A method and system for long-term monitoring of building settlement

Abstract

The application relates to the technical field of building safety monitoring and discloses a building settlement long-term monitoring method and a monitoring system.The method comprises the following steps: collecting historical settlement data of each building in a building colony and resampling the historical settlement data into standard time sequences; extracting local trend features of each sequence, constructing a feature similarity network with buildings as nodes and trend matching degrees as edge weights, and automatically dividing out building colonies with similar settlement behaviors through community discovery; identifying settlement mutation points shared in each colony, fusing and screening to obtain global significant time nodes; and training a time sequence prediction model by using settlement amount data corresponding to the nodes and performing settlement prediction on a target building.The method identifies common settlement rules through group behavior analysis, overcomes the defect that single-point analysis is easily disturbed by noise, and improves the accuracy and robustness of long-term prediction.

Description

Technical Field

[0001] This invention relates to the field of building safety monitoring technology, specifically to a method and system for long-term monitoring of building settlement. Background Technology

[0002] In long-term settlement monitoring of building clusters, conventional techniques primarily rely on independent analysis of historical settlement data for individual buildings. These methods directly model and predict individual time series, neglecting the inherent correlations between the settlement behaviors of building clusters as a whole. Building settlement is the result of the combined effects of its structural characteristics and external environmental factors. The limitations of isolated analysis become particularly pronounced when the monitoring scope expands from individual buildings to entire groups. Simply grouping buildings statically based on geographical proximity or design drawings fails to accurately reflect the inherent patterns of dynamic settlement processes. Geographically adjacent buildings may exhibit drastically different settlement responses to the same external factors due to differences in foundation type, load, or underground conditions. This coarse-grained grouping based on static attributes leads to a mismatch between the cluster division results and actual settlement dynamics, failing to provide effective support for regional settlement mechanism analysis.

[0003] Identifying key event points from settlement data is a crucial preliminary step in building predictive models. Current techniques typically detect abrupt change points independently on the time series of individual buildings. This method is highly susceptible to interference from factors specific to each building. Abnormal fluctuations in the data resulting from these individual events can be misinterpreted as significant settlement abrupt changes. The resulting set of abrupt change points contains a large amount of non-universal, sporadic local information. The model struggles to focus on real, global settlement patterns that affect the entire region and are driven by common factors, severely limiting the model's accuracy and generalization ability in long-term predictions. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for long-term monitoring of building settlement, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a method for long-term monitoring of building settlement, the method comprising:

[0006] Collect historical settlement datasets for each building within the building complex. The historical settlement datasets consist of timestamps and corresponding settlement values.

[0007] The historical settlement dataset of each building is resampled at equal intervals to generate a standard time series, the time axis of which has uniformly distributed time nodes.

[0008] Calculate the local trend characteristics of the standard time series for each building, and extract the settlement change rate near each time node using a sliding window method;

[0009] Based on the local trend features of all buildings, a feature similarity network is constructed between buildings. The nodes of the feature similarity network represent buildings, and the edge weights represent the trend feature matching degree.

[0010] Community discovery is performed based on feature similarity networks, dividing buildings into multiple clusters with similar settlement behaviors;

[0011] For each community, the standard time series of buildings within the community are analyzed to identify the set of settlement abrupt change points shared by the community;

[0012] By integrating the set of sedimentation mutation points from all communities, globally significant time nodes are selected.

[0013] A time series prediction model is trained using the settlement values ​​corresponding to globally significant time nodes.

[0014] The settlement data of the building to be monitored is input into the time series prediction model to obtain the settlement prediction results.

[0015] Preferably, the step of resampling the historical settlement dataset of each building at equal intervals to generate a standard time series includes:

[0016] Set a resampling time interval, and align the timestamps in the historical settlement dataset according to the resampling time interval;

[0017] For each time point, the missing settlement value is calculated using linear interpolation, so that the settlement value of each building is defined at the same time point.

[0018] The interpolated settlement values ​​are standardized to remove the influence of dimensions and generate a standard time series.

[0019] Preferably, the calculation of the local trend characteristics of the standard time series for each building includes:

[0020] Define the size of the sliding window and extract a subsequence of settlement values ​​within the window, centered at each time point;

[0021] For each subsequence, perform linear fitting and obtain the slope of the fitted line as the local rate of change at that time point;

[0022] Arrange the local rates of change at all time points in chronological order to form a sequence of local trend characteristics.

[0023] Preferably, the feature similarity network constructed between buildings includes:

[0024] Calculate the dynamic time-normalized distance between the local trend feature sequences of each pair of buildings;

[0025] The dynamic time-normalized distance is converted into a similarity weight, and the similarity weight is negatively correlated with the distance value;

[0026] A fully connected feature similarity network is constructed using buildings as nodes and similarity weights as boundary values.

[0027] Preferably, the community detection based on the feature similarity network includes:

[0028] The modularity optimization algorithm is applied to segment the feature similarity network to maximize the edge weights within the community;

[0029] The community division is iteratively adjusted until the modularity index converges, resulting in a stable building cluster division.

[0030] Preferably, the set of shared sedimentation abrupt change points identified in the community includes:

[0031] Calculate the first-order difference sequence for the standard time series of each building in the community;

[0032] Detect the extreme points in each first-order difference sequence and use the time nodes corresponding to the extreme points as candidate mutation points;

[0033] The frequency of occurrence of all buildings in the community at each candidate mutation point is counted, and candidate mutation points with frequencies exceeding the threshold are included in the community-shared set of settlement mutation points.

