments, the outputs of the combined models can be fused by weighted summation. For example, assuming that the outputs of LSTM and CNN are h LSTM and h CNN , the weighted summation can be performed by the following formula: h final =α·h LSTM +(1-α)·h CNN ; Among them, α is the weighting coefficient, which can be adjusted according to actual needs, h final The final output is obtained by weighted fusion of the results of LSTM and CNN models. LSTM is the output obtained by LSTM (Long Short-Term Memory Network), h CNN is the output obtained by CNN (Convolutional Neural Network).

[0068] In this embodiment, the constructed deep learning prediction model is trained through supervised learning. During the training process, a gradient descent algorithm (such as the Adam optimizer) is used to minimize the model's loss function. The loss function generally uses the mean square error (MSE) to measure the difference between the predicted result and the true value: Among them, y i is the true settlement value, is the settlement value predicted by the model, n is the number of samples, and L is the loss function value, which represents the error between the predicted result and the actual value.

[0069] The back-propagation algorithm updates the weights in the network, and the model gradually approaches the optimal prediction effect. Hyperparameters during training (such as learning rate and batch size) can be tuned through methods such as cross-validation.

[0070] The deep learning prediction model building module is closely connected to the data preprocessing module. This module cleans, denoises, and fills in missing values ​​for multiple sensor data types to ensure input data quality and consistency, providing reliable training data for the deep learning prediction model. Furthermore, the standardized data after data preprocessing facilitates convergence of LSTM and CNN models, improving training efficiency.

[0071] As technology advances, the structure of deep learning models can be further optimized. For example, the Convolutional LSTM (ConvLSTM) network can be introduced, combining the strengths of convolutional neural networks with those of LSTM networks to further enhance the model's ability to model spatiotemporal features. Furthermore, with the collection of larger, multi-source data, the scale and complexity of the model can be gradually expanded, further improving prediction accuracy.

[0072] S4. Model optimization and dynamic update: Using incremental learning to continuously optimize the prediction model based on newly collected multimodal data; In the practical application of building settlement prediction models, data distribution may change over time, leading to a decline in model performance. Therefore, an incremental learning approach is employed to continuously optimize the prediction model based on newly acquired multimodal data. This approach adapts to the dynamically changing data environment and ensures the stability and reliability of the prediction model in the long term. This optimization process not only involves the introduction of new data but also involves the rational consideration of historical data to avoid catastrophic forgetting. The implementation of incremental learning must be tailored to the specific scenario of building settlement prediction, taking into account factors such as data update frequency, model complexity, and computing resources to achieve optimal results.

[0073] In this embodiment, an incremental learning approach is used to dynamically optimize the settlement prediction model, enabling the model to adapt to newly acquired multimodal data while maintaining existing knowledge. Specifically, the model optimization process includes steps such as data acquisition, data preprocessing, incremental training, and model updating and verification.

[0074] Typically, building settlement monitoring data comes from a variety of sensors, such as foundation settlement sensors, tilt sensors, and accelerometers. These data have significant variations in sampling frequency and noise levels. Therefore, during the data collection phase, data fusion methods are necessary to ensure the temporal and spatial consistency of the data. Alternatively, a sliding time window strategy can be employed to batch new data according to time intervals, ensuring temporal continuity during incremental training.

[0075] Specifically, let X t represents the multimodal sensor data set collected at time t, X t ={x t,1 ,x t,2 ,...,x t,N}, where x t,i represents the data collected by sensor i at time t, and N is the number of sensors. To align the data, linear interpolation or Kalman filtering is used to fill missing values ​​so that the data meets the requirements of the input model.

[0076] In one possible implementation, incremental learning mainly adopts a strategy based on weight adjustment to avoid retraining the entire model and reduce computational overhead. Assume that the parameter set of the original model is θ old , the updated gradient after new data training is Δθ, then the new model parameters can be expressed as: θ new =θ old +ηΔθ; Among them, η is the learning rate, which determines the influence of new data on model parameters, θ new is the updated model parameter, θ old is the current model parameter, Δθ is the gradient of the model parameter, indicating the direction and magnitude of the parameter update. In order to prevent the model from overfitting, a regularization term can be introduced during the training process, such as elastic network regularization: Among them, λ1 and λ2 control the influence weights of L1 and L2 regularization respectively to ensure the stability of model update, L reg is a regularization term that aims to prevent the model from overfitting by penalizing excessive values ​​of model parameters, θ j is the jth component of the model parameters.

