Intelligent evaluation method for stability of deep foundation pit based on multi-source monitoring data
Patent Information
- Application Number
- CN202611092863.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-22
- Publication Date
- 2026-10-20
- Estimated Expiration
- 2046-07-22
Smart Images
Figure CN122595270B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of geotechnical engineering monitoring technology, and in particular to an intelligent assessment method for the stability of deep foundation pits based on multi-source monitoring data. Background Technology
[0002] With the continuous development of urban construction, the excavation depth, scale and surrounding environment complexity of deep foundation pit projects are constantly increasing. The spatiotemporal variability of geotechnical mechanical parameters and the dynamic disturbance of the construction process are becoming increasingly prominent. When using numerical simulation methods with preset fixed parameters, there is often a certain deviation between the calculation results and the actual state of the project, which puts forward higher requirements for the dynamic adaptability of the evaluation method.
[0003] In recent years, the rapid development of IoT sensing technology and data science has provided new technical approaches for the stability assessment of deep foundation pits. Field monitoring networks based on multi-source sensors can acquire real-time physical information such as stress, displacement, and water level in the foundation pit. Data fusion technologies, such as aggregation and data integration, also provide effective methods for dynamically updating the parameters and state of the soil and rock mass.
[0004] In practical engineering applications, different types of sensors have different sampling frequencies and transmission delays. Monitoring data often has asynchronous and noise interference characteristics, requiring more efficient preprocessing methods to achieve a unified representation of the global spatiotemporal state. At the same time, when dealing with high-dimensional geotechnical engineering problems, the covariance matrix is prone to non-physical spurious correlations during the assimilation process, which will affect the accuracy and stability of data assimilation to some extent.
[0005] Current methods for assessing the instability risk of deep foundation pits are mostly based on classical mechanical strength criteria, using indicators such as safety factors to determine stability and providing important evidence for early warning of engineering safety. However, as complex heterogeneous media, soil and rock masses exhibit significant uncertainties in their mechanical evolution, and a single mechanical indicator is insufficient to comprehensively characterize the multidimensional features of instability risk.
[0006] Furthermore, with the continuous growth of foundation pit monitoring data, the computational efficiency and real-time performance of assessment methods also need to be further improved to meet the needs of online monitoring and rapid response at engineering sites. Summary of the Invention
[0007] The purpose of this invention is to propose an intelligent assessment method for the stability of deep foundation pits based on multi-source monitoring data in order to solve the above-mentioned problems.
[0008] To achieve the above objectives, the present invention adopts the following technical solution: Intelligent assessment methods for deep foundation pit stability based on multi-source monitoring data include: The physical signals collected by the field sensors are converted into discrete observation column vectors. Interpolation alignment and spatial mapping are used to eliminate asynchrony and generate global spatiotemporal observation column vectors. The original state vector and evolution parameter column vector of the discrete model of the foundation pit are obtained and concatenated to expand the dimension. The preset error covariance matrix is decomposed and combined with white noise sampling to generate the initial augmented state evolution matrix. A linear difference evolution equation is constructed using an autoregressive model, and a state transition operator is constructed by combining the finite difference equation. The initial augmented state evolution matrix is then derived to generate the prior prediction augmented set matrix and the prior error covariance matrix. By using a spatial localization function to eliminate spurious elements in the prior error covariance matrix, calculating the cross covariance matrix and Kalman gain matrix, and assimilating the global spatiotemporal observation column vector with the prior prediction augmented set matrix, the posterior updated augmented set matrix is obtained. The mean of the posterior updated augmented set matrix is extracted and inversely mapped. The principal stresses are analyzed using the stress tensor. The global instability risk index is calculated by combining Shannon entropy and local reliability, and the corresponding control command is triggered.
[0009] Preferably, the step of converting the physical signals acquired by the field sensors into discrete observation column vectors includes dimensional transformation of the physical signals output by the vibrating wire stress monitoring sensor: The zero-crossing point of the AC signal from the vibrating wire stress monitoring sensor is captured using a hardware timer and an edge-triggered interrupt mechanism to calculate the real-time frequency. The floating-point arithmetic unit is then used to substitute the real-time frequency into the following physical dimension conversion formula to generate the actual physical stress observation value: ; In the formula, Indicates the first A stress sensor in time The actual physical stress observation value, Indicates the calibration coefficient. Indicates real-time frequency. Indicates the initial calibration frequency. This represents the coefficient of thermal expansion. Indicates real-time temperature. This indicates the initial temperature.
[0010] Preferably, the step of eliminating asynchronicity by using interpolation alignment and spatial mapping to generate a global spatiotemporal observation column vector includes: Establish a data structure based on hash mapping and doubly linked lists in computer memory to cache sensor data payloads with inconsistent arrival times; The daemon process is started to calculate the derivative value using forward and backward differential logic, and executes polynomial interpolation logic to resample the sensor data payload to a unified evaluation time node to splice and generate a global spatiotemporal observation column vector. A first-in-first-out circular queue is established as a sliding time window. The sliding mean and sliding standard deviation are calculated. Pulse anomaly elements that exceed the set statistical threshold are identified and forcibly replaced. The global spatiotemporal observation column vector is output.
[0011] Preferably, the step of obtaining the original state vector and evolution parameter column vector of the discrete model of the foundation pit and concatenating them to expand the dimension specifically includes: The computational nodes are indexed and arranged according to the graph theory topology mapping rules of spatial grid generation, and the apparent mechanical morphology of the nodes that ensures spatial locality is represented as the original state vector. Extract the effective elastic modulus, effective cohesion, and effective internal friction angle, and call the mathematical function interface to perform logarithmic nonlinear mapping to construct an evolution parameter column vector in the logarithmic domain; Allocate contiguous physical memory blocks and use the direct memory access controller to vertically concatenate and stack the byte streams of the original state vector and the evolution parameter column vector to generate an augmented state column vector.
[0012] Preferably, the step of decomposing the preset error covariance matrix and combining it with white noise sampling to generate the initial augmented state evolution matrix includes: Read the preset error covariance matrix stored in a compressed sparse format, and call the linear algebra subroutine to perform lower triangular decomposition to obtain the lower triangular matrix; Extract the hardware state parameters of the central processing unit motherboard as high-entropy seeds, and combine them with a transformation algorithm to generate a white noise vector that follows a standard normal distribution; The initial augmented state sample column vector is generated by: ; In the formula, Indicates the first An initial augmented state sample column vector, This represents the optimal estimated augmented state mean vector. Represents a lower triangular matrix. Represents a white noise vector; Arrange all generated initial augmented state sample column vectors horizontally in a column-major, contiguous memory arrangement to generate the initial augmented state evolution matrix.
[0013] Preferably, the construction of the linear difference evolution equation using the autoregressive model includes execution time-series deduction for the evolution parameter column vector: ; In the formula, Indicates time step The evolution parameter column vector, Indicates time step The evolution parameter column vector, This represents the diagonal matrix of autoregressive coefficients. Represents a constant offset vector. Represents the process evolution noise vector; During execution, the autoregressive coefficient diagonal matrix is compressed and stored as a one-dimensional floating-point array, and the single-instruction multiple-data extended instruction set is invoked to adjust the time steps. The adjacent elements of the evolution parameter column vector are fused with the corresponding elements of the one-dimensional floating-point array in parallel for multiplication and addition operations.
[0014] Preferably, the step of constructing a state transition operator by combining finite difference equations to deduce the initial augmented state evolution matrix and generate a priori prediction augmented set matrix and a priori error covariance matrix includes: The updated parameters are derived using the linear difference evolution equation. The logarithmic domain variables in the evolution parameter column vector are then subjected to exponential inverse mapping to restore the physical parameter values and assemble the global stiffness matrix. The state increment column vector is obtained by using a numerical algorithm based on incomplete Cholsky decomposition preprocessing and conjugate gradient. The nonlinear transition operator containing state deduction updates is applied independently to the column vectors of each set of samples, and then concatenated to obtain the prior prediction augmented set matrix; Extract the column vector of the predicted mean of the prior prediction augmented set matrix to construct the perturbation prior set matrix, and multiply the perturbation prior set matrix with its transpose to obtain the prior error covariance matrix.
[0015] Preferably, the step of eliminating spurious elements in the prior error covariance matrix using a spatial localization function and calculating the cross covariance matrix includes: A sparse mask array composed of Boolean numbers is constructed as the observation operator matrix, and the model prediction column vector is extracted from the high-dimensional state memory paging by using pointer indirect addressing. A static lookup table based on the Euclidean distance and compact support function of state nodes is pre-established in the read-only storage area; the normalized distance is extracted in the assimilation loop as the memory offset address, the decay coefficient is read from the static lookup table, and the localized constraint matrix is generated. Perform an element-wise aligned Hadamard product operation on the localized constraint matrix and the cross covariance matrix calculated based on the perturbation set, and force zero out any spurious covariance elements that exceed the cutoff correlation radius.