[0034] Preferably, the filtering of globally significant time nodes includes:

[0035] Calculate the weight of time nodes in the set of settlement abrupt change points for each community, with the weights normalized based on the number of buildings within the community;

[0036] Merge the set of settlement mutation points of all communities and weight the repeated time points;

[0037] The top k time points with the highest total weights are selected as the global significance time points, where k is a preset integer.

[0038] Preferably, the trained time series prediction model includes:

[0039] The settlement values ​​of each building at the globally significant time points are organized into feature vectors;

[0040] A deep neural network structure is constructed using feature vectors as input and the final settlement of the building as output.

[0041] The parameters of the deep neural network are optimized using the backpropagation algorithm until the loss function stabilizes.

[0042] Preferably, the method for obtaining the predicted settlement amount includes:

[0043] Extract the settlement values ​​of the building to be monitored at globally significant time points to form the input feature vector;

[0044] Input the input feature vector into the trained time series prediction model;

[0045] The model output is the predicted settlement of the building to be monitored.

[0046] Preferably, the present invention also includes a long-term building settlement monitoring system, the system including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method described above.

[0047] Compared with the prior art, the beneficial effects of the present invention are:

[0048] By calculating the local trend characteristics of the settlement time series of each building and constructing a feature similarity network based on this, dynamic settlement behavior patterns are used as the basis for community division. This allows buildings with similar settlement evolution patterns to be automatically grouped together, without being absolutely limited by their geographical distance or static type. Grouping based on dynamic behavior similarity can more realistically reflect the consistency of buildings' responses to common external environmental factors. Subsequent analysis of each community can further reveal their shared settlement mechanisms, providing a more homogeneous and representative data foundation for model training and improving the accuracy of settlement prediction.

[0049] By first identifying a shared set of abrupt change points within each community with similar settlement behavior, and then using community consensus to filter out globally significant time nodes, noise caused by individual building-specific factors is eliminated. A time node is considered a significant event with regional influence only if it is abruptly observed in multiple buildings. This consensus-based screening mechanism enhances the robustness and representativeness of the identified key nodes. Training the model with data corresponding to these globally significant nodes allows the model to focus on learning widely existing settlement patterns triggered by common factors, reducing the interference of local special events, thereby improving the generalization ability of the prediction model and the reliability of its long-term trend prediction. Attached Figure Description

[0050] Figure 1 This is a schematic diagram illustrating the working principle of the long-term building settlement monitoring method described in this invention.

[0051] Figure 2A flowchart for generating standard time series using equal-interval resampling;

[0052] Figure 3 A flowchart for calculating local trend characteristics;

[0053] Figure 4 Feature similarity network and community partitioning graph;

[0054] Figure 5 A statistical chart of shared settlement mutation points in the community. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] Please see Figure 1 This invention provides a method for long-term monitoring of building settlement. The method includes: collecting historical settlement datasets for each building within a building cluster, the dataset consisting of timestamps and corresponding settlement values; resampling the historical settlement datasets of each building at equal intervals to generate a standard time series, ensuring a uniform distribution of time nodes on the time axis; calculating the local trend features of the standard time series for each building, where the feature extracts the settlement change rate near each time node using a sliding window method; constructing a feature similarity network between buildings based on the local trend features of all buildings, where nodes represent buildings and edge weights represent the trend feature matching degree; performing community detection based on the feature similarity network to divide the buildings into multiple clusters with similar settlement behaviors; analyzing the standard time series of buildings within each cluster to identify a set of shared settlement abrupt change points; fusing the settlement abrupt change point sets of all clusters to filter out globally significant time nodes; training a time series prediction model using the settlement values ​​corresponding to the globally significant time nodes; and inputting the settlement data of the building to be monitored into the time series prediction model to obtain the settlement prediction result.

[0057] Example 1: See Figure 2In practical implementation, the resampling step of the long-term building settlement monitoring method begins with setting the resampling time interval, a predefined fixed time unit used to unify the timeline of all historical settlement datasets. The choice of the resampling time interval depends on the actual monitoring scenario and data analysis needs. For example, in long-term monitoring, the resampling time interval may be set monthly or quarterly to balance data granularity and computational efficiency. The timestamps in the historical settlement dataset are aligned according to the resampling time interval. This means mapping the historical settlement data points of each building to a common set of time nodes that are evenly distributed along the timeline. The alignment process involves examining the timestamps of each historical settlement dataset and determining which timestamps are closest to the nodes defined by the resampling time interval, adjusting for mismatched timestamps to ensure consistency. In some embodiments, the resampling time interval can be set via user input or a configuration file, allowing for flexible adaptation to the monitoring cycles of different building clusters. Optionally, the resampling time interval can be automatically calculated based on the average sampling rate of historical data, for example, by analyzing the interval distribution between timestamps to select the most suitable resampling time interval. It is understandable that timestamp alignment is the basis for generating standard time series. It eliminates the bias caused by the inconsistency in the data collection time of different buildings and provides a unified time frame for subsequent analysis.

