A construction site safety monitoring and early warning method and system

Abstract

The application discloses a kind of construction site safety monitoring early warning method and system, belong to building safety monitoring technical field.The method includes: in response to the elastic wave signal collected by the distributed vibration sensor array deployed in construction site, construct the space-time tensor data field with time window as unit;According to the Lie algebra derivative between each tensor slice in the space-time tensor data field and the tensor slice of adjacent time window, generate the tangent vector manifold representing the evolution path of vibration source;The dimensionality reduction mapping of tangent vector manifold is carried out based on local retention projection, and the characteristic trajectory curve in low-dimensional embedding space is obtained;Identify the curvature singular point on characteristic trajectory curve, and mark the time stamp corresponding to curvature singular point as potential risk time;With potential risk time as center, intercept dangerous segment in elastic wave signal, and output safety warning level by calculating mutual information entropy matrix of multiple channel signals in dangerous segment.

Description

Technical Field

[0001] This invention relates to the field of building safety monitoring technology, specifically to a method and system for safety monitoring and early warning of building construction sites. Background Technology

[0002] Construction sites are complex environments where vibrations from deep foundation pit excavation, tower crane operation, and various heavy machinery operations intertwine, easily triggering safety accidents such as foundation pit collapse and structural instability. Existing safety monitoring and early warning methods largely rely on the time-domain or frequency-domain characteristic analysis of single or multiple vibration sensors, setting alarms based on thresholds for statistical quantities such as vibration amplitude and frequency. This approach simplifies vibration signals to a few scalar indicators, ignoring the rich geometric structure inherent in the multi-channel elastic wave signals captured by distributed sensor arrays in both space and time. The evolution of vibration sources is essentially a continuous deformation process in a high-dimensional state space. Existing technologies using Euclidean space feature representations struggle to characterize the inherent geometric continuity of this deformation, resulting in insufficient ability to identify dangerous vibration modes under strong noise backgrounds and multi-source aliasing conditions, leading to high false alarm and false negative rates. Furthermore, existing early warning methods often rely on fixed windows or threshold lines set manually based on experience when determining risk moments, failing to automatically capture abrupt changes in the vibration behavior's evolution trajectory. During construction, the transition of vibration modes from a stable state to a precursor to instability often manifests as a rapid, nonlinear migration. This migration corresponds to a dramatic change in the trajectory geometry on the intrinsic manifold of the observed data. The key to improving the real-time performance and accuracy of early warning lies in constructing a continuous path from distributed sensor data that reflects the non-Euclidean evolution of the vibration source, and automatically identifying singular inflection points representing abrupt state changes along this path. Summary of the Invention

[0003] This invention provides a method for safety monitoring and early warning of construction sites. It utilizes elastic wave signals collected by a distributed vibration sensor array and employs an analysis method combining Lie algebra derivatives on the manifold with local preservation projections to automatically identify abrupt changes in the vibration state evolution and output a safety warning level. This enables early detection and graded alarm of regional instability precursors at construction sites.

[0004] The objective of this invention can be achieved through the following technical solutions: This invention provides a method and system for safety monitoring and early warning of construction sites. By deploying a distributed vibration sensor array, it can highly sensitively sense weak elastic waves such as foundation vibration and mechanical impact during the construction process. The multi-channel signals are constructed as a spatiotemporal tensor data field on a time window. The dynamic migration characteristics of the vibration source are characterized by the evolution path of the tangent vector on the Lie group manifold, so as to realize early identification and graded early warning of risk moments.

[0005] The method specifically includes the following steps: At least four three-component accelerometers are deployed in a non-uniform spiral pattern around the foundation pit and tower crane foundation at the construction site, forming a distributed vibration sensing array. This array can simultaneously acquire elastic wave signals including longitudinal wave components, transverse wave components, and surface wave components. In response to the elastic wave signals acquired by the array, a spatiotemporal tensor data field is constructed using time windows as the basic unit. As a preferred embodiment of the invention, the elastic wave signals acquired by each sensor node within a single time window are arranged sequentially to form node vectors. Then, the node vectors of all nodes are stacked into a two-dimensional matrix according to spatial coordinates. After concatenating the two-dimensional matrices of three consecutive time windows along the time axis, and performing centering and amplitude normalization processing, a tensor field reflecting the spatiotemporal correlation characteristics of vibration is obtained. This process unifies multi-source, asynchronous vibration information into a structured tensor representation, which is beneficial for suppressing instantaneous noise interference and highlighting the inherent spatiotemporal structure of vibration events.

[0006] Based on the spatiotemporal tensor data field, tensor slices from the current time window and the previous adjacent time window are extracted, and the logarithmic mapping difference between them is calculated to obtain the tangent vector in the tangent space of the Lie group manifold. Preferably, the tangent vector is separated into rotational and scaling components through singular value decomposition. The rotational component is constructed by the product of the left and right singular matrices, and the scaling component is determined by the difference between the diagonal elements of the singular values ​​and the unit vector. The direct sum of the two constitutes the Lie algebraic derivative. Using the time window index as the independent variable, the Lie algebraic derivatives under each time window are concatenated in the direct product space of a special orthogonal group and a symmetric positive definite matrix to form a tangent vector manifold. This tangent vector manifold completely encodes the continuous evolution law of the vibration source in the rotational and scaling degrees of freedom, and provides a more accurate directional description of micro-variable processes such as construction impact and landslide creep.

[0007] A dimensionality reduction mapping based on local-preserving projection is performed on the tangent vector manifold to obtain feature trajectory curves in the low-dimensional embedding space. Preferably, the Euclidean distance between every two tangent vectors in the manifold is first calculated and an adjacency graph is constructed. Then, the local reconstruction weight coefficients are obtained by minimizing the weighted reconstruction error and a sparse weight matrix is ​​formed. The sparse weight matrix is ​​then subjected to generalized eigenvalue decomposition, and the eigenvector corresponding to the smallest non-zero eigenvalue is taken as the projection basis. Each tangent vector is projected onto this basis to obtain low-dimensional coordinate values, which are then connected in time sequence to form the feature trajectory curves. This dimensionality reduction mapping can both preserve the local neighborhood structure of the manifold and compress high-dimensional tangent vectors into intuitive low-dimensional trajectories, significantly reducing the complexity of subsequent calculations.