[0077] In some embodiments, incremental training can also be combined with a replay mechanism, that is, mixing new data with some historical data during training to reduce the risk of catastrophic forgetting. new , the historical data set is D old , the training data set can be expressed as: D train =D new ∪Sample(D old ,k); Among them, Sample(D old ,k) represents randomly sampling k samples from the historical data set to maintain the balance of data distribution, D train is a new training dataset, which contains the latest data and samples extracted from the old dataset. new is a new data set, representing the newly acquired sample data, D old It is a historical data set, representing the data that existed before.

[0078] In one possible implementation, the model update adopts the knowledge distillation method, that is, the prediction results of the new model are used to approximate the output of the old model to ensure the continuity of the model. Specifically, the output of the old model is defined as y old , the output of the new model is y new , then the knowledge distillation loss function can be expressed as: Where τ is the temperature parameter, and Represents the predicted probability distribution of the old model and the new model for category i, L KD is the knowledge distillation loss function, which is used to measure the difference between the predictions of the old model and the new model.

[0079] To ensure the performance of the updated model, the validation phase uses cross-validation to evaluate the model's generalization ability. In some embodiments, rolling validation can be used, using the most recent data as the validation set to simulate actual prediction scenarios to evaluate the adaptability of the incremental learning model.

[0080] The incremental learning optimization module in this embodiment is closely integrated with the data acquisition and deep learning prediction modules. First, this module relies on standardized data provided by the data acquisition and preprocessing modules to ensure the quality of incremental training data. Second, this optimization process is combined with parameter updates of the deep learning prediction model during the training phase, enabling the model to continuously improve its prediction accuracy during the continuous learning process. Furthermore, the module's output model will serve as a core component of the subsequent building settlement monitoring system, providing more accurate settlement prediction results.

[0081] In some embodiments, federated learning methods can be incorporated to enable monitoring data from different regions to share model parameters without directly transmitting data, thereby improving the efficiency of incremental learning while protecting data privacy. Furthermore, reinforcement learning mechanisms can be introduced to enable the model to dynamically adjust its learning strategy as it learns new data, improving training efficiency and adaptability.

[0082] S5. Intelligent early warning: Dynamically calculate the settlement risk threshold based on the prediction results. When the predicted value reaches the set safety warning range, different levels of early warning signals are triggered to provide settlement risk prompts and decision support; In the building settlement prediction and monitoring system, the intelligent early warning module dynamically calculates settlement risk thresholds based on the output of the prediction model and adjusts the safety warning range based on actual settlement trends. When the predicted value reaches the set warning range, different levels of early warning signals are triggered to provide settlement risk warnings and decision support. This early warning mechanism not only identifies abnormal settlement in the short term but also analyzes the overall stability of the building based on long-term trends, thereby improving the safety and reliability of the monitoring system.

[0083] In this embodiment, the intelligent early warning module mainly includes multiple links such as risk threshold calculation, warning range setting, early warning signal triggering, and decision support. This module is closely connected with the prediction model and data analysis module to ensure the accuracy and timeliness of early warning results.

[0084] Generally, building settlement safety thresholds are dynamically adjusted based on engineering design standards, historical monitoring data, and predictive model output. The calculation of risk thresholds involves both static safety standards and dynamic risk assessment. Static safety standards are typically determined by engineering specifications. For example, in some scenarios, cumulative settlement exceeding a certain value is considered abnormal. Dynamic risk assessment, on the other hand, combines actual monitoring data with statistical analysis and machine learning methods to calculate appropriate risk thresholds.

[0085] In one possible implementation, the sliding window method can be used to calculate the short-term settlement change rate to determine whether there is an abnormal change trend. Suppose the settlement prediction value of a building at time t is S t, the sliding window size is w, then the short-term sedimentation rate can be expressed as: Among them, V t Represents the short-term settlement rate. If its value exceeds the mean of the historical statistical range plus the set standard deviation multiple, it can be determined that the current settlement trend is abnormal. t To express the settlement value at time t, S t-w To express the settlement value at time tw, w is the time window for calculating the rate.

[0086] Alternatively, a Bayesian dynamic updating strategy can be used to continuously revise the risk threshold based on newly acquired data, making it more relevant to the current building settlement state. In this approach, the weight of historical data is gradually decayed to ensure that the latest data has a greater impact on the threshold adjustment.