[0016] Preferably, the calculation of the Kalman gain matrix involves assimilating the global spatiotemporal observation column vector with the prior prediction augmented set matrix to obtain the posterior updated augmented set matrix, which includes the following correlation features: The singular value decomposition algorithm is called to calculate the inverse matrix of the sum of the observation space autocovariance matrix and the diagonal observation error covariance matrix under spatial localization constraints, and then multiplied by the cross covariance matrix to obtain the Kalman gain matrix. The posterior updated state column vector is derived using the following set-congruent update equation: ; In the formula, This represents the posterior update state column vector. This represents the column vector of predicted samples in the prior prediction augmented set matrix. Represents the Kalman gain matrix. This represents the perturbed observation set vector generated after perturbation of the global spatiotemporal observation column vector. This represents the column vector predicted by the model; During the write phase of executing the set-combinatorial update equation, a mutex lock mechanism is acquired to prevent concurrent writes from generating dirty data in memory, and then the a posteriori update augmented set matrix is generated.
[0017] Preferably, the extraction of the mean of the posterior-updated augmented set matrix is used for inverse mapping, the principal stresses are analyzed using the stress tensor, and the global instability risk index is calculated by combining Shannon entropy and local reliability, triggering corresponding control commands, including: The linear algebra package is used to perform Haushold transformation and implicit matrix root-finding algorithm on the third-order stress symmetric matrix to analyze the principal stresses; A histogram counter is constructed by dividing the physical memory into equidistant digital intervals, and the probability distribution function of the statistical state variable is calculated by using atomic addition operations. The global instability risk index is derived using the following bi-objective weighted inner product formula: ; In the formula, This indicates an index representing the overall risk of instability. and This represents the balance constant of the risk dimension. Represents a column vector weighted by the spatial importance of nodes. This represents a one-dimensional column vector consisting of the local failure probabilities of all evaluated nodes. A one-dimensional column vector representing the information entropy of all nodes; The global instability risk index is compared with a preset incremental risk trigger threshold, which in turn triggers the graphics rendering pipeline to draw a digital model or sends a shutdown command message to the field programmable logic controller.
[0018] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. This invention employs an augmented state set assimilation method to incorporate the mechanical evolution parameters of soil and rock mass and structural state variables into a unified update system, enabling real-time reflection of the changing patterns of soil and rock mechanical properties during construction. The introduction of a spatial localization constraint mechanism effectively suppresses non-physical spurious correlations in the high-dimensional covariance matrix, improving the stability and accuracy of the data assimilation process. This allows numerical simulation results to better reflect the actual engineering conditions, providing a state-based basis for the dynamic assessment of deep foundation pit stability.
[0019] 2. This invention establishes a multi-dimensional instability risk assessment system that integrates mechanical mechanisms and information theory. It comprehensively considers the deterministic characteristics of the mechanical state of soil and rock and the uncertain characteristics of its evolution process, enabling a more comprehensive quantification of the overall instability risk of deep foundation pit engineering. At the algorithm implementation level, it is fully adapted to modern computer hardware architecture. Through optimization techniques such as parallel computing, direct memory access, and atomic operations, it improves the efficiency of large-scale numerical calculations and data processing. At the same time, it constructs a hierarchical risk response mechanism, which can trigger corresponding visualization and control commands based on the assessment results, meeting the actual needs of online monitoring and rapid early warning at the engineering site, and providing support for the refined management and control of deep foundation pit construction safety. Attached Figure Description
[0020] Further details, features, and advantages of this application are disclosed in the following description of exemplary embodiments in conjunction with the accompanying drawings, in which: Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation
[0021] Several embodiments of this application will now be described in more detail with reference to the accompanying drawings to enable those skilled in the art to implement this application. This application may be embodied in many different forms and for various purposes and should not be limited to the embodiments set forth herein. These embodiments are provided to make this application thorough and complete, and to fully convey the scope of this application to those skilled in the art. The embodiments described do not limit this application.
[0022] Unless otherwise defined, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It will be further understood that terms such as those defined in commonly used dictionaries shall be interpreted as having a meaning consistent with their meaning in the relevant field and / or the context of this specification, and shall not be interpreted in an idealized or overly formal sense unless expressly defined herein.
[0023] Example 1
[0024] Its specific implementation method is combined with the appendix Figure 1 Please provide a detailed explanation.
[0025] Appendix Figure 1 The flowchart of the intelligent assessment method for deep foundation pit stability based on multi-source monitoring data provided in the embodiments of the present invention shows the complete steps from multi-source heterogeneous data acquisition and registration calculation to joint assessment quantification and early warning threshold determination.
[0026] In this embodiment, it includes: Step 1, Multi-source heterogeneous data acquisition and registration calculation: Convert the physical signals acquired by the field sensors into discrete observation column vectors, use interpolation alignment and spatial mapping to eliminate asynchrony, and generate global spatiotemporal observation column vectors; This step establishes a three-dimensional spatial coordinate system for the deep foundation pit. The raw physical signals collected by the multi-source heterogeneous sensors deployed on site are transmitted across the physical hardware boundary to the physical memory segment of the central computer node through a specific industrial serial communication protocol and a high-precision analog-to-digital conversion hardware interface, and are transformed into discrete double-precision floating-point observation column vectors in a unified mathematical space.
[0027] By leveraging the multi-core, multi-threaded concurrent architecture and advanced cache queue scheduling mechanism of the central processing unit, timestamp interpolation alignment algorithm and spatial projection transformation matrix are executed to eliminate the asynchronicity and dimensionality differences of multi-source data in spatiotemporal distribution.
[0028] The specific process is as follows: Establish a global Cartesian rectangular coordinate system at the deep foundation pit construction site: Set a fixed corner point of the foundation pit as the origin of the coordinate system, set the direction parallel to the long side of the foundation pit as the X-axis, the direction parallel to the short side of the foundation pit as the Y-axis, and the direction perpendicularly downwards pointing to the center of the earth as the Z-axis.
[0029] Each sensor node deployed in and around the foundation pit is assigned a globally unique three-dimensional spatial coordinate vector.
[0030] In order to realize the mapping from physical entities to digital objects in the digital system, the system establishes a spatial topology mapping table in the relational persistent database (such as PostgreSQL) on the central server.
[0031] The primary key of the mapping table is the universally unique identifier of the sensor node. The corresponding fields include double-precision floating-point values of the X, Y, and Z coordinates, the sensor's Media Access Control (MAC) address, and the network protocol (IP) identifier of the node device.
[0032] The system maintains a pre-read cache of this mapping table in memory to accelerate the subsequent data matching process. Three main types of monitoring sensors were deployed on-site: a displacement monitoring sensor cluster, a stress monitoring sensor cluster, and a groundwater level monitoring sensor cluster.
[0033] For displacement monitoring sensor clusters (such as automated total stations, fixed MEMS inclinometers, etc.), let the number be... indivual.
[0034] No. The three-dimensional spatial coordinates of a displacement sensor are defined as vectors. .
[0035] The raw data collected by the displacement sensor at discrete time points are usually angles or relative offsets in different directions. The built-in microelectromechanical system (MEMS) converts these weak voltage changes into high signal-to-noise ratio digital baseband signals through a high-precision 24-bit differential analog-to-digital converter.
[0036] The on-site data acquisition gateway controller sends request data frames to each displacement sensor in a polling manner via an RS485 serial bus with a physical layer configuration of fixed baud rate (e.g., 9600bps), no parity bit, and single stop bit.
[0037] After the sensor responds, it returns a hexadecimal data packet containing three-dimensional displacement components. The gateway's edge computing microprocessor parses the received data packet header, slave address, and function code according to the preset Modbus RTU communication protocol, and removes the cyclic redundancy check code through bit-by-bit XOR shift operations, thereby verifying the integrity of the data packet.
[0038] After extracting the effective load, the output is a three-dimensional displacement component. Defined at any discrete time point... , No. The output observation vector of each displacement sensor is ; The vector is then encapsulated into a custom datagram based on Transmission Control Protocol / Internet Protocol via Industrial Ethernet and pushed to the receive socket ring buffer of the central computer system kernel.