[0058] In practical implementation, for each time point, linear interpolation is used to calculate the missing settlement value. Linear interpolation is a simple numerical method that assumes that the settlement changes linearly between adjacent time points. The calculation process first identifies the historical settlement data points corresponding to each time point. If a time point does not have a directly corresponding settlement value, the closest timestamp before and after that time point and its settlement value are found. Then, the linear interpolation formula is applied to estimate the missing value based on the settlement values ​​and time differences between the preceding and following time points, ensuring that a defined settlement value exists at each time point. The implementation of linear interpolation needs to handle boundary cases, such as missing values ​​at the beginning or end of the time series. In such cases, extrapolation or default values ​​may be used to fill the missing values, but this invention prioritizes interpolation to maintain reliability. In some embodiments, linear interpolation can be combined with data quality checks, such as ignoring outliers or using robust interpolation variants, but the core assumption remains linear. Optionally, for cases with severe data loss, multiple interpolation or trend-based interpolation can be introduced, but this embodiment uses linear interpolation as the standard method. It is understandable that the use of linear interpolation ensures data continuity, reduces noise caused by uneven sampling, and is computationally efficient, making it suitable for large-scale datasets.

[0059] In practice, the interpolated settlement values ​​are standardized to eliminate the influence of dimensions and make the settlement values ​​of different buildings comparable. Standardization typically employs the Z-score method or the min-max scaling method. The Z-score method calculates the mean and standard deviation of the settlement values, converting each value into a standard score, resulting in a mean of zero and a standard deviation of one for the processed data. The min-max scaling method linearly transforms the settlement values ​​to a specified range, such as between zero and one, scaling based on the minimum and maximum values ​​of the dataset. Standardization requires first calculating statistics for the entire historical settlement dataset, such as the mean, standard deviation, minimum, or maximum value, and then applying the transformation formula to each settlement value. During standardization, it is crucial to ensure that the statistical calculations are based on the complete interpolated dataset to avoid bias, and that the processed data retains the original trend. In some embodiments, other methods such as decimal scaling standardization can be used, but this embodiment primarily uses Z-score and min-max scaling. Optionally, standardization can be performed in clusters, i.e., first dividing the buildings into clusters and then standardizing. However, this embodiment of the invention prioritizes global standardization to maintain consistency. It can be understood that standardization makes the settlement values ​​dimensionless, which facilitates subsequent feature extraction and model training, and avoids numerical instability caused by differences in magnitude.

[0060] In practice, the complete process of generating standard time series includes setting the resampling time interval, timestamp alignment, linear interpolation calculation, and standardization. These steps are performed sequentially to ensure data quality. After setting the resampling time interval, timestamp alignment is achieved programmatically, for example, by using a time series library to convert irregular data into equally spaced sequences. Linear interpolation is performed on the aligned time nodes, executing the interpolation algorithm for each missing point and verifying the reasonableness of the interpolation results, such as checking whether the interpolated values ​​are within a reasonable physical range. Standardization, as the final step, applies the selected method to generate the final standard time series, which is stored in the form of an array or list, with each element corresponding to the processed settlement value of a time node. During implementation, data verification may include visual checks or statistical tests to confirm that resampling and standardization have not introduced significant errors. In some embodiments, the step of generating standard time series can be parallelized to accelerate data preprocessing for large-scale building clusters. Optionally, the standard time series can be cached or stored in a database for reuse in subsequent steps. It can be understood that generating standard time series is a key preliminary step in the monitoring method, providing clean and consistent input data for calculating local trend characteristics.

[0061] In practice, the specific value of the resampling time interval needs to be determined based on the application scenario. For example, the resampling time interval might be set to days for short-term monitoring and years for long-term monitoring. Timestamp alignment algorithms may involve time parsing and rounding to ensure that each resampling time node accurately corresponds to a calendar time. The implementation details of linear interpolation include handling the assumption of equal time intervals; if the original data timestamps are not uniform, timestamps may need to be sorted before interpolation. During standardization, the Z-score method needs to calculate global statistics, which may introduce the risk of data leakage. Therefore, in implementation, it is necessary to ensure that the statistics are calculated only based on training data. The resampling step can support dynamically adjusted intervals, such as adaptively selecting the resampling time interval based on the data change rate. Optionally, standardization can be skipped if the data units are consistent, but this embodiment recommends always performing it to enhance robustness. It is understood that the goal of the entire resampling process is to generate high-quality time series and reduce the uncertainty of subsequent analysis.

[0062] In practical implementation, linear interpolation assumes smooth settlement changes, which holds true in most settlement monitoring scenarios but may not be applicable to abrupt events. Therefore, verification using domain knowledge is necessary during implementation. The choice of standardization processing affects the performance of subsequent algorithms. The Z-score method is suitable for normally distributed data, while min-max scaling is more robust to outliers. The optimal method can be selected experimentally during implementation. After generating standard time series, the data format is usually converted to a floating-point array, with time nodes serving as indices for subsequent sliding window operations. When setting the resampling time interval, computational resources and storage limitations must be considered. Too dense an interval increases the data volume, while too sparse an interval may lose details. Timestamp alignment algorithms may use linear search or binary search to match time nodes to improve efficiency. In programming implementations, linear interpolation often involves looping through each time node and calling the interpolation function to calculate missing values. The code implementation of standardization processing may use mathematical library functions.

[0063] In practice, the resampling time interval may be set based on the temporal resolution of historical data. For example, if the original data is in weekly units, the resampling time interval may remain weekly or be adjusted to monthly. When aligning timestamps, the average or earliest value may be taken for overlapping timestamps to simplify processing. Linear interpolation can be extended to piecewise linear interpolation to handle non-linear trends, but this embodiment maintains a simple linear form. During standardization, if the dataset is too large, statistics can be calculated in batches to avoid memory issues. The steps for generating the standard time series can be logged to track processing status and errors. Optionally, metadata such as processing parameters and timestamps can be appended to the standard time series.