[0008] Multiple discrete points are sampled at equal arc length intervals on the characteristic trajectory curve, and the first and second derivatives at each point are calculated to obtain the curvature value. Preferably, the statistical characteristics of the curvature of all discrete points are calculated, and discrete points whose curvature values ​​exceed the sum of the mean and three times the standard deviation are identified as curvature singularities, and their corresponding timestamps are marked as potential risk moments. Through curvature singularity detection, geometric features representing abrupt changes in the vibration evolution path can be automatically captured, effectively eliminating common stable background vibration interference and reducing the false alarm rate.

[0009] Centered on a potential risk moment, a pre-defined first time period is extended forward and a pre-defined second time period is extended backward to extract the dangerous segment from the original elastic wave signal. The dangerous segment signal is then divided into multiple channels, and the mutual information entropy value between any two channels is calculated. Preferably, the mutual information entropy is obtained by first estimating the marginal probability distribution of a single signal and the joint probability distribution of two signals, and then taking the double integral of the logarithm of the ratio of the product of the joint distribution and the marginal distribution. All mutual information entropy values ​​are arranged into a symmetric mutual information entropy matrix, the matrix norm of this matrix is ​​calculated, and the safety warning level is determined and output based on the numerical interval to which the norm belongs. The mutual information entropy matrix can quantify the nonlinear correlation strength of multi-channel signals during the risk period, and the matrix norm comprehensively reflects the range and degree of hazard expansion, making the warning level sensitive to local hazards while also possessing global assessment capabilities.

[0010] As a preferred solution, after outputting the safety warning level, the system matches the corresponding alarm strategy from the warning rule base based on the level. This strategy includes the frequency of audible alarms, the frequency of light flashing, and a list of remote notification terminals. Simultaneously, it packages five trajectory points before and after the dangerous segment and the potential risk moment into a warning evidence package, which is sent to the cloud server and on-site management terminal via wireless network. The warning level is also displayed in real-time on the on-site monitoring screen. This response output mechanism constructs a complete information chain from risk identification to response, facilitating retrospective review and remote supervision.

[0011] This invention also includes a construction site safety monitoring and early warning system. The system includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the aforementioned construction site safety monitoring and early warning method. This system can be embedded in a field industrial control computer or edge computing node, receiving sensor array signals in real time and completing the entire process from tensor construction and manifold analysis to hierarchical early warning. It does not rely on long-distance transmission or a centralized computing center, significantly improving response time and on-site adaptability.

[0012] The beneficial effects of this invention are: Elastic wave signals within continuous time windows are organized into a spatiotemporal tensor data field. The logarithmic mapping difference between tensor slices of adjacent time windows is calculated to obtain the tangent vector in the tangent space of the Lie group manifold. This tangent vector is further decomposed into a direct sum of rotational and scaling components, serving as the Lie algebraic derivative. Using the time window index as the independent variable, the Lie algebraic derivatives corresponding to each time window are concatenated to form a tangent vector manifold. This tangent vector manifold resides in the direct product space of a special orthogonal group and a symmetric positive definite matrix, enabling the expression of the vibration source's evolution as a continuous path on the manifold. The rotational component records the dynamic changes in the vibration propagation direction in space, while the scaling component captures the diffusion and contraction of vibration energy in the medium. Compared to traditional Euclidean space eigenvector sequences, this representation strictly preserves the intrinsic structure of the vibration field in non-Euclidean geometry, allowing for a more separable description of the essential differences between different vibration modes in the tangent space. When abnormal load movement or local yielding occurs at the construction site, the evolution path of the vibration source is characterized by a disruption of continuity or abrupt change in direction on the tangential manifold, thus providing input features that are more sensitive to geometric changes for subsequent abrupt change detection.

[0013] When performing local-preserving projection dimensionality reduction mapping on a tangent vector manifold, the Euclidean distance between tangent vectors is calculated and an adjacency graph is constructed. A sparse weight matrix is ​​obtained by minimizing the weighted reconstruction error between each node and its neighboring nodes. Generalized eigenvalue decomposition is performed on the sparse weight matrix, and the eigenvectors corresponding to the smallest non-zero eigenvalues ​​are extracted as projection bases. The tangent vectors of each time window are projected onto the low-dimensional embedding space and connected in chronological order to form feature trajectory curves. Local-preserving projection preserves the local neighborhood structure of the manifold during dimensionality reduction, allowing the trajectory curves of the vibration source evolution path in the low-dimensional space to faithfully inherit the local geometric relationships on the high-dimensional manifold, avoiding the destruction of the nonlinear manifold structure by traditional global linear dimensionality reduction methods such as principal component analysis. The curvature values ​​of sampling points of equal arc length are calculated on the feature trajectory curves. By statistically analyzing the mean and standard deviation of the curvature, points with curvature exceeding the sum of the mean and three times the standard deviation are identified as curvature singularities, and their corresponding timestamps are marked as potential risk moments. This curvature singularity detection strategy is based entirely on the geometric features of the trajectory itself, without the need to pre-set thresholds or refer to historical baselines. It can adaptively detect the moment when the vibration mode undergoes a fundamental change in the complex construction vibration background, effectively reducing false alarms caused by persistent strong noise or slow trend changes. Attached Figure Description

[0014] The invention will now be further described with reference to the accompanying drawings.