[0039] For a stress monitoring sensor cluster (such as a vibrating wire axial force gauge, a rebar gauge, etc.), let its number be... indivual; No. The three-dimensional spatial coordinates of a stress sensor are defined as a vector. ; Stress sensors typically output frequency signals, which must be converted to physical dimensions at the data acquisition front-end or central computing node. When a vibrating wire sensor is subjected to force and deforms, the induced electromotive force generated by its internal excitation coil exhibits a specific alternating frequency.
[0040] The on-site data acquisition terminal uses the microcontroller's internal hardware high-frequency timer and edge-triggered interrupt mechanism to accurately capture the timestamp difference of the AC signal voltage zero-crossing point, thereby calculating an extremely high-resolution frequency value.
[0041] According to the fundamental equations of vibrating string physics, strain is proportional to the square difference of the vibration frequencies of the steel string. Let the first... A stress sensor in time The output real-time frequency is Its initial calibration frequency is The calibration coefficient of the stress sensor is defined as follows: ; To eliminate stress measurement errors caused by temperature expansion in the metal steel wire, a temperature compensation operator is introduced, assuming the real-time temperature is... The initial temperature is The coefficient of thermal expansion is .
[0042] Then the first A stress sensor in time The actual physical stress observation value can be obtained through the following calculation: ; The calculation process of the above formula is strictly performed by the floating-point unit (FPU) of the central processing unit in accordance with the IEEE 754 double-precision floating-point standard.
[0043] The calculation program reads 64 bits of frequency, calibration coefficients, and temperature variables from the heap memory and loads them into the processor's floating-point register set.
[0044] After undergoing low-level fusion multiply-accumulate microinstruction stream operations, the output stress scalar is saved to a specified physical memory address using a write-back strategy.
[0045] This operation is performed concurrently on all stress sensor nodes using multiple threads, and the output scalars are aggregated to obtain the stress vector.
[0046] For a groundwater level monitoring sensor cluster (e.g., a piezometer), let its number be... indivual.
[0047] No. The spatial coordinates of each water level sensor are defined as follows: .
[0048] Piezometers operate based on the principle of hydrostatic pressure, and the physical quantity they collect is pore water pressure. The piezoresistive sensitive bridge element inside the piezometer converts minute changes in water pressure into standard industrial analog current signals (e.g., 4-20 mA).
[0049] The signal conditioning circuit at the data acquisition front end uses a high-precision shunt resistor to linearly map the signal into a low-voltage analog voltage signal. This signal then enters the digital control system via the microcontroller's 12-bit analog-to-digital sampling peripheral, where it is further multiplied by the factory-calibrated linearity coefficient. This process is then set at time... , No. The pore water pressure output scalar value of each water level sensor is .
[0050] After acquiring the aforementioned multi-source heterogeneous raw data, spatiotemporal registration calculations must be performed. Because the data acquisition frequencies of various sensors differ significantly, directly assembling data from different times would lead to logical inconsistencies in the temporal dimension.
[0051] An advanced dynamic data structure based on a combination of hash mapping and doubly linked lists is established in the main memory of the central computing system to cache multi-channel sensor data streams with inconsistent arrival times without blocking.
[0052] The system pre-sets a globally unified digital assimilation evaluation cycle. And set the discrete time step sequence Each time step corresponds to a unified evaluation time node.
[0053] For any evaluation time point The central computer's operating system allocates a separate background daemon process specifically for handling time synchronization tasks.
[0054] The daemon traverses the hash table nodes in memory to find all sensors at the specified time nodes. The nearby cached data payload with an absolute UNIX timestamp.
[0055] To achieve smooth data resampling with continuity of the first derivative, the system calls the built-in numerical computation core dynamic link library, uses forward and backward difference logic to calculate the derivative values at discrete time nodes, and starts piecewise cubic Hermitian polynomial interpolation logic.
[0056] At the execution level of the specific processor microarchitecture, the compiler hardcodes the basis function coefficients required for interpolation and loads them into the processor's L1 data cache.
[0057] The processor utilizes the Advanced Vector Extensions (AVX-2) single instruction multiple data stream mechanism to simultaneously perform wide-bit floating-point multiplication and accumulation pipeline operations on multiple data channels such as displacement, stress, and water level.
[0058] Through this high-speed spatiotemporal transformation scheduling based on pure hardware instruction parallelism, the system eliminates data misalignment caused by clock drift in the physical system of the device and asynchronous communication within the machine cycle, generating data at a unified evaluation time node. The entire concurrent data set is stored in the program's heap memory pages as contiguous memory blocks.
[0059] After time alignment is completed, spatial correlation mapping and global spatiotemporal observation column vector assembly calculation are performed. Due to the extremely large degree of freedom of the deep foundation pit physical model, a spatial distance weighted correlation matrix is constructed according to the topological distribution of sensors in three-dimensional space.
[0060] Define a global spatiotemporal observation column vector Its total dimension is ; The aligned displacement components, stress components, and pore water pressure components are then vertically spliced in a fixed order: ; In terms of operating system memory layout management, the above splicing process does not require slow deep data copying; The system uses pointer arithmetic operations to perform a linear and continuous mapping of the starting physical addresses of different sensor data buffers, ensuring... In virtual memory space, it is represented as a one-dimensional contiguous array of double-precision floating-point numbers.
[0061] This underlying optimization improves the cache hit rate of the microprocessor in the subsequent matrix multiplication stage, avoiding cache miss penalties caused by the discrete distribution of data.
[0062] To eliminate random pulse noise caused by sudden electromagnetic interference during data acquisition and transmission, the assembled global spatiotemporal observation column vector was modified. Apply adaptive statistical threshold filtering; The system allocates a fixed-length First-In-First-Out (FIFO) circular queue in memory as the sliding time window storage structure. For each element in the global spatiotemporal observation column vector, the microprocessor's registers traverse all historical data corresponding to its circular queue, and the moving average is calculated using an accumulator. and sliding standard deviation .
[0063] If the current time Corresponding observation elements satisfy Upon receiving the logical condition, the instruction execution unit of the central processing unit will trigger the pipeline branch predictor to respond and execute the exception handling subroutine. The program determines that the value is an anomaly caused by a hardware channel fault. It directly modifies the value at the address of the element using a memory overwrite instruction, forcibly replacing it with the arithmetic mean of the previous value and the mean. At this point, the multi-source heterogeneous raw physical signals are aligned and transformed into a global spatiotemporal observation column vector with digital coherence. It is locked in a protected memory page of the computing system, preventing other low-priority processes from modifying it.
[0064] Step 2, Construction of augmented state parameter matrix: Obtain the original state vector and evolution parameter column vector of the discrete model of the foundation pit and concatenate them to expand the dimension. Decompose the preset error covariance matrix and combine it with white noise sampling to generate the initial augmented state evolution matrix. During the excavation of a deep foundation pit, the surrounding soil undergoes continuous time-varying degradation due to stress release, groundwater seepage, and soil creep.
[0065] This step involves high-dimensional splicing of the external geometric physical quantities representing the system state with the internal constitutive equation parameters representing material properties to construct an ultra-high-dimensional "augmented state vector".
[0066] In this ultra-high-dimensional state space, a set of Gaussian random floating-point numbers with a specific error covariance structure is generated, thereby completely transforming the complex problem of physical degradation tracking into parallel iterative computation of matrix sample arrays in a multidimensional probability space.
[0067] The deep foundation pit and surrounding soil are divided into a finite number of computational nodes using a spatial discretization algorithm, assuming that they include... There are 12 computational nodes, each with three translational degrees of freedom in a three-dimensional coordinate system. Therefore, the global displacement state column vector of the foundation pit is set as follows: , dimension .
[0068] The nodes not only have displacements, but also have stress field state components that are highly coupled to them. and seepage field state components ; The column vectors that purely represent the current mechanical state of the system are assembled into a single data vector, which is defined as the original state vector. ; set up The total dimension is ; In the virtual addressing space of the computing system, the above node states are abstracted and mapped as a very large one-dimensional dynamic continuous array.
[0069] The underlying addressing mechanism of the system uses graph theory topology mapping rules generated by the spatial grid to index and arrange the nodes, ensuring that adjacent pit nodes in physical space are as continuous as possible in the memory address distribution of their corresponding state variables. This is a spatial locality optimization strategy designed specifically for computer memory access mechanisms.
[0070] This vector objectively describes the apparent deformation mechanical morphology of the foundation pit within a certain time slice.
[0071] Identify the soil material parameters that require joint evolution tracking. In different soil layer distribution areas of deep foundation pits, select key resistance mechanical parameters that have a significant impact on the stability of the foundation pit.
[0072] Assume the excavation pit profile is divided into: For each soil layer region, the effective elastic modulus is extracted. Effective cohesion and effective internal friction angle .