[0064] Example 2: See Figure 3In practice, defining the sliding window size is the first step in calculating local trend characteristics. The sliding window size is an integer representing the number of consecutive time points used to extract a subsequence from the standard time series. The choice of sliding window size directly affects the accuracy of capturing local rates of change. For example, in a ten-year standard time series with monthly intervals, the sliding window size can be defined as twelve months to analyze annual-scale settlement trend changes. Setting the sliding window size requires a trade-off between the locality and stability of the trend. A sliding window size that is too small may be sensitive to data noise, while a sliding window size that is too large may over-smooth out short-term real changes. In some embodiments, the sliding window size can be preset based on the expected settlement cycle or structural characteristics of the building. For example, a larger sliding window size can be used for buildings on soft soil foundations to smooth out short-term fluctuations. Optionally, the sliding window size can also be optimized using data-driven methods, such as determining the most representative window scale by analyzing the autocorrelation function of the standard time series.

[0065] In practice, a subsequence of settlement values ​​is extracted from a window centered on each time node. This operation involves systematically extracting a series of fixed-length subsets from the standard time series. For each time node in the standard time series, the window is extended forward and backward by two time nodes, centered on that node, to form a complete subsequence. When a time node is too close to the start or end of the standard time series, the sliding window cannot extend symmetrically on both sides. In this case, boundary conditions need to be handled. A common method is to extract only a local window starting from the start or ending at the end, even though the subsequence length may be smaller than the preset sliding window size. The extraction operation is implemented through a programmed loop, traversing each index position of the standard time series and calculating the start and end indices according to the sliding window size, thereby extracting the corresponding array of settlement values. In some embodiments, a filling strategy can be used to handle boundaries when extracting subsequences, such as filling missing positions with the first or last value of the sequence. However, this embodiment prefers asymmetric extraction to maintain the authenticity of the data. Optionally, the subsequence extraction can be parallelized, that is, multiple time points can be processed simultaneously to improve computational efficiency, especially when the standard time series is very long.

[0066] In practice, a linear fit is performed on each subsequence, and the slope of the fitted line is taken as the local rate of change at that time point. The linear fit is performed using the least squares method. The least squares method finds a straight line that minimizes the sum of the squares of the vertical distances (residuals) from all data points in the subsequence to this line, thus best approximating the trend of the subsequence. The linear fit yields two parameters: the slope (local rate of change) and the intercept. The slope characterizes the average rate of change of the settlement value with time within the time range corresponding to the sliding window size. The slope can be calculated using the following formula:

[0067]

[0068] Where: symbol The slope of the fitted line represents the local rate of change; the symbol is... Indicates the actual number of time points contained in the current subsequence; symbol Indicates the first subsequence The relative time index of each time node relative to the starting point of the subsequence; symbol Indicates the first subsequence Standardized settlement values ​​corresponding to each time point.

[0069] In practice, the local rates of change at all time points are arranged in chronological order to form a local trend feature sequence. This sequence constitutes a compact representation of the trend dynamics of the original standard time series. The local trend feature sequence is a one-dimensional array of the same length as the original standard time series, with each element corresponding to a local rate of change value at a given time point. Arranging them in chronological order means strictly following the time point order of the standard time series, ensuring that the time axis of the local trend feature sequence is completely consistent with the original data. After forming the local trend feature sequence, post-processing operations can be performed, such as applying a smoothing filter to suppress any abnormal fluctuations that may exist in the local rate of change sequence. However, in this embodiment, the original fitting result is preserved to reflect instantaneous changes.

[0070] In practical implementation, the robustness of the linear fitting step is crucial. When there are outliers in the subsequences that significantly deviate from the trend, the results of ordinary least squares (OLS) may be greatly affected. To enhance robustness, robust linear fitting methods can be introduced, which can tolerate outliers to a certain extent. This embodiment preferentially uses the standard least squares method because it is computationally simple, efficient, and assumes that the standard time series has undergone prior resampling and standardization, resulting in high data quality. The linear fitting process needs to be executed independently for each subsequence, and the computational cost is proportional to the length of the standard time series and the sliding window size. Optimizing the algorithm implementation or utilizing matrix operations can improve efficiency. In some embodiments, intermediate calculation results, such as the sum of subsequences and sum of squares, can be cached to avoid redundant calculations.

[0071] In practical implementation, the size of the sliding window needs to be considered in relation to the specific application scenario. For example, for a slow, uniform settlement process, a larger sliding window size helps to identify long-term trends. When extracting subsequences, ensuring the correct relative order of time nodes is crucial; any misalignment will distort the linear fitting results. The least squares calculation for linear fitting involves floating-point operations, so attention must be paid to numerical precision, especially when the time index span of the subsequence is large or the settlement values ​​differ significantly. The formation of the local trend feature sequence is the final output of feature extraction, which will be used for subsequent similarity calculations between buildings. In some embodiments, multiple different sliding window sizes can be tried for the same building to generate multiple sets of local trend feature sequences to capture trend information at different time scales, but this embodiment uses a single preset sliding window size by default. Optionally, the local trend feature sequence can be visualized to intuitively determine the points of change in the settlement trend.