[0015] Figure 1 This is a flowchart of a safety monitoring and early warning method for construction sites. Figure 2 This is a flowchart of the method for constructing the tangent vector manifold of distributed vibration sensing signals; Figure 3 This is a flowchart of feature trajectory curve generation based on tangent vector manifold learning; Figure 4 This is a flowchart for detecting singularities in the curvature of characteristic trajectory curves and determining potential risk moments; Figure 5 This is a schematic diagram of the structure of a safety monitoring and early warning system for construction sites. Detailed Implementation

[0016] 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.

[0017] See Figure 1 This invention provides a method for safety monitoring and early warning at construction sites, comprising the following steps: responding to elastic wave signals collected by a distributed vibration sensing array deployed at the construction site, constructing a spatiotemporal tensor data field in units of time windows; generating a tangent vector manifold characterizing the evolution path of the vibration source based on the Lie algebra derivative between each tensor slice in the spatiotemporal tensor data field and tensor slices of adjacent time windows; performing a dimensionality reduction mapping based on local preservation projection on the tangent vector manifold to obtain a feature trajectory curve in a low-dimensional embedded space; identifying curvature singularities on the feature trajectory curve and marking the timestamps corresponding to the curvature singularities as potential risk moments; extracting dangerous segments from the elastic wave signals centered on the potential risk moments, and outputting a safety warning level by calculating the mutual information entropy matrix of multiple channel signals in the dangerous segments.

[0018] In specific implementation, please refer to Figure 2 The distributed vibration sensing array comprises multiple sensor nodes, each acquiring elastic wave signals within a single time window. The duration of each time window is set based on the transient characteristics of the vibration signals at the construction site, and the elastic wave signals acquired within the time window are discrete-time sequences. For each sensor node, the elastic wave signals acquired within the corresponding time window are arranged in chronological order of sampling time, forming a column vector, which serves as the node vector of the sensor node. The dimension of the node vector is determined by the number of sampling points within the time window.

[0019] The node vectors of all sensor nodes are stacked according to their spatial coordinates to obtain a two-dimensional matrix corresponding to the current time window. The spatial coordinates of the sensor nodes are determined based on the actual geometric layout of the distributed vibration sensor array. The spatial coordinate order is either along a non-uniform spiral from the inner circle to the outer circle and increasing circumferentially, or arranged in ascending order of the x-axis, y-axis, and z-axis in three-dimensional coordinates. During stacking, the node vector of each sensor node serves as a row of the two-dimensional matrix, and the row index is determined by the spatial coordinate order of the corresponding sensor node. The number of rows in the two-dimensional matrix equals the total number of sensor nodes, and the number of columns equals the dimension of the node vectors.

[0020] Three consecutive two-dimensional matrices from different time windows are concatenated along the time axis to form the initial tensor block of the spatiotemporal tensor data field. During sequential concatenation, the two-dimensional matrices of the previous time window, the current time window, and the next time window are stacked sequentially along the time dimension to form a third-order tensor. The three dimensions of the initial tensor block correspond to the spatial order of sensor nodes, the temporal order of sampling points within a single time window, and the temporal order between windows, respectively. The initial tensor block is then centered by calculating the mean of all elements and subtracting the mean from each element. After centering, the centered tensor block undergoes amplitude normalization by calculating the standard deviation of all elements and dividing each element by the standard deviation, thus obtaining the spatiotemporal tensor data field.

[0021] The spatiotemporal tensor data field extracts a tensor slice from the current time window as the first slice, and extracts a tensor slice from the previous time window adjacent to the current time window as the second slice. The spatiotemporal tensor data field contains multiple tensor slices along the time dimension, each tensor slice being a two-dimensional matrix. The first and second slices are both two-dimensional matrices, corresponding to the states at time t and time t-1, respectively. The first and second slices are considered as elements of a Lie group. The structure of the Lie group is constructed based on the spatial correlation of sensor nodes and the signal amplitude characteristics. The group operation rules between Lie group elements are defined by matrix multiplication. When calculating the logarithmic difference between the first and second slices, the inverse element of the second slice in the Lie group is calculated, and the first slice is multiplied by the inverse element to obtain a product matrix. The logarithmic mapping on the Lie group is then calculated on this product matrix to obtain a matrix belonging to the corresponding Lie algebra; this matrix is ​​the tangent vector. The tangent vector lies in the tangent space of the first slice on the Lie group manifold of the spatiotemporal tensor data field. The formula for calculating the tangent vector is:

[0022] in, This represents the tangent vector corresponding to the current time window. Indicates the first slice. Let represent the inverse element of the second slice in the Lie group. Represents the group multiplication operation on Lie groups. Let represent the logarithmic mapping from a Lie group to its Lie algebra.

[0023] After obtaining the tangent vector, it is decomposed into rotational and scaling components. In practice, extreme decomposition is performed on the tangent vector to obtain an orthogonal matrix as the rotational component and a symmetric positive definite matrix as the scaling component. Extreme decomposition is achieved through singular value decomposition: the tangent vector is decomposed using singular values, and the tangent vector is equal to the product of the left singular matrix, the singular value diagonal matrix, and the transpose of the right singular matrix. Then, the rotational component is composed of the product of the left singular matrix and the transpose of the right singular matrix, and the scaling component is composed of the product of the right singular matrix, the singular value diagonal matrix, and the transpose of the right singular matrix. The direct sum of the rotational and scaling components is used as the derivative of the Lie algebra. The direct sum operation is performed by combining the rotational and scaling component matrices in a block diagonal manner along the main diagonal direction to form a new block diagonal matrix, which is the matrix representation of the derivative of the Lie algebra.

[0024] Using the time window index as the independent variable, the derivatives of the Lie algebra corresponding to each time window are concatenated in the tangent space to form a tangent vector manifold. During concatenation, the derivatives of the Lie algebras are arranged as sequence elements in ascending order of the time window index. For each time window, the derivative of the Lie algebra is a block diagonal matrix composed of the direct sum of the rotation and scaling components. The entire tangent vector manifold is represented as an ordered set of block diagonal matrices along the time dimension. The tangent vector manifold lies in the Lie algebra corresponding to the direct product space of a special orthogonal group and a symmetric positive definite matrix. The structure of this direct product space is jointly determined by the special orthogonal group to which the rotation component belongs and the symmetric positive definite matrix group to which the scaling component belongs.