[0073] Since these parameters are all strictly non-negative physical quantities in actual engineering, in order to prevent negative values that violate physical common sense from appearing in subsequent high-dimensional Kalman matrix addition and subtraction operations, thereby causing numerical underflow anomalies or non-digital interruptions in the underlying computing system, the system must call the processor's mathematical core library at the operation front end to perform logarithmic nonlinear mapping processing on them. Define the mapped parameter variables as follows: ; ; ; Arrange the mapped parameters of all soil regions in index order to construct an evolution parameter column vector. Let its total dimension be .
[0074] This vector describes the sequence of intrinsic physical properties that determine the deformation resistance of the foundation pit and is stored by the system memory allocator in a protected array of pointers to independent parameter structures within the heap area.
[0075] The original state vector With evolution parameter column vector Vertically concatenated and stacked, defined as an augmented state column vector. : ; Augmented State Column Vector The dimension is expanded to Under the scheduling of the operating system's underlying memory management module, the process of splicing this high-dimensional vector is not a simple element-by-element data copying.
[0076] The system requests a block of memory of exactly [size] from available contiguous physical memory pages by calling high-level memory reallocation system functions (such as the `realloc` operation in the C standard library or the virtual memory allocator of a specific operating system). A massive contiguous block of memory multiplied by the number of double-precision floating-point bytes (typically 8 bytes).
[0077] Through direct memory access controllers or low-level memory block move functions optimized for assembly language (such as memcpy), and The data is seamlessly spliced into the new physical memory address segment in the form of a high-speed byte stream.
[0078] Through this extended-dimensional algebraic splicing, all subsequent differential prediction and assimilation algorithms will no longer deliberately distinguish between "external state appearance" and "internal parameter essence," but will instead place them all in a unified high-dimensional computer memory space for equal-weighted large-scale matrix multiplication and addition floating-point operations, thereby achieving joint synchronous computation and evaluation of apparent deformation and underlying soil degradation.
[0079] After defining the memory for the individual augmented vector, it is necessary to strictly adhere to the core theory of stochastic probability in ensemble Kalman filtering to generate a vector containing... The initial augmented state evaluation matrix for each sample.
[0080] Number of samples in the set The number of concurrent computing power is usually set between 100 and 500 based on the concurrent computing power of the computer hardware cluster (such as the number of concurrent threads of the graphics processing unit (GPU) on the motherboard or the number of physical cores of the multi-core central processing unit (CPU)) in order to achieve a balance between computing power and computing time.
[0081] Let the optimal estimated augmented state mean vector at the initial time be... The vector consists of the initial constitutive parameters provided in the engineering geological survey report and the initial static displacement (usually initialized as a zero scalar) obtained from digital leveling before the foundation pit excavation.
[0082] Set the initial state error covariance matrix as , It is a dimension A giant symmetric positive definite matrix, whose numerical elements on the main diagonal represent the initial variance measure of each state variable and physical parameter, and whose off-diagonal elements represent the initial spatial topological correlation between any two variables.
[0083] because Often, these are enormous numbers in the hundreds of thousands. If the matrix is stored using a conventional two-dimensional dense array format, it will easily exhaust hundreds of gigabytes (GB) of conventional workstation main memory. The system kernel forces the use of advanced sparse matrix memory structures such as compressed sparse rows or compressed sparse columns for persistent storage and loading, retaining only the floating-point values of non-zero elements and their corresponding row offset indices and column pointer information in physical memory.
[0084] In order to generate a covariance matrix that satisfies the requirements in computer memory A statistically constrained set of Gaussian random samples performs purely analytical mathematical operations at the hardware execution layer by calling highly optimized underlying linear algebra subroutine libraries (such as the dpotrf routine in LAPACK).
[0085] The CPU's instruction scheduler reads binary instruction sets from dynamic link libraries for large symmetric positive definite covariance matrices. Initiate the Cholsky lower triangular decomposition operation to find a lower triangular matrix. .
[0086] In the specific chip-level transistor execution logic, the processor allocates a dedicated high-speed temporary register area to store tiny blocks of the matrix, and uses the advanced vector extension instruction set to package multiple independent floating-point multiplication and summation operations into a single long instruction cycle machine instruction.
[0087] This data parallelism mechanism accelerates the processor's computation through multiple iterative loops. The process of the diagonal and off-diagonal elements of the matrix makes The analytical mathematical conditions are tightly closed and absolutely satisfied in the hardware arithmetic logic unit.
[0088] The system needs to generate completely independent random white noise sequences. In order to ensure that the noise source has a sufficiently high random entropy distribution and is not limited by the inherent cycle of ordinary software pseudo-random number generation algorithms, the operating system kernel program extracts seeds with extremely high entropy values by reading state parameters at the level of computer motherboard hardware chips (such as thermal noise pulses generated by real-time temperature fluctuations of diodes inside the central processing unit, or tiny physical deviations in disk read / write head seek time).
[0089] After acquiring the hardware-level high-entropy seed, the operating system starts a set of dedicated parallel computing threads. These threads read the seed information and strictly follow the Box-Muller transformation logic to transform the uniformly distributed pseudo-random sequence into a floating-point data stream that follows a standard normal distribution.
[0090] During this process, the transcendental function hardware accelerator inside the FPU directly uses lookup tables in conjunction with a polynomial approximation algorithm to solve the time-consuming logarithmic and trigonometric function microinstructions, and writes the results into a pre-allocated noise column vector in contiguous memory order. In the physical memory pages. Repeat the above process, generating a total of... A white noise vector like this.
[0091] Based on mean vector Decomposing the lower triangular matrix and standard white noise vector The system's main control daemon uses a task allocation engine to schedule idle processing cores of the multi-core processor, concurrently executing massive matrix and vector multiplication and addition instructions using lock-free data structures to generate the first... Initial augmented state sample column vectors : ; All calculated sample column vectors Arranged horizontally in ascending order of column indexes, a dimension of [dimensional value] is ultimately constructed in the computer's main memory area. The initial augmented state evolution matrix (i.e., the augmented set matrix at time t=0):
[0092] The initial augmented state evolution matrix It serves as the numerical foundation for all subsequent system state assimilation calculations and error covariance matrix iterations. Each column in the matrix independently represents a possible parallel evolution path of the deep foundation pit in the ultra-high-dimensional parameter space and state physical space.
[0093] At the hardware storage arrangement level, the system mandates that the two-dimensional matrix must adopt a column-first, contiguous memory arrangement format. This allows the CPU's memory controller to quickly fill the CPU's cache lines at all levels through continuous burst read transactions when extracting independent sample columns during subsequent Kalman gain calculations and state update operations, thereby reducing the addressing latency of random memory addresses.
[0094] At this point, the entire spatiotemporal physical model of the deep foundation pit is completely mapped into a high-dimensional floating-point array that can be used by modern high-speed electronic computers for large-scale parallel multithreaded matrix operations.
[0095] Step 3, Differential Prediction and Error Covariance Iteration: A linear difference evolution equation is constructed using an autoregressive model, and a state transition operator is constructed by combining the finite difference equation. The initial augmented state evolution matrix is deduced to generate the prior prediction augmented set matrix and the prior error covariance matrix. This step constructs a purely nonlinear discrete state transition operator, which extrapolates the initial augmented state matrix generated in step two or the augmented state matrix updated by the previous time step forward by one computation cycle on the time axis.
[0096] In this process, it is necessary to simultaneously realize the finite difference calculation of the apparent mechanical state of the system and the time-varying degradation tracking of the physical parameters of the shear and compressive resistance of the bottom layer of the deep foundation pit, and complete the analytical propagation of the error covariance matrix and the memory-level concurrent derivation calculation in the probability set space.
[0097] An algebraic degradation tracking model is constructed for the "evolutionary parameter column vector" in the augmented state vector. The degradation of soil parameters in deep foundation pits exhibits typical temporal memory and Markov chain characteristics.
[0098] Since complex physical constitutive partial differential equations must be discarded here to improve the computing throughput of computer microprocessors, the system adopts a first-order autoregressive mathematical model (AR(1)) to characterize the time evolution of parameters after logarithmic mapping.
[0099] Let time start from discrete time steps Advance to time step For the vector block representing the physical parameters in the augmented vector. The following linear difference evolution equation is constructed: ; in, For dimension is The autoregressive coefficient diagonal matrix represents the memory retention ratio of parameter decay; It is a constant offset vector, representing the inherent deterministic degradation trend of engineering geology; The process evolution noise vector for physical parameters.
[0100] In the underlying storage and execution implementation of computers, due to It has strict mathematical diagonal matrix properties, and the system kernel will never allocate a complete portion of it in the heap. Two-dimensional and multi-dimensional arrays require memory (which leads to significant space waste and addressing penalties).