[0072] Example 3: In a specific implementation, calculating the dynamic time warping distance between the local trend feature sequences of each pair of buildings is the first step in constructing the feature similarity network. Dynamic time warping distance is an algorithm used to measure the similarity between two time series, capable of handling scaling and velocity changes along the time axis. The calculation of dynamic time warping distance does not require the two local trend feature sequences to be of the same length; it finds an optimal warping path that minimizes the cumulative distance between the two sequences. The calculation process requires constructing a distance matrix, where the rows and columns correspond to the data points of the two local trend feature sequences, and each element in the matrix stores the Euclidean distance or other distance metric between the corresponding points. The dynamic time warping algorithm fills this distance matrix using dynamic programming and backtracks to find the optimal path; the final cumulative path distance is the dynamic time warping distance. In some embodiments, to control computational complexity and avoid excessive distortion, a warping window constraint can be introduced to limit the maximum distance the warped path deviates from the diagonal. Optionally, an acceleration algorithm can be used to approximate the calculation of the dynamic time warping distance to handle longer local trend feature sequences.

[0073] In practice, the dynamic time-warped distance is converted into similarity weights. The similarity weights are negatively correlated with the distance values; that is, the smaller the dynamic time-warped distance, the more similar the sequences, and the larger the corresponding similarity weight. The transformation function needs to map the dynamic time-warped distance to a standardized weight range, such as between zero and one. A common transformation method is to use a Gaussian kernel function or an inverse proportional function. The specific transformation can be achieved through the following formula:

[0074]

[0075] Where: symbol Represents the similarity weight between building A and building B; symbol This represents the dynamic time-warped distance calculated between the local trend characteristic sequences of building A and building B; symbol Represents the natural exponential function; symbol It is a scaling parameter that controls how quickly the weights decay with distance. The choice of scale parameter affects the distribution of similarity weights. A larger scale parameter results in slower weight decay, while a smaller scale parameter makes weight differences more significant. After the similarity weights are calculated, extreme values ​​need to be checked and handled, for example, ensuring that all weights are non-negative and within a reasonable range.

[0076] In practical implementation, a fully connected feature similarity network is constructed using buildings as nodes and similarity weights as edge values. This feature similarity network is a weighted undirected graph structure. Each building represents a node in the network, and the total number of nodes equals the number of buildings in the building cluster. Between each pair of different nodes (buildings), there exists an edge with a weight calculated earlier, thus forming a fully connected network. The adjacency matrix of the feature similarity network is a symmetric matrix, with diagonal elements typically set to zero or ignored, indicating that the relationship between a node and itself is not considered. Constructing a fully connected network means that any two buildings are connected in the initial stage, even if the connection strength (similarity weight) may be low. This fully connected structure ensures that no potential similarity relationships are missed during subsequent community discovery. The data structure of the feature similarity network can be represented in computer memory using an adjacency list or adjacency matrix. For large-scale building clusters, sparse matrix representation can save storage space. In some embodiments, to reduce network complexity, threshold filtering can be applied after constructing the fully connected network to remove edges with similarity weights below a certain threshold, forming a sparse network. Optionally, node attribute information, such as building age and height, can be added to the feature similarity network, but this embodiment mainly relies on trend similarity for community discovery.

[0077] In practical implementation, a modularity optimization algorithm is applied to segment feature-similar networks, maximizing the edge weights within communities. Modularity is an indicator of the quality of network community partitioning. The modularity optimization algorithm explores network partitions that maximize modularity by iteratively changing the community affiliation of nodes. The modularity value measures the concentration of edge weights within a community relative to the expected weights in a random network; a high modularity value indicates tight connections within communities and sparse connections between communities. The first step of the Louvain method is local optimization: initializing each node as an independent community, then attempting to move each node to the community of its neighboring nodes, calculating the modularity gain from the move, and accepting moves that improve modularity. The second step is network cohesion: shrinking each community partitioned in the first step into a new supernode, where the edge weights between supernodes are the sum of the weights of all edges between the original communities. These two steps are iterated until the modularity no longer significantly improves. The implementation of the modularity optimization algorithm requires defining a formula for calculating modularity, which considers the total edge weights and node degree distribution of the network. In some embodiments, different modularity optimization algorithms can be used, such as feature vector-based methods or label propagation algorithms, but the Louvain method is widely adopted due to its efficiency and ability to handle large-scale networks. Optionally, the modularity optimization process can set a resolution parameter to control the size of the discovered communities; a smaller resolution parameter tends to discover larger and fewer communities. It can be understood that modularity optimization algorithms can automatically identify building clusters with similar settlement behavior from feature similarity networks without pre-specifying the number of clusters.

[0078] In practice, the community division is iteratively adjusted until the modularity index converges, resulting in a stable building cluster division. Convergence means that the increase in the modularity value is below a preset tolerance threshold. The iterative process typically involves running the modularity optimization algorithm multiple times, as the algorithm may get stuck in local optima due to initial conditions. To obtain stable and robust division results, the algorithm can be randomly initialized multiple times, and the division result with the highest modularity can be selected, or a consensus clustering method can be used to integrate the results of multiple runs. A stable building cluster division result means that each building is assigned to a community, and buildings within the same community have high similarity in their local trend feature sequences. The division result can be represented as a list or dictionary, where the key is the building identifier and the value is the label of its community. After obtaining the cluster division, the results can be evaluated, for example, by checking whether the average similarity within a community is significantly higher than the average similarity between communities.

[0079] See Figure 4This diagram illustrates the topology and community segmentation results of a network based on the similarity of building settlement characteristics. Each node represents a building, and the lines connecting nodes indicate the similarity in settlement trends. The thickness and transparency of the lines reflect the strength of this similarity. Different colored node groups represent building clusters with similar settlement behavior, automatically identified by the algorithm. Buildings of the same color exhibit a high degree of consistency in their settlement patterns. The network structure clearly demonstrates the natural grouping characteristics within building clusters. Closely connected node groups indicate that these buildings are affected by similar geological conditions or external environmental factors, while sparse connections between clusters reflect the differences in settlement characteristics between different clusters. This data-driven community discovery method provides an intuitive visualization tool for understanding the spatial distribution patterns of building cluster settlement behavior.