[0087] Specifically, the setting of early warning alert zones requires comprehensive consideration of building structural characteristics, geological environmental factors, and long-term monitoring data. Alert zones can be divided into multiple levels, such as normal (Green), concern (Yellow), warning (Orange), and severe (Red). Each level corresponds to a different settlement threshold and requires different treatment measures.

[0088] In some embodiments, the quantile method (QuantileMethod) can be used to set the warning range, that is, to calculate the settlement values ​​corresponding to different quantiles based on historical data. 10 , Q 50 and Q 90 Represent the 10th, 50th, and 90th percentiles of historical data respectively, and the warning threshold can be defined as follows: Normal range: sedimentation value is less than Q 50 .

[0089] Focus: Q 50 ≤S t 90 .

[0090] Warning range: Q 90 ≤S t 95 .

[0091] Severity range: S t ≥Q 95 .

[0092] Among them, S t is the current predicted settlement value, Q 95 is the 95% quantile, corresponding to the extreme case in historical data, Q 50 ​​is 50% of the historical data, that is, the median settlement value of the historical data, Q 90 It is the 90th percentile of the historical data, that is, the maximum value of the settlement value below 90% in the historical data.

[0093] As a possible implementation approach, fuzzy logic methods can be combined to dynamically adjust the warning ranges of different levels based on environmental factors, load changes, and other conditions. For example, when rainfall increases, it may be necessary to lower the warning thresholds at all levels to increase warning sensitivity.

[0094] In one possible implementation, when the predicted value enters the set warning range, the intelligent early warning system automatically triggers the corresponding warning signal and takes different levels of response measures. The triggering mechanism of the warning signal can be based on the principle of temporal continuity. That is, if the predicted value reaches the warning threshold at a certain moment but then returns to normal, the warning will not be triggered. Only when the predicted value exceeds the threshold multiple times in a row will a formal alarm be triggered to reduce false alarms.

[0095] As an option, early warning signals can be presented in a multimodal manner, including visual alarms (such as flashing warning lights), SMS notifications, and automatically generated risk reports. For example, when the predicted settlement value reaches the critical range, the system can automatically send an alarm SMS to the management personnel, along with a detailed settlement trend analysis report.

[0096] In some embodiments, an adaptive alarm threshold adjustment strategy can be introduced. This means that during system operation, if the warning threshold is found to be oversensitive or undersensitive, its triggering conditions can be dynamically adjusted. For example, if a large number of false alarms occur within a short period of time, the time window for triggering the warning can be appropriately increased.

[0097] In this embodiment, the intelligent early warning system not only provides subsidence alerts but also offers decision support to managers. Based on historical data, real-time monitoring data, and forecast results, the decision support system generates risk analysis reports and proposes appropriate measures.

[0098] Alternatively, a Bayesian decision network can be used to calculate the optimal choice among different treatment options. For example, if settlement reaches the warning range, the system can evaluate the risk-benefit ratio of "support and reinforcement" versus "continue observation" and make a recommended decision.

[0099] Specifically, the system can calculate the expected value of risk loss under different intervention measures, such as: E Loss =P failure ×C failure +P repair ×C repair ; Among them, P failureis the probability that the building settlement exceeds the critical value, C failure is the loss cost of structural damage caused by settlement, P repair is the probability of taking repair measures, C repair is the repair cost, E Loss is the total loss, which represents the loss of the building caused by excessive settlement, including failure loss and repair cost.

[0100] In some embodiments, the decision support system can also be combined with a geographic information system (GIS) to visually present the spatial distribution of building settlement and provide annotations of high-risk areas so that managers can make accurate decisions.

[0101] The intelligent early warning module in this embodiment is closely integrated with the prediction model and incremental learning module. First, the module relies on the settlement trend analysis results of the prediction model to calculate the risk threshold and set the warning range. Second, the module incorporates incremental learning methods to continuously optimize the warning threshold to adapt it to the dynamic changes in building settlement. Furthermore, the intelligent early warning results can be further fed back to the incremental learning module to adjust model parameters and improve prediction accuracy.

[0102] In some implementations, artificial intelligence algorithms, such as deep reinforcement learning, can be incorporated to enable the early warning system to automatically optimize its strategies based on historical data. Furthermore, swarm intelligence methods can be employed to combine settlement data from multiple buildings to establish a global risk assessment system, enhancing the overall intelligence of the monitoring system.

[0103] The multi-source data-artificial intelligence fusion building correction and settlement prediction system described below and the multi-source data-artificial intelligence fusion building correction and settlement prediction method described above can be referenced to each other.