[0101] Instead, the memory manager stores it in a highly compressed manner as a contiguous one-dimensional single-precision or double-precision floating-point array.
[0102] When performing computational tasks, the arithmetic logic unit (ALU) of the central processing unit invokes the Single Instruction Multiple Data (SIMD) extended instruction set (such as AVX-2 or AVX-512) to... Multiple adjacent elements are simultaneously loaded into the processor's wide-bit-width data register, along with the compressed one-dimensional... The corresponding elements of the array execute hardware-level parallel fused multiply-accumulate instructions.
[0103] Process evolution noise It uses a true random number generator chip bridged to the computer motherboard to sample in real time using hardware thermal noise or avalanche breakdown effect, and writes it directly into the memory's mapped cache through the high-speed peripheral component interconnect bus.
[0104] Construct a nonlinear physical control operator for the "original state vector" (i.e. displacement, stress, and seepage state) in the augmented vector.
[0105] Extract the updated physical parameters Since these parameters are in the logarithmic field (e.g. , (etc.), the system control program passes the memory address pointers of these logarithmic variables to the execution pipeline of the floating-point coprocessor.
[0106] The processor calls internal microcode-level transcendental functions to accelerate the hardware, executing high-precision natural exponent inverse mapping instructions (such as fexp or similar instructions in low-level assembly) to quickly restore them to the true physical effective elastic modulus, cohesion, and other values.
[0107] This hardware-level mapping avoids the function stack frame overhead associated with calling upper-level software libraries. Using these restored real parameter values, the current time step is assembled in computer memory. global stiffness matrix .
[0108] Based on the principle of equilibrium in finite element computation, construct the system of finite difference algebraic equations for the current time step: ; in, Let be the desired state increment vector; For time steps The applied external load increment vector (such as the equivalent nodal force vector generated by excavation and unloading).
[0109] At the data structure and addressing level, the global stiffness matrix The dimensionality is extremely high (up to millions) and exhibits extremely strong banded sparsity. To absolutely avoid memory overflow exceptions, the system forces a compacted sparse column (CSC) format for its simplified construction.
[0110] The CSC format consists of three independent one-dimensional dynamic arrays: An array for storing all non-zero floating-point values contiguously, an integer array for storing the corresponding row indices, and an array of pointers for indicating the starting offset of each column in the numerical array.
[0111] To solve this giant system of equations, the system invokes a parallel numerical algorithm based on conjugate gradients preprocessed by incomplete Cholsky decomposition.
[0112] The underlying task scheduler of the computer operating system will start a computationally intensive daemon thread pool for this operation. In a non-uniform memory access architecture, the thread is bound to a specific physical core to maximize the use of the local node's memory bandwidth.
[0113] The iterative computation is performed in the Krylov subspace, and each iteration involves the multiplication of an extremely large sparse matrix and a dense vector.
[0114] In order to break the inherent memory bandwidth bottleneck of the von Neumann architecture, the memory controller hardware will start an aggressive prefetching mechanism at this stage, loading the memory data blocks required for the next cycle into the L2 cache in advance across the bus.
[0115] The compiler eliminates pipeline bubbles caused by conditional jump instructions through loop unrolling, ultimately solving for a high-precision state increment column vector with extremely low tolerance threshold. .
[0116] After calculating the state increment, the node's physical state vector at the current moment is accumulated and updated: ; in, The state noise vector is generated by the evolution of the system state within discrete time steps. It follows a Gaussian distribution with zero mean and is independently sampled and generated by the computer's pseudo-random number generator.
[0117] The parametric evolution equation and the state finite difference equation are combined in pure logic and control flow, and defined as a nonlinear state transition global operator. , used to The augmented set matrix at time t is mapped as The augmented set matrix of prior predictions at time step; To ensure dimension alignment and logical consistency in subsequent unified matrix operations, the process noise of the aforementioned physical parameters is... With state process noise Vertically concatenated and stacked, defined as the overall process noise column vector. ; The above nonlinear transfer operator Independent application Augmented set matrix at time Each column vector sample in ( ); Among them, when When =1, ;when When >1, ; This step is a typical parallel computing task in computer science theory, as there is no data read / write dependency between individual samples.
[0118] The operating system's task scheduling module utilizes the shared memory multithreading standard library to... The prediction and inference tasks for each sample are evenly distributed among all physical CPU cores on the motherboard.
[0119] Each computing node performs the following forward prediction operation within its process, which has an independent address space: ; When all After all sample column vectors have completed prediction calculations, the main control thread enters a blocked waiting state through a global synchronization barrier.
[0120] The system employs a spinlock mechanism based on atomic operations to poll the status flags of each child thread in real time.
[0121] After all worker threads have reported completion signals, the memory manager uses pointer offset mapping technology to horizontally align the new vectors scattered in the various high-speed work area caches to obtain the time step. Prior prediction augmented set matrix .
[0122] The numerical distribution characteristics of the predicted state must be analyzed through set statistical properties. The arithmetic mean of the prior prediction set in each dimension is calculated to form a column vector of predicted means. Construct the perturbation prior set matrix. The method is to Subtract the mean vector from each column in the data. This makes the column mean of the perturbation matrix zero.
[0123] Using this perturbation matrix, calculate the prior error covariance matrix. This matrix directly characterizes the uncertainty of the prediction results and the degree of coupling correlation between its internal physical quantities: ; To achieve extremely high floating-point processing efficiency, the system kernel links to and calls the General Matrix Multiplication (GEMM) Level 3 interface in the Basic Linear Algebra Subpackage (BLAS), which is highly optimized for hardware microarchitecture.
[0124] This interface program automatically detects the current processor's cache level capacity and applies the ultra-large-dimensional perturbation matrix. The data is dynamically divided into tiny rectangular blocks that fit the size of the processor's L1 cache.
[0125] By residing data in cache and register files for extended periods, page faults and latency caused by main memory access are minimized, thereby reducing... Massive multiply-accumulate operations with time complexity are compressed into an extremely short physical time.
[0126] Thus, the massive physical evolution process of the deep foundation pit was precisely transformed into high-speed matrix set memory transformation driven by underlying nonlinear operators.
[0127] Step 4, Assimilation Update Based on Kalman Gain: Use the spatial localization function to eliminate spurious elements in the prior error covariance matrix, calculate the cross covariance matrix and Kalman gain matrix, and assimilate the global spatiotemporal observation column vector with the prior prediction augmented set matrix to obtain the posterior update augmented set matrix. The core logic of this step is: The multi-source real sensor global spatiotemporal observation column vector, registered and cleaned in step one and stored in a high-priority memory queue, is used. For the prior prediction augmented set matrix derived from the discrete difference equation and reassembled in memory in step three... This involves performing "measurement updates" and "state assimilation" at the algebraic computation level.
[0128] To address the "false long-range covariance" defect caused by excessively long physical distances between sensors in complex physical systems of deep foundation pits, as well as the floating-point "numerical divergence" disaster caused by high-dimensional matrix inversion operations, this step incorporates highly optimized operator mapping logic and a spatial topology localization operation mechanism based on memory lookup tables into the standard Kalman filter framework.
[0129] Constructing an observation operator matrix that connects the high-dimensional augmented state data space and the low-dimensional sensor observation data space. (or nonlinear observation function) ); The observation of physical quantities such as displacement and pore water pressure is essentially just extracting the calculated values of the corresponding spatial coordinate nodes from the high-dimensional state node vector.
[0130] In low-level computer memory operations, the system pre-constructs a sparse mask array composed of Boolean numbers "0" and "1" to equivalently represent... ; The processor's execution unit, through calling bitwise AND logic control instructions and pointer indirect addressing operations, can quickly "extract" the corresponding target value block directly from the physical memory page where the augmented state vector is located with extremely low nanosecond clock cycle consumption, and store it in a dedicated low-dimensional buffer.
[0131] For vibrating wire stress observation data, since the relative displacement or internal force vector of discrete nodes is usually calculated in the state space, while the stress gauge measures the axial scalar stress of local rigid components (such as the support shaft of a steel support), it is necessary to introduce a rigorous geometric deformation transformation algebraic relationship.
[0132] Let the geometric cross-sectional area of the local component be... The material's elastic modulus parameter is obtained by inverse solving using pointers at the corresponding positions of the augmented vector. The displacement difference between the two local nodes is the evaluation variable. The initial fixed length is ; The formula for calculating the force dimension is: ; The above-mentioned linear memory extraction based on Boolean masks and nonlinear physical mapping based on algebraic computation are combined and defined as the global unified observation operator. .