[0080] Example 4: In a specific implementation, a first-order difference sequence is calculated for the standard time series of each building within the community. The first-order difference sequence is obtained by subtracting the settlement value at each time point in the standard time series from the settlement value at the previous time point, resulting in a new sequence one time shorter than the original sequence. The purpose of calculating the first-order difference sequence is to capture the instantaneous changes in settlement values ​​between consecutive time points. A positive first-order difference value indicates an increase in settlement, a negative value indicates a decrease in settlement, and a zero value indicates no change. The calculation of the first-order difference sequence can be implemented through a programmed loop, traversing the standard time series from the second time point to the last time point, calculating the difference between the current value and the previous value at each position. Before the difference calculation, it is necessary to ensure that the time points of the standard time series are strictly arranged in chronological order and that there are no missing values; otherwise, the difference results will be incorrect. In some embodiments, higher-order difference sequences can be calculated to capture acceleration changes, but this example focuses on the first-order difference because it is simple and can effectively identify abrupt changes. Optionally, the first-order difference sequence can be smoothed to suppress noise, but this embodiment prefers to use the original difference values ​​to preserve the abrupt signal.

[0081] In practical implementation, extreme points are detected in each first-order difference sequence, and the time nodes corresponding to these extreme points are used as candidate mutation points. An extreme point is defined as a value in the first-order difference sequence whose absolute value is significantly greater than that of surrounding points. Extreme point detection can employ threshold-based methods, such as setting an absolute threshold and considering points whose absolute first-order difference value exceeds this threshold as extreme points; or statistical methods can be used, such as considering points that deviate from the mean by a certain standard deviation as extreme points. The time node corresponding to an extreme point refers to the next time node in the original standard time series corresponding to that difference value, because the first-order difference represents the change between adjacent time nodes. The detection process requires traversing each position in the first-order difference sequence, comparing the magnitude of the current difference value with that of neighboring values, and recording the timestamp of the extreme point. After extreme point detection, deduplication may be necessary, such as merging multiple consecutive extreme points into a single representative point; however, this embodiment retains all detected extreme points as initial candidates.

[0082] In practice, the frequency of occurrence of all buildings within the community at each candidate mutation point is counted. Candidate mutation points with frequencies exceeding a threshold are included in the community-shared set of settlement mutation points. Frequency calculation requires iterating through each building in the community, checking its candidate mutation point set, and counting how many buildings share each unique time point. The frequency is calculated as the ratio of the number of buildings sharing a particular time point to the total number of buildings in the community, or a direct count. A frequency threshold is set, for example, 0.5; only candidate mutation points with frequencies exceeding the threshold are considered community-shared. The community-shared set of settlement mutation points is a list of time points representing significant settlement change events experienced by a majority of buildings within the community. The statistical process can be implemented by constructing a hash table or using database grouping queries to efficiently calculate the occurrence count of each time point. The choice of frequency threshold affects the size of the set; a higher threshold produces a smaller but more consistent set, while a lower threshold is more inclusive. In some embodiments, the frequency threshold can be dynamically adjusted according to the community size; smaller communities use lower thresholds to avoid missed detections.

[0083] In practice, the weights of time points within the set of settlement abrupt change points for each community are calculated, with the weights normalized based on the number of buildings within the community. The purpose of weight calculation is to quantify the importance of each abrupt change point within the community and to ensure comparability between communities of different sizes. Weight calculation can be achieved using the following formula:

[0084]

[0085] Where: symbol Indicates time node In the community Weights in; symbols Indicates time node In the community The frequency of occurrence, i.e., the community How many buildings contain time nodes in the candidate mutation point set? ;symbol Represents a community Total number of buildings within the area. Weight The value ranges from zero to one; a larger value indicates a higher time point. In the community The more prominent the internal aspect.

[0086] In practice, the set of settlement mutation points from all communities is merged, and repeated time nodes are weighted and summed. The merging operation involves aggregating time nodes from the settlement mutation point sets of all communities into a single global set, while retaining the weight information of each time node within its original community. For repeated time nodes—the same time node contained in multiple community sets—its weight values ​​in different communities are added together to obtain the global weight sum for that time node. The weight summation reflects the cumulative significance of a time node across multiple communities; the more communities that share a time node, and the higher its weight within each community, the larger its global weight sum. The merging process can be achieved by traversing the mutation point sets of all communities, using a dictionary or mapping structure to record the cumulative weight of each time node. Identification of repeated time nodes is based on precise timestamp matching, requiring consistency in time node representation. After merging, each unique time node in the global set has a corresponding global weight sum value.

[0087] In practice, the top k time nodes with the highest total weights are selected as globally significant time nodes, where k is a preset integer. The selection process is based on sorting the global weights; the higher the global weight sum, the stronger the global significance of the time node. The parameter k determines the number of globally significant time nodes finally selected. The value of k can be set by expert experience, such as k=10, or determined based on the cumulative proportion of the total weights, such as selecting nodes with a cumulative weight sum of 80%. The sorting operation arranges the time nodes in the global set in descending order of global weight sum, and then takes the top k nodes as the output. These globally significant time nodes represent the time points when the most shared and influential subsidence mutation events occur in the entire building cluster. The selection of the k value needs to be balanced between coverage and accuracy; a k value that is too large will introduce noise, while a k value that is too small may miss important events. The result of the selection process is an ordered list of time nodes, which serves as the feature time points for subsequent training of the prediction model. In practice, the mutation point detection and global screening processes can record intermediate results in a table for verification. See Table 1, which shows the frequency statistics of candidate mutation points for a hypothetical community A.