[0025] It should be noted that the decomposition of the tangent vector to obtain rotation and scaling components can be performed using either of the following two specific implementation methods. The first implementation method is based on extreme decomposition, and the second implementation method is based on extracting the symmetric part after logarithmic mapping. In practical applications, the specific method can be selected based on the numerical condition number of the tangent vector matrix: when the condition number of the tangent vector matrix is ​​lower than a preset threshold, extreme decomposition is preferred to ensure computational efficiency; when the condition number is high, logarithmic mapping is preferred to enhance numerical stability. Only one method should be used; both should not be used simultaneously.

[0026] In practice, the rotation component is obtained by performing extreme decomposition of the tangent vector, as detailed below: (See attached document). Figure 3The extreme decomposition operation is performed on the tangent vector matrix to obtain an orthogonal matrix and a symmetric positive definite matrix. The orthogonal matrix is ​​used as the rotation component, and the symmetric positive definite matrix is ​​used as the scaling component. The extreme decomposition operation is as follows: the tangent vector matrix is ​​decomposed into singular values, resulting in a left singular matrix, a singular value diagonal matrix, and a right singular matrix; the left singular matrix is ​​multiplied by the transpose of the right singular matrix, and the product matrix is ​​the rotation component; the right singular matrix, the singular value diagonal matrix, and the transpose of the right singular matrix are multiplied sequentially, and the product matrix is ​​the scaling component.

[0027] Alternatively, as an alternative implementation, the extreme decomposition method described above can be omitted. Instead, the scaling component can be obtained by extracting the symmetric portion after performing a logarithmic mapping on the tangent vector. This method and the aforementioned extreme decomposition method are mutually exclusive and cannot be used simultaneously. When this method is chosen, the rotation component is obtained by extracting the antisymmetric portion after performing a logarithmic mapping on the tangent vector. Specifically: perform a logarithmic mapping on the tangent vector matrix to obtain a logarithmic mapping matrix; extract the antisymmetric portion of this logarithmic mapping matrix, which is equal to half the difference between the logarithmic mapping matrix and its transpose. The resulting antisymmetric matrix is ​​used as the rotation component.

[0028] The scaling component is obtained by extracting the symmetric part after performing a logarithmic mapping on the tangent vector. Specifically, the tangent vector matrix is ​​logarithmically mapped to obtain a logarithmic mapping matrix. The logarithmic mapping process is based on the inverse operation of the exponential mapping from a Lie group to a Lie algebra. The logarithmic mapping matrix is ​​located in the Lie algebra space. The symmetric part of the logarithmic mapping matrix is ​​extracted. The symmetric part is equal to the sum of the logarithmic mapping matrix and the transpose of the logarithmic mapping matrix multiplied by one half. The resulting symmetric matrix is ​​used as the scaling component.

[0029] Optionally, in another implementation, the tangent vector is input to the singular value decomposition operator to obtain a left singular matrix, a singular value diagonal matrix, and a right singular matrix. A rotation component is constructed based on the product of the transpose of the left and right singular matrices, and the rotation component matrix is ​​equal to the left singular matrix multiplied by the transpose of the right singular matrix. A scaling component is constructed based on the difference between the diagonal elements of the singular value diagonal matrix and the unit vector. The construction process is as follows: each diagonal element of the singular value diagonal matrix is ​​subtracted by the value 1 to obtain a difference diagonal matrix. The scaling component matrix is ​​obtained by multiplying the right singular matrix by the difference diagonal matrix and then by the transpose of the right singular matrix. All components of the unit vector are 1.

[0030] In practice, when calculating the Euclidean distance between any two tangent vectors in the tangent vector manifold, each tangent vector matrix is ​​stretched column-wise to form the corresponding stretched tangent vector. For any two stretched tangent vectors, the Euclidean distance is calculated, which is equal to the L2 norm of the difference between the two stretched tangent vectors. The process of constructing an adjacency graph based on the Euclidean distance is as follows: each tangent vector is treated as a node in the adjacency graph. For each node, the k other nodes with the smallest Euclidean distance to that node are selected as adjacent nodes. The value of k is determined based on the total number N of tangent vectors in the tangent vector manifold. When N is less than 100, k is 5; when N is between 100 and 500, k is 10; and when N is greater than 500, k is 15. Edges are established between nodes and adjacent nodes to form the adjacency graph, and the weight of the edges in the adjacency graph is temporarily set to 1.

[0031] The local reconstruction weight coefficient for each node is calculated based on the adjacency graph. This local reconstruction weight coefficient is obtained by minimizing the weighted reconstruction error between each node and its neighboring nodes. For the tangent vector corresponding to the i-th node, the stretching vector is... Its adjacent node index set is The local reconstruction weight coefficients are obtained by solving the following optimization problem:

[0032] in, This represents the stretching vector of the tangent vector corresponding to the i-th time window. Let represent the stretching vector of the tangent vector of the j-th adjacent node. This represents the local reconstruction weight coefficient of the i-th node with respect to the j-th neighboring node. Describes the Euclidean norm of a vector. This represents the set of node indices adjacent to the i-th node. Additional constraints are applied when solving the optimization problem. The local reconstruction weight coefficients are obtained by constructing the local covariance matrix of the i-th node and solving the constrained linear equations. The elements of the local covariance matrix are the inner products between the tangent vectors and stretching vectors of adjacent nodes.