[0133] To eliminate the overhead of stack frame pushing and popping during regular function calls, the compiler encapsulates the observation operator as an inline function.
[0134] In a multi-threaded concurrent environment, the system directly expands the inline function and applies it to the prior prediction set matrix. For each independent column vector sample, the model prediction column vector in the low-dimensional observation space is obtained. ; Due to the set number of set samples Compared to the vast number of degrees of freedom in soil mechanics, the calculated prior error covariance matrix is often extremely limited. Non-zero statistical noise digital elements will inevitably be generated between spatial nodes that are far apart, causing numerical interference between physically unrelated regions during assimilation calculations.
[0135] To suppress this mathematical error, the system pre-builds a fast static lookup table based on the Gaspari-Cohn fifth-order piecewise rational compact support function in the read-only memory area or in protected memory pages locked by the operating system.
[0136] Suppose there are any two evaluation nodes in the state space. and The three-dimensional Euclidean distance scalar between them is And a cutoff correlation radius constant determined based on the attenuation law of geological characteristics is preset. The normalized dimensionless distance was calculated. ; When entering the computationally intensive assimilation loop, the main program no longer calls the arithmetic logic unit in real time to calculate the extremely resource-intensive fifth power operations and high-precision floating-point division operations. Instead, it directly uses the normalized distance. Multiply by the scaling factor to convert to an integer memory offset address, and directly extract the corresponding covariance decay coefficient value from the high-speed lookup table.
[0137] If distance is found in the conditional judgment Exceeding the cutoff radius The processor's branch predictor will directly direct the instruction exit that returns a Boolean false or floating-point zero value.
[0138] Through this typical low-level computer data structure processing strategy of "trading pre-stored space for extremely fast execution time," a dimension-1 data structure is generated extremely quickly in computer memory. Dense localization constraint matrix ; Based on similarity logic, a cross-dimensional localization matrix is generated to represent the distance between nodes in the state space and nodes in the observation space. ; Construct the observation space perturbation matrix using basic linear operations: ; Using the augmented state perturbation matrix calculated in step three With the above-mentioned observation space perturbation matrix Calculate the cross covariance matrix with spatial localization mask constraints. : ; Among them, mathematical symbols Represents the Hadama product.
[0139] In the actual instruction execution pipeline of the processor, the Hadamard product corresponds to element-wise aligned multiplication based on the memory head address pointers of two matrices with the same dimension.
[0140] The system compiler uses deep loop unrolling technology when generating machine code to eliminate the overhead of branch prediction failure caused by loop counter increment and judgment. It directly forces the tiny spurious covariance that exceeds the effective cutoff radius to zero from the lowest level of algebraic execution logic.
[0141] Similarly, the autocovariance matrix of the observation space is calculated in parallel. ; Let the diagonal observation error covariance matrix established based on the sensor hardware factory calibration parameters be... .because and calculated All are square matrices, and their dimensions are relatively small. To ensure numerical stability during matrix inversion and avoid division-by-zero errors caused by minimal pivots, the system's underlying algorithm library does not use ordinary Gaussian elimination. Instead, it calls a high-precision singular value decomposition (SVD) algorithm or LU decomposition logic with partial pivot selection to safely compute the inverse matrix.
[0142] During the solution process, if the floating-point value enters the denormalized number range, the system will safely refresh the register state according to the IEEE 754 standard.
[0143] The final algebraic calculation instruction stream of the Kalman gain matrix is represented by the following formula: ; The gain matrix It accurately measures the relative information reliability of digital model predictions and real-time hardware measurement data across all degrees of freedom.
[0144] The computer system performs ensemble posterior assimilation update computation based on Kalman gain. To prevent excessive variance shrinkage of the ensemble samples during multiple update iterations, the single-time true global spatiotemporal observation column vector read in step one and stored in the cache must be processed. Apply Gaussian random distribution perturbation.
[0145] set up Based on the observation covariance matrix The sampled independent Gaussian white noise vectors are used to construct a virtual set of perturbation observations: ; For the augmented set matrix of prior predictions Each column vector sample data block Apply assimilation update operation code one by one to obtain the posterior update state column vector. : ; This algebraic equation is the set homogenization update equation, and the arithmetic control unit utilizes a high-dimensional gain matrix. The calculated low-dimensional residual error is mapped back to the ultra-high-dimensional augmented state space with extremely high fidelity.
[0146] During this data write-back phase, all concurrent worker threads must request and acquire a mutex lock from the operating system or use a compare-and-swap atomic instruction to absolutely prevent concurrent writes from generating dirty data.
[0147] Following cache coherence protocols such as MESI, each computing core writes the updated floating-point sequence accurately into the specified posterior memory block.
[0148] The memory addresses of all the computed posterior column vectors are logically rearranged and combined to generate the final posterior updated augmented set matrix. : ; This matrix fully absorbs all the effective physical digital information from the multi-source observation equipment on site during the current period, corrects the cumulative error of the pure differential extrapolation in the previous calculation cycle, and completes the data assimilation calculation flow of the entire deep foundation pit digital evaluation system from "prior theoretical prediction" to "posterior measured alignment".
[0149] Step 5, Joint evaluation quantification and early warning threshold determination: Extract the mean of the posterior updated augmented set matrix for inverse mapping, use the stress tensor to analyze the principal stress, combine Shannon entropy and local reliability to calculate the global instability risk index, and trigger the corresponding control command; When the large-scale spatiotemporal extrapolation calculation of the deep foundation pit reaches the current discrete time node, the posterior updated augmented set matrix generated in step four... The memory block contains a wealth of statistical information, including displacement and stress scalar values that reflect apparent physical deformation, as well as parameters that reflect the degradation of soil shear physical properties.
[0150] This step, from the perspectives of computer algebraic statistics, computational mechanics strength criteria, and digital information theory, updates the augmented set matrix posteriorly. Parallel data mining and feature fusion are performed on the sample column vectors to quantify the overall stability risk index, and when the safe memory preset threshold is triggered, it is directly converted into physical control instructions for the underlying industrial hardware equipment and high-level calls to the visualization graphics computing interface.
[0151] Update the augmented set matrix posteriorly. Perform algebraic dimensionality reduction and inverse mapping of physical quantities. Calculate the posterior arithmetic mean column vector of the set. .
[0152] In modern multi-core computer system architectures, the operating system kernel scheduler will have... A giant matrix of floating-point elements The virtual memory is strictly divided into multiple contiguous data blocks according to the page size. Each active physical worker thread is responsible for reading and locking a specific data block, and using the CPU's advanced vector extension instructions (such as the AVX-512 instruction set) to perform pipelined unhindered accumulation operations on multiple floating-point numbers in high-speed registers simultaneously.
[0153] After the local data blocks are accumulated, a parallel reduction algorithm based on a binary tree structure is used to safely merge the local sums output by each worker sub-thread in the main control daemon thread. Finally, the sum is divided by the total number of integer samples. This allows for the generation and storage of a posterior state mean vector containing precise numerical values in the heap memory. and the mean vector of posterior evolutionary parameters .
[0154] for For each parameter component in the logarithmic field, the system calls the hardware interface of the exponential function, which is highly optimized by assembly language in the standard mathematical core function library (e.g., calling the exp() macro instruction inside the FPU), to reverse map the underlying pure binary logarithmic data back to the real physical numerical space.
[0155] Set the extracted first The effective cohesion of each spatial assessment node is: The effective internal friction angle is .
[0156] This inverse mapping operation ensures the physical rationality of subsequent algebraic mechanics calculations and the absolute consistency of the internal dimensions of the calculator.
[0157] For each digital evaluation node in the spatial mesh of the foundation pit, perform analytical solution of the eigenvalues of the stress tensor.
[0158] From the posterior state mean vector in memory Extracting the first value through pointer offset Each node corresponds to six independent floating-point components of the stress tensor: Normal stress and shear stress To algebraically determine the mechanical yield state of the node in the digital model, it is necessary to calculate the principal stress values under this three-dimensional stress state. In the main computer program, a cache is dynamically allocated to construct a third-order scalar symmetric matrix as the stress matrix. .
[0159] The system prohibits the use of traditional geometric analytical formulas for finding roots, which are inefficient and prone to causing floating-point rounding errors.
[0160] The program directly calls the symmetric matrix eigenvalue solving routines in the basic linear algebra package (such as the dsyev function written in FORTRAN) through static linking.
[0161] At the microinstruction execution level, this routine first uses the Haushold transform to quickly compress the third-order stress-symmetric matrix into a tridiagonal matrix. Then, it uses the QR algorithm with implicit shifts to converge and solve for the three real eigenvalues within a very short system clock cycle through a very limited number of memory register iterations.