[0088] Table 1: Frequency statistics of candidate mutation points in community A

[0089] Time Node Number of buildings frequency 2020-01 8 0.80 2020-06 5 0.50 2021-03 6 0.60

[0090] The "Time Node" column represents the timestamp of the candidate mutation point; the "Number of Buildings Found" column indicates how many buildings in cluster A experienced a mutation at that time point; the "Frequency" column is the number of buildings found divided by the total number of buildings in cluster A (assumed to be 10); and the "Weight" column is the calculated normalized weight. Table 1 helps visualize the frequency distribution and weight calculation results. In practice, such statistics can be automatically generated through database queries or scripts. The data in Table 1 can be used to check the effectiveness of the threshold application; for example, if the frequency threshold is set to 0.5, points such as 2020-01 and 2021-03 would be included in the shared set.

[0091] See Figure 5 This paper presents the statistical characteristics and global significance analysis results of settlement abrupt change points in various building communities. The charts comprehensively illustrate the distribution of abrupt change points across different communities using three key indicators: the average significance weight of each abrupt change point, the weight of the most influential single event, and the distribution of the number of abrupt change points in each community. Different communities exhibit significant differences in abrupt change characteristics; some communities show frequent but low-intensity abrupt change patterns, while others show sparse but far-reaching significant abrupt change events. The labels at the top of the charts indicate the globally significant time points selected from all communities; these points represent the most shared and influential settlement abrupt change events across the entire building community. This abrupt change point analysis method based on group behavior can effectively identify regional common problems, providing important evidence for settlement prediction and risk management.

[0092] Example 5: In specific implementation, the settlement values ​​of each building at globally significant time nodes are organized into a feature vector. The dimension of the feature vector is equal to the number of globally significant time nodes, and each element corresponds to a standardized settlement value at a globally significant time node. The organization process requires traversing every building in the building cluster and querying its standard time series to extract the settlement observation values ​​at the selected globally significant time nodes. These observation values ​​are then arranged in chronological order to form a fixed-length numerical array, i.e., the feature vector. If a building is missing a settlement value at a certain globally significant time node, it needs to be filled in using interpolation methods, such as linear interpolation using values ​​from adjacent time nodes, to ensure the integrity of the feature vector. The organization of the feature vector compresses the settlement history data of each building into a multi-dimensional point, the dimension of which is determined by the number k of globally significant time nodes. This allows the settlement patterns of different buildings to be compared and modeled in a unified feature space. In some embodiments, the feature vectors can be further normalized, for example, by Z-score standardization of each feature dimension, to eliminate the influence of differences in the distribution of settlement values ​​at different time points.

[0093] In practical implementation, a deep neural network structure is constructed using feature vectors as input and the final settlement of the building as output. A deep neural network is a feedforward neural network containing multiple hidden layers. The number of neurons in the input layer is strictly equal to the dimension of the feature vector, i.e., the number of globally significant time points, k. Each neuron receives one element value from the feature vector. The output layer consists of one neuron, whose output value represents the model's prediction target, i.e., the final settlement of the building. This is a continuous value, therefore the problem is set as a regression task. The number of hidden layers and the number of neurons in each layer are hyperparameters that need to be set. For example, a deep neural network structure with two hidden layers can be constructed, with the first hidden layer containing 64 neurons and the second containing 32 neurons. Each neuron in the hidden layer performs a linearly weighted sum of its inputs, and then transforms it through a non-linear activation function. Commonly used activation functions are ReLU, which is recommended in this embodiment because it effectively alleviates the gradient vanishing problem. The layers in the deep neural network structure are fully connected, meaning that each neuron in each layer is connected to all neurons in the next layer, and each connection has a trainable weight parameter. The mathematical expression of a deep neural network structure can be viewed as a complex composite function that maps the input feature vector to the output predicted settlement value.

[0094] In practice, the backpropagation algorithm is used to optimize the parameters of the deep neural network until the loss function stabilizes. The backpropagation algorithm guides parameter updates by calculating the gradient of the loss function with respect to the parameters of each network layer. The optimization process requires defining a loss function to quantify the difference between the final settlement predicted by the deep neural network and the actual final settlement. For regression problems, mean squared error (MSE) is typically used as the loss function. The formula for calculating the MSE loss function is:

[0095]

[0096] Where: symbol Represents the value of the loss function; symbol Indicates the number of buildings in the current training batch; symbol Indicates the first The actual final settlement of a building; symbol This indicates that the deep neural network represents the first... The predicted final settlement of a building. The backpropagation algorithm starts from the output layer, calculates the gradient based on the loss function, and propagates the gradient layer by layer backward, using the chain rule to calculate the gradient of the weights and biases of each layer. Parameter updates use gradient descent or its variants, such as stochastic gradient descent, momentum descent, or the Adam optimizer, adjusting parameter values ​​based on the calculated gradients to minimize the loss function. The training process typically requires multiple iterations, each traversing the entire training dataset until the value of the loss function no longer decreases significantly on the validation set, indicating a stable state. During training, both training and validation losses need to be monitored to prevent overfitting.