[0104] Please see the attached Figure 2 The present invention also provides a building correction settlement prediction system integrating multi-source data and artificial intelligence, including: The data acquisition module is used to collect multi-source data related to building settlement, including data from level instruments, total stations, groundwater level sensors, and soil pressure sensors, and transmit the collected data to the data processing module; The data processing module is connected to the data acquisition module and is used to receive multi-source data and clean, denoise, fill missing values, normalize and time-align the data to form standardized data input, and transmit the processed data to the deep learning prediction module; The deep learning prediction module is connected to the data processing module and is used to receive standardized data input and build a settlement prediction model based on long short-term memory neural network and convolutional neural network to predict the future settlement trend of buildings. The prediction results are then transmitted to the model optimization module and the intelligent early warning module; The model optimization module is connected to the deep learning prediction module to receive the prediction results and continuously optimize the deep learning prediction model using incremental learning, adaptive adjustment and reinforcement learning mechanisms to adapt it to long-term sedimentation changes. The optimized model parameters are fed back to the deep learning prediction module to update the prediction model. The intelligent early warning module is connected to the deep learning prediction module and the user interaction module respectively. It is used to receive the settlement prediction results and calculate the dynamic early warning threshold based on historical data. When the settlement prediction result exceeds the warning threshold, it triggers an early warning signal according to different risk levels and sends an early warning information to the user interaction module. The user interaction module is connected to the intelligent early warning module to receive early warning information and provide users with real-time monitoring of settlement status, historical data query and early warning feedback functions. Users can adjust the early warning threshold or input additional information through the user interaction module. The input information can be transmitted to the intelligent early warning module to optimize the early warning mechanism.

[0105] The system of this embodiment can be used to execute the above method embodiments, and its principles and technical effects are similar, so they will not be repeated here.

[0106] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A building correction settlement prediction method based on multi-source data and artificial intelligence fusion, characterized by: The following steps are involved: Multi-source data acquisition: Acquire multiple types of sensor data related to building settlement, including levels, total stations, groundwater level sensors, and soil pressure sensors, to form multimodal data input; Data preprocessing: performing data cleaning, missing value filling, denoising, time alignment, and normalization on the multi-category sensor data to obtain high-quality standardized data; Deep learning prediction model construction: A deep learning model combining long short-term memory neural networks and convolutional neural networks is used to train preprocessed multi-class sensor data and build a building settlement prediction model; Model optimization and dynamic updating: Using incremental learning to continuously optimize the prediction model based on newly collected multimodal data; Intelligent early warning: Dynamically calculate the settlement risk threshold based on the prediction results. When the predicted value reaches the set safety warning range, different levels of early warning signals are triggered to provide settlement risk prompts and decision support.

2. The building correction settlement prediction method based on multi-source data and artificial intelligence fusion according to claim 1 is characterized in that: The data preprocessing includes: Missing value filling: fill in missing data through time series interpolation method; Denoising: Use signal filtering methods to filter out random noise in the data; Time alignment: Align the data using interpolation methods based on the data sampling time intervals of different sensors; Normalization processing: Normalize the data of each sensor to keep its value range consistent.

3. The building correction settlement prediction method based on multi-source data and artificial intelligence fusion according to claim 1 is characterized in that: The deep learning prediction model construction includes: Long short-term memory neural network is used for time series modeling to extract the long-term dependencies of building settlement data; Use convolutional neural networks to extract spatial features and analyze the correlation between different sensor data; The outputs of the long short-term memory neural network and the convolutional neural network are integrated to form a comprehensive settlement trend prediction model.

4. The building correction settlement prediction method based on multi-source data and artificial intelligence fusion according to claim 1 is characterized in that: The model optimization and dynamic update include: Dynamic optimization based on incremental learning automatically adjusts model parameters when new data is input, so that the prediction model can adapt to the new settlement environment; Through the adaptive adjustment mechanism, the learning weights of the model are automatically corrected according to the changes in the building settlement trend; A feedback learning mechanism is used to continuously optimize the model structure and parameters based on historical prediction errors.

5. The building correction settlement prediction method based on multi-source data and artificial intelligence fusion according to claim 1 is characterized in that: The intelligent early warning includes: Calculate dynamic subsidence risk thresholds based on historical data and adjust warning trigger conditions in real time; Early warning signal classification includes three levels: low risk, medium risk, and high risk, each corresponding to different safety alert measures; Early warning information is visualized, showing the building settlement status based on digital twin technology, to assist managers in making decisions.