[0162] The three real roots are sorted in descending order of their values using the quicksort algorithm, and defined as the maximum principal stresses. Intermediate principal stress and minimum principal stress ; A discrete set concurrent computation logic based on the first-order moment method (FORM) is used to quantify the reliability index of each local node. For the first... For each node, a failure function in pure algebraic form is constructed within the algebraic framework of Mohr-Coulomb strength theory. .
[0163] For the a posteriori updated augmented set matrix The first in The program obtains the next independent sample column vector using a cache prefetching technique. Parameters corresponding to each node , and the maximum and minimum principal stresses , .
[0164] Calculate the first The sample at the th The scalar value of the failure function at each node : ; When performing the above intensive calculations involving trigonometric functions, the system allocates a contiguous, page-locked high-speed main memory area specifically for storing the calculation result array.
[0165] when When the condition code is true, it means that the sample has crossed the mechanical failure boundary at the level of this algebraic operation deduction; otherwise, it is within the safe numerical limit.
[0166] Statistics using accumulators Extract the mean of the discrete function values of each sample. and standard deviation Calculate the local reliability index of this node. ; Introducing the cumulative distribution numerical operator of the standard normal distribution The reliability index is nonlinearly mapped to a scalar of local failure probability. .
[0167] Considering The operator contains extremely time-consuming numerical integration operations. During the initialization phase, the system kernel pre-compiles the Chebyshev polynomial approximation algorithm into a read-only instruction stream residing in the instruction cache.
[0168] When the processor receives a negative parameter At that time, the arithmetic logic unit only needs to perform a few simple multiplication and addition pipeline instructions to output a failure probability value with double precision and high fidelity, thereby ensuring the real-time performance of the digital system.
[0169] To comprehensively measure the degree of spatial information disorder remaining in the system after multi-source data fusion, the Shannon entropy calculation mechanism from digital information theory is introduced.
[0170] For the The physical displacement state variables at each node are updated in the posterior augmented set matrix. of Fit a discrete sequence of probability mass functions from the n sample column vectors. Suppose the state variable falls on the nth... The probability of a defined range of numbers is And the sum of the probabilities of each interval is equal to the digital constant 1.
[0171] To statistically analyze this probability distribution in physical memory without conflict, the system strictly divides the range of the displacement variable into upper and lower extreme values. It has equidistant floating-point number intervals and requests the operating system to initialize a space of length [length missing]. A continuous integer array is used as a histogram counter.
[0172] The program uses multiple threads to traverse and update the augmented set matrix posteriorly. Each sample column vector is used to quickly and accurately locate the corresponding array index subscript by using floating-point division between the current displacement value and the interval step size, as well as hardware-level floor function instructions.
[0173] The processor ensures that the count value at the corresponding index memory address is safely incremented by forcibly calling the atomic addition operation locked by the memory bus, thereby completely avoiding race conditions and data overwrite disasters caused by multi-threaded concurrent writes from the hardware level.
[0174] The local state information entropy of this node The algebraic formula is: ; The magnitude of the information entropy scalar directly reflects the objective numerical uncertainty of the current state of the deep foundation pit system.
[0175] After completing the matrix analysis of local spatial nodes, perform global spatial topology weighted fusion to calculate the global instability risk index of the foundation pit; Define a node spatial importance weighted column vector Its total dimension is strictly equal to the total number of computation nodes partitioned in step two. (This dimensional constraint is crucial for preventing out-of-bounds access to computer arrays.)
[0176] The program assigns corresponding dimensionless weight values to each node based on topological parameters such as the relative absolute depth of each node in the three-dimensional coordinate system of the foundation pit and the geometric Euclidean distance from the rigid support structure, and stores them in the above column vector.
[0177] Organize the local failure probabilities of all evaluation nodes into a one-dimensional column vector. Organize the information entropy of all nodes into a one-dimensional column vector. ; Calculate the global instability risk index The core evaluation indicator is derived using a weighted inner product formula with two objectives: ; In the formula, and The risk dimension balance constant is pre-programmed into the motherboard's read-only register based on the engineering design and survey report.
[0178] Because the formula involves two extremely large sets of long vector inner product operations during execution, the system main control program directly calls the series dot product standard interface of the basic linear algebra subroutine library.
[0179] This interface directly controls the register to perform vectorized loop unrolling, and uses the processor's multiply-accumulate instruction pipeline to output the global risk scalar result in a handful of clock cycles.
[0180] The digital system has entered a multi-level automated early warning and judgment phase, along with underlying hardware control execution. A sequence of incrementally increasing risk-triggered digital thresholds, including a yellow warning threshold, is established within the computer's main memory protected area. Orange alert threshold and red alert threshold .
[0181] The calculated real-time global instability risk index is determined through an advanced conditional jump logic mechanism. Perform a step-by-step alignment with the above constant sequence: If logical condition Upon successful completion, the system determines that the physical foundation pit is in a routine inspection state.
[0182] At this time, the central processing unit sends rendering instructions to the graphics processing unit through the peripheral component interface bus. The GPU loads the matrix state data into the video memory and uses the vertex shader and fragment shader in the graphics rendering pipeline to draw a three-dimensional topological digital twin model of the pit with a safe green coloring in real time on the LCD monitoring screen in the control room. At the same time, the matrix snapshot log at the current moment is serialized and stored in the non-volatile database block of the solid-state drive in an asynchronous non-blocking input / output (AsyncI / O) manner.
[0183] If logical condition The system determines that the deformation has intensified locally.
[0184] The background communication daemon is awakened and automatically assembles a control data packet containing the hexadecimal physical address of the acquisition device, the write register function code, the configuration payload area, and the cyclic redundancy check code (CRC16), and injects it into the network protocol stack of the Transmission Control Protocol / Internet Protocol (TCP / IP) in the operating system kernel layer.
[0185] Data packets are sent to the network interface card (NIC) hardware and routed at high speed through the field industrial Ethernet switch to each front-end data acquisition gateway. After the gateway microcontroller parses the instructions, it rewrites the sensor's internal configuration registers via the bus, forcibly shortening the sensor's sleep cycle, thereby significantly increasing the multi-source data sampling frequency for the next communication cycle at the physical hardware level.
[0186] If logical condition Upon successful activation, the system generates a high-priority alarm trigger signal; The state of a specific pin in the general-purpose input / output interface register on the computer motherboard is rewritten by the main control program, outputting a continuous 5V high-level signal to the external pin. This DC digital signal is then electrically isolated from the weak current through an optocoupler on the circuit board, driving a small solid-state electromagnetic relay in the field power control cabinet to engage, directly activating the 220V AC power supply circuit of the audible and visual alarm lighthouse, emitting a loud physical audible and visual intervention alarm.
[0187] If logical condition Upon successful execution, the system determines that a critical instability danger state has been reached. The computer control program immediately triggers a non-masked, highest-priority hardware interrupt request (NMI) to the central processing unit core. The processor immediately suspends all current background low-priority task threads and uses the industrial real-time communication bus protocol dedicated to the programmable logic controller (PLC) to send an emergency stop digital command message to the power distribution box of the heavy excavating machinery at the excavation site.
[0188] After parsing the message, the PLC directly cuts off the power supply to the control coil of the AC contactor of the mechanical equipment. Simultaneously, the digital signal drives the frequency converter (VFD) of the groundwater recharge system pump to operate at full load, controlling the digital electric proportional valve to open immediately, and using external hydraulic fluid physical means to forcibly establish reverse water pressure to balance the seepage pressure difference inside and outside the foundation pit.
[0189] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.
[0190] It should be noted that, in this document, the use of relational terms such as "first" and "second" is merely for distinguishing one entity or operation from another, and does 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 a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.
[0191] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.
[0192] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.
[0193] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A method for intelligent assessment of deep foundation pit stability based on multi-source monitoring data, characterized in that, include: The physical signals collected by the field sensors are converted into discrete observation column vectors. Interpolation alignment and spatial mapping are used to eliminate asynchrony and generate global spatiotemporal observation column vectors. The original state vector and evolution parameter column vector of the discrete model of the foundation pit are obtained and concatenated to expand the dimension. The preset error covariance matrix is decomposed and combined with white noise sampling to generate the initial augmented state evolution matrix. A linear difference evolution equation is constructed using an autoregressive model, and a state transition operator is constructed by combining the finite difference equation. The initial augmented state evolution matrix is then derived to generate the prior prediction augmented set matrix and the prior error covariance matrix. By using a spatial localization function to eliminate spurious elements in the prior error covariance matrix, calculating the cross covariance matrix and Kalman gain matrix, and assimilating the global spatiotemporal observation column vector with the prior prediction augmented set matrix, the posterior updated augmented set matrix is obtained. The mean of the posterior updated augmented set matrix is extracted and inversely mapped. The principal stresses are analyzed using the stress tensor. The global instability risk index is calculated by combining Shannon entropy and local reliability, and the corresponding control command is triggered.