[0033] The local reconstruction weight coefficients corresponding to all nodes are combined into a sparse weight matrix. The number of rows and columns of the sparse weight matrix are both equal to the total number N of tangent vectors in the tangent vector manifold. For the i-th row of the sparse weight matrix, if index j belongs to... Then, fill in the local reconstruction weight coefficient in the i-th row and j-th column. If index j does not belong to If the value is less than 0, then fill in 0. After assigning values ​​to the sparse weight matrix, construct the matrix. ,in, It is the identity matrix. This is a sparse weight matrix. For the matrix... Generalized eigenvalue decomposition (GEM) is used to obtain all eigenvalues ​​and their corresponding eigenvectors. The goal of GEM is to find solutions that satisfy... The eigenvalues ​​and eigenvectors of , where For the constructed real symmetric positive semi-definite matrix, It is the identity matrix. For eigenvalues, For eigenvalues The corresponding eigenvectors. By solving this generalized eigenvalue problem, a set of non-negative real eigenvalues ​​and their corresponding eigenvectors can be obtained. Then, the smallest eigenvalue is extracted from all non-zero eigenvalues. 1 eigenvalue, The preset dimension for the low-dimensional embedding space. The value can be 2 or 3. This... The eigenvectors corresponding to the smallest eigenvalues ​​are arranged in ascending order of eigenvalues ​​to form a projection basis matrix. Each column of this projection basis matrix is ​​a projection direction vector.

[0034] The tangent vector corresponding to each time window is stretched and its inner product is performed with each projection direction vector in the projection basis to obtain the low-dimensional coordinate value of that time window in the corresponding projection direction. All low-dimensional coordinate values ​​in all projection directions constitute the low-dimensional coordinate vector of that time window, with dimension d. The low-dimensional coordinate vectors of all time windows are connected in chronological order to form the feature trajectory curve. The feature trajectory curve is a continuous polyline in d-dimensional space, with the vertices of the polyline corresponding to the low-dimensional coordinate vectors of each time window.

[0035] In specific implementation, please refer to Figure 4 Multiple discrete points are sampled at equal arc length intervals on the feature trajectory curve. The starting point of the sampling is the vertex corresponding to the first time window of the feature trajectory curve, and the ending point is the vertex corresponding to the last time window of the feature trajectory curve. For each discrete point on the feature trajectory curve, the first-order derivative and the second-order derivative are calculated using the difference method in the neighborhood of that discrete point. The first-order derivative is calculated as follows: taking the current discrete point as the reference, take one adjacent sampled discrete point forward and one adjacent sampled discrete point backward. Subtract the low-dimensional coordinate vector of the forward adjacent sampled discrete point from the low-dimensional coordinate vector of the backward sampled discrete point. Divide the resulting difference vector by twice the arc length interval to obtain the first-order derivative at the current discrete point. The second-order derivative is calculated as follows: add the low-dimensional coordinate vector of the forward adjacent sampled discrete point to the low-dimensional coordinate vector of the backward sampled discrete point, then subtract twice the low-dimensional coordinate vector of the current discrete point. Divide the resulting difference vector by the square of the arc length interval to obtain the second-order derivative at the current discrete point. For the first and last sampled discrete points, the first and second derivatives are calculated using one-sided difference, respectively.

[0036] After obtaining the first and second derivatives at each discrete point, the curvature value at each discrete point is calculated based on these derivatives. The curvature value is calculated according to the definition of curvature of a space curve in differential geometry. For a characteristic trajectory curve in d-dimensional space, d takes the value of 2 or 3, and the curvature value... The calculation formula is:

[0037] in, This represents the curvature value at the p-th discrete point, where p is the index number of the discrete point, and the value of p ranges from 1 to P, where P is the total number of discrete points. Let represent the first derivative vector at the p-th discrete point, which is a d-dimensional vector; Let represent the second derivative vector at the p-th discrete point, which is a d-dimensional vector; This represents the cross product operation of vectors. When d equals 2, and The two-dimensional vector is extended into a three-dimensional vector by padding it with zeros at the end before performing the cross product operation. The Euclidean norm of a vector is denoted by .

[0038] After obtaining the curvature values ​​corresponding to all discrete points, calculate the mean and standard deviation of the curvature values ​​for all discrete points. (Meaning of curvature values) By using all P curvature values The standard deviation of the curvature values ​​is obtained by summing and dividing by P. The result is obtained by first calculating the square of the difference between each curvature value and the average value, then summing the squared values, dividing by P, and taking the square root.

[0039] Discrete points whose curvature values ​​are greater than the sum of three times the mean and standard deviation are considered curvature singularities. The criteria for this determination are... Discrete points that satisfy this condition are marked as curvature singularities. If no discrete point satisfies this condition, then there are no curvature singularities on the current characteristic trajectory curve, and there are no potential risk moments in the current time period.

[0040] The parameter values ​​corresponding to the singularity of curvature on the feature trajectory curve are extracted. These parameters are the arc length parameters corresponding to the singularity of curvature on the feature trajectory curve, and the arc length parameters are accumulated starting from the starting point of the feature trajectory curve. The mapping method for converting the arc length parameters back to the time window index is as follows: based on the correspondence between the arc length parameters corresponding to each discrete point on the feature trajectory curve and the time window index, a linear interpolation method is used to determine the time window index corresponding to the arc length parameter of the singularity of curvature. The specific process of linear interpolation is as follows: locate the interval in which the arc length parameter of the singularity of curvature falls in the pre-stored arc length parameter sequence. The q-th element in the arc length parameter sequence corresponds to the discrete point arc length parameter of the q-th time window. Based on the time window indices corresponding to the two endpoints of the interval, linear interpolation is performed according to the arc length ratio to obtain the time window index value corresponding to the singularity of curvature. The starting timestamp corresponding to the time window index is marked as the potential risk moment, and the starting timestamp is the moment corresponding to the first sampling point of the time window.