[0097] In practice, the settlement values ​​of the building to be monitored at globally significant time points are extracted to form the input feature vector. This process is completely consistent with the process of organizing the feature vector during the training phase. The extraction of the input feature vector requires that the settlement data of the building to be monitored must include the observations corresponding to all globally significant time points. If any are missing, interpolation must be performed using the same method as in the training phase to ensure consistency between the input data and the data distribution during model training. The resulting input feature vector is a k-dimensional numerical array that encapsulates the settlement status information of the building to be monitored at historical key time points. The extraction operation can be performed in real time or in batches, depending on the needs of the monitoring system. The extraction results are organized into the input format desired by the model, typically a one-dimensional array or matrix.

[0098] In practice, the input feature vector is fed into a trained time-series prediction model, and the model output is the predicted settlement amount of the building to be monitored. The input process involves passing the prepared input feature vector data to a deep neural network model with optimized parameters. Internally, the model performs forward propagation calculations, transforming the input data through various layers of the deep neural network structure, ultimately producing a specific numerical value at the output layer. This value is the model's prediction of the final settlement amount of the building to be monitored. The prediction result is a continuous scalar, representing the final settlement amount that the building to be monitored may achieve based on its key historical settlement patterns, under the patterns learned by the model. The model can be applied in an offline system, periodically predicting the settlement of batches of buildings; or it can be integrated into an online monitoring platform to achieve near real-time settlement trend early warning.

[0099] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0100] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for long-term monitoring of building settlement, characterized in that, The method includes the following steps: Collect historical settlement datasets for each building within the building complex. The historical settlement datasets consist of timestamps and corresponding settlement values. The historical settlement dataset of each building is resampled at equal intervals to generate a standard time series, the time axis of which has uniformly distributed time nodes. Calculate the local trend characteristics of the standard time series for each building, and extract the settlement change rate near each time node using a sliding window method; Based on the local trend features of all buildings, a feature similarity network is constructed between buildings. The nodes of the feature similarity network represent buildings, and the edge weights represent the trend feature matching degree. Community discovery is performed based on feature similarity networks, dividing buildings into multiple clusters with similar settlement behaviors; For each community, the standard time series of buildings within the community are analyzed to identify the set of settlement abrupt change points shared by the community; By integrating the set of sedimentation mutation points from all communities, globally significant time nodes are selected. A time series prediction model is trained using the settlement values ​​corresponding to globally significant time nodes. The settlement data of the building to be monitored is input into the time series prediction model to obtain the settlement prediction results; The set of shared settlement abrupt change points identified in the community includes: Calculate the first-order difference sequence for the standard time series of each building in the community; Detect the extreme points in each first-order difference sequence and use the time nodes corresponding to the extreme points as candidate mutation points; The frequency of occurrence of all buildings in the community at each candidate mutation point is counted, and candidate mutation points with frequencies exceeding the threshold are included in the community-shared set of settlement mutation points. The time nodes for selecting globally significant data include: Calculate the weight of time nodes in the set of settlement abrupt change points for each community, with the weights normalized based on the number of buildings within the community; Merge the set of settlement mutation points of all communities and weight the repeated time points; The top k time points with the highest total weights are selected as the global significance time points, where k is a preset integer.

2. The method for long-term monitoring of building settlement according to claim 1, characterized in that, The process of resampling the historical settlement dataset for each building at equal intervals to generate a standard time series includes: Set a resampling time interval, and align the timestamps in the historical settlement dataset according to the resampling time interval; For each time point, the missing settlement value is calculated using linear interpolation, so that the settlement value of each building is defined at the same time point. The interpolated settlement values ​​are standardized to remove the influence of dimensions and generate a standard time series.

3. The method for long-term monitoring of building settlement according to claim 1, characterized in that, The calculation of the local trend characteristics of the standard time series for each building includes: Define the size of the sliding window and extract a subsequence of settlement values ​​within the window, centered at each time point; For each subsequence, perform linear fitting and obtain the slope of the fitted line as the local rate of change at that time point; Arrange the local rates of change at all time points in chronological order to form a sequence of local trend characteristics.

4. The method for long-term monitoring of building settlement according to claim 1, characterized in that, The feature similarity network constructed between buildings includes: Calculate the dynamic time-normalized distance between the local trend feature sequences of each pair of buildings; The dynamic time-normalized distance is converted into a similarity weight, and the similarity weight is negatively correlated with the distance value; A fully connected feature similarity network is constructed using buildings as nodes and similarity weights as boundary values.

5. The method for long-term monitoring of building settlement according to claim 1, characterized in that, The community detection based on feature similarity networks includes: The modularity optimization algorithm is applied to segment the feature similarity network to maximize the edge weights within the community; The community division is iteratively adjusted until the modularity index converges, resulting in a stable building cluster division.

6. The method for long-term monitoring of building settlement according to claim 1, characterized in that, The trained time series prediction model includes: The settlement values ​​of each building at the globally significant time points are organized into feature vectors; A deep neural network structure is constructed using feature vectors as input and the final settlement of the building as output. The parameters of the deep neural network are optimized using the backpropagation algorithm until the loss function stabilizes.

7. The method for long-term monitoring of building settlement according to claim 1, characterized in that, The obtained settlement prediction results include: Extract the settlement values ​​of the building to be monitored at globally significant time points to form the input feature vector; Input the input feature vector into the trained time series prediction model; The model output is the predicted settlement of the building to be monitored.

8. A long-term building settlement monitoring system, the system comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 7.

Images