6. The multi-source data-artificial intelligence fusion building correction settlement prediction method according to claim 1 is characterized in that: The dynamic update includes: Adaptive learning rate adjustment: Dynamically adjust the learning rate based on the prediction error generated by the model in real-time applications; Incremental learning technology: When new data arrives, the existing model is optimized using incremental learning methods, allowing the model to continuously adapt to new data changes without complete retraining; Performance feedback mechanism: Each feedback of new data will monitor the performance of the model, and the reduction of prediction error will guide the direction of model parameter adjustment.

7. The building correction settlement prediction method based on multi-source data and artificial intelligence fusion according to claim 1 is characterized in that: The long short-term memory neural network comprises: Multi-layer LSTM network structure: The long short-term memory neural network is composed of multiple stacked LSTM layers; Bidirectional LSTM design: uses a bidirectional long short-term memory network to consider both future and past information; Forget Gate Optimization: Based on the traditional LSTM network, the forget gate is optimized so that the model can flexibly adjust the memory length according to changes in input data and optimize the information retention process; Time-weighted mechanism: A time-weighted mechanism is introduced to dynamically adjust the model's predictive ability according to the weights of different time steps during model training.

8. The multi-source data-artificial intelligence fusion building correction settlement prediction method according to claim 1 is characterized in that: The convolutional neural network includes: Convolutional layer design: A multi-layer convolutional neural network structure is used to gradually extract local and global features from settlement data, helping the model identify the relationship between building settlement and changes in the external environment; Adaptive convolution kernel adjustment: By adaptively adjusting the size and step size of the convolution kernel, the network can automatically select the most appropriate convolution configuration based on the spatial distribution characteristics of different data; Combining pooling layer and convolution layer: Set the pooling layer after the convolution layer, using the maximum pooling or average pooling method to reduce the feature dimension and avoid overfitting; Combining convolutional networks with LSTM networks: Convolutional neural networks are combined with long short-term memory neural networks to simultaneously extract temporal and spatial features from data.

9. The multi-source data-artificial intelligence fusion building correction settlement prediction method according to claim 1 is characterized in that: The data cleaning includes: Missing data detection and filling: Statistical methods are used to detect and fill missing values ​​in sensor data; Outlier detection: Use Z-Score and IQR methods to detect outliers in the data and correct or delete them; Data denoising: Apply signal processing methods such as wavelet transform or Kalman filtering to filter out noise in the data; Deduplication of data: Deduplication of duplicated data that is likely to exist in sensor data is performed to avoid repeated use of data at the same time; Data time synchronization: For data from different sensors, time alignment technology is used to ensure that all data is analyzed within a unified time dimension to avoid errors caused by time differences.

10. A building rectification and settlement prediction system integrating multi-source data and artificial intelligence, according to any one of claims 1 to 9, characterized in that: include: A data acquisition module is used to collect multi-source data related to building settlement, including data from a level, a total station, a groundwater level sensor, and a soil pressure sensor, and transmit the collected data to a data processing module; A data processing module, connected to the data acquisition module, is used to receive the multi-source data, clean, denoise, fill missing values, normalize and time-align the data to form standardized data input, and transmit the processed data to the deep learning prediction module; The deep learning prediction module is connected to the data processing module and is used to receive standardized data input and build a settlement prediction model based on long short-term memory neural network and convolutional neural network to predict the future settlement trend of buildings. The prediction results are then transmitted to the model optimization module and the intelligent early warning module; A model optimization module is connected to the deep learning prediction module, and is used to receive the prediction results and continuously optimize the deep learning prediction model using incremental learning, adaptive adjustment and reinforcement learning mechanisms to adapt it to long-term sedimentation changes, and feed the optimized model parameters back to the deep learning prediction module to update the prediction model; The intelligent early warning module is connected to the deep learning prediction module and the user interaction module respectively. It is used to receive the settlement prediction results and calculate the dynamic early warning threshold based on historical data. When the settlement prediction result exceeds the warning threshold, it triggers an early warning signal according to different risk levels and sends an early warning information to the user interaction module. The user interaction module is connected to the intelligent early warning module, and is used to receive early warning information and provide users with real-time monitoring of settlement status, historical data query and early warning feedback functions. Users can adjust the early warning threshold or input additional information through the user interaction module. The input information can be transmitted to the intelligent early warning module to optimize the early warning mechanism.

Images