[0041] In practice, the step size for equal arc length interval sampling is determined by the ratio of the total arc length of the feature trajectory curve to the preset number of sampling points. The total arc length of the feature trajectory curve is obtained by calculating the sum of the Euclidean distances between all adjacent vertices on the feature trajectory curve. The preset number of sampling points is proportional to the number of time windows, with a scaling factor set to 2, meaning the preset number of sampling points equals the total number of time windows multiplied by 2. Setting the number of sampling points according to this scaling factor ensures that the arc length spacing corresponding to the step size of equal arc length interval sampling is less than the average arc length spacing between adjacent vertices on the feature trajectory curve, thereby completely preserving the local geometric variation features of the feature trajectory curve through dense sampling.

[0042] In practice, the potential risk moment is used as the time center point. A first preset time period is extended forward, and a second preset time period is extended backward to extract the hazardous segment signal from the original elastic wave signal. The first preset time period is determined based on the average duration from the start of an event to the potential risk moment in historical hazardous event records at the construction site, and is set to 5 seconds. The second preset time period is determined based on the average duration from the potential risk moment to the end of an event in historical hazardous event records at the construction site, and is set to 10 seconds. The original elastic wave signal is a continuous record of all sensor nodes of the distributed vibration sensing array over all time. During extraction, based on the time center point and the time intervals determined by the first and second time periods, the multi-channel signal within the corresponding interval is extracted from the original elastic wave signal to form the hazardous segment signal.

[0043] The hazardous segment signal is divided into multiple channels according to the channel numbering of the distributed vibration sensor array, with each channel corresponding to a sensor node. The channel number and the spatial coordinates of the sensor node are sequentially linked. After division, each channel signal is a time series of elastic wave signals from a single sensor node within the intercepted time interval.

[0044] For the multi-channel signals obtained by partitioning, calculate the mutual information entropy value between any two channels. The calculation process is as follows: The probability density is estimated for all sampling points of the first channel signal to obtain the first marginal probability distribution. The probability density estimation adopts the kernel density estimation method, with a Gaussian kernel as the kernel function. The bandwidth is calculated using the Scott rule, and the bandwidth value is equal to the standard deviation of the sampling values ​​of the first channel signal multiplied by the negative fifth power of the total number of sampling points.

[0045] The probability density is estimated for all sampling points of the second channel signal to obtain the second marginal probability distribution. The probability density estimation also uses the Gaussian kernel function, and the bandwidth is calculated using the Scott rule. The bandwidth value is equal to the standard deviation of the second channel signal sampling values ​​multiplied by the negative fifth power of the total number of sampling points.

[0046] The joint probability density is estimated for all sampled pairs of the first and second channel signals to obtain the joint probability distribution. The joint probability density estimation uses a two-dimensional Gaussian kernel function, and the bandwidth matrix is ​​determined by combining the covariance matrix of the sampled values ​​from the two channel signals with the total number of sampled points according to Scott's rule.

[0047] Arrange the mutual information entropy values ​​between all pairs of channel signals into a symmetric mutual information entropy matrix. For a distributed vibration sensing array with a total of M channels, the mutual information entropy matrix is ​​an M x M square matrix, where the element in the m-th row and n-th column is... Diagonal elements Defined as zero, and the matrix satisfies symmetry, i.e. .

[0048] The matrix norm of the mutual information entropy matrix is ​​calculated using the Frobenius norm. The Frobenius norm is calculated by summing the squares of all elements in the matrix and then taking the square root. Specifically, the off-diagonal elements of the mutual information entropy matrix are included in the summation of squares, while zero diagonal elements do not affect the result.

[0049] The corresponding safety warning level is determined and output based on the numerical range of the Frobenius norm of the mutual information entropy matrix. The safety warning levels are divided into three levels: low risk, medium risk, and high risk. The boundary values ​​of the numerical ranges are determined by statistical clustering of the Frobenius norm of the mutual information entropy matrix of historical hazardous segments at the construction site. Specifically, historical hazardous segment samples with labeled risk levels are collected, and the Frobenius norm of the mutual information entropy matrix for each sample is calculated. The goal is to minimize the misclassification rate between adjacent risk levels, and the boundary thresholds are searched and determined. Based on the optimization results, when the Frobenius norm is less than 2, it is classified as low risk, and a low-risk safety warning level is output; when the Frobenius norm is greater than or equal to 2 and less than 5, it is classified as medium risk, and a medium-risk safety warning level is output; when the Frobenius norm is greater than or equal to 5, it is classified as high risk, and a high-risk safety warning level is output. The output safety warning level is in the form of a warning level identifier, which is linked to a preset alarm strategy.

[0050] In specific implementation, please refer to Figure 5 The distributed vibration sensing array consists of at least four three-component accelerometers. A three-component accelerometer is a sensing device capable of simultaneously measuring vibration acceleration in three orthogonal directions in space. The three output components correspond to the east-west horizontal vibration acceleration, the south-north horizontal vibration acceleration, and the vertical vibration acceleration at a specific location on the construction site. The three-component accelerometers are arranged in a non-uniform spiral configuration. The center of the non-uniform spiral is set at the geometric center of the foundation pit or the center point of the tower crane foundation at the construction site. The polar coordinate equation of the non-uniform spiral is given by the following formula:

[0051] in, Indicates at the polar angle The radial distance from a point on the helix to the center point of the helix, in meters; This represents the initial radial distance of the spiral, with a value of 2 meters. This represents the radial distance at the end of the spiral, and is taken as 80% of the radius of the area affected by the foundation pit or tower crane foundation. This represents the polar angle variable, with units in radians, and a value range from 0 to... ; This represents the terminal polar angle of the helix, with a value of 6π. This represents the non-uniform distribution index, with a value of 1.5. When the radial spacing of the spiral is 1.5, it gradually becomes denser as the polar angle increases, resulting in a higher density of sensor nodes near the periphery of the foundation pit or tower crane foundation than in the inner ring area.