2. The intelligent assessment method for deep foundation pit stability based on multi-source monitoring data according to claim 1, characterized in that, The physical signals acquired by the field sensors are converted into discrete observation column vectors, including dimensional transformation of the physical signals output by the vibrating wire stress monitoring sensor: The zero-crossing point of the AC signal from the vibrating wire stress monitoring sensor is captured using a hardware timer and an edge-triggered interrupt mechanism to calculate the real-time frequency; The floating-point arithmetic unit is used to substitute the real-time frequency into the following physical dimension conversion formula to generate the actual physical stress observation value: ； In the formula, Indicates the first A stress sensor in time The actual physical stress observation value, Indicates the calibration coefficient. Indicates real-time frequency. Indicates the initial calibration frequency. This represents the coefficient of thermal expansion. Indicates real-time temperature. This indicates the initial temperature.
3. The intelligent assessment method for deep foundation pit stability based on multi-source monitoring data according to claim 1, characterized in that, By eliminating asynchronicity through interpolation alignment and spatial mapping, a global spatiotemporal observation column vector is generated, including: Establish a data structure based on hash mapping and doubly linked lists in computer memory to cache sensor data payloads with inconsistent arrival times; The daemon process is started to calculate the derivative value using forward and backward differential logic, and executes polynomial interpolation logic to resample the sensor data payload to a unified evaluation time node to splice and generate a global spatiotemporal observation column vector. A first-in-first-out circular queue is established as a sliding time window. The sliding mean and sliding standard deviation are calculated. Pulse anomaly elements that exceed the set statistical threshold are identified and forcibly replaced. The global spatiotemporal observation column vector is output.
4. The intelligent assessment method for deep foundation pit stability based on multi-source monitoring data according to claim 1, characterized in that, Obtain the original state vector and evolution parameter column vector of the discrete model of the foundation pit, and concatenate and expand the dimension, specifically including: The computational nodes are indexed and arranged according to the graph theory topology mapping rules of spatial grid generation, and the apparent mechanical morphology of the nodes that ensures spatial locality is represented as the original state vector. Extract the effective elastic modulus, effective cohesion, and effective internal friction angle, and call the mathematical function interface to perform logarithmic nonlinear mapping to construct an evolution parameter column vector in the logarithmic domain; Allocate contiguous physical memory blocks and use the direct memory access controller to vertically concatenate and stack the byte streams of the original state vector and the evolution parameter column vector to generate an augmented state column vector.
5. The intelligent assessment method for deep foundation pit stability based on multi-source monitoring data according to claim 1, characterized in that, The preset error covariance matrix is decomposed and combined with white noise sampling to generate an initial augmented state evolution matrix, including: Read the preset error covariance matrix stored in a compressed sparse format, and call the linear algebra subroutine to perform lower triangular decomposition to obtain the lower triangular matrix; Extract the hardware state parameters of the central processing unit motherboard as high-entropy seeds, and combine them with a transformation algorithm to generate a white noise vector that follows a standard normal distribution; The initial augmented state sample column vector is generated by: ； In the formula, Indicates the first An initial augmented state sample column vector, This represents the optimal estimated augmented state mean vector. Represents a lower triangular matrix. Represents a white noise vector; Arrange all generated initial augmented state sample column vectors horizontally in a column-major, contiguous memory arrangement to generate the initial augmented state evolution matrix.
6. The intelligent assessment method for deep foundation pit stability based on multi-source monitoring data according to claim 1, characterized in that, A linear difference evolution equation is constructed using an autoregressive model, including execution time-series deduction for the evolution parameter column vector: ； In the formula, Indicates time step The evolution parameter column vector, Indicates time step The evolution parameter column vector, This represents the diagonal matrix of autoregressive coefficients. Represents a constant offset vector. Represents the process evolution noise vector; During execution, the autoregressive coefficient diagonal matrix is compressed and stored as a one-dimensional floating-point array, and the single-instruction multiple-data extended instruction set is invoked to adjust the time steps. The adjacent elements of the evolution parameter column vector are fused with the corresponding elements of the one-dimensional floating-point array in parallel for multiplication and addition operations.
7. The intelligent assessment method for deep foundation pit stability based on multi-source monitoring data according to claim 1, characterized in that, A state transition operator is constructed using finite difference equations to deduce the initial augmented state evolution matrix, generating the prior prediction augmented set matrix and the prior error covariance matrix, including: The updated parameters are derived using the linear difference evolution equation. The logarithmic domain variables in the evolution parameter column vector are then subjected to exponential inverse mapping to restore the physical parameter values and assemble the global stiffness matrix. The state increment column vector is obtained by using a numerical algorithm based on incomplete Cholsky decomposition preprocessing and conjugate gradient. The nonlinear transition operator containing state deduction updates is applied independently to the column vectors of each set of samples, and then concatenated to obtain the prior prediction augmented set matrix; Extract the column vector of the predicted mean of the prior prediction augmented set matrix to construct the perturbation prior set matrix, and multiply the perturbation prior set matrix with its transpose to obtain the prior error covariance matrix.
8. The intelligent assessment method for deep foundation pit stability based on multi-source monitoring data according to claim 1, characterized in that, The spatial localization function is used to eliminate spurious elements in the prior error covariance matrix, and the cross covariance matrix is calculated, including: A sparse mask array composed of Boolean numbers is constructed as the observation operator matrix, and the model prediction column vector is extracted from the high-dimensional state memory paging by using pointer indirect addressing. A static lookup table based on the Euclidean distance and compact support function of state nodes is pre-established in the read-only storage area; the normalized distance is extracted in the assimilation loop as the memory offset address, the decay coefficient is read from the static lookup table, and the localized constraint matrix is generated. Perform an element-wise aligned Hadamard product operation on the localized constraint matrix and the cross covariance matrix calculated based on the perturbation set, and force zero out any spurious covariance elements that exceed the cutoff correlation radius.
9. The intelligent assessment method for deep foundation pit stability based on multi-source monitoring data according to claim 8, characterized in that, Calculate the Kalman gain matrix, assimilate the global spatiotemporal observation column vectors with the prior prediction augmented set matrix to obtain the posterior updated augmented set matrix, which contains the following associated features: The singular value decomposition algorithm is called to calculate the inverse matrix of the sum of the observation space autocovariance matrix and the diagonal observation error covariance matrix under spatial localization constraints, and then multiplied by the cross covariance matrix to obtain the Kalman gain matrix. The posterior updated state column vector is derived using the following set-congruent update equation: ； In the formula, This represents the posterior update state column vector. This represents the column vector of predicted samples in the prior prediction augmented set matrix. Represents the Kalman gain matrix. This represents the perturbed observation set vector generated after perturbation of the global spatiotemporal observation column vector. This represents the column vector predicted by the model; During the write phase of executing the set-combinatorial update equation, a mutex lock mechanism is acquired to prevent concurrent writes from generating dirty data in memory, and then the a posteriori update augmented set matrix is generated.
10. The intelligent assessment method for deep foundation pit stability based on multi-source monitoring data according to claim 1, characterized in that, The mean of the posterior updated augmented set matrix is extracted and inversely mapped. The principal stresses are analyzed using the stress tensor. The global instability risk index is calculated by combining Shannon entropy and local reliability, triggering corresponding control commands, including: The linear algebra package is used to perform Haushold transformation and implicit matrix root-finding algorithm on the third-order stress symmetric matrix to analyze the principal stresses; A histogram counter is constructed by dividing the physical memory into equidistant digital intervals, and the probability distribution function of the statistical state variable is calculated by using atomic addition operations. The global instability risk index is derived using the following bi-objective weighted inner product formula: ； In the formula, This indicates an index representing the overall risk of instability. and This represents the balance constant of the risk dimension. This represents a column vector weighted by the spatial importance of nodes. This represents a one-dimensional column vector consisting of the local failure probabilities of all evaluated nodes. A one-dimensional column vector representing the information entropy of all nodes; The global instability risk index is compared with a preset incremental risk trigger threshold, which in turn triggers the graphics rendering pipeline to draw a digital model or sends a shutdown command message to the field programmable logic controller.
Citation Information
Patent Citations
Tunnel surrounding rock stability evaluation method and system based on multi-physics field parameter inversion
CN121522770A
Continuum robot state estimation method based on Gaussian process regression
CN122432468A