[0052] The first of the four three-component accelerometers is installed at the starting point of the helix, the second at a polar angle of 2π, the third at 4π, and the fourth at 6π. The four accelerometers are mounted on the top of the reinforced concrete slope surrounding the excavation pit and on the surface of the reinforced concrete foundation of the tower crane. The accelerometer bases are secured to the foundation surfaces using stainless steel expansion bolts, with epoxy resin filling the gap between the bases and the foundations to eliminate contact gaps. The elastic wave signal includes longitudinal wave components, transverse wave components, and surface wave components. The longitudinal wave component is a compression wave generated by the compression and rarefaction motion of medium particles in the propagation direction; the transverse wave component is a shear wave with the vibration direction perpendicular to the propagation direction; and the surface wave components are Rayleigh waves and Love waves propagating along the ground surface or excavation face. The longitudinal wave component, transverse wave component, and surface wave component are extracted from the three orthogonal components output by the three-component accelerometers after polarization analysis and wave field separation processing.

[0053] The tangent vector manifold is specifically the manifold structure obtained by concatenating the derivatives of the Lie algebras corresponding to each time window in the direct product space of a special orthogonal group and a symmetric positive definite matrix. The special orthogonal group consists of all orthogonal matrices with a determinant of 1, and the group operation is matrix multiplication; the symmetric positive definite matrix group consists of symmetric matrices with all positive eigenvalues, and the group operation is matrix multiplication. The direct product space of the special orthogonal group and the symmetric positive definite matrix is ​​of the form... The binary composition, in which It belongs to a special orthogonal group. Belonging to the symmetric positive definite matrix group, group multiplication in the direct product space is defined as performing group multiplication on each component. In the direct product space, the derivative of the Lie algebra corresponding to each time window is represented as a block diagonal matrix. The upper left block of the block diagonal matrix contains the elements of the special orthogonal group Lie algebra corresponding to the rotation component, represented by an antisymmetric matrix, and the lower right block contains the elements of the symmetric positive definite matrix group Lie algebra corresponding to the scaling component, represented by a symmetric matrix. Arranging the block diagonal matrices corresponding to each time window in ascending order of time window index forms a tangent vector manifold. The tangent vector manifold is a parameterized curve in the Lie algebra corresponding to the direct product space, with the time window index as the parameter.

[0054] After outputting the safety warning level, the corresponding alarm strategy is matched from the warning rule base based on the safety warning level. The warning rule base is a pre-built query table containing three records. The first record has a low-risk safety warning level, and the matched alarm strategy is set to sound once every 30 seconds for 1 second each time, and flash the light once every 10 seconds. The remote notification terminal list only includes the on-site safety officer's terminal. The second record has a medium-risk safety warning level, and the matched alarm strategy is set to sound once every 10 seconds for 2 seconds each time, and flash the light once every 3 seconds. The remote notification terminal list includes the on-site safety officer's terminal and the project manager's terminal. The third record has a high-risk safety warning level, and the matched alarm strategy is set to continuous sounding of the sound alarm and continuous flashing of the light. The remote notification terminal list includes the on-site safety officer's terminal, the project manager's terminal, the supervision unit's terminal, and the construction unit's on-site representative's terminal.

[0055] The warning evidence package is formed by packaging five trajectory points before and after the potential risk moment in the hazardous segment and the characteristic trajectory curve. The hazardous segment is a segment of the original elastic wave signal extracted with the potential risk moment as the center. The five trajectory points before and after the potential risk moment in the characteristic trajectory curve are obtained as follows: locate the trajectory point corresponding to the potential risk moment on the characteristic trajectory curve, take the five adjacent trajectory points in the forward time reverse direction, and take the five adjacent trajectory points in the backward time sequence direction. Combine the low-dimensional coordinate vectors and corresponding curvature values ​​of a total of eleven trajectory points, including the trajectory point corresponding to the potential risk moment, to form a trajectory point snapshot. The data structure of the warning evidence package consists of a packet header, a hazardous segment data segment, and a trajectory point snapshot data segment. The packet header contains the unique identifier of the warning evidence package, the generation timestamp, and the safety warning level identifier. The hazardous segment data segment contains the vibration acceleration sample values ​​of all channels within the intercepted time interval. The trajectory point snapshot data segment contains the low-dimensional coordinate vector sequence and curvature value sequence of the eleven trajectory points.

[0056] The early warning evidence package is sent to the cloud server and the on-site management terminal via a wireless communication network. The wireless communication network uses a power frequency wireless local area network, and the communication protocol is an IoT message protocol based on the Transmission Control Protocol. After receiving the early warning evidence package, the cloud server performs persistent storage by writing the early warning evidence package into a distributed file system and creating an index. The index key is the generation timestamp of the early warning evidence package and the safety warning level identifier. The on-site management terminal is an industrial-grade tablet computer. After receiving the early warning evidence package, the on-site management terminal pops up an alarm interface, which displays the safety warning level, the potential risk time, and the time-domain waveform of the dangerous segment. The safety warning level is displayed in real time on a large on-site monitoring screen. The large on-site monitoring screen is an LCD video wall installed in the construction site command center. The displayed content includes the text label of the safety warning level and the background color corresponding to the safety warning level: green background for low risk, yellow background for medium risk, and red background for high risk. It also displays the timestamp text of the potential risk time.

[0057] A construction site safety monitoring and early warning system includes a memory, a processor, and a computer program stored in the memory and running on the processor. The memory is a non-volatile semiconductor memory device with a storage capacity sufficient to store elastic wave signal data and intermediate processing results collected by a distributed vibration sensor array throughout the entire construction cycle. The processor is a multi-core central processing unit that executes all the aforementioned steps when running the computer program.

[0058] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for safety monitoring and early warning at construction sites, characterized in that, Includes the following steps: In response to elastic wave signals collected by a distributed vibration sensor array deployed at a construction site, a spatiotemporal tensor data field is constructed in units of time windows. Based on the Lie algebra derivative between each tensor slice in the spatiotemporal tensor data field and the tensor slice of the adjacent time window, a tangent vector manifold representing the evolution path of the vibration source is generated. The tangent vector manifold is subjected to a dimensionality reduction mapping based on local preservation projection to obtain the feature trajectory curves in the low-dimensional embedding space; Identify curvature singularities on the characteristic trajectory curve and mark the timestamps corresponding to the curvature singularities as potential risk moments; Centered on the potential risk moment, a dangerous segment of the elastic wave signal is extracted, and the safety warning level is output by calculating the mutual information entropy matrix of multiple channel signals in the dangerous segment.

2. The method for safety monitoring and early warning at construction sites according to claim 1, characterized in that, The construction of the spatiotemporal tensor data field with time windows as units specifically includes the following steps: The elastic wave signals collected by each sensor node in the distributed vibration sensing array within a single time window are arranged in chronological order to form a node vector for each sensor node. Stack the node vectors of all sensor nodes in the order of their spatial coordinates to obtain a two-dimensional matrix corresponding to each time window; The two-dimensional matrices of three consecutive time windows are concatenated along the time axis to form the initial tensor block of the spatiotemporal tensor data field; The initial tensor block is centered and its amplitude is normalized to obtain the spatiotemporal tensor data field.

3. The method for safety monitoring and early warning at construction sites according to claim 2, characterized in that, Based on the Lie algebraic derivatives between each tensor slice in the spatiotemporal tensor data field and the tensor slices of adjacent time windows, a tangent vector manifold representing the evolution path of the vibration source is generated, specifically including the following steps: Extract the tensor slice of the current time window from the spatiotemporal tensor data field as the first slice, and extract the tensor slice of the previous time window adjacent to the current time window as the second slice. Calculate the logarithmic mapping difference between the first slice and the second slice to obtain the tangent vector of the first slice in the tangent space on the Lie group manifold of the spatiotemporal tensor data field; The tangent vector is decomposed into rotational and scaling components, and the direct sum of the rotational and scaling components is taken as the derivative of the Lie algebra. Using the time window index as the independent variable, the derivatives of the Lie algebras corresponding to each time window are concatenated in the tangent space to form the tangent vector manifold.

4. The method for safety monitoring and early warning at construction sites according to claim 3, characterized in that, The rotation component is obtained by performing polar decomposition on the tangent vector, and the scaling component is obtained by extracting the symmetric part after performing logarithmic mapping on the tangent vector.

5. The method for safety monitoring and early warning at construction sites according to claim 3, characterized in that, The dimensionality reduction mapping based on local preservation projection is performed on the tangent vector manifold to obtain the feature trajectory curve in the low-dimensional embedding space, specifically including the following steps: Calculate the Euclidean distance between every two tangent vectors in the tangent vector manifold, and construct an adjacency graph based on the Euclidean distance, wherein each node in the adjacency graph corresponds to a tangent vector; The local reconstruction weight coefficient of each node is calculated based on the adjacency graph. The local reconstruction weight coefficient is obtained by minimizing the weighted reconstruction error between each node and its neighboring nodes. The local reconstruction weight coefficients are combined into a sparse weight matrix, and the sparse weight matrix is ​​subjected to generalized eigenvalue decomposition to extract the eigenvector corresponding to the smallest non-zero eigenvalue as the projection basis. Each tangent vector is inner-producted with the projection basis to obtain the low-dimensional coordinate values ​​corresponding to each time window. All low-dimensional coordinate values ​​are then connected in chronological order to form the feature trajectory curve.

6. The method for safety monitoring and early warning at construction sites according to claim 5, characterized in that, Identifying curvature singularities on the characteristic trajectory curve and marking the timestamps corresponding to these curvature singularities as potential risk moments specifically includes the following steps: Multiple discrete points are sampled at equal arc length intervals on the characteristic trajectory curve, and the first-order derivative and second-order derivative at each discrete point are calculated. The curvature value at each discrete point is calculated based on the first-order derivative and the second-order derivative. The curvature value is equal to the modulus of the cross product of the first-order derivative and the second-order derivative divided by the cube of the modulus of the first-order derivative. Calculate the mean and standard deviation of the curvature values ​​of all discrete points, and identify discrete points whose curvature values ​​are greater than three times the sum of the mean and the standard deviation as curvature singularities; Extract the parameter values ​​corresponding to the curvature singularities on the feature trajectory curve, convert the parameter values ​​back to the time window index, and mark the starting timestamp corresponding to the time window index as the potential risk moment.

7. A method for safety monitoring and early warning at construction sites according to claim 6, characterized in that, The step size of the equal arc length interval sampling is determined by the ratio of the total arc length of the feature trajectory curve to the preset number of sampling points, and the preset number of sampling points is proportional to the number of time windows.

8. A method for safety monitoring and early warning at construction sites according to claim 6, characterized in that, Centered on the potential risk moment, a dangerous segment of the elastic wave signal is extracted, and a safety warning level is output by calculating the mutual information entropy matrix of multiple channel signals in the dangerous segment. This process specifically includes the following steps: Using the potential risk moment as the time center point, extend it forward by a preset first time period and backward by a preset second time period to extract the dangerous segment signal from the original elastic wave signal; The dangerous segment signal is divided into multiple channel signals according to the channel number of the distributed vibration sensing array, and each channel signal corresponds to a sensor node; Calculate the mutual information entropy value between any two channel signals, and arrange all mutual information entropy values ​​into a symmetric mutual information entropy matrix; Calculate the matrix norm of the mutual information entropy matrix, determine the corresponding security warning level based on the numerical range of the matrix norm, and output the security warning level.

9. A method for safety monitoring and early warning at construction sites according to claim 1, characterized in that, The distributed vibration sensing array is specifically composed of at least four three-component accelerometers arranged in a non-uniform spiral pattern around the foundation pit and on the tower crane foundation at the construction site. The elastic wave signal includes longitudinal wave components, transverse wave components, and surface wave components.

10. A safety monitoring and early warning system for construction sites, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the construction site safety monitoring and early warning method as described in any one of claims 1 to 9.

Images